The United States is not underinvesting in AI. Amazon, Microsoft, Alphabet and Meta are guiding to roughly $725 billion of combined capital expenditure in 2026, up about 77% in a year. That is private corporate capex, not government spending — and the distinction is most of the problem.

Nearly all of it funds closed frontier models and the infrastructure that serves them: an asset class a foreign ministry cannot hold, inspect, host on its own soil, or adopt on its own terms. Meanwhile the layer ministries can adopt has moved. China accounted for roughly 41% of geographically attributed Hugging Face downloads over the trailing year. Qwen reached 69% of new monthly derivative models in February 2026, while Meta's share peaked at 44% in August 2024 and has since fallen to around 11%.

The capability cushion is thinning fast. Stanford's AI Index put the gap between the top American and top Chinese model at 2.7% as of March 2026, and Epoch AI measures the average gap between leading closed models and open equivalents at roughly four months since January 2026.

And the gap is not where most proposals assume. Frontier capability has moved into post-training — reinforcement learning, verifiable rewards, agentic environments — which American frontier labs do best and keep closed, and which Chinese open labs ship. The non-Chinese open ecosystem sits between the two with strong base models and a fragmented reinforcement loop. The structural reason is not ideological. For an American lab the model is the revenue: publish the weights and the serving margin goes to whoever hosts it. The open layer has no clear owner on an American balance sheet. Public support exists but is small and science-scoped — NSF and NVIDIA committed a combined $152M to Ai2's fully open OMAI programme, against $725B of private capex aimed elsewhere. Congress has not passed a comprehensive AI statute at all.

The rest of this brief is the evidence for that sentence: how the board split, what each layer looks like now, why more capital cannot fix it on the timelines available, and how much of the American economy is currently levered to the outcome.

For twenty years the enterprise technology fight was a share fight. On-premise to cloud. General-purpose CPUs to accelerated computing. Different winners each time, but always the same board: every vendor could sell to every buyer, and no ministry had to ask whose law the infrastructure sat under.

That mattered more than it looked, because the thing being sold stopped being optional along the way. Cloud is not a productivity add-on. It runs hospital records, core banking, airline scheduling, election administration and government services. It became critical infrastructure while everyone was still arguing about migration timelines.

AI is repeating that arc faster, and this time the board is different. The third fight is not about share. It is about jurisdiction.

None of this was a single decision. American technology policy, Chinese industrial policy and a run of sovereignty shocks moved on separate clocks and arrived at the same place: a world where buying a stack is a statement about whose jurisdiction you accept. The sovereignty track is the one most often left out of the chip debate, and it is the one doing the most work.

A stack is not a shopping list. Each layer sets the switching cost of the layer above it. Kernels written for one accelerator port at real expense. Models tuned to one memory profile need rework. Developer habits, once formed, are slow and costly to change. Porting is possible — compiler and framework layers exist precisely to make it so — but choosing a column is a decade-length commitment that a procurement officer makes in a quarter.

What follows walks the stack from the bottom. At each layer the same question: which column is winning the decision, and why. The answer is uncomfortable in four places and reassuring in one, and the reassuring one has a 2028 expiry date.

Nvidia went from near-total dominance of Chinese AI compute in 2023 to roughly 40% in 2025, with Bernstein forecasting about 8% for 2026. Huawei went the other way, from close to nothing to an estimated 50–60%. No product failure caused this. Export controls removed the incumbent, and Beijing converted the vacuum into a captive market.

The mechanism is a captive market rather than a cheque. A reported draft plan would commit roughly $295B (¥2 trillion) over five years to a national AI computing grid, targeting at least 80% domestic core technology. It remains in planning. In May 2026, nine categories of domestic accelerators — Huawei, Alibaba, Biren, Moore Threads — cleared a state security review qualifying them for government and security-sensitive deployment. Demand on that side is not being left to the market.

Private capital then does the rest. ByteDance committed $5.6B in orders for Huawei's Ascend 950PR — roughly 350,000 chips, close to half of Huawei's 2026 production target. The customer is simultaneously validator and co-financier of the silicon it depends on. And the 950PR ships with a CUDA-compatible software path, which attacks the one moat export controls never touched.

