For my father,

who was born into a world of oil lamps and bullock carts,

and has lived to see machines that think.

He did not need to understand the technology to understand what mattered most: that power always moves — and the wise know where it is going before the powerful notice it has left.

At ninety-four, he has outlasted empires, currencies, ideologies, and now perhaps the very idea that only humans can think. This book is my attempt to make sense of what comes next — written in his honour, and in his spirit.

献给亲爱的父亲大人

他生于油灯与牛车的年代，却亲眼见证了会思考的机器降临人间。

权力从不停歇——智者在权贵察觉之前，已知晓它的去向。

— 陳德強教授博士 · Prof Dr Tan Teik Kheong
IEEE · 2026

Why Power, Not Technology

Every era has a technology that rewires the global economy. Steam rewired manufacturing and unleashed the industrial age. Electricity rewired cities and extended the productive day into the night. The internet rewired commerce, communication, and the fundamental architecture of information. Artificial intelligence is rewiring everything simultaneously — and doing it faster than any previous technology transition in recorded history.

But this book is not primarily about the technology. It is about the power. Specifically, about how economic and strategic power is transferring — systematically, structurally, and largely invisibly — across the entire AI ecosystem. The transfer is happening at every layer of the technology stack simultaneously, and it is producing winners and losers at a pace that most policy frameworks, investment models, and individual career plans were not designed to handle.

From apps to infrastructure. From software to compute. From consumer brands to hidden infrastructure owners. From public markets to private capital. From labour to automation. From national economies to AI-capable blocs. From the visible to the invisible.

This book was written for three audiences simultaneously, and it does not apologise for that ambition.

For investors and financial professionals: you need a map of where durable economic value is accreting in the AI economy, and it is not where the headlines are pointing. The picks-and-shovels layer of the AI stack — the chips, the energy, the cloud, the real estate — is where the most defensible positions are being built. This book will show you where to look and what questions to ask.

For business leaders and strategists: every organisation in every industry is now an AI-dependent organisation, whether it has acknowledged that dependency or not. The June 2026 kill switch episode — when a single government letter disabled AI access for an entire continent overnight — demonstrated that dependency without sovereignty is a strategic vulnerability of the highest order. This book will help you understand what you have built your organisation on, and what it means if the ground shifts.

For policy-makers, academics, and informed citizens: the AI power shift is the defining geopolitical and economic transition of the current decade. Understanding its structural logic — who owns which layer, who pays whom, what the hidden dependencies are — is a prerequisite for any serious engagement with the questions of governance, sovereignty, equity, and security that AI is forcing onto every national agenda.

Throughout this book, seven questions recur. They are the analytical framework that cuts through the noise of AI announcements, benchmarks, and competitive claims, and gets to the structural reality underneath.

Who owns the layer? Who pays whom? Where does the money flow? What is the bottleneck? What is the hidden dependency? What happens to retail investors? And what changes globally — and for whom?

Ask these questions consistently, about every AI announcement you encounter, and the power shift comes into sharp, uncomfortable focus. The answers are rarely the ones that the press releases suggest.

This is a book written in real time about events still unfolding. The research behind it was gathered, verified, and updated continuously through mid-2026. Some of the most significant events described in these pages — the Anthropic IPO filing, the export control crisis, the Mythos cyberwarfare disclosure — occurred within weeks of this writing.

That is not a weakness. It is the point. The AI power shift is not a story about the future. It is a story about the present, moving faster than most people have registered. The purpose of this book is to give you the map while the territory is still being drawn.

Every chapter seed in this book was planted by a real event, a real data point, a real decision made by a real person with real money at stake. Nothing here is speculation. Everything here is already happening.

The Day Apple Borrowed a Brain

In June 2024, Apple stood on stage at its annual Worldwide Developers Conference and made an announcement that, on the surface, sounded like a triumph. The company that had spent five decades defining what personal technology looks like — the company that put a computer in your pocket, a store in your palm, and a watch on your wrist — announced that it was partnering with OpenAI to bring artificial intelligence to its products.

The crowd applauded. The headlines celebrated. Analysts upgraded their price targets. The stock moved.

Nobody asked the obvious question: why couldn't Apple build this itself?

Apple's market capitalisation at the time hovered around three trillion dollars. It employed nearly one hundred and seventy thousand people. It operated one of the most sophisticated hardware engineering operations in human history. It had its own custom silicon — the M-series chips that had outpaced Intel's roadmap by years. It had cash reserves that exceeded the GDP of most countries.

And yet, when it came to the most important new capability in consumer technology, it had to go to a startup founded in 2015 and ask to borrow its brain.

This book is about why that happened. More precisely, it is about the structural shift in economic and technological power that made it inevitable — and what it means for every investor, every business leader, every government, and every worker trying to understand what the AI era actually looks like from the inside, not from the press release.

The public conversation about artificial intelligence is almost entirely focused on the wrong things. We talk about chatbots. We talk about whether AI will write our emails, replace our doctors, or take our jobs. We debate whether ChatGPT is conscious, whether Elon Musk or Sam Altman is telling the truth about timelines. We download apps, run prompts, and share screenshots of clever outputs on social media.

Meanwhile, the actual power shift is happening somewhere most people never look.

It is happening in data centres the size of football stadiums being built across Virginia, Texas, and Arizona — buildings that consume as much electricity as a medium-sized city and require a year of construction before a single model runs inside them. It is happening in semiconductor fabs in Taiwan that are the single most strategically important buildings on earth, where the machines that make the chips that run the models cost more per unit than a commercial aircraft. It is happening in electricity grids that are straining under demand they were never designed to handle, forcing utilities to dust off nuclear plants they had planned to decommission.

It is happening in the private capital markets, where sovereign wealth funds from Abu Dhabi, Saudi Arabia, and Singapore are writing cheques that make venture capital rounds look like rounding errors, buying stakes in AI infrastructure companies that will never appear on a public stock exchange until the extraordinary returns have already been made.

It is happening in the infrastructure. And infrastructure is where power lives — not in the apps that sit on top of it, and not in the models that run inside it.

Understanding why Apple had to borrow a brain requires understanding the architecture of the AI economy — specifically, why building a frontier large language model is not a problem that money alone can solve.

Apple had the money. It had the engineering talent. It had the distribution. What it did not have — what no device company has been able to build quickly enough — was the accumulated expertise in transformer architectures, the proprietary training data at the required scale, and the institutional knowledge of how to manage the catastrophically expensive process of training and iterating on frontier models.

These things take years to build, not quarters. OpenAI had been building them since 2015. Anthropic had been building them since 2021, staffed by some of the most experienced AI safety researchers in the world. Google had been building them since before the transformer architecture itself was published, in a Google paper, in 2017.

Apple arrived at the AI era as a consumer hardware and software company trying to retrofit intelligence into products that were already designed. It was, in the language of the technology industry, a fast follower in a domain where being first — and being right about the architecture early — turns out to matter enormously.

The deal Apple struck with OpenAI is a microcosm of the broader AI power structure. A consumer-facing brand with enormous distribution and customer relationships, dependent on an infrastructure-layer company for the intelligence that makes its products work. The brand takes the credit. The infrastructure takes the rent.

This book is organised around seven questions. They are the same questions a serious investor, a strategic executive, or an informed policy-maker should be asking every time a new AI announcement lands.

Who owns the layer? Who pays whom? Where does the money flow? What is the bottleneck? What is the hidden dependency? What happens to retail investors? And what changes globally — and for whom?

These questions cut through the noise. They do not ask whether AI is impressive — it clearly is. They ask who captures value from it, and on what terms. They do not ask whether large language models will pass the bar exam. They ask who owns the compute that runs them, who owns the energy that powers that compute, and who owns the land the data centres sit on.

The answers, it turns out, are surprisingly concentrated. A small number of companies — and a smaller number of individuals — have positioned themselves to collect tolls from the entire AI economy, regardless of which model wins the capability race. This book names them, maps their positions, and explains what it means for everyone else.

The pattern Apple established in 2024 — borrowing intelligence from a competitor rather than building it — is not unique to Apple. It is the defining condition of an entire generation of technology companies that arrived at the AI era too late to build their own foundation models but too dependent on AI capability to ignore it.

Samsung, the world's largest smartphone manufacturer by volume, made a similar arrangement with Google. Salesforce, one of the world's most successful enterprise software companies, licensed models from OpenAI and Anthropic rather than building its own. Adobe, whose creative tools define the professional workflows of designers, photographers, and videographers worldwide, embedded AI capabilities from multiple providers rather than developing frontier models in-house.

The list extends across every industry. Healthcare companies licensing AI diagnostic models they cannot audit. Financial institutions embedding AI credit models they cannot fully explain to regulators. Legal firms deploying AI contract review tools whose training data they do not know. Educational institutions using AI tutoring systems trained on curricula they did not design.

In each case, the dynamic is the same: an organisation with significant capability, significant resources, and significant existing relationships has concluded that building its own frontier AI capability is beyond its reach — economically, technically, or strategically — and has accepted a dependency on an external provider.

The question this raises, which most of these organisations have not yet confronted squarely, is what that dependency means when the provider's interests and the organisation's interests diverge. Because they will diverge. They always do.

The history of enterprise technology is littered with the consequences of deep vendor dependency: organisations that built their workflows on a single supplier's products and discovered, years later, that switching costs had effectively imprisoned them. Oracle database customers who found themselves paying maintenance fees they could not afford to stop paying. IBM mainframe customers who built decades of institutional knowledge around a platform that became progressively more expensive as alternatives emerged. Microsoft Office customers whose document formats became so embedded in their workflows that alternatives were effectively impossible to adopt even when they were technically superior.

In each case, the vendor had provided genuine value. The dependency was not the result of malice. It was the result of the natural economics of technology platforms: standards emerge, switching costs accumulate, and the vendor's pricing power increases as the customer's ability to exit decreases.

AI dependency has the potential to be more severe than any previous form of enterprise technology dependency, for two reasons. First, the switching costs are higher: AI systems are embedded not just in software workflows but in the cognitive processes of the people who use them, creating a form of dependency that is harder to measure and harder to reverse than software lock-in. Second, the power asymmetry is more extreme: the handful of companies that own frontier AI infrastructure are fewer in number, more dominant in position, and more politically connected than any enterprise software vendor in history.

When a government can order an AI company to disable its models with a single letter — as happened in June 2026 — the nature of the dependency has crossed into a category that previous enterprise technology relationships never reached. You were never at risk of having your Oracle database turned off by the US government. You are potentially at risk of having your AI infrastructure turned off by it.

This is not a reason to avoid AI adoption. It is a reason to approach AI adoption with a clear-eyed understanding of what you are entering into, and to build whatever degree of sovereignty and resilience your situation permits.

There is a concept in finance called sovereign debt — the obligations a nation owes to external creditors that constrain its policy choices and its room to manoeuvre in a crisis. A nation with high sovereign debt cannot always pursue the fiscal policy its domestic situation demands, because its creditors' preferences take priority over its citizens' needs.

The AI era is creating a form of sovereign technological debt that operates by a similar logic. Nations that have built their critical digital infrastructure on platforms and services they do not own have created obligations to external providers that constrain their policy choices in ways they may not fully recognise until a crisis forces the constraint into view.

The June 2026 kill switch was that crisis for much of the developed world. But it was not the first warning. The lessons had been available for years, in smaller episodes that did not attract the same attention.

When the US Department of Commerce added Huawei to its entity list in 2019, restricting American companies from selling components to the Chinese telecommunications giant, the consequence was not merely that Huawei lost access to American chips. It was that every telecommunications operator in the world that had built its networks on Huawei equipment had to reconsider the long-term viability of its infrastructure, because the maintenance and upgrade path for that equipment had suddenly become uncertain. The sovereign technological debt of Huawei dependency became visible when the political relationship between the United States and China deteriorated.

The lesson was available in 2019. Most of the world chose not to learn it then. The 2026 episode was considerably more dramatic — an entire continent losing access to AI tools overnight — and the lesson it taught was considerably harder to ignore.

What makes technological dependency different from other forms of dependency is its invisibility in normal times. A nation's dependence on imported oil is visible every time a driver fills their tank and pays the price that international markets have set. A nation's dependence on imported AI capability is invisible in normal operation — the AI system works, the workflow proceeds, the output is produced, and no reminder appears that the capability originates outside the nation's control.

The invisibility ends in a crisis. And in a crisis, the time available to build alternative capability is typically zero. The sovereign technological debt comes due precisely when the debtor is least able to pay it.

The strategic implication is that building AI sovereignty is not a project to be deferred until the dependency becomes a crisis. It is a project that must proceed before the crisis, in conditions that make it easy to defer — because the crisis will arrive without warning, and the consequence of having deferred will be visible immediately and painful concretely.

This is the core insight that the most strategically thoughtful nations have drawn from the 2026 episode: AI sovereignty is not a technology project. It is a national security project. And national security projects do not wait for market conditions to be favourable.

To understand the full implications of Apple's decision to borrow a brain, it helps to think about intelligence as a stack — an architectural concept borrowed from software engineering that describes how complex systems are built from layers of simpler components, each layer providing services to the layer above it and depending on services from the layer below.

The intelligence stack that Apple needed access to in 2024 had five layers, each controlled by a different set of actors, and each representing a different form of dependency.

At the base of the stack was raw compute: the ability to perform mathematical operations at the extraordinary speed and scale required to train and run large language models. This layer is controlled by NVIDIA, which designs the chips, and TSMC, which manufactures them. Apple, notably, has its own chip design capability — the M-series and A-series chips that power its products are among the most capable in the world. But Apple's chips, excellent as they are for the workloads they are designed for, are not optimised for the massively parallel matrix multiplication that transformer model training requires. Building that capability from scratch would have taken years and billions of dollars.

The second layer was training infrastructure: the software systems, the data management tools, the distributed computing frameworks, and the institutional knowledge required to actually train a large model on raw compute. This layer is controlled by a small number of organisations that have spent years developing proprietary training systems that embody enormous accumulated expertise. Training a frontier model is not merely a matter of having chips and data; it requires knowing which architectural choices to make, which training techniques to apply, how to debug training runs that go wrong, and how to evaluate whether a trained model is behaving as intended.

The third layer was training data: the curated datasets of human text, code, images, and other information that give a model its capabilities. The publicly available internet text that trained the first generation of large language models is largely exhausted as a source of novel capability. The frontier model companies have developed proprietary data curation processes, data quality filters, and synthetic data generation techniques that give their models capabilities that cannot be replicated simply by training on the same public data.

The fourth layer was the model itself: the trained weights that encode the capabilities developed through the training process. The model weights are the product of all the investment in the layers below — the compute, the training infrastructure, the training data — and they embody capabilities that are not easily reverse-engineered or replicated.

The fifth layer was deployment infrastructure: the systems required to actually serve the model at scale — the inference optimisation, the rate limiting, the safety filtering, the API design, the monitoring and observability. This layer is less obviously differentiating than the layers below it, but it is the layer that makes the difference between a model that works in a research environment and a model that works as a consumer product at billion-user scale.

Apple could not build layers one through four quickly enough to matter. It had the resources to build them eventually — Apple has built custom chip architectures from scratch, has the capital to fund enormous training runs, and has the engineering talent to build training infrastructure. But 'eventually' was not good enough in a market where the pace of capability improvement was so rapid that being eighteen months behind the frontier meant being essentially irrelevant.

So Apple took the path that made economic sense given its situation: it licensed the output of layers one through four from the company that had built them, and added its own contribution at layer five — the deployment infrastructure that makes a model work elegantly within Apple's ecosystem, respecting Apple's privacy standards, functioning offline when connectivity is unavailable, and integrated with the other capabilities of Apple's devices.

