Editor’s note: This essay continues the Synthetic Civilization political economy series. The first essay, “Output Without Income,” argued that AI may preserve production while weakening the wage-based social bargain. The second, “The Market Becomes an Interface,” argued that allocation is moving upstream into pre-market systems that determine eligibility before buyers and sellers ever meet. The third, “The Wage Was a Legitimacy Machine,” argued that employment did more than pay people; it explained them. The fourth, “Tenants of Intelligence,” argued that the next class divide is ownership versus dependency inside rented intelligence environments. This essay turns from the tenant to the ground beneath her: the compute infrastructure that makes synthetic production possible, rent-producing, and politically consequential.
The history of political economy is largely a history of what counts as the ground.
Land was the original ground. Whoever owned it received rent. Whoever rented it owed service. The material substrate of production, including soil, water, pasture, forests, and mineral deposits, concentrated wealth and power in the hands of those who held title. The peasant could cultivate. The lord collected. The feudal order was not only an ideology. It was a consequence of what the ground was and who controlled it.
Industrial capitalism shifted the ground. Factories, machines, railways, canals, ports, and coal mines replaced land as the primary substrate of production. Capital became the thing that organized economic life. The worker could sell labor. The capitalist collected surplus. The logic was different from feudalism, but the question remained recognizable.
Who owns the substrate from which value flows?
The AI economy is asking that question again.
Compute is becoming the new ground.
Not merely as metaphor. In political-economic terms, compute is the material substrate of synthetic production. Intelligence does not exist in the abstract. It runs on chips, inside data centers, through energy grids, cooled by water, connected by fiber, financed by long-duration capital, and protected by law.
The model that generates value is not floating freely.
It is sitting on estate.
And estate, as history shows, produces rent.
Compute Is Not a Factory. It Is Estate.
The distinction matters because factories and estates behave differently.
A factory is productive capital. It creates value by transforming inputs into outputs. Its value derives from what it makes. Factories depreciate, become obsolete, require labor, face competition, and can be replicated when the economics justify it. A factory makes its owner a producer.
Estate is different.
Estate is territorial capital. It is productive, but also positional. It does not merely generate output. It controls the conditions under which other actors can produce at all. Its value derives not only from what it makes, but from what others must pay to access it.
Land earns rent not because the landowner works harder than the cultivator, but because the land is where cultivation must happen.
Compute is acquiring the properties of estate, not factory.
A frontier AI system requires advanced semiconductors that only a narrow set of firms can design, fabricate, package, and deploy at scale. It requires data centers whose construction costs rise as facilities are redesigned for AI workloads. It requires enormous and reliable electricity supply. It requires cooling systems, land, fiber connectivity, grid interconnection, water access, regulatory approval, specialized labor, security, and continuous capital expenditure.
These are not ordinary software inputs. They are conditions of existence.
You cannot spin up a competing compute estate the way you spin up a competing website. The barriers are physical before they are financial. They sit in fabs, power grids, data-center campuses, water systems, transmission lines, export-control regimes, procurement contracts, and capital markets.
That is the first property of estate: it cannot be easily replicated.
The second property is positional power.
Estate produces rent because others must pass through it to produce. The AI compute estate produces rent for the same reason. Intelligence can be accessed through APIs, cloud subscriptions, inference services, enterprise licenses, hosted models, and managed platforms. But that access comes at prices, on terms, under policies, and within availability constraints set by the estate owner.
The estate owner may also be productive. Hyperscalers, chipmakers, cloud providers, and model companies are not passive landlords. They build, engineer, maintain, secure, and improve the systems they control.
But their deeper power comes from owning the ground on which others become productive.
The tenant rents access to the ground.
The estate owner controls the lease.
The Physical Ground of Synthetic Production
It is easy to describe the AI economy as software, models, algorithms, and intelligence. That abstraction is partly real. AI does move through digital channels in ways that land, coal, steel, and railways did not.
But the abstraction hides the physical base.
Synthetic intelligence remains anchored to material substrate. That substrate creates scarcity. Scarcity creates concentration. Concentration creates rent.
Start with chips.
