The $690 Billion Lock-In Reshaping AI’s Architecture

Capital, once committed, defends itself. That is the lesson a $660 billion to $690 billion AI infrastructure commitment is about to teach the technology industry, whether the industry is ready to learn it or not. The five largest US hyperscalers have committed to that range for capital expenditure in 2026, nearly doubling their 2025 levels and channelling the bulk of it into one architectural assumption: more parameters, more clusters, more GPUs, more data centers. The decision is being made at a scale that does not bend easily when new research arrives, and the new research is arriving.

Layered on top of corporate capex sits a project at a different size class. The Stargate joint venture, announced in January 2025, targets $500 billion in AI infrastructure investment by 2029 with an initial $100 billion deployment and roughly 7 GW of capacity planned across five US sites. Against that wall of capital, the pure-play AI vendors that are supposed to consume it look small. OpenAI closed 2025 with about $20 billion in annual recurring revenue, and Anthropic’s run rate passed $9 billion by January 2026. The entire cohort of pure-play AI vendors, including Cohere, Mistral, and Perplexity, likely accounts for less than $35 billion in projected 2026 revenue.

The build-out is the size of a permanent architectural choice. Futurum’s February 2026 breakdown of hyperscaler AI spending lays out the per-company plans: Amazon at $200 billion, Alphabet at $175 billion to $185 billion, Microsoft tracking toward $120 billion or more, Meta at $115 billion to $135 billion, and Oracle at $50 billion. Every one of those lines assumes the same kind of hardware at the bottom of the stack. That is what a commitment becomes once it is large enough.

Inside the Locked-In Supply Chain

The hardware underneath that $660 billion to $690 billion has its own bottleneck, and the bottleneck is a single company. ASML is the only maker in the world of extreme ultraviolet lithography systems, the only class of machine capable of patterning the smallest blueprints on the most advanced AI chips, per CNBC’s 2025 profile of the company’s newest High-NA system. The newest machine, the High-NA Twinscan EXE, carries a price tag of more than $400 million, weighs more than a double-decker bus, and takes seven partially loaded Boeing 747s to ship. Only five have ever been delivered. The buyers are the same three names every quarter: Intel, TSMC, and Samsung.

ASML’s older EUV systems are not cheap either, with prices starting at $220 million, and the company sold 44 of them in 2024. ASML’s CEO Christophe Fouquet told CNBC that the company is the exclusive maker of EUV and that Moore’s law is the discipline that keeps pricing in check. That is a comfortable position for the supplier. It is a brittle one for everyone who depends on the supply.

Re-entry into the leading edge is not a quarterly decision for any foundry that has stepped away. In 2018, GlobalFoundries walked away from the leading-edge race, a move its CEO Tom Caulfield explained to CNBC with a single line: “the expenses for a leading edge fab were doubling every two or three years. And right now we’re looking at investments of campuses upwards of $100 billion.” GlobalFoundries now makes only 12-nanometer chips and above, what it calls essential chips, and has built a multi-billion-dollar US defense business around staying out of the leading-edge capital race. The 2018 decision was rational at the time. Six years later it is also permanent: a foundry cannot simply buy its way back into the leading edge without committing the same $100 billion-plus a new entrant would face today. That is how lock-in works in capital-heavy chains. Each rational choice narrows the next one, and the consumer hardware that depends on those foundries pays the bill, as the cost of a flagship phone in 2027 now shows. CNBC documented the exit in the 2018 decision Caulfield walked CNBC through, and the picture has not softened since. The High-NA lithography machine CNBC profiled in 2025 shows what the locked-in end of that supply chain looks like. How rising chip costs are already reshaping consumer hardware shows the other end.

How Electricity’s War Repeated Itself

The pattern is older than AI and older than semiconductors. In the late 1880s, Thomas Edison had built his company around Direct Current, a working electrical system with installed plants, financiers, and trained crews. George Westinghouse and Nikola Tesla pushed Alternating Current, which transmitted power over longer distances at lower cost. The two sides converged in what became known as the war of the currents.

Alternating Current won on transmission economics, not on the preferences of the dominant incumbent. By the end of 1887 Westinghouse had 68 alternating-current power stations operating against Edison’s 121 DC stations, according to the standard histories of the contest. Edison did not yield on the merits of the case. He campaigned against AC on safety grounds, staged public demonstrations designed to frighten the public about AC’s lethality, and resisted adoption because the shift threatened the value of the system he had already financed. The engineering case for AC was strong. The economic case for preserving DC was strong too, for the people already invested in it. The final outcome was decided by the cost of staying with the existing system, not by the engineering merits alone.