This is the honest part of the brief, and it should be said before an opponent says it. Chinese silicon is a generation behind and constrained where it is hardest to fix. One caveat on the comparison below: the 950PR is optimised for prefill and recommendation rather than general training, so it is not a like-for-like substitute for a B200 in a training cluster — Huawei's training-oriented part is the 950DT.

So the Eastern column does not win on peak capability and will not for at least two years. It wins on availability, price, and policy alignment — three things a ministry buying inference capacity in 2026 weighs more heavily than peak FLOPS.

Treating a two-year hardware lead as the strategy is the error. The lead is real. It is also the only thing on this page trending in the West's favour, and it does not transfer upward: a buyer who standardises on the Eastern column at the model and developer layers will not switch back when the hardware gap closes. They will already have rebuilt around it.

Alibaba Cloud is the fourth-largest cloud provider in the world, operating in about 30 regions as of August 2026. Huawei Cloud runs 101 availability zones across 34 regions, concentrated in Asia, the Middle East and South America. New capacity keeps landing in exactly the markets that are deciding: Johor, Bangkok, the Gulf, Latin America.

The commercial engine underneath is the part policymakers routinely miss. Alibaba's Cloud Intelligence revenue reached ¥158.1B in FY2026, up 34%, with AI product revenue posting triple-digit growth for eleven consecutive quarters. Open weights are the customer acquisition cost; paid inference is the margin. Giving the model away is not generosity. It is the funnel.

US providers compete hard here, but they arrive selling a closed model on a metered API, into governments whose stated requirement is control. Chinese providers arrive with weights the ministry can hold, hosted on infrastructure the ministry can site domestically, at a price that clears a treasury review. On the specific criterion sovereign buyers say they are using, the Eastern column has the better offer.

This is the layer where the argument is decided, because this is the layer that sets defaults. China accounted for roughly 41% of geographically attributed Hugging Face downloads over the trailing year. In June 2026, Chinese open models held five of the top ten trending slots.

Capability was supposed to be the West's insurance. It is thinning. Stanford's AI Index put the gap between the top American and top Chinese model at 2.7% as of March 2026. The gap between leading open models and the best closed systems has compressed from six-to-nine months to roughly two-to-three.

Then August. Zhipu revealed that the anonymous model topping coding leaderboards was GLM-5.3-Flash — 320B parameters, 18B active, MIT licence — scoring 57 on the Artificial Analysis intelligence index, level with Claude Opus 4.8. It processed 62 trillion tokens before formal release, the largest launch OpenRouter has recorded. Zhipu states the whole thing ran on a cluster of more than 100,000 domestically produced accelerators, at hardware efficiency and cost per token it claims are comparable to mainstream Nvidia GPUs.

A precision note worth keeping: Z.ai did not name the accelerator vendor in its GLM-5.3-Flash launch material, so the defensible claim there is production-scale inference on domestic silicon. The separate training claim belongs to GLM-5, released 11 February 2026, which Reuters reported was trained entirely on Huawei Ascend chips using the MindSpore framework with no NVIDIA hardware. Specific cluster sizes and fab nodes circulating for that model are secondary reporting, not primary disclosure.

DeepSeek is the more structural signal, and the one that does not reverse. V4 shipped with day-zero adaptation across Ascend, Cambricon, Hygon and Moore Threads. More importantly, its kernels are no longer written purely in CUDA but in a domain-specific language, TileLang, that compiles to different hardware. The decoupling is happening at the compiler layer, not the chip layer. Hardware leads are temporary. A portable software layer is not.

Open-weight API access runs at roughly one-eighth the list price of closed — about $0.23 per million tokens against $1.86. Capability follows within months. For most production workloads, the rational choice is the cheap model that will be good enough by the time the project ships.

Every developer who fine-tunes on a Chinese base model this year produces the tooling, the evaluation harnesses and the tutorials that make the next developer's choice easier. Defaults are not opinions. They are accumulated infrastructure, and they are being accumulated now, in the open, at a rate no closed ecosystem can match by pricing.

Almost every proposal to help American open models assumes the problem is a big training run. It is not. Frontier capability has moved decisively into post-training — reinforcement learning, verifiable rewards, agentic environments, evaluation harnesses — and that is the one part of the pipeline the non-Chinese open ecosystem does not have.