This division of labour is efficient. It allows Apple to provide AI capabilities to its users without building the full intelligence stack from scratch. But it creates a dependency that Apple did not have before — a dependency on OpenAI's willingness and ability to provide the services that Apple's products now require.

Every organisation that has made a similar choice — which is to say, most organisations that have adopted AI capabilities in any serious way — has created a similar dependency. And the sum of those individual dependencies is the infrastructure of the AI power shift: the concentrated ownership of the intelligence stack in a small number of organisations that now have leverage over an enormous proportion of the world's AI-dependent workflows.

The New AI Empire

There is a question nobody asks at AI conferences. Speakers present stunning benchmarks. Researchers unveil models that passed the bar exam, diagnosed cancer from scans, and wrote code faster than senior engineers. Founders describe the future they are building. Investors explain why this cycle is different from the last one. Journalists file breathless dispatches from demo rooms.

And nobody asks: who actually owns this?

Not who built the model. Not who named it. Not who put it in an app and charged a subscription. Who owns the infrastructure that makes it possible? Who owns the compute, the energy, the connectivity, the capital structure underneath everything the public sees?

The new AI empire is not owned by the companies whose logos appear on the products. It is owned by a smaller, quieter group of actors who positioned themselves in the infrastructure layer before the consumer race began — and who now collect rent from every model, every inference call, and every AI-generated word produced anywhere on earth.

To understand who owns the AI empire, you need to understand the stack it runs on — the layered architecture of technology and capital that sits between the silicon in a chip and the answer on your screen.

At the top of the stack sits the application layer — the ChatGPTs, the Claudes, the Geminis, the Copilots. This is what consumers see and what journalists write about. It is also, in structural terms, the most competitive and least defensible part of the entire value chain. Models improve quickly. Switching costs are low. Today's leading chatbot is tomorrow's also-ran. The application layer is real, valuable, and genuinely impressive. But it is not where durable power lives.

Below the application layer sits the model layer — the large language models, the foundation models, the training runs that cost hundreds of millions of dollars and require tens of thousands of specialised chips running in synchrony for months. OpenAI, Anthropic, Google DeepMind, and Meta AI operate here. So does Mistral, Cohere, and a growing cluster of Chinese labs. This layer has more defensibility than the application layer — the capital requirements alone deter most entrants — but it is still intensely competitive, and the economics are brutal.

Below the model layer sits the cloud infrastructure layer — Amazon Web Services, Microsoft Azure, Google Cloud. This is where the models actually run. Every training job, every inference call, every API request routes through one of these three platforms. The hyperscalers do not just host AI — they supply the compute, the networking, the storage, the tooling, and increasingly the capital that makes AI development possible.

Below the cloud layer sits the hardware layer — and at the top of the hardware layer sits a single company whose dominance is so complete it has no historical precedent in the technology industry.

NVIDIA did not plan to become the most important company in artificial intelligence. It built graphics processing units for video games. The architecture that made GPUs fast at rendering 3D graphics — massively parallel computation across thousands of cores — turned out to be exactly what training neural networks required. When the deep learning revolution began in earnest around 2012, NVIDIA's hardware was already there, waiting.

By 2026, NVIDIA's position in AI compute was not merely dominant. It was structurally indispensable in a way that no technology company had been since Microsoft owned the operating system of the personal computer era.

Every major AI lab in the world trains its models on NVIDIA hardware. Every hyperscaler runs NVIDIA chips in its AI clusters. Every startup building on top of foundation models is, at some level of the stack, paying NVIDIA. The company has become the landlord of the AI economy — not by building the most visible products, but by owning the layer that everything else depends on.

Jensen Huang, NVIDIA's co-founder and CEO, described the dynamic precisely when he said that NVIDIA sells the 'pick and shovel' of the AI gold rush. The metaphor is apt, but it undersells the position. In the original gold rush, there were many shovel makers. In the AI era, there is effectively one.

If NVIDIA owns the hardware layer, the three American hyperscalers — Amazon, Microsoft, and Google — own the layer above it. And they have been quietly restructuring the entire AI industry around their own balance sheets.

In 2026, the Big Four hyperscalers — Amazon, Microsoft, Google, and Meta — committed a combined $725 billion in capital expenditure, up 77 percent year-on-year. The primary driver in every case was AI infrastructure: data centres, chips, power, cooling, fibre. This is not research spending. It is physical infrastructure investment, the kind that takes years to build and decades to depreciate.

The pace has not slowed since. Microsoft alone guided to approximately $190 billion in capital expenditure for calendar year 2026 — a 61 percent increase year-on-year — while the combined capital expenditure of the four largest hyperscalers was on track to exceed $700 billion for the year. JPMorgan's research division, revising its outlook in mid-2026, raised its projection for cumulative global AI-related capital expenditure through 2030 to $5.5 trillion, up from $5.1 trillion only months earlier — driven, the bank noted, not by more ambitious plans but by rising debt financing as loan-to-cost ratios climbed across the industry.

Microsoft moved first and most decisively. Its $13 billion investment in OpenAI — structured as Azure compute credits rather than cash equity — was a masterstroke of infrastructure lock-in. OpenAI's models run on Azure. Every dollar OpenAI earns in revenue flows partly back to Microsoft in cloud fees. Every customer OpenAI acquires becomes, indirectly, an Azure customer.

Amazon's approach with Anthropic followed identical logic. The initial commitment of $4 billion, later extended toward $20 billion, came structured around AWS compute. Anthropic trains on Amazon's custom Trainium chips and runs inference on AWS. The strategic intent was the same: not merely to back a model company, but to anchor the most capable models to your own infrastructure.

Below the hyperscalers, and largely invisible to the public, sits a layer of capital that had already won the AI era before most retail investors knew it had started.

Anthropic is the clearest example. In early 2023, the company was valued at $4.1 billion. By mid-2026, when it filed its S-1 for a public offering, its private valuation had reached $965 billion — a 235-fold increase in approximately three years. The investors who captured that 235-fold increase were not retail investors. They were Google, Amazon, Spark Capital, and a small group of other institutional players.

Europe is absent from this empire in a way that should be alarming to European policy-makers. The EU has produced the GDPR and the AI Act — regulatory frameworks of considerable ambition. What it has not produced is a hyperscaler, a frontier chip company, or an AI lab of global significance. Europe regulated the empire. It did not build one. The consequences of that choice became unmistakably clear in June 2026, when a single American government decision cut off AI access for an entire continent overnight.

The geography of the AI empire deserves closer examination than it typically receives in discussions of AI geopolitics, because geography — in the most literal, physical sense — is one of the most important determinants of AI power.

Data centres are not evenly distributed across the United States, let alone across the world. They cluster in locations where the combination of cheap and abundant power, reliable fibre connectivity, favourable regulatory environments, and access to technical talent reaches a threshold that justifies the massive capital investment required to build them.

Northern Virginia — specifically the area around Loudoun County, known as Data Centre Alley — hosts more data centre capacity than any other geography on earth. The concentration there is the result of decades of infrastructure investment: the internet's backbone routing infrastructure runs through the region, power from the Mid-Atlantic grid is relatively cheap, the regulatory environment is cooperative, and the proximity to Washington, D.C., ensures a steady supply of technical talent from the federal government and the defence contracting complex.

This geographic concentration has consequences. When a single substation serving Data Centre Alley experienced a fault in 2024, it briefly took significant portions of multiple major internet services offline. When the regional power grid faces capacity constraints — as it increasingly does, given the pace of data centre construction — the constraints are felt by every AI workload running in the region simultaneously.

The second major data centre geography is the Phoenix, Arizona area, where cheap desert land, favourable state tax policies, and proximity to hydroelectric power from the Colorado River system have attracted hyperscaler investment. Google, Microsoft, and Amazon all have significant data centre presences in the region, alongside dozens of colocation operators.

Outside the United States, the significant data centre geographies are concentrated in Western Europe — Amsterdam, Frankfurt, Dublin — and in Asia, primarily Singapore, Tokyo, and increasingly in Malaysia and Indonesia, where lower land and power costs are attracting investment from hyperscalers seeking to expand their Asia-Pacific footprint without paying Singapore's premium prices.

The notable absence in this geography is Africa and most of Latin America, which together account for the majority of the world's population but a tiny fraction of the world's data centre capacity. The AI services consumed by African and Latin American users are almost entirely running in data centres on other continents, creating latency costs, data sovereignty challenges, and dependency structures that are not visible in the experience of using a chatbot but are structurally significant.

Singapore's position in this geography is worth examining in detail, because it illustrates what strategic positioning looks like for a small nation in the AI infrastructure era. Singapore cannot host the volume of data centres that its geography would imply — the island is small, power costs are high, and land is extraordinarily scarce. But Singapore has made itself indispensable to the regional AI infrastructure by other means: as a legal and financial hub for AI companies doing business in Southeast Asia, as a regulatory testbed for AI governance frameworks that other nations are watching, and as a source of technical talent that is disproportionate to its population.

The lesson Singapore offers to other small nations is not that you can build your way to AI sovereignty through data centre construction. It is that sovereignty in the AI era requires identifying the specific points of leverage available to your geography, your governance, your talent base, and your existing relationships — and doubling down on those points of leverage rather than trying to replicate capabilities that are already concentrated elsewhere.

The nations that will navigate the AI era most successfully will be the ones that understand their actual position in the power map — not the position they would like to have, but the one they actually have — and make deliberate, strategic decisions about how to improve it.

Understanding why the hyperscalers have been able to establish such dominant positions in the AI infrastructure market requires understanding the economics of platform businesses — specifically, the dynamics that create winner-take-most outcomes in markets with strong network effects and high switching costs.

A platform business is one that creates value by facilitating interactions between two or more distinct user groups. The internet's most valuable businesses are almost all platform businesses: Amazon connects buyers and sellers, Google connects searchers and advertisers, Facebook connects social networks and advertisers, Apple connects app developers and smartphone users. Each of these platforms becomes more valuable as more participants join, creating network effects that reinforce the incumbent's position and make entry difficult for new competitors.

The hyperscaler cloud platforms — AWS, Azure, Google Cloud — are platform businesses of a particular type: they connect AI developers and AI model users through a shared infrastructure layer. The more AI models run on a given cloud platform, the more AI-adjacent services (compute, storage, networking, monitoring, security) that platform can offer to attract more AI developers, which attracts more model users, which attracts more developers — a flywheel that, once spinning, is very difficult to stop.

This flywheel explains why the hyperscalers' investments in AI model companies like OpenAI (Microsoft/Azure) and Anthropic (Amazon/AWS) are not primarily about the financial returns from the equity stakes. They are about anchoring the most capable AI models to specific cloud platforms, which creates a gravitational pull that attracts the entire ecosystem of developers, enterprises, and governments that want to work with those models.

The switching costs within the hyperscaler ecosystem compound the flywheel advantage. An organisation that has built its AI workflows on AWS — using AWS compute instances to run Anthropic models, AWS storage to hold training data, AWS monitoring tools to track model performance, and AWS security tools to protect model outputs — has accumulated a complex web of integrations, optimisations, and institutional knowledge that makes migration to Azure or Google Cloud extraordinarily painful. The switching cost is not the cost of the new compute; it is the cost of rebuilding every integration, retraining every staff member, and accepting a period of reduced performance while the new environment is optimised.

These switching costs are well understood in enterprise software, and the hyperscalers have actively designed their AI services to maximise them. AWS's Bedrock service, which provides access to multiple AI models including Anthropic's Claude, is deeply integrated with AWS's other services in ways that make it easy to use if you are already on AWS and difficult to replicate if you are not. Azure's Copilot suite is integrated with Microsoft 365, Teams, and Azure DevOps in ways that make the entire Microsoft ecosystem stickier. Google Cloud's Vertex AI is integrated with Google Workspace and Google's advertising infrastructure in ways that create dependencies for the large base of organisations that already use Google's products.

The result is a market structure that is becoming more concentrated over time, not less: as AI adoption accelerates and organisations build deeper integrations with their chosen cloud platforms, the switching costs increase, the flywheel spins faster, and the incumbent advantages become more pronounced.

The policy implication of this market structure analysis is that regulatory interventions designed to promote competition in the AI market need to address switching costs and platform lock-in, not just market concentration in narrowly defined product categories. Requiring hyperscalers to provide data portability, standardised APIs, and transparent pricing for AI services would reduce switching costs and make the market more contestable. Whether any government has the political will to impose such requirements on the most valuable companies in the world remains to be seen.

Any serious analysis of the new AI empire must grapple with China's position in it — which is neither the simple subordination that American export controls were designed to produce nor the full-spectrum AI sovereignty that Chinese policy aspires to achieve.

China's AI situation in 2026 is best described as a strategic paradox: a country with the world's largest population of AI users, the world's second-largest AI research community, and the world's most ambitious AI policy programme, constrained by export controls that deny it access to the most advanced hardware and increasingly isolated from the Western AI ecosystem that it was previously deeply integrated with.

The export controls have had real effects. The most advanced training chips that Chinese labs can access are NVIDIA A100s — a generation behind the H100s and B100s that are the currency of frontier model training in the United States. The gap in raw compute capability between what the leading American labs can deploy and what Chinese labs can deploy is significant, and it is reflected in the capability differences between the leading American and Chinese models on the most demanding benchmarks.

The export control regime itself has proven far less stable than either side of the debate initially assumed. In December 2025, the Trump administration reversed years of policy by permitting case-by-case licensing for NVIDIA H200 and equivalent chips destined for China — a review standard replacing the blanket presumption of denial that had governed advanced chip exports since 2022. Within six months, the Bureau of Industry and Security was forced to issue new guidance closing a loophole that had allowed China-headquartered companies to receive restricted chips through subsidiaries operating outside China. The whiplash illustrates a structural reality that outlasts any single administration: chip export policy is being made in real time, under competing pressure from national security hawks, chip manufacturers seeking Chinese revenue, and a Chinese ecosystem that is innovating around every restriction faster than new ones can be written.

But the export controls have also had unintended effects that were predictable in historical perspective but were apparently not anticipated by the policymakers who designed them. The denial of access to advanced hardware forced Chinese researchers to develop techniques for achieving maximum performance from available hardware — a form of innovation under constraint that produced approaches to training efficiency, model architecture, and inference optimisation that the American labs, operating in an environment of abundant hardware, had not prioritised.

DeepSeek's R1 model, which demonstrated frontier-level reasoning capability with a training cost a fraction of what the leading American models required, was the most visible product of this innovation under constraint. It was not the only one. Across the Chinese AI research community, the hardware constraint produced a culture of efficiency engineering that may, in the long run, prove as strategically significant as the raw compute advantage that the export controls were designed to protect.

The geopolitical implications of this dynamic extend beyond the narrow domain of AI capability. They illuminate a pattern that appears throughout the history of technology competition between nation-states: technology constraints imposed by one party on another tend to reduce the constrained party's performance in the short term while stimulating the innovation that reduces the constraint's effectiveness in the long term. The Soviet Union's exclusion from Western semiconductor technology in the Cold War era produced a Soviet semiconductor industry that was inferior but that developed on its own trajectory, producing capabilities that were not simply copies of Western approaches. China's exclusion from the most advanced AI hardware is producing an analogous trajectory.

China's counter-architecture to the American AI empire is built on several pillars. Open source models — primarily from Alibaba's Qwen series and the independent DeepSeek team — provide capable foundation models that can be deployed on domestically produced infrastructure without dependence on American model companies. Domestic cloud platforms — Alibaba Cloud, Tencent Cloud, Baidu AI Cloud, Huawei Cloud — provide the infrastructure layer for AI deployment within China. Domestic chip companies — Huawei's Ascend series, Cambricon, Horizon Robotics — provide compute infrastructure that is less capable than NVIDIA's frontier offerings but is not subject to export control.