Advanced semiconductor manufacturing is one of the most concentrated industrial systems in the world. TSMC accounted for roughly 70 percent of pure-play foundry revenue in 2025, and Counterpoint Research put its share of that market near 72 percent in the third quarter of 2025. Industry estimates also place TSMC’s share of the most advanced process nodes around 90 percent or higher. Those are the nodes that matter most for AI accelerators, frontier models, and high-performance compute. [1]
This concentration did not arise because chips are easy to monopolize in the ordinary sense. It arose because leading-edge manufacturing is extremely hard to reproduce. It requires extreme ultraviolet lithography, specialized equipment, advanced packaging, deep process knowledge, high yields, trusted customer relationships, and decades of accumulated manufacturing discipline.
Building a leading-edge fab takes years and billions of dollars. Even state-backed attempts to rebuild domestic semiconductor capacity move slowly. Intel’s Ohio semiconductor project, originally expected to begin production in 2025, has been pushed toward 2030 to 2031. Micron’s New York megafab has also faced a two-to-three-year construction delay, with the first facility’s operation pushed toward the end of the decade. [2] [3]
Political will can subsidize the chip estate.
It cannot instantly reproduce it.
And reproduction, when it comes, does not break the estate. Land was never scarce because the planet ran short of dirt. There is plenty of dirt. Land was scarce because position was scarce: the field had to be where it was, and acquiring more of it required already holding the means to acquire it. Compute is the same. The capital now pouring into fabs and data centers does not dissolve the estate. It builds more of it. Every increment of leading-edge capacity is raised by the same narrow set of actors who could afford the last one, financed on terms only they can meet, on ground only they can secure. Supply expands. Ownership does not. The buildout is not the estate eroding. It is the estate compounding.
Below chips sits energy.
AI computation does not only require electricity. It requires dense, stable, predictable electricity at locations where data-center capacity can be built and connected. The International Energy Agency projects that global data-center electricity consumption could more than double by 2030, reaching roughly 945 terawatt-hours, comparable to Japan’s current electricity consumption. The same IEA analysis notes that capital expenditure by five major technology companies surged to more than $400 billion in 2025 as they raced to build the infrastructure needed for frontier AI. [4]
This matters because energy is not a marginal input to AI. It is the binding condition of synthetic production.
Intelligence does not run where power cannot reach.
Below energy sit land and water.
AI data centers require large sites near power infrastructure, cooling capacity, fiber networks, and permissive regulatory environments. Cooling systems require water or energy-intensive alternatives. Local governments increasingly face conflicts over grid pressure, land use, water consumption, tax incentives, noise, environmental impact, and public benefit. In 2023, US data centers directly consumed about 17 billion gallons of water, with hyperscale and colocation facilities accounting for most of that direct use. Hyperscale data centers alone are expected to consume between 16 billion and 33 billion gallons of water annually by 2028. [5]
A model may appear weightless to the user.
The estate beneath it is heavy.
And below all of this sits law.
Data centers require permits, zoning approval, environmental review, energy contracts, and regulatory clearance. Export controls govern who can obtain the most advanced chips. Procurement rules shape who can sell AI into governments, hospitals, banks, schools, and corporations. National AI strategies increasingly ask where computation happens, who owns it, under whose jurisdiction it runs, and who can be excluded from it.
Law does not merely regulate compute.
Law helps make compute scarce.
Scarcity is what allows estate to produce rent.
The physical ground is not a metaphor. It is where the intelligence actually lives.
Why Compute Produces Rent
Rent is not the same as profit.
Profit is the return to productive enterprise. It rewards risk, labor, capital, invention, and execution. Rent is the return to positional control over something others must access in order to produce.
Compute produces rent for four reasons.
First, scarcity at the leading edge.
The most capable models require the most advanced chips. The most advanced chips are produced through supply chains that are narrow, specialized, and difficult to replicate. Scarcity at the foundational layer propagates upward through the entire AI stack. Whoever controls access to leading-edge chips does not merely run a factory. They hold a choke point on the envelope within which frontier AI can exist.
Nvidia’s fiscal 2026 full-year revenue reached $215.9 billion, up 65 percent from the prior year, while data-center revenue rose 68 percent to a record $193.7 billion. In the first quarter of fiscal 2027, data-center revenue hit a further record of $75.2 billion, up 92 percent year over year. Gross margins held in the 71 to 75 percent range across the full year. Margins like that, at that scale, in the enabling hardware of a new production system, are not merely evidence of a strong product cycle. They reveal the pricing power that appears when one layer becomes the bottleneck through which the rest of the economy must pass. [7]
Second, scale thresholds.