The shift itself is not in dispute. OpenAI's Dan Roberts said in May 2025 that the lab was prioritising large-scale RL over pre-training and expected RL compute to exceed what had been spent on pre-training. Cursor disclosed that Composer 1.5 used more compute on RL post-training than on pre-training its base model — the only plainly quantified lab disclosure of the inversion. The direction of travel below is an analyst estimate, but the endpoint matches what the labs describe.

Now the part that should worry a policymaker. Chinese open labs have built real post-training pipelines — DeepSeek, Alibaba, Moonshot, MiniMax and Z.ai have all invested heavily in post-training, coding capability and permissive release as a distribution strategy. American frontier labs have the best pipelines of all, and keep them closed. The non-Chinese open ecosystem sits in the gap between the two.

Ai2's Nathan Lambert, who led post-training on Olmo and Tülu, puts the asymmetry precisely: open pre-training infrastructure is strong — labs raising hundreds of millions simply pick up NVIDIA's Megatron-LM and make it work — while the RL revolution brought a large jump in post-training infrastructure complexity that the open side is still fragmented across many competing libraries. Prime Intellect's INTELLECT-3 report says the same thing from the builder's seat: open frameworks, RL environments, evaluations and stable recipes all have notable shortcomings against the proprietary pipelines inside frontier labs.

The four largest US hyperscalers are guiding to roughly $725 billion of capital expenditure in 2026, up about 77% from $410B in 2025, with consensus pointing above $900B in 2027. The problem is not the size of the American bet. It is where the bet cannot go.

For an American lab, the model is the revenue. Publish the weights and anyone can serve it: the serving margin goes to the host, the product margin goes to whoever builds on top, and the lab keeps the goodwill. The open layer has no owner on any American balance sheet. It is not that the market refuses to fund it. It is that few firms can justify it commercially.

Note what this means for the standard American rebuttal. "We out-invest them ten to one" is true and irrelevant, because the investment is going into an asset class the rest of the world cannot adopt on its own terms. Capex does not travel. Weights do.

Start with the fact that cuts against the panic. The United States is not short of compute. It holds about 32 GW of operational data-centre capacity against roughly 21.7 GW for the next fourteen countries combined, and has another 70.6 GW under construction or planned against 29.2 GW for all of them together. The Americas host 43.4 GW of operational capacity, 93.6% of it inside the US. On installed base, nobody is close.

So the honest framing is not that America cannot build. It is that America already built, and is now throttled on the margin. The instinct in Washington is to answer a competitive threat with more volume — it is being tried, at $725B a year — and the reason that does not settle this is that the constraint has moved from capital to electrons, transformers, skilled trades and local consent, none of which respond to a larger appropriation. Everything below is about the rate of new additions, not the installed base.

The numbers underneath are stark. Roughly 2,000 gigawatts of generation and storage are actively seeking connection to the US grid — about double the capacity of the entire existing US power plant fleet. Historically, only 13% of the capacity that entered the queue between 2000 and 2019 had reached commercial operation by the end of 2024, while 77% was withdrawn. ERCOT was tracking a large-load queue near 410 GW as of April 2026, with data centres accounting for around 87% of it. PJM missed its reliability target by 6,625 megawatts in its December 2025 auction — the first shortfall in the operator's history.

This is already biting the largest buyers. On Microsoft's January 2026 earnings call, CFO Amy Hood told investors that demand continues to exceed available supply and that Azure capacity constraints — driven by power and space rather than chip procurement — would persist at least through June 2026, against a commercial backlog that had more than doubled to $625 billion. Of the US data-centre capacity expected in 2026, only about a third is under active construction. A campus joining the Northern Virginia queue in mid-2026 cannot realistically expect utility power before 2030 to 2033, regardless of how much capital stands behind it.

The political ceiling arrived early

Even where the engineering can be solved, consent is running out first. Utilities requested a record $31 billion in rate hikes in 2025, more than double the prior year. Residential electricity prices rose about 7% over 2025, and roughly 7.3% in the twelve months to April 2026. A March 2026 Gallup survey found seven in ten Americans oppose data centre construction in their own communities, and community pushback blocked or delayed at least 75 projects worth roughly $130 billion in the first quarter of 2026 alone.