The same dynamic shows up in software formats and keyboard layouts. W. Brian Arthur, the Santa Fe Institute economist, named the underlying tendency path dependence in his 1994 book Increasing Returns and Path Dependence in the Economy, documented on the institute’s own page as a foundational text on how an early lead compounds into a permanent one. Arthur’s 1994 framework for technology lock-in describes how an early technology does not have to remain the best to keep winning once the surrounding ecosystem has organised around it. Suppliers, training programs, regulators, and financiers each have a stake in the existing path. The argument is not about conspiracies. It is about incentives.

The 2026 AI buildout is repeating the move at a much larger scale. Capital, supply chains, and procurement are consolidating around one architectural assumption. The alternative architectures are not absent from the research record. They are absent from the procurement pipeline. That gap is what gives the present moment its edge.

The Reactor the United States Walked Away From

The molten salt reactor story is the cleanest second example. Oak Ridge National Laboratory’s Molten Salt Reactor Experiment operated from January 1965 through December 1969, logging more than 13,000 hours at full power, per ORNL’s own history of the project. On October 8, 1968 it became the first reactor ever to run on uranium-233, the isotope bred from thorium. The reactor was designated a nuclear historic landmark in 1994.

Alvin Weinberg, who had directed ORNL since 1955, championed the molten salt line because the design runs at low pressure, produces less long-lived waste, and resists weapons proliferation more effectively than the conventional water-cooled reactor. He kept advocating the molten salt path against the prevailing liquid-metal fast-breeder direction pushed by the Nixon administration. The Children’s Museum of Oak Ridge biographical record on Weinberg states that he was fired in 1973 after 18 years as director, removed for continuing to back the technology he had built. The MSR programme itself ran 1957-1976, according to the World Nuclear Association’s molten salt reactor history, and was wound down. The technology did not lose on its merits. It lost the institutional contest.

China picked up what the US put down. The Chinese Academy of Sciences’ Shanghai Institute of Applied Physics launched the Thorium Molten Salt Reactor programme in 2011 with a start-up budget of $350 million, per the World Nuclear Association. On November 4, 2025, SINAP announced the first-ever thorium-to-uranium fuel conversion in an operating molten salt reactor, a milestone confirmed in the academy’s own English-language release. The 2 MW experimental reactor at Wuwei is currently the only operating molten-salt reactor in the world loaded with thorium fuel. Oak Ridge’s record of the 1965-1969 reactor documents the original US success, and SINAP’s November 2025 thorium-uranium conversion announcement documents what happened when capital followed.

What the AI Researchers Now Say Out Loud

The reckoning is not arriving only as outside commentary. Ilya Sutskever, co-founder of OpenAI and now head of Safe Superintelligence, told a Reuters interviewer that “the 2010s were the age of scaling, now we’re back in the age of wonder and discovery,” framing the 2020-to-2025 period as one in which scaling alone carried the field forward and announcing that the next breakthroughs will require research paths that do not run on the same hardware. A January 2026 paper from MIT’s Initiative on the Digital Economy put numbers on the same intuition, concluding that “as more compute is invested in training AI models, the marginal gains in performance decrease substantially.” Stanford’s 2026 AI Index is now debating whether pre-training itself is exhausting its headroom, with the panel split between those who see capability continuing to accelerate and those who see the curve bending. The signal is no longer one outsider’s view. It is the field’s own internal debate.

Capital has not yet caught up to the signal. The $660 billion to $690 billion in 2026 capex is still flowing into the architectural assumptions of the age of scaling: more parameters, more clusters, more GPUs. The phrase “age of wonder and discovery” means research paths that do not run on the same equipment. A budget that buys only the previous era’s hardware cannot, by itself, fund the next one.

The Architectures That Don’t Fit the Bill

The alternative paths are not hypothetical. A short list of them sits in the published research record, each one a different bet on how machine intelligence should run.