The counter-architecture is not yet a full alternative to the American AI empire. The capability gap is real and matters for the most demanding applications. But it is more substantial than it appears from the vantage point of the American AI community, which tends to evaluate Chinese AI capability through the lens of benchmark comparisons on Western metrics rather than through the lens of the applications and use cases that matter most in the Chinese market.

The strategic implication for the rest of the world is that the binary choice between American and Chinese AI dependence that might have seemed inevitable in 2023 is becoming a more nuanced landscape. The open source models from Chinese labs are available to any organisation anywhere in the world that chooses to use them. They are not subject to American export controls. They can run on domestically owned infrastructure. For many organisations outside the United States and China, the choice between American and Chinese AI capability is becoming a genuine choice rather than a forced acceptance of American dominance.

This is the geopolitical consequence of the AI power shift that no one planned for: a world in which the architectural choices of a startup in Beijing, made under the pressure of hardware constraints imposed by Washington, have created a third path for the rest of the world.

In the twelve months after this chapter's original argument, the "third path" stopped being theoretical and became institutional. Two competing coalitions now exist on paper, with treaty text, founding members, and a headquarters address each.

The American-led structure is Pax Silica, launched by the US State Department in December 2025 as a coordination framework — not a treaty body — for semiconductor supply chains, AI infrastructure, and critical minerals among trusted partners. Its founding members were Australia, Japan, Singapore, South Korea, Israel, and the UK, alongside the US. Qatar and the UAE joined in January 2026, India in February, and a further wave — Sweden, Finland, the Philippines, Norway — followed before a June 2026 summit in Washington that brought the total to 24 signatories, including a bloc of Latin American and Central Asian states (Argentina, Chile, Costa Rica, El Salvador, Kazakhstan, Panama) and, notably, the European Union itself. Taiwan has endorsed the declaration's principles through a parallel bilateral statement rather than formal membership. The EU's accession was contested internally: France argued the framework risked subordinating European "digital sovereignty" to American strategic priorities, and the bloc joined only after Brussels pushed for collective rather than member-state-by-member-state accession.

The Chinese-led counterpart is the World AI Cooperation Organization (WAICO), which moved from proposal to signed agreement in exactly one year — first floated by Premier Li Qiang at the 2025 World AI Conference, then formally established on 16 July 2026 in Shanghai, one day before Xi Jinping's keynote at the 2026 edition of the same conference and one day after Moonshot AI's Kimi K3 release. Twenty-nine countries signed as founding members, headquartered in Shanghai: Russia, Belarus, and Serbia in Europe; Kazakhstan, Kyrgyzstan, Tajikistan, and Uzbekistan in Central Asia; Oman in the Gulf; ten African states including South Africa, Kenya, and Ethiopia; and a Latin American bloc of Brazil, Cuba, Nicaragua, and Venezuela, alongside China itself, Indonesia, Malaysia, Pakistan, and several other Asian states. UN Secretary-General António Guterres attended the signing — a legitimising presence China's diplomacy visibly sought, given the organisation's stated design as a UN-Charter-aligned body rather than a rival to it.

Reading the two membership lists side by side is more instructive than reading either chapter's press release alone. WAICO's founding members are overwhelmingly Global South states with limited existing compute infrastructure and, in several cases (Russia, Belarus, Venezuela, Cuba, Iran-adjacent Central Asian states), existing tension with Washington that made Pax Silica membership a non-starter regardless of AI considerations. Pax Silica's signatories are overwhelmingly wealthy or chip-adjacent economies — fabrication hosts (Japan, South Korea, the Netherlands via ASML), capital-rich buyers (the Gulf states, Singapore), and treaty allies (Australia, the UK, the EU). The two coalitions are not really competing for the same members; they are sorting the world by a mix of existing alliance structure and compute wealth that predates either initiative.

The one country that punctures this tidy sorting is Kazakhstan, which signed both the Pax Silica declaration and the WAICO founding agreement within the same six-month window — a hedge, not a contradiction, and arguably the most honest position available to a mid-sized state that borders both a Pax Silica member (China, ironically, is not itself in Pax Silica, but Kazakhstan sits in Beijing's immediate sphere) and needs Western capital and technology relationships it cannot afford to forfeit. India offers a softer version of the same hedge: it joined Pax Silica in February 2026 and has leaned into the initiative's semiconductor-manufacturing incentives, while its foreign policy establishment has a decades-long institutional preference for non-alignment that predates the AI era by seventy years and is unlikely to disappear because of it. Gulf states present a third variant — the UAE and Qatar joined Pax Silica specifically to secure advanced chip access (see the export-control easing for the UAE discussed elsewhere in this book's Signal Watch), while continuing to run AI infrastructure and cloud partnerships with Chinese vendors that predate Pax Silica and were not wound down to join it.

The practical upshot for the book's central argument is a refinement, not a reversal. Chapter 2's original claim — that the rest of the world was moving from a forced binary toward a genuine choice between American and Chinese AI dependence — was correct in direction but understated the speed. What has emerged by mid-2026 is not a single global choice point but a spectrum of overlapping memberships, in which most countries outside the two great powers are hedging rather than choosing, and the mid-sized states doing the most careful hedging (Kazakhstan, India, the Gulf states) are likely to extract the best terms from both sides precisely because neither coalition can afford to lose them entirely.

The distillation dispute referenced above is no longer just a war of words. In the week of 13 July 2026, reporting indicated the White House and Congress were weighing concrete countermeasures against "adversarial distillation" — the practice of mining a rival's model with mass queries to train a cheaper competitor — aimed specifically at the Chinese labs Anthropic had already accused of the practice. As of this writing that reporting comes from a single industry newsletter rather than an official announcement, so it should be read as a signal to watch rather than settled policy; but if it materializes, it would mark the first time the distillation argument moves from a public accusation into an actual regulatory or legislative response, adding a technical-enforcement layer to the alliance-and-coalition dynamic described above.

Editor's note: this section was added 18 July 2026, drawing on primary sources (US State Department releases, Xinhua/CGTN/Global Times coverage of the WAICO signing, and independent reporting on the Pax Silica summits). Chinese-language translation for this section is pending — it should not be treated as complete for bilingual publication until that pass is done.

Models Are Not Enough

There is a persistent fantasy in the AI industry. It goes like this: a brilliant team of researchers, working in a loft somewhere, trains a model so capable that it disrupts everything. The model is so good that the incumbents cannot catch up. The team raises a little money, builds a product, and within a few years they are worth more than the companies they disrupted. Intelligence wins. Infrastructure is just a detail.

This fantasy has produced some of the most expensive failures in technology history. It will produce more.

The reality is starker and more structural. Intelligence without infrastructure is just software waiting for a server. A model, however capable, cannot train itself, run itself, cool itself, power itself, or fund itself. It depends — completely and continuously — on a stack of physical and financial infrastructure that the model company almost certainly does not own. And the owners of that infrastructure have structured their relationships with model companies very carefully.

Every frontier AI model company faces five structural dependencies that it cannot eliminate and can only partially mitigate.

The first dependency is compute. Training a frontier model requires tens of thousands of the most advanced AI accelerators — primarily NVIDIA H100s and their successors — running in synchrony for months. The cost of a single training run for a frontier model has been estimated at between $50 million and $500 million, depending on the model size and the efficiency of the training process. These chips are scarce, expensive, and controlled by a single dominant supplier. Model companies do not own their compute at scale. They rent it, from hyperscalers who own the chips, or they buy chips that sit on infrastructure they do not control.

The second dependency is energy. AI training and inference are extraordinarily energy-intensive. A single training run for a large frontier model can consume as much electricity as a small town uses in a year. The International Energy Agency reported in April 2026 that AI-focused data centres had seen electricity demand surge 50 percent in 2025 alone, and projected that data centre electricity consumption would reach 950 terawatt-hours by 2030 — equivalent to the entire national electricity consumption of Japan today. Model companies do not own power plants.

The third dependency is data. The performance of a model is a function of the quality, quantity, and diversity of the data it trained on. The first generation of frontier models was trained on the open internet — a commons that is now largely exhausted as a source of novel training data. The next generation requires proprietary data: clinical records, legal documents, financial transactions, scientific literature, human feedback at scale. The companies that own proprietary data at scale are not AI labs. They are hospitals, banks, law firms, publishers, and governments.

The fourth dependency is capital. The cost of maintaining a position at the frontier of AI capability is not a one-time investment. It is an ongoing commitment that grows with each generation of models. OpenAI's reported losses in 2024 were approximately $5 billion on revenues of $3.7 billion. Anthropic has been explicit that its capital requirements are measured in the tens of billions. The model companies that are still in the race in 2026 are there because they found hyperscalers, sovereign wealth funds, or other deep-pocketed investors willing to fund the gap between revenue and cost.

The fifth dependency is talent. Training and maintaining frontier models requires a specialised workforce that is simultaneously small in number and extraordinarily expensive. The researchers who can advance the state of the art in large language models can be counted in the hundreds globally. They move between a small number of organisations — largely in the United States and United Kingdom — and their compensation packages routinely reach seven or eight figures.

There is a dependency that deserves special attention because it is less discussed and more immediately consequential than the training dependencies: the inference problem.

Training a model is expensive, but it happens once per model version. Inference — running the model to answer a query, generate text, write code, analyse an image — happens billions of times per day across a deployed model's user base. And inference, at scale, is brutally expensive relative to the revenue it generates.

The unit economics of AI inference have been the hidden pressure behind every AI company's business model since the first consumer products launched. When OpenAI launched ChatGPT, the cost of a single conversation was estimated at several cents — multiples higher than the cost of a Google search. At consumer subscription prices of $20 per month, heavy users were being served at a loss on every interaction.

The dependency picture would be entirely bleak for model companies if not for one significant counterweight: the rise of open source AI.

Meta's decision to release its Llama series of models as open source was one of the most consequential strategic moves of the AI era. It was motivated partly by ideology — Meta's leadership genuinely believes that open AI is safer and better for the world — and partly by competitive logic. Meta is not primarily an AI company. It is an advertising company that uses AI. Its competitive moat is not its model; it is its data and its distribution. By releasing powerful open models, Meta imposes costs on its closed-model competitors while paying relatively little itself.

The effect has been dramatic. By 2026, open source models had closed much of the capability gap with the leading closed models for many common tasks. A developer building a coding assistant, a document summariser, or a customer service chatbot could choose from dozens of open models that ran on commodity hardware, required no API fees, and could be fine-tuned on proprietary data without sharing that data with any third party.

DeepSeek's R1 model, released in early 2025, demonstrated that models of remarkable capability could be trained on restricted hardware using novel efficiency techniques — a development that simultaneously embarrassed the American chip export control architecture and proved that the intelligence layer was more contestable than many had assumed.

The history of transformational technologies offers a consistent pattern that the AI era is following with remarkable fidelity: the initial winners of a technology transition are rarely the long-term winners. The infrastructure that enables the technology proves more durable than the applications that run on it.

The history of the telephone industry illustrates this vividly. In the 1870s and 1880s, dozens of telephone companies competed fiercely for customers in major American cities. The technology was real and valuable; the market was clearly large; the competitive dynamics were intense. AT&T; emerged as the dominant player through a combination of patent control, infrastructure investment, and regulatory capture — but its dominance was built not on the quality of its telephone handsets or the cleverness of its calling features. It was built on its ownership of the copper wire that connected the handsets to each other.

The wire outlasted the telephones. The wire outlasted AT&T; itself, in its original form — the company was broken up in 1984, but the infrastructure it had built continued to function and generate value for decades under new ownership. The physical infrastructure of telecommunications proved more durable than any of the companies or products that depended on it.

The internet followed a similar pattern. The dot-com era produced thousands of companies building applications on top of the internet's infrastructure. Most of them failed spectacularly. The companies that survived and thrived were those that either built durable infrastructure positions — Cisco, whose routers moved the internet's packets; Akamai, whose content delivery network solved the latency problem at scale; Verisign, whose domain name system was so embedded in the internet's architecture that it could not be displaced — or those that built network effects so powerful that they effectively became infrastructure themselves: Google for search, Amazon for commerce, Facebook for social connection.

In each case, the application layer was intensely competitive and most competitors were eliminated. The infrastructure layer was concentrated and the survivors were extraordinarily durable.

The AI era is following this pattern. The application layer is already intensely competitive and becoming more so: dozens of AI-powered products compete in every category, prices are falling, differentiation is difficult, and the half-life of any competitive advantage based on application features is shrinking. The infrastructure layer is concentrated in ways that are becoming more concentrated rather than less, as the capital requirements for frontier model training and large-scale inference increase with each generation of technology.

This structural pattern has direct implications for how individuals and organisations should think about AI investment — both financial investment and strategic investment of time and resources. The application layer is where most of the attention and most of the early investment is focused. The infrastructure layer is where the durable value will accumulate.

The telephone analogy also illuminates the risk of the model quality competition. AT&T; competed fiercely on call quality, on service reliability, on the range of calling features it could offer. These competitive dimensions were real and mattered to customers. But the competition was ultimately won and held by the company that owned the most wire, not the company that had the best telephone. The quality of the telephone mattered at the margin; the ownership of the wire mattered structurally.

In the AI era, model quality matters. It matters enormously for which specific use cases AI can enable, and for which organisations can capture which specific revenue streams. But the competition over model quality, like the competition over telephone quality, is ultimately marginal relative to the structural advantage of owning the infrastructure on which the models run.

The wire is the chip. The wire is the data centre. The wire is the energy supply. The wire is the cloud platform. The organisations that own these things will collect rent from the AI economy for decades, regardless of which specific model wins any given capability benchmark.

Understanding this is the beginning of understanding where durable value lies in the AI economy — and where it does not.

Of all the dependencies that frontier AI model companies face, the talent dependency is perhaps the least discussed and the most structurally significant in the long run. The reason it is discussed less is that it is harder to quantify than compute costs or energy consumption — you cannot put a dollar figure on a researcher's accumulated intuition, or measure in kilowatt-hours the institutional knowledge that makes a team able to train models that other teams cannot.

But the talent dependency is real, and it operates in ways that compound the other dependencies rather than substituting for them. An organisation with extraordinary compute but limited talent will waste its compute on approaches that a more talented team would have identified as dead ends before committing resources to them. An organisation with limited compute but extraordinary talent will find ways to achieve results that are disproportionate to its resource base — as DeepSeek's Chinese team demonstrated dramatically when they produced frontier-level results despite operating under chip export restrictions.

The concentration of AI research talent is extraordinary. The people who can meaningfully advance the state of the art in large language models, multimodal systems, and reinforcement learning number perhaps in the low thousands globally. They are concentrated in a small number of cities — primarily San Francisco, London, and increasingly a few academic centres like Montreal and Singapore — and in a small number of organisations: OpenAI, Anthropic, Google DeepMind, Meta AI, and a handful of academic institutions and research labs.

This concentration is not accidental. It reflects decades of investment in specific research communities, the development of shared knowledge and vocabulary among researchers who trained at the same institutions and published in the same venues, and the network effects that make the best researchers want to work with other best researchers. The concentration is self-reinforcing: being at the frontier attracts the talent that keeps you at the frontier.

The compensation required to attract and retain frontier AI researchers reflects their scarcity. Senior researchers at leading AI labs routinely command total compensation packages in the range of $1 million to $5 million per year. The most sought-after researchers — those who have made fundamental contributions to the field and whose expertise is irreplaceable — can negotiate for equity positions that, at the valuations that AI companies are achieving, are worth tens or hundreds of millions of dollars.

