After DeepSeek, another Chinese model has rattled the United States.
On July 22, four things happened in a single day. Michael Kratsios, head of the White House Office of Science and Technology Policy, publicly accused Moonshot AI on X of “large-scale distillation” of American models with its Kimi K3. OpenAI president Greg Brockman told Bloomberg that K3 “is a very good model, no doubt about it.” Nearly 200 Silicon Valley startups signed a letter warning the Trump administration that banning Chinese open-source models would mean “hundreds of companies dying overnight.” And Nvidia CEO Jensen Huang told Axios that American firms “absolutely should be allowed” to use Chinese models.
Four voices, four directions, one trigger: Kimi K3, released by Moonshot AI on July 16. If DeepSeek’s shockwave in early 2025 mainly hit Wall Street’s faith in AI infrastructure spending, K3 strikes at something more fundamental — what should the US AI industry do when China can open-source model weights at close to frontier quality? In a single week, K3 tore open a rift across both Silicon Valley and Washington.
Not just another Chinese model
K3 itself: 2.8 trillion parameters, a mixture-of-experts architecture with 16 of 896 experts active per pass, and a one-million-token context window. It is the largest open-weight model in the world, with full weights due July 27. On Artificial Analysis’s intelligence index it scores 57, ranking third — behind only Anthropic’s Fable 5 (60) and OpenAI’s GPT-5.6 Sol (59), and ahead of Claude Opus 4.8 (56). On front-end coding it hit number one on the Arena leaderboard within 24 hours of launch.
Technically it brings real innovations — a hybrid linear attention mechanism (Kimi Delta Attention) delivering 6.3x decode speedups at million-token scale, selective cross-layer retrieval (Attention Residuals), and quantization-aware training from the SFT stage — for roughly 2.5x better scaling efficiency over its predecessor K2.
But what really unnerved America is not the benchmarks. It is that performance, price and openness arrived at once. K3’s API costs $15 per million output tokens — expensive by Chinese standards (DeepSeek V4 is $0.87, Zhipu’s GLM-5.2 is $4.4), but less than a third of Fable 5. A near-frontier model, far cheaper than rivals, with fully open weights. And this did not come from nowhere: Cursor, the coding tool SpaceX is reportedly buying for around $60 billion, runs its Composer 2 core on Kimi K2.5; DoorDash’s CTO says it has handed “lower-level work” to Kimi K2.6. K3’s release was less a debut than a public confirmation that Chinese open-source AI is already embedded in Silicon Valley’s production stack.
A fractured Silicon Valley
Had this just been “a Chinese model with high benchmarks,” the story would end there. Instead, the week after launch produced America’s first openly factional fight over whether open source is good or bad.
The first camp is the White House and policy hawks. Kratsios accused Moonshot of building “a sophisticated internal platform to distill American models at scale,” while Treasury Secretary Bessent floated sanctions if US model “watermarks” were found in Chinese models, and Congress advanced anti-distillation legislation.
The second camp is America’s closed labs. OpenAI’s Dean Ball, head of strategic futures, called open models “inherently decelerationist” — arguing they erode frontier labs’ profits and thus investment — and urged the administration to manufacture “regulatory risk” to scare US firms away, warning of a dystopian “AI communism.”
The third camp — startups and open-source advocates — rejected that outright. A new group, the Little Tech Association (Proton, Replit, Y Combinator among them), organised the letter to the White House: ban Chinese open models and it is American founders, not Chinese firms, who die. Nvidia’s Huang was the heaviest voice against the hawks: Wall Street misread DeepSeek and is misreading K3, he argued — free AI is good for chips, data centres and hardware, because cheaper open models expand demand for compute rather than shrink it. “There is zero chance of China driving US companies out,” he said, adding that openness makes AI safer, not more dangerous, because outsiders can audit models and build defences.
The decelerationist debate
Ball’s “open source is decelerationist” argument does touch a real industry logic. Frontier models cost billions; if a Chinese lab offers a near-equal open alternative at rock-bottom cost, closed labs’ margins compress, capital markets mark down their terminal value, and frontier progress slows. AI researcher Nathan Lambert concedes the economic logic holds — but says it is not enough to stop OpenAI and Anthropic from becoming the world’s most valuable companies, because the pie is growing and closed models retain moats in enterprise trust and reliability.
Critics note the irony: modern AI, OpenAI included, was built on open foundations — the Transformer paper, PyTorch, shared research — before OpenAI “closed the door behind it.” To call open source decelerationist now, they say, is to build a tower with others’ open tools and then bar the door. Some brand Ball’s rhetoric “digital McCarthyism”: closed labs using a national-security narrative to entrench their market position.
China’s own tempo
Markets reacted fast: the week K3 launched, the Philadelphia Semiconductor Index fell 12.5%, its worst drop in 15 months, dragging down Nvidia, AMD and Broadcom. Even Chinese AI stocks fell — Zhipu down 28% in Hong Kong, MiniMax down 16%.
Step back and Silicon Valley’s fights trace to the goal America chose: extreme accelerationism — AGI as the single top priority, funded by a profit-maximising closed model and extreme capital density. Ball’s fear of “open-source deceleration” is really a fear that open models drain that profit engine’s fuel.
China’s labs do not have the same capital density, so they grew their own path. Moonshot founder Yang Zhilin is the clearest example: in 2023 he called closed source “the only path to a super app”; by GTC 2026 he was the only independent Chinese model founder invited to speak, laying out a route that replaces the Transformer era’s three foundations — the Adam optimiser, attention, and residual connections — with open alternatives that drew praise from Elon Musk and Andrej Karpathy. K3 delivered on that roadmap four months later.
Asked once about the gap with US peers, Moonshot’s team said: “We have fewer GPUs than our American peers, but we squeeze every card to the limit.” Co-founder Zhou Xinyu, asked how he views OpenAI’s spending, answered: “We don’t know either — only Sam knows. We have our own tempo.” Those five words may be the most important footnote to China’s whole AI path. DeepSeek, Zhipu and others are releasing in a dense, multi-front rhythm. Brockman puts China roughly four months behind on overall capability — a number that signals not backwardness but two paths converging on the frontier from different directions. This will not be a winner-take-all story.
Source: Sohu IT. Translated and adapted from Sohu IT (it.sohu.com).