New York has ordered the first statewide moratorium on new hyperscale facilities. At least eighteen states have introduced special rate-class legislation, with moratorium bills filed in several more. Rising electricity rates are now expected to be a live issue in the November midterms. The buildout has become an affordability question, which means its pace is set by voters rather than by capital markets.

Fabrication is no easier. A single leading-edge fab costs $15–20B before tools arrive. TSMC's first Arizona fab slipped from 2024 to 2025 and the second from 2026 to 2027–28, with construction running 30–50% above the cost of equivalent facilities in Taiwan. McKinsey and SEMI project a US shortfall of roughly 157,000 skilled semiconductor workers. Reshoring the physical base is correct policy and it is a decade-long project.

The conclusion a policymaker should draw is not despair about concrete. It is that the binding constraint and the available lever are in different places. Every physical layer of the Western column is on a four-to-ten-year clock and politically contested at home. The model layer ships over a network — no queue, no substation, no zoning hearing — and can move globally in a weekend.

And this is where the installed-base fact becomes the whole argument. Closing the post-training gap does not require a single new data centre. The world's largest fleet of accelerators is already operating on American soil, most of it privately held, overwhelmingly pointed at closed models and commercial inference. What the open layer needs is an allocation of capacity that exists, not the construction of capacity that does not. That is a far cheaper ask than a build programme, and it is the only version of the ask that can be delivered inside the window in which these decisions are being made.

The usual objection to funding an open stack is that it is a niche concern. The exposure numbers say otherwise. The United States has concentrated an extraordinary share of its growth and its equity market value in a single vertically integrated stack, at exactly the moment international demand for that stack is being competed away.

Put those together and the shape of the risk is clear. Roughly a third of the index, roughly a third of the country's measured growth, and the earnings engine of the whole market are levered to the same capex cycle — and the demand curve underneath it is being contested by a column that ships for free at the model layer.

Set against that, the cost of standing up a funded open frontier programme is a rounding error — a low single-digit fraction of one quarter's hyperscaler capex, against an exposure measured in tens of trillions. This is cheap insurance on one of the most concentrated bets in the history of American capital markets.

Take Singapore, the most sophisticated buyer in Southeast Asia and the one with the most freedom to choose. In the 2026 ISEAS survey — which polls regional opinion-makers in policy, academia, business, civil society and media rather than a representative public sample — 66.3% of Singaporean respondents picked China over the United States in a forced choice, up from 47.1%. AI Singapore's latest SEA-LION release is built on Qwen as its foundation model.

The route around the question is also closing, though from the American side rather than the Chinese one. In September 2025 the US Bureau of Industry and Security adopted an Affiliates Rule extending Entity List and Military End-User restrictions to any foreign entity 50% or more owned by a listed party, wherever it is incorporated — a look-through that would end the practice of routing a China-exposed venture through a Singapore entity. BIS stayed the rule in November 2025; it is currently scheduled to take effect on 10 November 2026, which gives the region a deadline rather than a reprieve.

The structural trap is regional. ASEAN supplies the land, the electricity, the connectivity and the market for the AI economy, while depending on external powers for chips, foundation models, cloud platforms and standards. States keep formal sovereignty and lose operational control of the systems running critical functions. Every government in the region can see this. It is why the money is moving.

These buyers are not asking to be told which superpower to trust. They are asking for an arrangement in which that question does not have to be answered permanently. Right now no one is offering one, so they are answering it by default — layer by layer, procurement by procurement, in the direction of whoever shipped weights they could hold.

This is not a call for neutrality between Washington and Beijing. It is the opposite. A third stack is the only instrument that keeps the countries currently choosing from defaulting into the Eastern column — because it gives them the thing they actually want, which is not an American vendor but a seat at the table.

The distinction that makes it work: governance is shared, supply chain is not. Participating states hold real authority over data, evaluation and release. The silicon, the interconnect and the security review stay inside allied jurisdictions. A country gets sovereignty over how the thing is run without acquiring dependency on Chinese hardware — which is precisely the trade the Eastern column cannot match, because Beijing will not share governance and cannot offer allied supply.

The cost of not doing this is not measured in market share. It is measured in the number of governments whose public services, defence tooling and national models are running on a stack the United States does not supply, cannot inspect, and will not be able to price out once the switching costs have set.