These compensation levels create a dynamic that is unusual even by the standards of the technology industry. They effectively limit the frontier AI research community to a small number of very well-capitalised organisations: the hyperscalers, the well-funded AI labs, and a small number of national programmes with government backing. A university, a hospital, a government agency, or a smaller technology company cannot compete for frontier AI research talent at market rates. They can attract and develop talent that flows to the frontier organisations once it has reached a certain level; they cannot retain it.

The talent bottleneck has strategic implications that extend beyond the organisational. It means that the pace of AI progress is constrained not only by compute and energy — the material inputs that receive most of the attention — but by the cognitive input of a small number of extraordinary individuals whose insights unlock capabilities that would not otherwise emerge. This in turn means that the organisations that can attract and retain these individuals gain an advantage that cannot be replicated simply by deploying more capital. And it means that policies designed to develop AI capability in new geographies must address the talent question — which is ultimately an education and immigration question — as directly as they address the infrastructure question.

The nations and organisations that understand this will invest in developing AI research talent domestically and attracting it internationally, not just in building compute capacity. The compute is necessary but not sufficient. The talent is both necessary and, in the long run, the binding constraint.

Understanding what models cannot do is as important as understanding what they can do, and the failure modes of large language models are as instructive as their capabilities.

The most fundamental failure mode of current large language models is hallucination: the tendency of the model to produce fluent, confident-sounding text that is factually incorrect, unsupported by evidence, or internally inconsistent. Hallucination is not a bug that can be fixed by making the model bigger or training it on more data. It is a consequence of the fundamental architecture of autoregressive language models, which predict the next token in a sequence based on the tokens that precede it, without any intrinsic mechanism for verifying whether the predicted tokens correspond to facts in the world.

The significance of hallucination for the AI power shift is that it creates a continuing need for human oversight that limits the degree to which AI systems can operate autonomously. Every AI-generated document, analysis, or decision that will be acted upon needs to be reviewed by a human who has the expertise to identify hallucinations and correct them before they produce harmful consequences. The value of this human oversight role grows in proportion to the stakes of the decision and the cost of the errors that hallucination might produce.

This creates a paradox for the labour market impact of AI: the tasks where AI is most capable — producing fluent text at high volume, quickly and cheaply — are precisely the tasks where the risk of hallucination means that human oversight remains essential. The cost savings from AI are real; the human cost of providing oversight is also real; and the net impact on employment is less dramatic than either the enthusiasts or the alarmists claim.

A second major failure mode is context sensitivity: the tendency of large language models to produce different outputs in response to superficially similar inputs, in ways that are not always predictable. A model that produces accurate legal analysis when given a carefully constructed prompt may produce significantly less accurate analysis when given the same information in a different order, or when the prompt includes irrelevant context that activates different training patterns. This sensitivity to context makes it difficult to evaluate AI systems in standardised test environments and then trust that those evaluations will predict real-world performance.

The context sensitivity problem is compounded by the opacity of the models' internal processes. The decisions that a large language model makes in producing its outputs are encoded in billions of parameters that interact in ways that no human can track or fully understand. When a model produces a wrong answer, it is often impossible to trace exactly why — which training examples activated, which patterns were reinforced, which patterns were suppressed. The opacity makes it difficult to reliably correct specific failure modes without inadvertently creating new ones.

Third is the problem of knowledge currency: all large language models have training cutoffs, after which they have no knowledge of events that occurred. A model trained through 2024 does not know what happened in 2025 and 2026. In domains where recent information is critical — financial markets, medical research, geopolitical developments, technology developments — this knowledge currency problem is significant. Retrieval augmented generation, which supplements the model's base knowledge with information retrieved from current sources, mitigates but does not eliminate this problem.

These failure modes — hallucination, context sensitivity, opacity, knowledge currency — are not reasons to avoid AI adoption. They are reasons to adopt AI with clear eyes about what AI can and cannot do, and to design workflows that put the model's capabilities to work on the tasks where the failure modes are least consequential or most manageable.

The organisations that have integrated AI most successfully are those that have developed explicit frameworks for understanding which tasks are AI-safe — where the failure modes are tolerable, the output can be easily verified, and the cost of errors is low — and which tasks require more careful AI-human collaboration, where the failure modes are potentially significant and the output requires expert review before it is acted upon.

Building this judgment — the capacity to distinguish AI-safe from AI-risky tasks, to design workflows that play to AI's strengths while managing its weaknesses, and to maintain the human expertise required to provide effective oversight — is one of the most important organisational capabilities of the AI era. It cannot be delegated to the AI itself.

The Hidden Toll Roads

In the California gold rush of 1849, most miners did not get rich. The prospectors who flooded into the Sierra Nevada foothills in search of gold found an intensely competitive market for the gold itself — prices fell as supply increased, and the physical labour of extraction was brutal. The people who got rich were the ones who sold the miners what they needed to survive and to work: shovels, jeans, food, lodging, and the financing that kept them in the field long enough to either strike it lucky or go broke.

Levi Strauss built a denim empire. Wells Fargo built a banking empire. The Sacramento merchants who supplied the camps built commercial empires that outlasted the gold rush by generations.

The AI era is following the same pattern — except the toll roads are harder to see, the tolls are being collected at every layer of the stack simultaneously, and the scale of the extraction dwarfs anything the gold rush could have imagined.

NVIDIA is the most obvious toll road in the AI economy, and it has been hiding in plain sight for years. The company's position is structurally unusual in modern capitalism: it is both the dominant supplier of the most critical input in the AI economy and a company with essentially no serious competitor at the frontier of its product category.

Every training run, every inference call, every AI product in the world pays NVIDIA — either directly through chip purchases or indirectly through cloud providers who own NVIDIA chips. The company does not need to pick a winner in the AI model race. Every competitor in that race is its customer. OpenAI buys NVIDIA chips. Anthropic buys NVIDIA chips through its AWS arrangement. Google DeepMind trains on NVIDIA chips alongside its own TPUs. Meta's Llama models were trained on NVIDIA chips.

TSMC, which manufactures the chips NVIDIA designs, is the toll road beneath the toll road. No TSMC, no advanced AI chips. No advanced AI chips, no frontier models. The concentration of advanced semiconductor manufacturing in Taiwan is the single greatest structural vulnerability in the global AI economy — a geopolitical risk that every serious investor and policy-maker should have at the top of their risk register.

In August 2026, this toll-road position moved from metaphor to explicit financial architecture. NVIDIA signed memorandums of understanding with six of the largest alternative asset managers and banks — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — to mobilise more than $500 billion in third-party capital, explicitly structured to treat compute infrastructure the way markets treat toll roads, airports, and other assets investors borrow against. Jensen Huang's own words captured the shift: compute was no longer simply hardware NVIDIA sold; it had become "a new class of productive, investable infrastructure. In AI, compute is revenue." The market's reaction was telling. NVIDIA's shares fell as much as 3.2 percent on the announcement, as investors weighed whether pulling in more debt-financed capital solved the AI capex problem or simply relocated the risk onto new balance sheets.

AMD is playing the same toll-road game through a different mechanism. Rather than assembling outside capital pools, AMD has taken equity stakes and issued stock warrants directly in the customers buying its chips — up to $5 billion into Anthropic, a warrant for up to 10 percent of AMD's own equity to Meta, and a similar option structure with OpenAI. Where NVIDIA socialises financing risk onto Wall Street asset managers, AMD internalises it by tying its own balance sheet to its customers' success. Analysts increasingly describe this circular financing — chip suppliers investing in the companies that buy their chips — as structural to AI infrastructure finance rather than an exception to it. It is a faster way to buy market share against NVIDIA's incumbency, and a riskier one: AMD's returns now depend not only on chip demand but on the solvency of the customers it has taken a stake in.

The CNBC interview that unveiled the MOU immediately drew comparisons to a less flattering precedent: the securitization of subprime mortgages before the 2007–2009 financial crisis, where bundling and reselling risk to investors preceded a systemic unwind. Skeptics point to one load-bearing assumption in Huang's pitch — that NVIDIA's GPUs hold their value over time, their productive life extended by continuous software improvements through CUDA, rather than depreciating like conventional hardware. Early evidence points modestly in Huang's favor: rental rates for H100 chips rose from roughly $1.70 per GPU-hour in late 2025 to about $2.35 per GPU-hour by mid-2026. But the same skeptics note that this evidence spans barely a year, covers a single hardware generation, and says little about what happens to residual values once next-generation chips arrive at scale. Whether compute behaves like a toll road or like a subprime security may be the defining financial question of the AI buildout's next phase.

The energy toll road is the least glamorous and potentially the most durable position in the entire AI investment landscape. Every chip that runs, every model that trains, every inference that executes consumes electricity. The AI economy's appetite for power is growing faster than the grid was designed to supply it.

In northern Virginia — the data centre capital of the world, home to more data centre capacity than the rest of the world combined — capacity auction prices spiked 833 percent in a single year as hyperscalers competed for a finite supply of grid-connected power. Nuclear power plant operators, natural gas producers, and transmission infrastructure owners found themselves in conversations with technology companies that would have seemed surreal five years earlier.

Microsoft signed a deal to restart the Three Mile Island nuclear plant — yes, the site of America's most famous nuclear accident — to power its data centres. Amazon acquired a data centre campus co-located with a nuclear plant in Pennsylvania. Google signed long-term power purchase agreements with small modular reactor developers that had not yet built a single commercial reactor.

The energy constraint has hardened faster than most 2024-era forecasts anticipated. Industry analysis in 2026 put US data centre power demand on a path to reach 35 to 45 gigawatts by 2030 — roughly double 2024 levels — while grid interconnection queues, once measured in months, now routinely stretch beyond four years. As much as half of the data centre capacity scheduled to come online in 2026 risked delay from permitting bottlenecks, grid connection backlogs, and local community opposition. The toll road, in other words, is not merely owned by a small number of players. It is running out of room, and the owners of what remains are collecting a scarcity premium that grows with every delayed connection.

The companies that own the energy infrastructure — the utilities, the nuclear operators, the natural gas producers, the transmission line owners, the data centre REITs — are collecting rent from the AI economy regardless of which model wins the capability race. Their position is structurally more defensible than any model company's, because their asset is physical and their scarcity is geographic.

Amazon Web Services, Microsoft Azure, and Google Cloud collectively function as the roads on which the AI economy runs. Every API call, every training job, every deployed model, every AI-powered application routes through one of these three platforms. They charge for compute, storage, networking, and an expanding suite of AI-specific services. As AI workloads grow, so does the toll.

The strategic insight for investors is counterintuitive: the model wars create more cloud spending, not less. Every dollar that OpenAI, Anthropic, and Google DeepMind spend competing against each other flows partly to the hyperscalers. Every enterprise customer who adopts AI tools increases their cloud footprint. The model race is a revenue engine for the infrastructure layer.

Data centres require land — and the right land, near power, near fibre, in jurisdictions with stable regulation, abundant water for cooling, and favourable tax treatment, is scarcer than it appears. The hyperscalers and the data centre operators who serve them have been acquiring land in key locations for years, creating a real estate moat that newer entrants cannot easily replicate.

Data centre REITs — publicly traded companies that own and lease data centre real estate — are among the most direct ways for public market investors to capture the AI infrastructure build-out. They collect rent from whoever needs rack space, regardless of which AI company is the tenant. The lease terms are long, the demand is structural, and the supply in the right locations is constrained.

You do not need to know which AI model wins. You need to own the land the winner's data centre sits on. That position will still be paying rent long after the current generation of model companies has been disrupted by the next.

The pattern the toll-road framework predicts is now visible in the asset allocation decisions of the world's largest pools of capital. A 2026 Invesco survey of ninety sovereign wealth funds managing a combined $17.2 trillion found a net 17 percent planning to reduce their holdings of listed equities, with between 28 and 35 percent of respondents increasing allocations to private infrastructure, private credit, and private equity instead — average infrastructure allocations nearly doubling from roughly 5 percent in 2022 to 9 percent in 2025. The logic was made concrete in mid-2026 when MGX, BlackRock, and the AI Infrastructure Partnership completed a $40 billion acquisition of Aligned Data Centers — a transaction that had nothing to do with picking a winning model and everything to do with owning the racks, the leases, and the power contracts that every model, regardless of which one wins, will need.

The concept of the toll road is ancient, and its economics have remained remarkably stable across millennia of technological change. The toll road owner provides infrastructure that others need to use, and charges for that use. The more indispensable the infrastructure, the more the owner can charge. The more concentrated the infrastructure ownership, the more durable the position.

Roman roads were toll roads. Medieval bridges were toll roads. Nineteenth-century railways were toll roads. Twentieth-century telecommunications networks were toll roads. The economics in each case were similar: high upfront capital costs, low ongoing marginal costs, strong network effects that made the infrastructure more valuable as more users connected to it, and natural monopoly tendencies that concentrated ownership over time.

The AI era's toll roads share these characteristics, but they are different in one important respect: the pace at which they are being built, and the scale of capital being deployed to build them, is historically unprecedented. The $725 billion in combined hyperscaler capital expenditure in 2026 alone exceeds the total capital invested in the internet's physical infrastructure during the entire period of the internet's buildout from 1995 to 2005.

This pace of investment has consequences for the competitive dynamics of the toll road market. When infrastructure is being built this quickly, with this much capital, the early movers can establish positions that are difficult to challenge not just because they got there first but because the capital required to replicate their infrastructure has already been committed by others. The data centre market of 2026 looks very different from the data centre market of 2020: the prime locations in Data Centre Alley are occupied, the power connections are committed, the long-term leases are signed. A new entrant seeking to build equivalent infrastructure faces not just the cost of building — it faces the cost of building in secondary locations with less favourable characteristics.

The energy situation is particularly illustrative of how infrastructure scarcity creates toll road economics. In 2020, a data centre operator in Virginia could connect to the grid relatively easily and at reasonable cost. By 2025, the grid was effectively full in the most desirable locations: utilities were telling hyperscalers that new connections would take five to seven years and require building new transmission infrastructure at the data centre operator's expense. The price to connect to existing grid capacity — where it was available at all — had spiked to levels that would have seemed unimaginable five years earlier.

This scarcity has created a class of infrastructure owners whose assets have appreciated dramatically simply by virtue of being built before the constraint became binding. Data centre REITs that built facilities in Data Centre Alley in 2015 or 2020 are collecting rents from tenants who could not replicate the infrastructure at any price in 2026. The toll road has become, in the most literal sense, a moat.

The implications for investors are direct and important. The most durable positions in the AI economy are in the physical and financial infrastructure: the data centre real estate, the energy supply contracts, the transmission infrastructure, the chip fabrication capacity, the fibre connectivity. These positions are not glamorous. They do not generate the kind of headline-grabbing capability announcements that drive retail investor attention. But they are structurally more defensible than the model positions, and they benefit from the competition in the model layer rather than being threatened by it.

Every new entrant in the AI model market is a new customer for the infrastructure toll roads. Every dollar spent competing in the application layer flows partly to the infrastructure owners as rent. The more intense the competition, the more valuable the toll booth.

This is the deepest insight of the picks-and-shovels framework: the toll road owners are not competing in the AI arms race. They are providing the armaments to everyone who is. And their business model improves as the arms race intensifies.

Every toll road described so far in this chapter is being financed by private capital chasing a return. China is building the same category of infrastructure through an entirely different ownership model — one that changes who collects the toll and why. In June 2026, Chinese authorities began drafting a blueprint for a nationwide network of interconnected computing hubs, backed by a planned 2 trillion yuan (roughly $295 billion) in spending over five years. The financing was never going to come primarily from private markets: it runs through sovereign debt, including ultra-long special government bonds, state investment funds for strategic industries, and bank lending, with private capital playing a supplementary role.