• Neuromorphic chips, which emulate spiking neurons in silicon
• State-space models, which compete with transformers on long-sequence tasks
• Liquid neural networks, which adapt their behaviour after training
• Photonic accelerators, which compute with light rather than electrons
• Analogue computing, which trades digital precision for raw efficiency

Each of these designs competes for the same research dollars, the same graduate students, and the same procurement budget. The 2026 capex flows almost entirely toward transformer-era GPU clusters. Research money follows the dominant path, procurement follows the dominant path, and talent follows the dominant path.

The official advice on this risk has started to harden into a single recommendation: build model-agnostic architectures. Companies that keep the door open between model providers are hedged against a future in which one architectural line falls behind. That advice is sound for a buyer of AI services. It does not address the larger problem, which lives deeper in the stack. The ASML monopoly and the leading-edge foundry exit are constraints on the supply side, not the application side. The hardware pipeline is what is hardening. How rising chip costs are already reshaping consumer hardware is the visible downstream symptom.

China’s TMSR success offers the template for what an alternative path can do when capital gathers around it. The technology did not change between 1969 and 2011. The institutional decision to fund it did. The same pattern is available to AI. The risk is that by the time the research case becomes impossible to ignore, the sunk costs of the current path will have hardened into a generational lock-in. That is what history suggests happens when money freezes progress.

Frequently Asked Questions

What is AI infrastructure lock-in?

AI infrastructure lock-in is the convergence of capital, supply chains, and procurement on a single architectural path, such that switching to an alternative becomes prohibitively expensive even if a better one emerges. In 2026, that lock-in is taking the form of a $660 billion to $690 billion hyperscaler capital commitment directed almost entirely at transformer-era GPU clusters, plus a $500 billion Stargate infrastructure target layered on top. The numbers come from Futurum’s February 2026 analysis of US hyperscaler plans. Once committed at that scale, the capital base defends itself: research money, talent, and procurement all flow toward the path the money has already chosen.

Why is ASML’s EUV monopoly central to the lock-in?

ASML is the only company in the world that builds extreme ultraviolet lithography systems, the machines required to pattern the most advanced AI chips. Per CNBC’s 2025 profile, the newest High-NA Twinscan EXE carries a price tag above $400 million, weighs more than a double-decker bus, and has shipped only five units in total. The buyers are limited to Intel, TSMC, and Samsung. Older EUV systems start at $220 million, and ASML sold 44 of them in 2024. With one supplier, one buyer pool, and per-tool prices measured in hundreds of millions of dollars, the lithography layer sets a hard floor on who can build the hardware the AI build-out assumes.

How does the molten salt reactor story parallel today’s AI?

The molten salt reactor story shows what happens when a technically credible alternative loses the institutional contest. Oak Ridge’s MSRE operated from January 1965 through December 1969, logging more than 13,000 hours at full power and becoming the first reactor ever to run on uranium-233 on October 8, 1968, per ORNL’s own history. The director who championed the technology, Alvin Weinberg, was fired in 1973 after 18 years leading the lab, according to the Children’s Museum of Oak Ridge biographical record. The US programme wound down in 1976. China then launched the Thorium Molten Salt Reactor programme in 2011 with a start-up budget of $350 million, and SINAP announced the first-ever thorium-to-uranium fuel conversion in an operating MSR on November 4, 2025. The technology did not change. The institutional decision to fund it did.

Is AI actually hitting a scaling wall in 2026?

The most credible internal voices in the field now say yes, at least for raw pre-training scale. Ilya Sutskever, co-founder of OpenAI and now head of Safe Superintelligence, told a Reuters interviewer that the 2010s were the age of scaling and the field is now back in the age of wonder and discovery, a direct statement that more compute alone no longer produces proportional gains. MIT’s Initiative on the Digital Economy published a paper in January 2026 concluding that marginal gains in performance decrease substantially as more compute is invested. Stanford’s 2026 AI Index notes the debate is unresolved on capability more broadly, but the question is no longer whether the scaling curve is bending. The debate is what comes next.

What would it take to keep alternative AI architectures alive?

Two policy moves would matter most. The first is model-agnostic procurement, in which buyers of AI services commit to architectures that can swap model providers without rebuilding the surrounding infrastructure. The second is dedicated public R&D funding for non-transformer architectures at the hardware and model level, on the thorium reactor template where the technology did not change but the institutional decision to fund it did. The risk in 2026 is that both moves arrive after the lock-in has already hardened. The $660 billion to $690 billion in 2026 hyperscaler capex is being deployed right now, and every quarter it grows harder to fund the path that would make yesterday’s infrastructure look overbuilt.