Training frontier models, running inference for large user populations, supporting enterprise automation, and providing reliable AI services require infrastructure that most actors cannot own. The capital expenditure required exceeds the capacity of almost every startup, small firm, university, city, nonprofit, and mid-sized state. McKinsey estimates that AI-related data-center infrastructure alone could require roughly $5.2 trillion in cumulative capital expenditure by 2030. [6]
At the frontier, compute ownership is measured not in servers, but in campuses, power agreements, supply contracts, specialized teams, and multiyear financing.
This converts many potential competitors into renters before competition begins.
Third, integration costs.
A compute estate is not only hardware. It is the accumulated knowledge of how to operate hardware at scale, integrated with software systems, security architecture, compliance processes, latency optimization, energy management, enterprise support, and trust layers.
This is not a one-time build. It is an ongoing accumulation of integrated capacity.
The deeper the integration, the harder the estate is to reproduce.
Fourth, interface extraction.
Most tenants do not access compute directly. They access it through interfaces: model APIs, cloud contracts, inference endpoints, enterprise AI platforms, managed services, productivity suites, app stores, and procurement-approved vendors.
These interfaces allow the estate owner to price access, enforce usage policies, modify capabilities, deprecate models, bundle services, throttle usage, adjust safety rules, and capture value from the gap between the cost of running intelligence and the tenant’s dependence on the intelligence produced.
The tenant does not see the estate.
She sees the interface.
The gap between what it costs to run intelligence and what the tenant pays for access is where interface extraction lives. The estate owner prices not to the cost of the silicon but to the value of the dependency. A firm that has reorganized its operations around a model API does not face the price of electricity and amortized hardware. It faces the price of its own inability to function without access. That is the structural basis of interface pricing power, and it is what distinguishes the compute estate from ordinary productive capital. A factory charges for what it makes. The estate charges for what you cannot do without it.
The rent is collected before the ground becomes visible.
Open Weights Are Not Open Estate
The strongest objection to the compute-estate argument is open-source AI.
If foundation models can be released with open weights, then perhaps the estate argument fails. The intelligence becomes replicable. Anyone can download a model, fine-tune it, and deploy it without paying rent to a hyperscaler. The tenant condition dissolves when the land is free.
This objection contains a real insight.
It also misses the structural point.
Open weights are not open estate.
A powerful open model on a laptop is not the same as institutional AI capacity. It is not a substitute for a secure, audited, scalable, insured, compliant, supported, enterprise-grade system that a hospital, bank, government agency, defense contractor, school system, or multinational corporation can adopt.
The gap is not only model capability, although capability matters.
The gap is the full stack required to make intelligence economically and institutionally real.
Running an open model at scale still requires compute. The model weights may be free. The inference infrastructure is not. A startup serving millions of users still needs GPU access, hosting, uptime guarantees, security, latency management, observability, compliance, and capital.
Running an open model inside institutions requires trust. Healthcare, finance, government, and enterprise buyers do not adopt AI systems based on model weights alone. They need liability frameworks, audit trails, certifications, service-level agreements, vendor accountability, security reviews, and integration support.
Running an open model over time requires maintenance. Models degrade relative to newer systems. Security vulnerabilities appear. Fine-tuning requires data infrastructure. Evaluation requires human feedback. Deployment requires monitoring. Scaling requires operations. Each layer pushes the actor back toward compute providers, cloud platforms, managed services, and institutional interfaces.
Open-source AI reduces the barrier to intelligence.
It does not eliminate the estate on which intelligence must run when it becomes economically serious.
The tenant condition is not only about model access.
It is about the physical and institutional substrate required to make intelligence usable at scale.
Open source dissolves one gate.
The estate still stands.
The collision between open weights and institutional requirements is not theoretical. It is the operating reality of every regulated sector attempting to deploy AI. In the United States, federal agencies must satisfy FedRAMP authorization requirements for cloud AI procurement. No self-hosted open-weight deployment currently holds FedRAMP High authorization. The path to institutional AI runs through authorized cloud providers: AWS GovCloud, Microsoft Azure Government, and Google Cloud Government. The open model that cannot travel that path does not reach the hospital, the agency, the bank, or the defense contractor regardless of its capability. The compliance architecture is the gate through which the estate extracts rent from institutions that cannot operate outside it. [8]
Compute Produces Sovereignty
Estate is not only economic. It is political.
In agrarian orders, great estates were not merely sources of private wealth. They were the basis of military power, administrative capacity, and political authority. Control of land meant control of food, manpower, taxation, and obligation. Sovereignty was never fully separate from estate.