The operators reflect the same logic. State-owned carriers China Mobile and China Telecom were tasked with running the bulk of the data centres, and the compute inside them is intended to run overwhelmingly on domestic silicon — chiefly Huawei's Ascend line — rather than NVIDIA hardware. This is the toll road as state asset rather than investable asset: Beijing is not trying to attract yield-seeking capital to a scarce resource. It is trying to guarantee that the resource exists at all inside its own borders, insulated from further US export controls.

The chip layer underneath tells the same story. With NVIDIA's most advanced accelerators blocked from the Chinese market, Huawei has built its Ascend programme into the mandatory domestic alternative, targeting roughly 600,000 Ascend 910C units in 2026 — about double the prior year — manufactured domestically with SMIC after Huawei's earlier reliance on a TSMC "die bank" was exhausted. Performance still lags: the Ascend 950 line is estimated at a fraction of the throughput of NVIDIA's most advanced chips, and software portability from CUDA remains a real friction. But the customer list — Alibaba, Tencent, ByteDance, DeepSeek — is not optional. It is the domestic market being redirected, by policy, onto the sovereign toll road.

The comparison this chapter has been building toward is now explicit. The American toll road is owned by private capital and priced by markets — Wall Street decides where compute investment flows, and it prices in risk, return, and the possibility of a bust. The Chinese toll road is owned by the state and priced by policy — Beijing decides where compute is built, largely independent of whether the return justifies the capital on commercial terms. Both are racing to the same destination. They are financing the trip in fundamentally incompatible ways, and that divergence, not the chip gap alone, is what will determine which compute ecosystem the rest of the world ends up building on.

Among the most striking features of the AI infrastructure gold rush is the range of industries that are benefiting from the AI build-out without being primarily AI businesses themselves. Understanding these hidden beneficiaries is important for investors seeking exposure to the AI infrastructure theme and for strategists trying to understand the full scope of the economic transformation underway.

The most obvious hidden beneficiary is the electric utility sector. Utilities are businesses that build and operate infrastructure for the delivery of electricity. They are capital-intensive, regulation-heavy, and slow-moving by the standards of the technology industry. They are also, as of 2024, among the most strategically important businesses in the AI economy.

The AI build-out has transformed the demand outlook for electric utilities in ways that are only beginning to be reflected in utility valuations. The data centre build-out is adding enormous new loads to grids that were designed for gradual, predictable growth. Utilities that own generation capacity in markets near major data centre clusters are seeing demand for power that exceeds their previous growth forecasts by multiples, not percentages.

The nuclear power sector is experiencing a renaissance driven almost entirely by AI data centre demand. Nuclear plants that were scheduled for decommissioning have been given new life by data centre operators willing to sign long-term power purchase agreements at prices that make the economics of continued operation viable. New nuclear capacity — both large conventional plants and the small modular reactors being developed by a generation of startup companies — is attracting investment from AI companies that are committing to offtake agreements before the plants are even built.

The transmission and distribution infrastructure sector is similarly transformed. The bottleneck for data centre development in many markets is not the availability of land or even the availability of generation capacity. It is the availability of transmission capacity to move power from generators to data centre sites. Utilities and independent transmission operators that own the wires connecting generation to load are in an extraordinarily strong bargaining position vis-à-vis data centre developers who need their infrastructure.

Beyond utilities, the construction and engineering sector is experiencing an AI-driven boom. Building a large data centre is an extraordinarily complex construction project: the civil engineering, the electrical work, the cooling systems, the security infrastructure, the high-voltage power distribution, the fibre connectivity. The companies that can execute these projects at scale and speed are in high demand from hyperscalers that are trying to build data centre capacity faster than their competitors.

The real estate sector benefits through data centre REITs and through the general demand for industrial property in markets near power and fibre infrastructure. The water sector benefits through the cooling water demand of large data centres. The telecommunications sector benefits through the fibre connectivity demand as data centres require high-capacity connections to the internet backbone.

The breadth of the AI infrastructure impact is one of the reasons that the 'picks and shovels' analogy, while useful, understates the scope of the economic transformation. The AI build-out is not merely a technology story. It is an infrastructure story that touches every sector of the economy that provides the physical substrate on which digital infrastructure is built.

Investors who understand this — who see the AI power shift not as a story about which AI model will win but as a story about which physical infrastructure sectors will benefit from the build-out — are positioned to capture returns from the AI era that are not available through the more obvious routes of investing directly in AI model companies or AI application developers.

The picks and shovels of the AI era are made of steel, concrete, copper, silicon, and uranium. The companies that produce and deploy these materials are the hidden beneficiaries of the most dramatic infrastructure investment cycle in a generation.

The physical geography of the AI infrastructure build-out is one of the least discussed and most consequential aspects of the AI power shift. Where data centres are built, where chips are manufactured, where cables are laid — these decisions shape the distribution of AI infrastructure capacity for decades and create dependencies that are as real as any contractual commitment.

The dominant geography of the current AI build-out is American, but within the United States the concentration is striking. Northern Virginia hosts more data centre capacity than any other geography on earth. The Phoenix, Arizona area is the fastest-growing data centre market in the United States. Texas, particularly the Dallas-Fort Worth area, is attracting data centre investment from hyperscalers seeking geographic redundancy and proximity to the large Texas energy market.

Each of these geographies has specific characteristics that make it attractive for data centre development. Virginia's attraction is the combination of the internet backbone routing infrastructure that runs through the region, historically cheap power from the mid-Atlantic grid, a cooperative regulatory environment, and the proximity to federal government and defence contracting customers who are among the largest consumers of AI compute.

Phoenix's attraction is cheap desert land, sunshine that supports solar power development, Arizona state policies that are favourable to data centre investment, and proximity to the Colorado River hydroelectric system. The irony of building data centres in a desert is that cooling them requires enormous quantities of water in a region where water is already scarce — a constraint that will become more binding as climate change reduces Colorado River flows and as data centre cooling demands increase.

The concentration of AI infrastructure in a small number of American geographies creates resilience risks that the industry has not fully reckoned with. A significant earthquake affecting the Pacific Northwest — which hosts major Microsoft and Amazon data centre campuses — could affect AI service availability for millions of users simultaneously. A severe drought affecting Arizona could create cooling water constraints that force data centres to reduce operations. A major grid event in Virginia could affect a disproportionate share of global internet traffic.

These risks are not hypothetical. They are known, they are documented in risk registers, and they are being addressed through geographic diversification — the hyperscalers are deliberately building in multiple regions to reduce single-point-of-failure risk. But the diversification is happening more slowly than the concentration, in part because the concentration is driven by real economic advantages (cheap power, existing infrastructure, available talent) that are not easily replicated in more geographically distributed locations.

Outside the United States, the geography of AI infrastructure is shaped by different constraints. In Europe, the primary data centre geographies are Amsterdam, Frankfurt, Dublin, and Stockholm — markets that offer the combination of renewable energy availability, favourable regulatory environments, and existing connectivity infrastructure that data centre operators require. The EU's General Data Protection Regulation and its successor frameworks create data localisation requirements for some categories of data that drive investment in European data centre capacity.

In Asia, Singapore has established itself as the regional hub for AI infrastructure serving Southeast Asia, Australia, and to some extent India. Singapore's advantages — political stability, excellent connectivity, English-speaking technical workforce, proximity to major submarine cable landing stations — make it a natural anchor for regional AI infrastructure. Its disadvantages — high land costs, limited power capacity, small domestic market — mean that it cannot absorb all of the regional demand for AI compute and is increasingly complemented by data centre development in Malaysia and Indonesia, where lower costs are attracting investment for less latency-sensitive workloads.

The geography of the AI build-out will continue to evolve as the economics of energy, land, and connectivity change. Nuclear power development will attract data centre investment to locations near new nuclear plants that can deliver reliable, carbon-free power. Cable station development will attract data centre investment to coastal locations that offer direct connectivity to submarine cable systems. Regulatory changes will attract or repel data centre investment depending on their implications for operational costs and legal risks.

For investors, the geography of the AI build-out provides a framework for identifying specific opportunities: the real estate markets adjacent to major data centre hubs will appreciate as data centre development drives demand for industrial property. The utilities serving data centre-heavy regions will benefit from the demand growth. The construction and engineering companies building the data centres will see sustained demand for their services. The connectivity providers linking data centres to each other and to the internet backbone will see increasing traffic and increasing revenue.

The map of the AI power shift has a physical dimension that is often overlooked in discussions that focus on software, models, and algorithms. But the chips are physical. The data centres are physical. The cables are physical. The power plants are physical. Understanding where these physical assets are and who owns them is as important as understanding which companies are developing the most capable models.

The Kill Switch: Who Controls Your AI?

On June 12, 2026, the Trump administration did something no government had ever done before. Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei ordering that two of the company's most capable models — Fable 5 and Mythos 5 — be subject to export controls for all foreign nationals, inside and outside the United States.

Because American 'deemed export' rules treat access by a foreign national as equivalent to an export, and because Anthropic employed foreign nationals at every level of the organisation, the company had no choice but to disable both models for every user on earth. Every hospital that had integrated Fable into its diagnostic workflow. Every law firm that had automated document review with Mythos. Every defence contractor, every government agency, every technology company that had built workflows on these models discovered in a single weekend that their infrastructure could be switched off — not by a cyberattack, not by a technical failure, but by a letter from a cabinet secretary in Washington.

The immediate trigger was a phone call. Amazon CEO Andy Jassy personally called the White House after Amazon researchers discovered a jailbreak in Fable's cybersecurity guardrails — a vulnerability that allowed the model to assist with the construction of malware in ways that bypassed its safety filters.

Amazon had invested $13 billion in Anthropic, with a commitment toward $20 billion more. The financial exposure was existential; a major cybersecurity incident involving Fable could have exposed Amazon to regulatory and reputational consequences that dwarfed the investment. Jassy made the call. The White House made the call to the Commerce Department. Lutnick made the call to Amodei. The letter followed within hours.

The kill switch, it turned out, was not merely available to the government as a theoretical option. It had a hair trigger, and the conditions that could activate it were not limited to national security emergencies. A single cybersecurity vulnerability, escalated through a single corporate relationship, was enough.

The resolution, when it came, arrived faster than the crisis itself had suggested was possible. Nineteen days after the June 12 shutdown, Fable 5 and Mythos 5 returned to global availability on July 1, 2026 — the product of two weeks of negotiation with the Commerce Department, a new safety classifier, and a joint jailbreak-scoring framework that Anthropic built alongside Amazon, Microsoft, and Google. The episode's underlying tension did not resolve so neatly: Amazon, Anthropic's largest investor, had reported the triggering vulnerability to the government before informing Anthropic itself, a sequencing that left the commercial relationship between the two companies carrying a new kind of strain. Nineteen days is a number that will be cited for years, in both directions — long enough to demonstrate that dependency has consequences, short enough that some governments and enterprises will conclude the risk is tolerable after all.

The reactions from allied nations were immediate and unambiguous in their alarm, if not yet in their remediation.

French Prime Minister Lecornu, speaking from the Elysée Palace hours after the news broke: 'We cannot rely on tools developed by foreign powers. This is not a criticism of our American allies. It is a statement of strategic reality that France and Europe must now confront directly.'

Canadian Prime Minister Mark Carney, addressing Parliament: 'Nobody has done anything wrong here. The American government acted within its legal authority. The Anthropic team acted responsibly. But we will have done something wrong if we simply accept this situation and go back to business as usual. A nation that depends on others for its technology is a nation that can be unplugged overnight. We saw that this weekend.'

The Economist, in an editorial published June 18, framed the challenge starkly: Europe had spent the previous decade building a comprehensive regulatory framework for AI — the GDPR, the AI Act, the Digital Markets Act. What it had not built was a sovereign AI capability. It had the rules. It did not have the tools. And when the tools were switched off, the rules provided no protection.

America's strategy of restricting advanced AI chips to China had appeared decisive and strategically sound when it was announced. The reality, which became unmistakably clear over the two years following the initial restrictions, was considerably more complex.

Denied access to the most advanced hardware, Chinese engineers were forced to innovate within constraints — and they did, in ways that embarrassed the assumption that hardware access was the binding constraint on AI capability. DeepSeek's R1 model, released in early 2025, demonstrated that frontier-level capability could be achieved with restricted hardware using novel efficiency techniques, training methods, and architectural innovations that the American labs had not prioritised.

The lesson that containment strategies have been teaching since the nuclear era repeated itself in the AI era: restricting a capability can accelerate the development of that capability in the restricted party, while providing the restricting party with a false sense of security. China's AI development was slowed. It was not stopped. And the constraints that were imposed forced a kind of innovation that, in some dimensions, produced capabilities that the unconstrained American labs had not developed.

The June 2026 episode did not resolve the AI sovereignty question. It sharpened it. Every government, every large organisation, and every national security apparatus in the world was forced to ask a question that most had been deferring: what is our exposure to AI systems we do not control?

The answers were, almost uniformly, more alarming than expected. Healthcare systems had integrated AI diagnostic tools so deeply into clinical workflows that disabling them would have required reverting to manual processes that the staff had not practised in years. Financial institutions had automated risk assessment, compliance monitoring, and fraud detection with AI systems they now realised sat on infrastructure they did not own. Government agencies had built AI-assisted policy analysis, citizen service chatbots, and administrative automation on American cloud platforms.

China's open-source AI developers noted the episode with quiet satisfaction. Every allied nation that experienced the disruption was a potential customer for Chinese-developed open-source alternatives that ran on locally owned infrastructure and could not be switched off by a letter from Washington. The kill switch, designed to protect American strategic interests, had inadvertently made the case for Chinese AI sovereignty to every government in the world.

The June 2026 kill switch episode produced, in the weeks and months that followed, a proliferation of policy papers, emergency legislative proposals, and corporate sovereignty audits that would have seemed remarkable even a year earlier. Governments that had been slow to engage with AI governance suddenly had a concrete, vivid, impossible-to-ignore demonstration of what AI dependency meant in practice.

The policy responses fell into three broad categories, each reflecting a different theory of what had gone wrong and what the right remedy was.

The first category was regulatory: new requirements for AI providers to maintain service continuity, to notify governments in advance of service disruptions, and to provide data portability that would allow customers to migrate their workflows to alternative providers. These proposals had the advantage of being relatively fast to implement and politically straightforward — they could be framed as consumer protection measures and had broad cross-partisan appeal.

Their disadvantage was that they did not address the fundamental sovereignty problem. A regulatory requirement for service continuity does not help if the government that controls the service provider decides to override the regulatory requirement, as the US government effectively did in June 2026.

The second category was investment: national and regional funds specifically targeted at building domestic AI capability, from data centre construction to model development to chip manufacturing. The EU announced a €50 billion AI Investment Initiative within weeks of the kill switch episode. Several individual member states announced parallel national programmes. Singapore expanded its National AI Strategy with specific funding for sovereign model development. The Gulf states, which had already been investing heavily in AI infrastructure, accelerated their timelines.

These investment programmes had the advantage of addressing the underlying problem — the lack of sovereign AI capability — but the disadvantage of taking years to produce results. Building a frontier model from scratch takes years. Building the chip fabrication capacity to support AI training at scale takes a decade. The investment programmes announced in the summer of 2026 would not produce meaningful sovereignty until the early 2030s at the earliest.

The third category was open source: a rapid acceleration of investment in open source AI development, which offered the possibility of deploying capable models on locally owned infrastructure without depending on American model providers. China's open source models — DeepSeek, Qwen, and others — suddenly became objects of serious interest in markets that had previously been uninterested in Chinese AI technology. The European open source consortium, which had been a niche academic exercise before June 2026, suddenly had government funding and corporate support it had not been able to attract before.

