"China has all but caught up. The US is not going to 'win' the AI war. Here's what we should do instead," subtitled "Seven options, considered," is a July 20, 2026 essay by Gary Marcus on Marcus on AI, written after Moonshot AI's Kimi K3 release. It is an opinion piece and is recorded here as an argument.
The claim
Marcus argues that Kimi K3 — "largely on a par with the best American models," open-weight and downloadable free by anyone with sufficient hardware — "calls the business models of OpenAI and Anthropic into serious question, and may kill or greatly undermine their IPOs." He places it in a sequence with Z.ai's GLM 5.2 and Alibaba's forthcoming Qwen release rather than treating it as an isolated event, and notes the US stock market fell partly on the news.
He frames it as the culmination of an argument he dates to summer 2023: "no moat => more competitors => price wars => profits are scarce."
The prior-record claim
A substantial part of the essay is Marcus reasserting a January 2025 post-DeepSeek forecast. He lists what he says he predicted: that the US would not achieve decisive victory over China while focusing largely on LLMs; that the two would converge on a tie; that the absence of a technical moat would leave companies like OpenAI unprofitable; that Nvidia would be vulnerable; that the CHIPS Act would backfire without holding China back; that models would get more efficient and cheaper while hallucinations and reliability problems persisted; and that racing around LLMs would divert resources from more original ideas. His assessment: "That's basically what has happened."
He attributes the failure to heed this to a government "that lacked distance from Silicon Valley, and rarely consulted anyone for an outside opinion," and adds that "a bunch of cheap xenophobic dismissals of China got us nowhere." He also quotes Ryan Fedasiuk's July 17, 2026 observation that "the American moat in AI software is not as strong as many of us had hoped" and that the race is "an industrial systems competition to build and install compute."
Marcus calls for a congressional investigation into how the lead was squandered, listing questions on strategic error, IP protection, Chinese state subsidy, espionage, immigration restrictions causing talent drain (noting Kimi's founder attended Carnegie Mellon and returned to China), and lessons for quantum computing and fusion.
The seven options
| # | Option | Marcus's assessment |
|---|---|---|
| 1 | Do nothing — no subsidies or backdoor bailouts | Google, Microsoft, and Amazon can cover US needs including classified ones; but concedes much of the market to China, "so it's not ideal" |
| 2 | Outlaw open source | "Good luck with that; at least with LLMs, the cat is already out of the bag" — would persist on the dark web and abroad |
| 3 | Build a regulatory moat | Opposed. Quotes Will Manidis calling it "a proposal to use the informal, coercive power of the terminal, late-stage bureaucratic state to clear the American market of a cheaper frontier competitor," and Matt Stoller calling it "corporate communism"; argues it would crush US startups and raise prices |
| 4 | Bail out the big AI labs | "Just no." |
| 5 | Ban Chinese models altogether | Would raise consumer prices and stifle innovation; quotes Phil Libin that "we win by transforming games to non-zero-sum" |
| 6 | Buy out OpenAI and Anthropic "for two cents on the dollar" and convert them to national laboratories | Sees "poetic justice" since AI is built on uncompensated human IP, but warns of government abuse — mass surveillance, and an administration with "a tendency to prosecute critics and enemies on thin grounds" |
| 7 | Give up on winning and make AI a global public good | His preferred option |
On option 3 he notes Dean Ball, newly at OpenAI, appeared to argue a regulatory moat is where things are headed, and records in a footnote Ball's later claim that he was describing inevitability rather than proposing it — adding that "few people on X believe him" and leaving the intent open.
The CERN-for-AI proposal
Marcus's preferred option revives a proposal he made in a 2017 New York Times piece, based on a 2016 AI for Good talk, quoted in the essay: he looked "with envy" at CERN as "a huge, international collaboration, with thousands of scientists and billions of dollars of funding" pursuing tightly defined projects and sharing results with the world "rather than restricting them to a single country or corporation."
He argues the moment may finally be right, citing Xi Jinping's World AI Conference speech as "a tremendous opening for discussion" while declining to take it at face value. His concrete formulation is a decision to "outlaw closed source AI — requiring all manufacturers to share their weights, architectures and training data with accredited scientists and research organizations — combined with a deal with China to dedicate AI to the public good."
The framing rests on the same uncompensated-IP argument as option 6: turning AI, "built on so much of the internet that was sourced from hard work and thinking of the public, largely without compensation, into a true public good for the world."
Reception and relation to other positions
Marcus extended the argument in an open letter to David Sacks published July 24, 2026, writing that "China is closing the gap, but it's not because we have too much regulation on AI" (Source: garymarcus.substack.com).
The essay's economic premise — that Chinese open-weight releases threaten US frontier-lab business models — is directly contested by Ben Thompson's essay of the same day, which argues the apparent price advantage reflects a compute-shortage price umbrella rather than a marginal-cost advantage, and that frontier labs likely hold the best cost structure. The two pieces agree on opposing a regulatory moat and on the cybersecurity case for open availability, and diverge on whether the commercial threat is real. See US-China AI Competition: Different Races, Different Metrics, Open-Weight Frontier Models.
Provenance
Published on garymarcus.substack.com, dated July 20, 2026. Pulled and verified by gap scan; the seven-option structure and CERN-for-AI conclusion match the summary folded onto Gary Marcus from the July 2026 developments log.
Relationships
- contradicts: Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026) — on whether Chinese open-weight models threaten US frontier-lab economics.
- related: Gary Marcus — author.
- related: Kimi K3, GLM-5.2 — the releases prompting the essay.
- related: US-China AI Competition: Different Races, Different Metrics, Open-Weight Frontier Models, Regulatory Capture in AI Policy, AI Public Wealth Fund and Government Equity in AI.