Compute is acquiring a similar political character.
Not because AI systems already govern directly. The point is more basic. The capacity to operate at the frontier of intelligence, speed, automation, military planning, cyber defense, surveillance, logistics, research, and institutional execution increasingly depends on access to compute.
A state can remain formally sovereign while becoming computationally dependent.
It may have flags, courts, ministries, elections, borders, laws, budgets, police, and diplomatic recognition. But if its actual administrative, military, industrial, and scientific capacity depends on foreign chips, foreign cloud infrastructure, foreign models, foreign cybersecurity tools, and foreign technical expertise, its sovereignty becomes layered.
Legal sovereignty remains.
Operational sovereignty weakens.
The split matters because modern sovereignty was built on the assumption that legal authority could eventually command operational capacity. A state that could legislate, tax, procure, regulate, conscript, and build could translate formal authority into action. Compute weakens that assumption. A government may have the legal right to govern an AI-mediated institution while lacking the chips, cloud capacity, model access, cybersecurity stack, data-center capacity, and technical labor required to execute that authority independently. It can command in law while renting in practice.
That is not the disappearance of sovereignty.
It is the separation of sovereignty into a visible legal layer and a hidden operational layer.
This is why export controls on advanced chips matter. They are not ordinary trade policy. They are attempts to determine who can sit inside the frontier compute estate and who must remain outside it. Restricting access to advanced accelerators is a way of shaping the future boundary of AI capability, military modernization, industrial automation, and state capacity.
The geopolitics of AI is therefore not only a contest over models.
It is a contest over estate.
Who builds the fabs. Who secures the energy. Who controls the grid connections. Who owns the data centers. Who can finance the buildout. Who can sustain the capex cycle. Who can run training at frontier scale. Who controls inference infrastructure. Who gets access during scarcity. Who writes the laws around the estate.
These are not secondary questions that follow from AI capability.
They are the conditions that determine which capabilities can exist.
Singapore makes this condition visible with unusual clarity.
No state has spent more deliberately on the problem of being small. For six decades Singapore built sovereignty through institutional design, financial reserves, diplomatic positioning, legal infrastructure, and military investment calibrated precisely to the vulnerabilities of a city-state with no natural resources, no strategic depth, and no room for error. It is not naive about dependency. It has studied dependency as a condition of national existence.
And even Singapore cannot solve the compute estate problem.
Singapore launched its National AI Strategy 2.0 in December 2023 with over S$1 billion in committed investment across compute, talent, and industry development. The strategy is explicit about ambition: Singapore intends to be a global AI hub, not merely a regional one. Prime Minister Lawrence Wong described AI compute as a strategic necessity in his Budget 2024 speech, committing further investment over five years. [9]
The operational reality behind the strategy is instructive. Singapore’s National Supercomputing Centre runs ASPIRE 2B, its most advanced system, with over 1,500 Nvidia H200 GPUs delivering up to 115 petaFLOPS. At the launch, Singapore’s Minister for Digital Development and Information described the system’s capacity as “nowhere near the cluster sizes available to frontier model developers.” That phrase was not a complaint. It was an honest accounting of where Singapore sits in the compute hierarchy. [10]
Singapore’s government has migrated over 600 digital services and the bulk of its less sensitive government IT systems to commercial cloud infrastructure operated by AWS, Microsoft Azure, and Google Cloud, running through a centralized platform called the Government on Commercial Cloud. More sensitive workloads now run on AWS Dedicated Local Zones, which are physically located in Singapore but fully managed by AWS. The data stays in Singapore. The infrastructure management does not. [11]
AWS has committed over S$23 billion in Singapore cloud infrastructure through 2028. Microsoft has committed S$5.5 billion through 2029. These commitments are genuine investments in Singapore’s digital capacity. They are also the architecture of a dependency that no sovereign strategy can easily undo. Singapore signed the agreements, built the government services on top of them, and trained the workforce around them. The compute estate belongs to the firms that built it. Singapore’s government runs on lease. [12]
Singapore knows this. The updated NAIS priorities released in May 2026 acknowledge directly that “future supply dynamics remain uncertain” on compute access and commit to charting a path toward greater self-sufficiency. But the document also records what the path toward self-sufficiency looks like in practice: partnerships with Nvidia, OpenAI, Google, and Microsoft; agreements to expand data center capacity built and managed by foreign operators; and a national supercomputer that by Singapore’s own government’s account sits below the frontier threshold for the AI work that will define the next decade of institutional and strategic capacity. [9]
This is what the sovereign tenant looks like from the inside: not failure, not weakness, not ignorance. It is the condition of a capable, well-governed state operating at the frontier of institutional sophistication, still unable to own the ground on which its most consequential future capacities will run. If Singapore cannot close the gap between declared AI sovereignty and operational AI dependency, no state below the United States and China in the compute hierarchy can close it either. That is not a forecast. It is the current structure of the estate.