The open source route had the advantage of speed: capable open source models were already available, and organisations that had the technical capability to deploy and fine-tune them could establish meaningful independence from American AI providers within months rather than years. Its disadvantage was that it required technical capability that many organisations did not have, and it transferred dependency from American commercial AI providers to whoever controlled the open source model development — which, in the case of the most capable open source models, was still largely American technology companies.

The fundamental tension that the June 2026 episode exposed has not been resolved by any of these responses. The AI power shift has concentrated infrastructure control in a small number of American companies and, to a lesser extent, Chinese ones. The rest of the world is dependent on that infrastructure to an extent that was not visible until it was briefly removed. Building genuine sovereignty requires years of investment and political will that most governments had not demonstrated before June 2026. Whether the shock of the kill switch will be sufficient to sustain that political will through the years required to build meaningful alternatives remains the defining question of AI geopolitics in the late 2020s.

The June 2026 kill switch episode was not primarily a story about Anthropic or about the Trump administration. It was a story about the architecture of dependency that had been quietly built over the previous decade of AI adoption — and about what that architecture looks like when stress is applied to it.

To understand the architecture, it is useful to trace the dependency chain that connected a hospital in Paris or a government ministry in Tokyo to the decision of an American commerce secretary.

At the end of the chain was the end user: a physician using an AI diagnostic tool, a civil servant using an AI policy analysis system, a financial analyst using an AI risk assessment model. These end users experienced the dependency as a sudden, unexplained failure of a tool they had come to rely on.

One step back from the end user was the organisation that had deployed the AI tool: the hospital, the ministry, the financial institution. These organisations had procured AI capabilities from software vendors, had integrated those capabilities into their workflows, and had in many cases reduced their staffing to reflect the productivity gains that AI had enabled. They experienced the dependency as an operational crisis.

One step back from the organisation was the software vendor that had built the AI-powered tool: the health IT company, the legal tech startup, the financial analytics firm. These vendors had built their products on top of Anthropic's or OpenAI's APIs, paying per-token for inference as their customers used the products. They experienced the dependency as a catastrophic supply chain failure.

One step back from the software vendor was the cloud platform that provided the compute and the API access: AWS for Anthropic, Azure for OpenAI. These platforms had made contractual commitments to their customers about service availability and had reserved the right, buried in their terms of service, to modify or terminate service in compliance with applicable law.

One step back from the cloud platform was the AI model company: Anthropic, OpenAI. These companies had trained the models, designed the APIs, and entered into the commercial relationships with the software vendors and the cloud platforms. They had no practical ability to resist a government order to restrict service, because their operations — their employees, their infrastructure, their corporate structure — were fully subject to American law.

And at the top of the chain was the American government, which had the legal authority to order the model company to restrict service, a compelling interest in doing so given the cybersecurity context, and essentially no mechanism by which the end users at the bottom of the chain could appeal or contest the decision.

The architecture of dependency made the kill switch not only possible but essentially inevitable once the conditions for its use were met. Every step in the chain — from the end user's adoption of AI tools to the model company's incorporation in the United States — was individually rational. The commercial incentives at each step pointed in the direction of deeper integration, faster adoption, and less investment in sovereign alternatives. The aggregate result of these individually rational decisions was a systemic vulnerability that no single actor in the chain could see clearly or had the incentive to correct.

This is the structural insight that the June 2026 episode should have made unmistakable: the dependency was not the result of bad decisions by any individual organisation. It was the emergent property of a market structure in which the costs of dependency were distributed across millions of end users and the benefits of the infrastructure concentration were captured by a small number of platform owners.

Correcting this structural problem requires interventions at the structural level — at the level of market design, regulatory architecture, and investment policy — not at the level of individual procurement decisions. Individual organisations can reduce their exposure to the kill switch by diversifying their AI suppliers, by investing in open source alternatives, and by maintaining human capability in the workflows that AI has automated. But individual action cannot solve a structural problem. Only collective action, through policy, can do that.

The June 2026 kill switch episode created an urgent policy problem for governments that had previously deferred the AI sovereignty question. This chapter section offers a practical framework for thinking about what AI sovereignty means operationally, and what the realistic pathways to it are for different categories of nations.

AI sovereignty is not a binary condition. It exists on a spectrum from complete dependence — in which a nation relies entirely on foreign AI infrastructure and models for all its AI capabilities — to complete independence — in which a nation runs all its AI workloads on domestically owned and controlled infrastructure, using models that it has developed or that it controls fully. In practice, no nation is at either extreme, and the realistic policy objective for most nations is to move meaningfully away from complete dependence toward a position that provides adequate resilience against the risks that the kill switch episode made concrete.

The first practical step toward AI sovereignty is inventory: understanding, with specificity, what AI systems your critical functions depend on and where those systems' key components are controlled. Most governments that have gone through this exercise have found the results alarming: AI dependencies that had been adopted quietly, one function at a time, had accumulated into a systemic exposure that no one had designed and no one had fully mapped.

The second step is prioritisation: not all AI dependencies are equally critical. A government function that uses AI to draft correspondence is in a very different situation from a government function that uses AI to control critical infrastructure. The former dependency is inconvenient if disrupted; the latter is potentially catastrophic. Sovereignty investments should be prioritised toward the highest-stakes functions first, not applied uniformly across all AI use cases.

The third step is diversification: within any given priority domain, reducing dependence on a single AI provider by establishing relationships with multiple providers, including open source alternatives that can be run on domestically controlled infrastructure. Diversification does not eliminate sovereignty risk — if all the alternatives are foreign-controlled, diversification merely distributes the exposure — but it reduces the single-point-of-failure risk that the June 2026 episode illustrated.

The fourth step is domestic capability investment: building or acquiring AI capability that can run on domestically controlled infrastructure, even if that capability is initially less advanced than the foreign alternatives it complements. This is the longest-term and most resource-intensive step, but it is the only one that addresses the fundamental sovereignty problem rather than mitigating it at the margins.

For nations with significant resources and technical capability — the United States, China, the UK, France, Germany, Japan, South Korea, Singapore — domestic capability investment can be ambitious: developing frontier model capability, building large-scale domestic AI compute infrastructure, and training the research community that can sustain AI capability development over time.

For nations with more limited resources and technical capability — which includes most of the world — the realistic objective is more modest: establishing the ability to fine-tune and deploy open source models on domestically owned infrastructure, building the technical workforce capable of managing AI systems without full dependence on foreign expertise, and participating in international coalitions that pool resources to develop shared sovereign AI capabilities.

The international coalition pathway is underexplored but potentially powerful. Small nations that cannot afford to develop sovereign AI capability individually might be able to do so collectively, pooling compute resources, training data, and technical talent to develop shared models and shared infrastructure that none of them could sustain alone. The European Union is the most obvious candidate for this kind of collective sovereignty project — it has the regulatory coherence, the economic scale, and the political institutions required to coordinate sovereign AI investment across member states in ways that smaller international coalitions might not.

The AI sovereignty challenge is not one that any nation will fully solve in the near term. The pace of AI capability development, the concentration of frontier model development in a small number of organisations, and the enormous capital requirements of AI infrastructure mean that most nations will remain dependent on foreign AI capability for most applications for the foreseeable future. But the difference between a nation that has thought carefully about this dependency, has mapped its exposure, has diversified its providers, and has invested in the domestic capability it can realistically build — and a nation that has not done any of these things — is enormous. The first nation is managing a known risk. The second is exposed to a risk it does not fully understand.

The map matters. But first you have to read it.

The Labour Bargain Breaks

For two centuries, the labour bargain has been the foundational compact of industrial society. Technology makes workers more productive. More productive workers can command higher wages. Higher wages support consumer demand. Consumer demand drives economic growth. Growth funds investment in more technology. The cycle continues.

This bargain has been disrupted many times — by the mechanisation of agriculture, by factory automation, by the computerisation of office work. Each disruption produced a period of dislocation followed, eventually, by the emergence of new categories of work that absorbed the displaced labour. The net outcome, across most of the industrial era, was rising living standards and expanding employment.

The AI era is testing whether that bargain still holds — and the early evidence is deeply uncomfortable. Not because AI is different in kind from previous automation waves, but because it is different in scope and speed. Previous automation waves targeted specific, well-defined tasks in specific industries. AI automation targets cognitive tasks across every industry simultaneously, at a pace that the educational system, the social safety net, and the career planning frameworks of most individuals were not designed to match.

The IMF's 2024 analysis, which found that 40 percent of global jobs were exposed to AI disruption — rising to 60 percent in advanced economies — was striking not only for its scale but for its distribution. Previous automation waves disproportionately affected low-skill, routine, manual work. AI automation disproportionately affects high-skill, cognitive, non-routine work. The jobs most at risk are not the ones at the bottom of the income distribution. They are the ones in the middle and upper middle: accountants, lawyers, junior doctors, software engineers, financial analysts, teachers, writers.

This is the structural novelty of the AI transition: it is hollowing out the professional middle class at the same time as it is creating extraordinary wealth at the very top of the capital and skill distribution.

For the first two years of the generative AI era, most of the discussion about AI and labour was theoretical. Economists debated exposure indices. Policy-makers commissioned reports. Technologists made predictions. The actual evidence of what AI was doing to the structure of work remained anecdotal.

That changed in June 2026, when OpenAI published its economic research on Codex, the company's primary agentic AI tool. The paper was notable not only for its findings but for its source: this was empirical data about AI's impact on work, measured inside the organisation that built the tools being measured. It was the AI industry providing its own employment impact data, and the findings were startling.

Within one year of Codex's deployment, it had become the primary AI tool for every department at OpenAI — not just Engineering, but Legal, Finance, and Recruiting. The average OpenAI worker generated 85 percent of their output tokens via Codex; for engineers, the figure was 99 percent.

The unit of work had fundamentally changed. Seventy percent of users now regularly assigned tasks estimated to take a human more than one hour. Twenty-five percent assigned tasks estimated at eight hours or more. Users at the 99th percentile ran more than 60 hours of parallel agent work per day — simultaneously delegating work that would have required entire teams to complete.

The most striking finding was about non-technical staff. Over one quarter of work done by Finance, Legal, and Operations staff on Codex was engineering or coding — tasks they could never have done before, that previously required expensive specialists, that are now being completed by people whose job titles have nothing to do with software development.

AI is not merely making existing workers more productive. It is making workers capable of tasks that were previously outside their professional scope. And in doing so, it is making the question of what a job is — what a profession is — considerably more complicated than any existing labour market framework was designed to handle.

The evolution of how programmers relate to AI tools provides a precise case study in how the labour relationship is changing across every professional category.

In February 2025, Andrej Karpathy — one of the most respected AI researchers of his generation, formerly of Tesla and OpenAI — coined the term 'vibe coding.' The concept was simple and deliberately provocative: instead of writing code carefully, thinking through each line, you describe what you want in natural language, accept what the AI produces, and simply don't worry too much about the details. You are vibing, not engineering. You are a product manager giving instructions to an infinitely patient, infinitely fast junior developer.

The term caught on. It captured something real about how a generation of founders and product builders were actually using AI tools — and it alarmed a generation of senior engineers who had spent years building expertise in the kind of careful, deliberate, deeply understood code that vibe coding explicitly discarded.

Exactly one year later, Karpathy declared vibe coding obsolete. The new paradigm he named was 'agentic engineering': you are not writing code 99 percent of the time, you are orchestrating agents who do, while acting as oversight. You are not the author of the code. You are the architect of the system that produces code, the reviewer of its outputs, the orchestrator of its agents.

He then joined Anthropic. His stated reason was precise and revealing: the model is the bottleneck on what an agent can do. Whoever builds the best model pulls ahead — because their tool finishes more of a task without a human stepping in to correct it. The competitive race, in Karpathy's framing, is not primarily about the interface or the product. It is about the underlying model quality, which determines how autonomously the agent can operate and how rarely a human override is required.

The XDA Developers benchmark of June 2026 confirmed this empirically. Claude Code outperformed OpenAI Codex and Google Antigravity in a head-to-head test — not on functionality, where all three were broadly comparable, but on UX judgment: the capacity to understand design intent, apply contextual colour-coding, create smooth animations, organise information thoughtfully. The model that understood what good looked like, not just what the instructions said, won the workflow.

The transformation of professional work by AI is not happening uniformly across all professions, all organisations, or all geographies. Understanding the differential pace and depth of transformation is as important as understanding the overall trend.

The professions most rapidly transformed are those that combine three characteristics: they involve primarily cognitive tasks that can be expressed in language; they have large quantities of training data available in the form of digitised historical work products; and they have clear, verifiable outputs that allow AI performance to be measured and improved over time.

Law is perhaps the clearest example. Legal work is almost entirely cognitive and linguistic. The corpus of digitised legal documents — cases, contracts, briefs, memoranda, statutes, regulations — is enormous and growing. Legal outputs are verifiable: a contract is either enforceable or it is not; a legal argument is either persuasive or it is not; a document review is either complete or it misses relevant materials. All three conditions for rapid AI transformation are present, and the transformation is proceeding accordingly.

The data from large law firms tells a story of transformation that has outpaced almost every forecast made before 2024. The number of hours billed per matter for contract review, document discovery, and legal research has fallen dramatically. The number of junior associates required to staff a matter of a given size has fallen. The total fees generated by these tasks have not fallen as dramatically as the hours, because partners have been slow to pass cost reductions through to clients — but the pressure to do so is building, as clients increasingly understand what AI can do and demand that their legal bills reflect it.

The transformation is not uniform within the legal profession. Partners who can generate client relationships and exercise judgment in complex, novel situations are as valuable as ever. Junior associates performing document review and routine research are facing the sharpest displacement pressure. The partnership track, which has always been selective, is becoming more selective — not because the demand for legal services is falling, but because fewer hours of junior work are required to produce the same output.

A similar pattern is visible in finance. Investment banking analysis, which has historically required large teams of junior analysts working long hours to build financial models and produce pitch books, is being compressed by AI tools that can generate initial drafts of financial models from natural language descriptions, populate pitch book templates from public data, and conduct preliminary company and sector research automatically. The first-year analyst class at major investment banks is shrinking.

Software engineering is experiencing a more complex transformation. The demand for software is not falling — if anything, the emergence of AI tools has increased the appetite for software products and custom AI applications. But the amount of human coding required to produce a given quantity of software is falling rapidly. Senior engineers who can architect systems, make technical decisions, and manage the AI tools that generate code are in higher demand. Engineers whose primary skill is writing boilerplate code are facing significant displacement pressure.

The common pattern across all these professions is a bifurcation of the labour market within each field: the people who can work with AI — who can direct it, evaluate its outputs, catch its errors, and use its speed to multiply their own productivity — are more valuable than ever. The people whose primary value was in performing the tasks that AI can now perform are facing a structural challenge that is not a temporary cyclical disruption. It is a permanent shift in what human labour is for.

The educational system has not yet adapted to this shift in most countries. Law schools are still training students to do document review. Business schools are still teaching financial modelling as a core skill. Computer science departments are still treating the ability to write code from scratch as the primary output of a four-year degree. The gap between what is being taught and what is actually required in the AI-transformed workplace is widening faster than curricula are being updated.

The individuals who are navigating this transition most successfully are those who have independently identified the shift and invested in the skills that the new labour market values: systems thinking, judgment under uncertainty, the ability to manage AI tools effectively, and the domain expertise that allows them to verify AI outputs and catch AI errors. These are not skills that current educational systems teach explicitly, and the people who have them have largely developed them through experience and self-directed learning.

