Per query costs for running frontier AI models have collapsed, Chinese open source labs are closing the capability gap with US systems, and the hundreds of billions in announced US data center investments were sized for a premium pricing window that may not return.
The math on US AI's trillion-dollar data-center build-out is breaking in public. Per-query costs for running frontier large language models have cratered over the past year as Chinese open-source labs close the capability gap and US labs cut prices to defend share. The investment thesis underwriting hundreds of billions of dollars in announced compute infrastructure was sized for a world of durable premium pricing. That world is the one disappearing.
The case for treating that as a structural problem, not a passing cycle, comes most clearly from the AI critic Gary Marcus. In a recent Substack essay, Marcus lays out what he sees as three structural flaws in the current US AI paradigm. First, the underlying approach trains on the entire public internet, which makes it expensive to develop and operate. Second, the resulting systems are unreliable enough that the premium prices frontier labs have charged were never going to last once competitors arrived. Third, in Marcus's view, the basic approach has hard limits that more compute does not solve.
That is one informed position, not a measurement. But the competitive data behind it is now being reported independently. CNBC reported on June 26, 2026 that China's Zhipu is closing in on top US AI models even as Anthropic and OpenAI are constrained by US policy. The pattern Marcus named, replicable approach meets price competition, is no longer just a Substack thesis. It is showing up in benchmark coverage and in the kind of cost disclosures frontier labs would previously have kept quiet.
The US policy backdrop makes the squeeze sharper. According to a Washington Post report from June 15, 2026, a 90-minute White House deadline on AI policy sparked one of Silicon Valley's biggest recent fights over export controls and compute curbs. The fight matters here because the export-control regime is one of the few tools US policymakers have to slow Chinese frontier-AI progress. If the capex underwriting the US side of that race was sized for a premium-pricing world that is no longer arriving, the policy tools are propping up a thesis whose economics are already under pressure.
There is a real counter-argument, and it deserves air. In a June 26, 2026 Washington Post opinion piece, contributors argue that US caution on the most advanced AI may be strategically defensible even if it gives Chinese labs breathing room, because frontier-AI risk is itself a strategic variable, not just a competitive one. That view does not redeem the capex math, but it does puncture the assumption that "faster US frontier AI" and "better US AI economics" are the same objective. They may pull in opposite directions.
What follows from that is not a hand-wavy call to pivot. It is a separation of two bets that have been bundled together. Commodity inference, running general-purpose large language models at scale for chatbots, search, and productivity tools, is the layer where prices are collapsing, where Chinese open-source labs are catching up fastest, and where moats are shallow. The investor question there is whether any US lab can earn its cost of capital before prices reach a commodity floor.
Science-and-medicine AI is a different bet. Biomedical foundation models trained on curated scientific literature, protein structures, and clinical data, scientific-discovery agents that automate wet-lab cycles, FDA-relevant tooling for trial design and post-market surveillance: these require specialized training data, regulatory validation, and integration with clinical workflows that no open-source Chinese release can shortcut. The moats run through regulatory approval, proprietary datasets, and the slow, expensive work of clinical validation. That is the bet where the announced US compute investment has a plausible path to durable margins, and where the public-good case is also strongest.
Marcus's constructive redirect is to move capital and talent toward that deeper layer and away from the commodity chatbot race. The reporting does not prove his economic critique is correct in every particular. It does show that the question is now serious enough that the most expensive capex program in tech history should be treated as two bets, not one, and that the second bet is the one with a defensible moat.