Compute Produces Tenancy
The tenant condition is not only a platform effect.
It is a substrate effect.
When the ground is expensive to own, when ownership requires capital most actors cannot accumulate, when access to the ground is the precondition of productive activity, and when creating competing ground takes years of specialized effort, tenancy becomes the natural result.
Not as a conspiracy.
As a structural consequence of the substrate.
Most firms will not own the compute required to run frontier AI. They will rent access through cloud services, API contracts, enterprise platforms, hosted inference, managed model deployments, and productivity suites.
Most workers will not own the compute that increasingly mediates their productivity. They will use tools selected by employers, priced by vendors, governed by policies they did not write, and hosted on infrastructure they cannot inspect.
Most startups will not train frontier models. They will build on models trained by others, deploy on infrastructure owned by others, distribute through platforms owned by others, and sell into procurement systems that define institutional legibility before the product is judged.
Most states will not own the full stack required for frontier AI sovereignty. They will depend on alliances, vendors, imports, cloud partnerships, chip allocations, foreign expertise, and infrastructure they cannot fully reproduce.
This does not mean no one can build value as a tenant.
Tenants can be creative, profitable, powerful, adaptive, and innovative. They can build companies, products, workflows, media systems, services, and institutions on top of rented intelligence environments.
But they do not own the environment that makes their agency scalable.
That is the class condition emerging inside the AI economy.
The new divide is not only between labor and capital. It is between those who own the environments of synthetic production and those who rent the right to act inside them.
The Landlords of Intelligence
If compute is estate, the owners of compute occupy the structural position of landlords of intelligence.
That claim needs precision.
It does not mean hyperscalers are malicious. It does not mean compute ownership is automatically unjust. It does not mean the relationship is identical to historical landlordism. The major compute owners took enormous capital risk. They built infrastructure that produced real capabilities. They compete with one another. They employ engineers, operate complex systems, and solve difficult technical problems.
The analogy is structural, not moral.
A landlord of intelligence is an actor that owns the environment in which synthetic cognition becomes economically usable.
Cloud providers, AI infrastructure firms, chip designers, foundries, data-center operators, and model-platform companies do not merely sell products. They increasingly own the ground other actors must access to produce, automate, analyze, distribute, and decide.
The API price is not just a fee.
It is a form of rent.
The cloud subscription is not just an operating expense.
It is a lease on synthetic capacity.
The enterprise AI license is not just software procurement.
It is access to the estate where institutional intelligence runs.
The minimum usage fee that converts AI from possibility into operation is rent. The usage policy that determines what can be done with rented intelligence is lease law in platform form. The deprecation notice, throttling rule, safety update, quota limit, pricing tier, and compliance requirement are all part of the legal architecture of the estate.
Nvidia reveals one layer of this structure. It does not primarily rent compute services to end users. It sells the accelerators on which much of the AI compute estate runs. At that layer, the chip is not merely a component. It is a key to the estate.
The obvious objection is that margins like these invite their own destruction. Custom silicon arrives, AMD narrows the gap, inference costs fall every year, and the rent that looks structural today is competed away tomorrow. On that reading the estate is a product cycle wearing the costume of land.
It is the opposite, and the dynamism the objection points to is the proof. Watch who actually contests the frontier. Google fields its own tensor silicon, Amazon builds Trainium, the largest labs negotiate bespoke supply: the challengers to the bottleneck are not tenants breaking in, they are other landlords. The capital, the fabrication access, and the power required to stand up a competing cluster sit at the same order of magnitude that excluded everyone else to begin with, so the competition runs among the few who can own the ground, conducted over the heads of the many who can only rent the right to act on it. Falling unit costs do not change this. Cheaper cognition is met with more demand for it, and the cluster that confers advantage grows rather than shrinks, so the threshold does not disappear. It moves.