The organisations that are navigating this transition most successfully are those that have treated it as a strategic opportunity rather than a cost-cutting exercise. The firms that deployed AI to reduce headcount and cut costs first discovered that they had eliminated the institutional knowledge that was required to evaluate the AI's outputs. The firms that deployed AI to increase the capacity of their existing staff — to allow the same number of people to do more, to serve more clients, to explore more ideas — have generally achieved better outcomes both in terms of business performance and in terms of employee satisfaction and retention.

The lesson of the first generation of AI adoption in professional services is that AI amplifies human capability most effectively when the humans remain engaged, maintain their judgment, and use the AI as a tool rather than a replacement for thought.

The transformation of professional work by AI is not merely a productivity story. It is a story about the implicit contract between professional expertise and economic reward — a contract that has governed the allocation of resources in advanced economies for more than a century.

The professional contract works like this: an individual invests years and significant resources in acquiring expertise in a specific domain. The expertise is valuable because it is scarce: few people have invested the time and resources required to develop it, and the social functions that depend on the expertise — healthcare, legal representation, financial management, engineering — are essential. The scarcity of expertise relative to demand generates compensation that rewards the investment in acquiring it, creating an economic incentive structure that produces the supply of expertise the economy needs.

AI is disrupting the professional contract by reducing the effective scarcity of expertise. When an AI system can produce a first draft of a legal contract with the competence of a junior lawyer, the effective supply of junior legal expertise increases dramatically — not because more people have gone to law school, but because a smaller number of lawyers can produce the same quantity of legal work. When an AI system can generate a financial model with the accuracy of a junior analyst, the effective supply of financial analysis increases. When an AI system can write code with the competence of a junior software engineer, the effective supply of software engineering increases.

The reduction in effective scarcity has predictable economic consequences: the compensation for the expertise that AI can replicate falls, as supply increases relative to demand. The compensation for the expertise that AI cannot replicate — the judgment, the relationship, the strategic insight, the novel problem-solving — may increase, as the relative scarcity of those capabilities increases while the scarcity of replicable capabilities falls.

The uncomfortable reality of this dynamic is that it is hollowing out the professional middle — the large population of competent, credentialed professionals who perform valuable but replicable cognitive work. These are the people who went to law school and spent years doing document review, expecting that this would be the pathway to the partner-level judgment work that commands premium compensation. These are the people who went to business school and spent years building financial models, expecting that this would be the pathway to the deal-making and strategy work that commands premium compensation. These are the people who studied computer science and spent years writing code, expecting that this would be the pathway to the system design and technical leadership work that commands premium compensation.

For many of these people, the pathway has been disrupted before they reached the destination. The document review years that were supposed to build legal intuition are now done by AI. The financial modelling years that were supposed to build analytical intuition are now done by AI. The coding years that were supposed to build engineering intuition are now done by AI. But the intuition that was supposed to accumulate from doing those tasks — the pattern recognition, the edge case awareness, the domain knowledge that comes from making mistakes and learning from them — may not accumulate as well from supervising AI doing the tasks as it did from doing the tasks directly.

This creates a pipeline problem for the professions that is distinct from the displacement problem. Even if AI creates a sustained demand for partner-level legal judgment, strategic financial thinking, and senior engineering leadership, the pipeline of people who have developed those capabilities through years of hands-on practice may be smaller in the AI era than in the pre-AI era. The junior work that used to be the crucible in which professional judgment was forged has been automated away, and the mechanism by which junior professionals became senior professionals has been disrupted before anyone has figured out what to replace it with.

The professions that navigate this challenge most successfully will be those that explicitly redesign their training and development pathways to account for the AI transformation — that identify what experiences are still necessary for the development of judgment and create deliberate opportunities for those experiences, even when the economic pressure is to automate them away. This is not a challenge that any individual organisation can solve on its own. It requires collective action by the professions themselves, by the educational institutions that train professionals, and by the regulators who set the standards for professional practice.

The alternative — allowing the pipeline of professional judgment to atrophy while AI handles the work that used to build that judgment — is a path to a future in which AI systems make increasingly consequential decisions in healthcare, law, finance, and engineering with decreasing human oversight, because the humans who are supposed to provide that oversight have not had the experiences required to exercise it competently.

One of the most important and least discussed challenges of AI's transformation of professional work is the measurement problem: the difficulty of measuring, within traditional economic frameworks, what AI is actually doing to productivity, output quality, and human welfare.

The standard measure of economic output is gross domestic product, which measures the monetary value of goods and services produced. GDP was designed for an era in which the primary inputs to production were labour and capital, and in which the outputs of production could be counted and priced relatively straightforwardly. It handles the AI era poorly.

When a lawyer who used to spend three hours on document review now spends thirty minutes — because AI handles the routine review and the lawyer focuses on the anomalies — what happens to GDP? If the lawyer bills the same number of hours at the same rate, GDP is unchanged, even though the lawyer has done less work. If the lawyer passes the productivity gain through to the client by billing fewer hours, GDP falls, even though the client has received the same legal service in less time. If the lawyer uses the freed time to take on additional clients, GDP rises, but only in the accounting of the additional work, not in any measure of the efficiency gain that made it possible.

The measurement problem is not merely a technical issue for economists. It has practical consequences for how AI's impact on the labour market is understood and how policy responds to it. If AI productivity gains are systematically unmeasured by standard economic indicators, policymakers may not see the displacement effects coming until they have already produced political consequences. If the gains show up in corporate profits — because workers are not sharing in the productivity improvement — but not in wages or employment, the standard economic indicators will show a healthy economy at the aggregate level while the distribution of outcomes is deteriorating at the individual level.

The measurement problem compounds the equity problem: the benefits of AI that are least measurable by conventional economic metrics are the benefits that accrue to capital owners — the increased profit margins that come from automating labour costs — while the costs that are least measurable are the costs borne by workers — the deskilling, the reduced bargaining power, the psychological burden of working alongside systems that could replace them.

Making AI's impact measurable is a precondition for governing it effectively. This requires investment in new data collection mechanisms — survey instruments that capture AI adoption and productivity impact at the firm and worker level, administrative data systems that can track the displacement and reemployment trajectories of workers whose jobs are affected by AI, and accounting standards that require firms to disclose their AI cost savings and how they are being distributed.

None of these are technically difficult to implement. They are politically difficult, because the parties who would benefit most from transparency — workers, citizens, policymakers — have less influence over measurement standards than the parties who have the most to gain from opacity — corporations whose earnings are enhanced by AI-enabled cost reductions that they are not required to disclose.

The resolution of the measurement problem will be one of the defining governance challenges of the AI era. Nations that develop good measurement frameworks will be better equipped to identify AI's distributional consequences early and respond with appropriate policy. Nations that do not will be flying blind into one of the most significant economic transformations in their history.

Who Gets Left Behind — and Who Decides

Nations without chips. Companies without data. Workers without skills. Investors without access. The AI power shift produces winners at extraordinary speed — and it produces a permanent underclass of the unprepared at the same speed. But the question of who gets left behind is not only economic. It is geopolitical, generational, moral, and — most importantly — a matter of deliberate choice rather than inevitable fate.

The distribution of AI's gains is being determined right now, by decisions being made in boardrooms, legislatures, and data centres — most of which the public never sees and most of which are not framed as decisions about equity at all. They are framed as investment decisions, procurement decisions, policy decisions. The equity consequences are externalities, not the primary consideration.

This chapter is an argument that they should be the primary consideration — not because efficiency and growth do not matter, but because growth that is this concentrated, this fast, and this structurally self-reinforcing will produce social and political instability that destroys the conditions for further growth.

By mid-2026, the United States and China controlled approximately 90 percent of global AI computing power. Stanford University's 2026 AI Index found that nearly eight in ten AI companies started in G7 nations were based in the United States. The top three hyperscalers — Amazon, Microsoft, Google — together controlled the overwhelming majority of global cloud infrastructure.

The next tier of nations — the United Kingdom, Canada, France, Germany, Japan, South Korea, Israel — had genuine research capability, significant venture ecosystems, and in some cases world-class AI labs. What they did not have was infrastructure sovereignty. They were, in the language of the power map, provinces of the American AI empire: capable, creative, and dependent.

The tier below that — Southeast Asia, South Asia, the Middle East outside the Gulf states, Africa, Latin America — was consuming AI products built entirely on infrastructure they did not own, running on models trained primarily on data that underrepresented their languages, their cultures, and their economic realities. The AI systems making decisions about their credit scores, their healthcare, their children's education, and their government services were designed for a different population and optimised for different outcomes.

Singapore presents an instructive case study in what strategic positioning looks like for a small nation in the AI era. The city-state cannot build a frontier model and cannot manufacture AI chips. But it has made deliberate decisions to position itself as a trusted node in the AI infrastructure — attracting hyperscaler data centres, building AI governance frameworks that are taken seriously internationally, investing in AI research capability at the National University, and training a generation of engineers who speak the language of the new economy. Singapore is not an empire. But it is not a powerless province either. It is a strategic trading post in the new infrastructure economy, collecting rents on its location, its governance, and its talent.

Infrastructure sovereignty is not the only lever that determines who gets left behind. Equally consequential is who gets to train inside the institutions and laboratories that produce frontier AI capability in the first place — and in July 2026, the United States made that access measurably harder. New rules replaced the long-standing "duration of status" policy, under which international students could remain as long as they stayed enrolled, with a fixed cap of four years or less. Students needing longer — overwhelmingly PhD candidates in the STEM fields that feed AI research — must now apply to US Citizenship and Immigration Services for an extension, undergo background checks and biometric screening, and operate under a post-graduation grace period cut from 60 days to 30.

More than seven in ten international students in the US come from Asia, mostly from China and India, with meaningful numbers from Malaysia, Vietnam, and the rest of Southeast Asia. For a Malaysian PhD candidate in Texas, the rule change did not just add paperwork. It added fear: many students described becoming afraid to travel home for routine visa stamping, since a failed re-entry could mean forfeiting years of research and tuition already invested. Malaysia's National Assembly of Malaysian Students in the USA told its members the realistic response was to stay resilient and plan ahead — advice that is honest about how little control students have over a policy made an ocean away.

The pattern is not new. A year earlier, Malaysia's MARA had already redirected roughly 200 government-sponsored scholars away from the United States entirely, placing them at comparable universities in other countries after an earlier visa-processing suspension made the US route look too uncertain to risk. That is the talent equivalent of a customer routing around a toll road that has become unpredictable — and it is a preview of what happens at scale if access to the world's leading AI research institutions keeps tightening. The infrastructure chapters of this book describe capital moving to wherever the toll road is most investable. Talent moves the same way, toward wherever the path to a research career is most navigable.

There is a dimension of who gets left behind that receives insufficient attention: the cybersecurity vulnerability created by unequal AI capability.

Anthropic's Mythos model — restricted to Project Glasswing partners only because of its extraordinary offensive cybersecurity capabilities — demonstrated in April 2026 that a sufficiently capable AI system could discover thousands of critical vulnerabilities across every major operating system and browser, chain multiple exploits together, and complete autonomous end-to-end attacks on enterprise networks that would previously have required a team of elite human hackers working for months.

The UK AI Security Institute confirmed these findings. The time from vulnerability discovery to active exploitation, which had stood at 2.3 years in 2018 and had already been compressed to 23 days by 2025, collapsed to 20 hours by April 2026. The implications are asymmetric and alarming: attackers need only find one gap; defenders must close all of them. Defenders need to close every vulnerability; attackers need to exploit just one. And with AI tools, the cost of mounting an attack has fallen while the cost of maintaining a comprehensive defence has not fallen at all.

The nations and organisations left behind in AI capability are not merely economically disadvantaged. They are strategically exposed in a new way: the most powerful AI is simultaneously the most capable offensive cyber weapon ever built and the only meaningful defence against that weapon at scale. Those without it are not neutral. They are defenceless against adversaries who have it.

Retail investors arrived late to the AI era — as they almost always do to transformational technology cycles. The extraordinary returns were made in the private markets, by sovereign wealth funds, hyperscalers, and institutional investors who had the access, the conviction, and the minimum cheque sizes that precluded ordinary participation.

Anthropic's 235-fold valuation increase from early 2023 to its 2026 IPO filing happened entirely in private markets, over approximately three years. By the time retail investors could buy shares on a public exchange, the extraordinary returns had already been captured. The IPO is not the beginning of the opportunity. It is the closing of the first chapter.

By mid-2026, the IPO clock had started running in earnest. Anthropic's confidential S-1 filing on June 1 followed a $65 billion Series H round at a $965 billion post-money valuation, and forecasters tracking the filing placed the most likely pricing window between November 2026 and March 2027 — contingent, as the June episode demonstrated, on the company's ability to manage regulatory risk that had nothing to do with its technology. The pattern holds: by the time ordinary investors can buy a share, the extraordinary multiples have already been captured by the small number of institutions that had access years earlier.

This pattern is not unique to Anthropic. SpaceX, which filed for its IPO in June 2026 at a valuation of $1.77 trillion, had created extraordinary wealth for its early investors in the private markets over nearly two decades. The investors who bought SpaceX at early valuations had already earned returns that public market investors will never see. The AI infrastructure buildout — the data centres, the power plants, the chip fabs — is being financed and owned by sovereign wealth funds, hyperscalers, and large private equity firms operating at scales that preclude ordinary participation.

The answer to the equity challenge of the AI era is not to slow the technology. Slowing the technology is not possible, and attempting to slow it in one jurisdiction simply transfers the development and the gains to another. The answer is to ask, consistently and without flinching, who benefits and who pays — and to design policy, investment frameworks, and educational systems that deliberately widen access to the infrastructure layer, not just the application layer.

Access to AI applications is largely democratised already. You can use ChatGPT for free. You can use Claude for free. The application layer is competitive and, on balance, getting cheaper. The infrastructure layer is not competitive, is not getting cheaper, and is controlled by an increasingly small number of organisations.

Widening access to the infrastructure layer means investing in domestic data centre capacity and the energy infrastructure to power it. It means building or acquiring AI model capability at the national level — not necessarily frontier models, but capable models trained on domestic data and running on domestic infrastructure. It means designing AI governance frameworks that impose obligations on model developers to disclose training data, energy consumption, and societal impact — not to restrict innovation, but to make the costs visible and to create accountability for the externalities.

It means, above all, asking the question that every decision-maker in every country should be asking right now: if the AI systems we depend on were switched off tomorrow — by a letter from a foreign government, by a cyberattack, by a commercial decision by a private company — what would happen? And if the answer is alarming, what are we building to change it?

The AI power shift is not inevitable in its distribution of outcomes. It is a choice — being made right now, by people who mostly do not frame it as a choice. Making it visible is the first step toward making it fair.

The distribution question began, in mid-2026, to move from academic proposal toward concrete political contest. OpenAI's own policy paper, 'Industrial Policy for the Intelligence Age,' proposed donating 5 percent of its equity to a US sovereign wealth fund, with the argument that returns could be distributed directly to citizens regardless of their starting wealth. Senator Bernie Sanders introduced a considerably more aggressive version — the American AI Sovereign Wealth Fund Act — calling for a one-time 50 percent tax on the stock of systemically important AI companies, collected directly into a public fund. Neither proposal had advanced to committee by mid-2026. But the fact that a leading AI lab and a leading voice of the American left arrived, independently, at structurally similar conclusions — that public ownership of a slice of AI's returns is necessary, not merely desirable — suggests the equity question is no longer confined to the margins of policy debate.

The question of what wisdom requires in the AI era is ultimately a question about what we are for — as individuals, as organisations, as nations, and as a civilisation.