Nor does the competition among the owners reach the people renting from them. When Nvidia, AMD, Google, and Amazon contest the frontier, they are bidding against one another for the right to own the ground, not opening that ground to the tenants beneath it. Whatever the rivalry shaves off the price accrues to whichever landlord wins the contract, or to the few renters large enough to negotiate terms. It does not descend to the firm, the worker, or the state that can only take the rate it is offered. The competition is real. It is also sealed. A renter does not become an owner because the owners are fighting.
This is why the landlords may change while the estate logic remains. The rent was never a property of a particular chip or a particular vendor. It is a property of position: the conditions of synthetic production are held by a few and rented by everyone else, and that structure reconstitutes itself each time one owner is displaced by another who can meet its price. A factory can be out-competed into the ground. An estate only changes hands.
Compute does not merely create products.
It creates positions.
The political question is not whether today’s dominant companies remain dominant forever. The deeper question is whether compute itself will remain the kind of ground that produces landlords.
Every structural feature points in that direction: capital intensity, energy dependency, geographic concentration, technical scarcity, legal protection, institutional trust requirements, and scale advantages.
The State Comes Looking for the Ground
States have always tried to tax the ground.
In agrarian economies, land was the ideal fiscal base. It was visible, valuable, immovable, and impossible to hide. A field could not be shifted to a tax haven. Its value was tied directly to the productive capacity of the society around it.
As capital became mobile and intangible, taxation became harder. Profits could be shifted. Intellectual property could be located in favorable jurisdictions. Digital platforms could generate revenue across borders while locating taxable income elsewhere.
The platform era made value harder for states to locate.
Compute begins to reverse that.
The AI economy still produces intangible services, but the capacity to produce them is increasingly anchored in visible, permitted, energy-intensive, immovable sites. The surplus may flow through software, but the ground beneath it has an address.
Compute is heavy. It is energy-intensive. It requires fixed infrastructure. It depends on land, grid access, cooling systems, permits, and local political bargains. A data center cannot be moved to another jurisdiction as easily as an intellectual-property holding company.
This makes compute a future tax base.
Not easily. States will face tradeoffs. Tax too aggressively, and they may drive investment elsewhere. Tax too weakly, and they may subsidize the very estate that later collects rent from their citizens and firms. But structurally, compute looks more like taxable ground than the mobile capital of the platform era.
Data-center taxes, energy surcharges, cloud-service levies, infrastructure fees, compute-based fiscal instruments, and public claims on AI productivity will all become more politically thinkable as the estate becomes more visible.
This is not only a revenue issue.
It is a recognition issue.
The state eventually comes looking for the ground because the ground is where the surplus concentrates.
The compute estate does not only produce rent.
Over time, it invites taxation.
The Land Beneath Synthetic Production
The earlier transformations now acquire their ground.
Production can detach from labor income. Markets can move into interfaces. The wage can weaken as the central grammar of belonging. Independence can become tenancy. But these changes do not float above the world. They require a substrate.
Compute is that substrate.
It is the physical and economic ground that makes rented intelligence durable rather than temporary, systemic rather than incidental. The tenant is not surrounded only because of bad corporate behavior, regulatory failure, or platform ideology. She is surrounded because the ground on which synthetic intelligence runs is expensive to own, slow to replicate, concentrated among a narrow set of actors, and positioned as the precondition for participation in the AI-mediated economy.
That is the compute estate.
It is not ordinary productive capital.
It is territorial capital: positional, rent-producing, sovereignty-generating, and eventually taxable.
The factory produces output.
The estate produces the conditions under which output becomes possible.
In agrarian economies, the political question was who owned the land.
In industrial economies, it was who owned the machines.
In the AI economy, it is who owns the compute, and what everyone else must pay to think.
Compute is the land beneath synthetic production.
And land, once established as the ground, does not need to announce itself to govern.
Sources
[1] Counterpoint Research’s Q3 2025 foundry analysis put TSMC at roughly 72 percent of the pure-play foundry market; TrendForce separately placed TSMC’s pure-play foundry share at about 71 percent in Q3 2025, up from 70.2 percent in Q2 and near 70 percent for the full year. Industry estimates also commonly place TSMC’s share of the most advanced nodes around 90 percent or higher.
https://counterpointresearch.com/en/insights/global-foundry-2.0-market-Q3-2025-revenue
[2] Intel’s Ohio semiconductor project, originally expected to begin production earlier in the decade, has been delayed toward 2030 to 2031.
https://spectrumnews1.com/oh/columbus/news/2025/02/28/intel-delays-ohio-s-chip-plant-to-2030--2031
[3] Micron’s $100 billion New York megafab faced a two-to-three-year construction delay under the revised schedule in its Final Environmental Impact Statement (November 7, 2025); the first fabrication facility’s opening moved from 2028 to 2030, with construction now beginning in 2026 and the full four-fab buildout staggered through 2041.
https://www.constructiondive.com/news/micron-delay-construction-new-york-megafab/805622/
[4] The International Energy Agency projects global data-center electricity demand could more than double to roughly 945 TWh by 2030, comparable to Japan’s current electricity use. Its 2026 analysis also describes surging data-center investment, including more than $400 billion in 2025 capital expenditure by five large technology companies.
https://www.iea.org/reports/key-questions-on-energy-and-ai
[5] Pew Research Center, citing estimates from a 2024 Berkeley Lab report commissioned by the US Department of Energy, reports that US data centers directly consumed about 17 billion gallons of water in 2023, with hyperscale and colocation facilities using 84 percent of that total. It also reports that hyperscale data centers alone are expected to consume between 16 billion and 33 billion gallons annually by 2028.
https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/
[6] McKinsey & Company estimates that data centers equipped for AI processing loads could require roughly $5.2 trillion in capital expenditures by 2030, out of nearly $7 trillion in total data-center infrastructure spending.
https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers
[7] Nvidia reported fiscal 2026 full-year revenue of $215.9 billion, up 65 percent from the prior year, with data-center revenue rising 68 percent to a record $193.7 billion and full-year gross margin of 71.1 to 71.3 percent. In Q1 fiscal 2027 (ended April 26, 2026), data-center revenue reached a record $75.2 billion, up 92 percent year over year, with gross margin of 75 percent.
https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Fourth-Quarter-and-Fiscal-2026/default.aspx
[8] On FedRAMP authorization requirements for federal AI procurement: Office of Management and Budget, FedRAMP program documentation. No self-hosted open-weight AI deployment currently holds FedRAMP High authorization. Federal agencies procuring cloud-based AI must use FedRAMP-authorized providers, which in practice means AWS GovCloud, Microsoft Azure Government, and Google Cloud Government. https://www.fedramp.gov
[9] Singapore Ministry of Digital Development and Information, Update to Singapore’s National AI Strategy, May 2026. The update acknowledges that “future supply dynamics remain uncertain” on compute access and sets out refreshed priorities including expanded data center capacity and deepened partnerships with major AI firms. Prime Minister Lawrence Wong described AI compute as a strategic necessity in Budget 2024. https://www.mddi.gov.sg/newsroom/update-to-singapore-s-national-ai-strategy--refreshed-priorities-to-harness-ai-for-the-public-good-factsheet/
[10] Minister Josephine Teo, Opening Address at NSCC Launch of ASPIRE 2B Supercomputer, Singapore Ministry of Digital Development and Information. The minister stated that ASPIRE 2B’s capacity of up to 115 petaFLOPS with more than 1,500 Nvidia H200 GPUs is “nowhere near the cluster sizes available to frontier model developers.” https://www.mddi.gov.sg/newsroom/opening-address-by-minister-josephine-teo-at-national-supercomputing-centre--nscc--s-launch-of-aspire-2b-supercomputer/
[11] GovInsider, “Key lessons from the Singapore government’s ambitious whole-of-government cloud migration strategy.” GovTech’s Government on Commercial Cloud platform hosts over 600 government digital services running on AWS, Microsoft Azure, and Google Cloud. More sensitive workloads run on AWS Dedicated Local Zones, which are physically located in Singapore but fully managed by AWS. https://govinsider.asia/intl-en/article/key-lessons-from-the-singapore-governments-ambitious-whole-of-government-cloud-migration-strategy
[12] AWS press release: AWS to invest an additional S$12 billion in Singapore by 2028, bringing total planned investment to over S$23 billion. https://press.aboutamazon.com/sg/aws/2024/5/aws-to-invest-an-additional-sg-12-billion-in-singapore-by-2028-and-announces-flagship-ai-programme. Microsoft commitment of S$5.5 billion in Singapore AI and cloud infrastructure through 2029: Computer Weekly, “Microsoft to invest $5.5b in Singapore’s AI and cloud infrastructure,” April 2026. https://www.computerweekly.com/news/366641114/Microsoft-to-invest-55b-in-Singapores-AI-and-cloud-infrastructure