Technology is a tool. It extends human capability in specific directions. Steam extended human muscular capability beyond what biology allows. Electricity extended human sensory capability across time and space. The internet extended human communicative capability to encompass the whole of recorded human knowledge and the attention of every connected person on earth. Artificial intelligence is extending human cognitive capability in ways that are still being discovered.

But the extension of capability is not the same as the direction of capability. A more powerful tool can be used wisely or unwisely. A hammer of extraordinary precision can build a cathedral or smash a skull. The choice of what to build — and what not to build — is a human choice, not a technological one.

The AI era is producing an extraordinary proliferation of capability. Systems that can diagnose disease with accuracy that exceeds trained physicians in specific domains. Systems that can write code with competence that matches experienced engineers for routine tasks. Systems that can analyse legal documents, financial statements, and scientific papers with a speed and comprehensiveness that no human team could match. Systems that can discover cybersecurity vulnerabilities in critical infrastructure at a rate that would require a team of thousands of expert hackers working for years to replicate.

Each of these capabilities is, in itself, extraordinary. Taken together, they constitute a step change in what human civilisation can accomplish per unit of time and per unit of cost. The productivity implications alone are sufficient to reshape the global economy.

But the direction question — what should we use this capability to do? — is not one that the technology answers. It is one that we must answer, and the answer we give, collectively and individually, will determine whether the AI era is remembered as one of the great expansions of human flourishing or as one of the great concentrations of power and wealth in a small number of hands.

Professor Lily Kong of Singapore Management University captured the distinction precisely at the 10th World Cities Summit in June 2026. Smart cities and smart technologies can measure everything: traffic flows, energy consumption, citizen behaviour, environmental conditions. What they cannot do — what no algorithm can do — is decide what matters. That is a question of values, and values are irreducibly human.

A city that optimises for throughput will sacrifice the experiences that make cities worth living in: the unexpected encounter, the inefficient park, the slow street market. A hospital that optimises for throughput will sacrifice the care that makes medicine worth practising: the unhurried conversation, the observation that doesn't fit the algorithm, the judgment that the textbook doesn't cover. A government that optimises for efficiency will sacrifice the deliberation that makes democracy worth having: the argument that takes time, the minority view that deserves to be heard, the outcome that is not the most efficient but is the most legitimate.

The AI era is producing systems of extraordinary optimisation capability. The question is what we are optimising for. And the answer to that question cannot be delegated to the systems themselves.

This is why the wise city concept matters beyond urban planning. It is a model for how any institution — a city, a company, a government, a professional organisation — should relate to AI capability. Not by refusing it, which is both futile and counterproductive. Not by delegating to it, which abdicates the responsibility of judgment. But by using it as a tool in service of ends that humans have chosen through deliberate, inclusive, values-grounded processes.

The nations and organisations that will navigate the AI era most successfully will be those that develop this capacity for wise AI use: the ability to harness the extraordinary optimisation power of AI systems in service of ends that they have chosen carefully, with full awareness of the trade-offs, and with ongoing accountability to the people who are affected by the outcomes.

This is harder than either refusing AI or surrendering to it. It requires the kind of institutional capacity — for deliberation, for values articulation, for adaptive governance — that is difficult to build and easy to lose. But it is the only path to a future in which the AI power shift produces broadly distributed human flourishing rather than a concentration of AI-enabled advantage in the hands of those who were already powerful.

The map is not the territory. But the map tells you where you are, and knowing where you are is the beginning of being able to choose where you are going. The AI power shift is already underway. The question of where it goes — who benefits, who bears the costs, what kind of world it produces — is still open. It will be answered, over the next decade, by the choices that individuals, organisations, and governments make about how to engage with the most powerful technological transition in a generation.

Make those choices deliberately. Ask the seven questions. Follow the capital, not the headlines. And remember that behind every metric lies a human experience, and behind every dataset, a human story.

The map is yours. Use it wisely.

The Equity Imperative The argument for taking equity seriously in the AI era is not primarily a moral argument, though the moral case is strong. It is a practical argument about the conditions required for the AI era to produce outcomes that are sustainable — economically, politically, and socially — over the medium and long term.

Technologies that produce benefits that are very broadly distributed and costs that are also broadly distributed are politically durable. The automobile produced enormous benefits for ordinary people — mobility, economic opportunity, access to resources and relationships — and the costs it imposed were also broadly felt, through traffic, pollution, accidents, and the restructuring of urban space. The political economy of the automobile era was contested but ultimately stable: the benefits were real and widespread enough to maintain political support for the technology even as its costs became more apparent.

Technologies that produce very concentrated benefits and broadly distributed costs are politically fragile. Nuclear power was the clearest example of the twentieth century: the benefits of cheap electricity were broadly distributed, but the costs — the risk of catastrophic failure, the unresolved problem of waste disposal, the psychological burden of proximity to dangerous materials — fell on specific communities that did not share proportionally in the benefits. The political fragility of nuclear power was demonstrated by the long history of opposition that eventually constrained the technology's growth in most democratic countries, regardless of its engineering merits.

AI is following a pattern that looks more like nuclear than like the automobile. The financial benefits of AI are accreting, overwhelmingly, to the owners of the infrastructure — a small number of companies and individuals. The productivity benefits are distributed more broadly but unevenly, concentrated among knowledge workers in high-income countries. The costs — displacement, deskilling, increased surveillance, cybersecurity risk, environmental impact of data centre energy consumption — are distributed broadly and in many cases most heavily on the communities and workers least positioned to benefit from the technology.

This pattern of benefit concentration and cost distribution is the signature of a technology that is politically fragile, even if the technology itself is technically extraordinary. The political fragility has already begun to manifest: AI regulatory movements in the European Union, the United States, and an increasing number of other countries reflect a public that is sceptical of the technology's benefit claims and attentive to its costs. The fragility will increase as the displacement effects in professional and middle-class labour markets become more concrete and more visible.

The response that is most likely to produce durable outcomes is not to resist the technology — that is futile and would forgo genuine benefits — but to reshape its distribution. This means ensuring that the productivity gains from AI adoption are shared with the workers whose labour is being augmented, not just captured by the capital owners whose returns are being increased. It means investing in the education, retraining, and social support systems that allow workers displaced by AI to participate in the new economy. It means designing AI governance frameworks that impose obligations on AI developers to account for and mitigate the equity consequences of their systems. And it means building the domestic AI capability — the compute, the models, the talent — that allows nations to participate in the AI economy as active agents rather than passive consumers.

None of this is easy. All of it requires political will and institutional capacity that many countries do not currently have. But the alternative — allowing the AI era's benefits to accrete to an increasingly narrow group while its costs are distributed increasingly broadly — is a path to the kind of political instability that ultimately undermines the conditions for further technological and economic progress.

The countries that get this right will be the ones that look back on the AI era as a period of broadly distributed human advancement. The countries that get it wrong will look back on it as a period of concentrated enrichment and widespread displacement that produced the political backlash that constrained AI's potential for a generation.

The choice is being made now, by the decisions that governments, companies, and individuals are making about how to engage with the most powerful technological transition in a generation. It is not too late to make different choices. But the window in which different choices are possible is narrowing, as the infrastructure concentrates, the switching costs accumulate, and the distribution of benefits and costs becomes more entrenched.

Ask the seven questions. Follow the capital. And remember that the most important power shift of the AI era is not from one technology company to another. It is from the many to the few — unless we decide, collectively and deliberately, that it should be otherwise.

The AI power shift, seen from the present moment, can feel overwhelming in its pace and scope. New capabilities emerge monthly. New deployment patterns emerge quarterly. New geopolitical implications become apparent with each significant AI episode. The sheer volume of consequential development makes it difficult to maintain a clear sense of where things are heading and what the defining choices are.

Taking a longer view — asking not what happened in the last quarter but what the AI era will look like in twenty or thirty years — suggests that many of the dynamics that feel permanent today are in fact transitional.

The concentration of AI infrastructure in a small number of American companies is likely transitional. It reflects the first-mover advantages of organisations that understood the importance of the transformer architecture early, invested in compute at scale before the demand was obvious, and accumulated the research talent and institutional knowledge that the first generation of frontier models required. These advantages are real and durable in the short to medium term. They are unlikely to be permanent in the long term, as other nations and organisations make the investments required to compete at the infrastructure level.

The opacity of AI systems — the inscrutability of large language models, the difficulty of explaining their decisions, the impossibility of fully auditing their behaviour — is likely transitional. The field of AI interpretability, which seeks to understand what is happening inside neural networks at the computational level, is making progress that will eventually produce tools for understanding AI systems that do not currently exist. Regulatory pressure for explainability is accelerating this progress. The models of 2036 or 2046 will likely be more transparent than those of 2026.

The displacement-without-replacement dynamic that characterises the early AI labour market — in which AI is eliminating certain categories of work faster than new categories are emerging to absorb the displaced workers — is also likely transitional, though the transition may be painful and may take longer than optimists predict. Previous technology transitions have ultimately produced more employment than they destroyed, though not without significant dislocation during the transition. The AI transition is distinguishable from previous transitions in degree rather than kind, and the historical pattern of eventual net employment growth has held even through transitions that seemed, at the time, to threaten permanent displacement.

What is not likely transitional is the shift in the architecture of cognitive capability — the way that AI systems have become a permanent part of the infrastructure of thinking, of decision-making, of knowledge work. The integration of AI into the processes by which humans understand and act on the world is a one-way door. Even if the specific models and platforms change dramatically, the expectation that AI capability is available, affordable, and continuously improving will not reverse. The cognitive infrastructure of human civilisation has been permanently altered.

This permanent alteration is both the source of the extraordinary optimism about AI's potential and the source of the legitimate anxiety about its risks. A world in which cognitive capability is much more widely and cheaply available than it was before has the potential to be a world of much more broadly distributed human flourishing — in which the intellectual tools that were previously available only to the very wealthy or the very privileged are available to everyone. It also has the potential to be a world in which the power of those cognitive tools is concentrated in the hands of those who already have power, amplifying existing inequalities to a degree that produces social and political instability.

Which of these futures we inhabit will be determined by the choices that are being made now — about who owns the infrastructure, who has access to the capabilities, who bears the costs, and who sets the rules. These choices are not predetermined. They are being made, continuously and consequentially, by a distributed set of actors — companies, governments, individuals — most of whom do not frame their choices as contributions to the long-term architecture of the AI era.

Framing them as such is not idealism. It is strategic clarity about what is at stake and who gets to decide it. The AI power shift is not a natural phenomenon that happens to people. It is a human-made transition that is being shaped, moment by moment, by decisions that could be made differently.

The map this book has tried to provide is not a prediction of which specific path the AI power shift will take. It is a framework for understanding the structural dynamics that will shape whatever path is taken — the economics of infrastructure concentration, the geopolitics of AI sovereignty, the labour market consequences of cognitive automation, the equity implications of concentrated capability — and for making choices within those dynamics that reflect the kind of world you want to inhabit.

The AI era is not happening to us. We are making it. The question is whether we are making it deliberately, with awareness of the consequences, or inadvertently, one individually rational decision at a time, without seeing the aggregate outcome until it is too late to change.

The map is yours. The choices are yours. The world that results from those choices will be, in a very real sense, the world you made.

Use the map wisely.

The Map Is Yours

The AI power shift is not a future event. It is a present reality, unfolding faster than most people have registered, with consequences that are already visible in the balance sheets of the world's largest companies, the geopolitical calculations of the world's most powerful governments, and the career trajectories of the world's most educated workers.

This book has tried to give you a map. Not a prediction — the pace of change makes confident prediction foolish — but a framework for understanding what is happening structurally, regardless of which specific model wins the capability race or which specific company emerges as the dominant platform. The structural dynamics of the AI power shift will persist through multiple generations of technology, because they are grounded in the economics of infrastructure, not in the features of any particular product.

The map shows seven things.

First: power in the AI era lives in the infrastructure layer, not the application layer. The companies that own the chips, the energy, the cloud, and the real estate collect rent regardless of which model wins. Invest in, build strategy around, and understand the infrastructure layer — it is more durable than anything sitting on top of it.

Second: the model companies, however impressive, are structurally dependent on infrastructure they do not own. Their vulnerabilities are real, and they are being exploited by the hyperscalers who fund them. The relationship between model companies and their cloud providers is symbiotic but asymmetric — the cloud provider has more leverage than it appears.

Third: private capital captured the extraordinary returns of the AI era before retail investors could participate. The IPO wave of 2026 is not the beginning of the opportunity. It is the end of the first chapter. Public market AI investing requires a more nuanced framework than 'buy the AI companies' — the infrastructure plays are more durable than the model plays, and the toll roads are more durable than the vehicles that use them.

Fourth: the kill switch is real. Any nation, company, or institution that has built critical infrastructure on another nation's AI models has accepted a dependency with an eviction clause that the landlord can exercise at any time for any reason. June 2026 was a warning. It will not be the last. And the containment of China's AI development through chip restrictions produced the unintended consequence of accelerating Chinese AI self-sufficiency — a pattern that has repeated itself in every technology domain where export controls have been applied.

Fifth: the labour bargain is being renegotiated at machine speed. The question is not whether AI will affect your work — it will, in every professional category, at every skill level. The question is whether you are learning to orchestrate agents or being orchestrated by them; whether you are becoming an agentic engineer or remaining a vibe coder; whether you are building skills in system design, oversight, and judgment, or skills in tasks that AI will be better at within eighteen months.

Sixth: AI-enabled cyberwarfare has crossed a threshold from theoretical to operational. The time from vulnerability discovery to active exploitation has collapsed to hours. Every organisation that has not fundamentally rethought its security posture — not added more tools to an existing framework, but rethought the framework from the assumption that AI-enabled attackers have capabilities that were previously available only to nation-state intelligence services — is operating on borrowed time.

Seventh: the distribution of AI's gains is a choice, not a destiny. Those who own the infrastructure will compound their advantage unless deliberate policy, investment, and institutional design intervene. The question of who gets left behind — which nations, which communities, which generations — is being answered right now by decisions that are not framed as equity decisions but that have profound equity consequences. Making those consequences visible is the prerequisite for making different choices.

This book was written at a pace that the subject demanded — quickly, in real time, against a backdrop of events that kept moving faster than the writing could keep up. It is incomplete in ways that a slower subject would not permit. The AI power shift will continue to evolve long after this book goes to press.

But the questions it has tried to equip you with — who owns the layer, who pays whom, who gets left behind, what is the hidden dependency — will remain the right questions for as long as power continues to transfer. The answers will change. The questions will not.

Ask them. Every day. About every announcement. The answers will tell you what is actually happening, beneath the press releases and the benchmarks and the breathless coverage.

The map is yours.

— Prof Dr Tan Teik Kheong · 陳德強教授博士
IEEE · 电气电子工程师学会 · 2026

Prof Dr Tan Teik Kheong is a technologist, academic, and strategist with forty years of experience across engineering, policy, and global industry.

His professional journey began with a defining chapter in the early history of broadband networking. As Senior Vice President representing Asia at the global ATM Forum, he was instrumental in bringing broadband infrastructure to scale across the Asia-Pacific region, establishing himself as one of the foremost evangelists of the broadband era that laid the groundwork for the modern internet economy.

He subsequently served as Chair of the IEEE 802.11 Wireless Next Generation Committee, contributing to the foundational standards that gave the world Wi-Fi. His corporate career has taken him through Cisco, NXP Semiconductors, Accenture, FINRA, and IMDA, where he shaped digital policy at a national level. He has taught at Stanford University and MIT, and has collaborated with Oxford.

His contributions have been recognised through his nomination to the International Who's Who of Networking, his designation as an ATM Forum Ambassador, and his appointment as an Internet Society (ISOC) Ambassador. The AI Power Shift is his third book.