Starved of the best chips and outspent on compute, China's model builders have closed the frontier gap to months, not years. Their answer to every constraint has been the same: give the models away.
In January 2025, a Hangzhou lab most Western investors had never heard of wiped roughly $600 billion off Nvidia's market value in a single trading day, the largest one-day loss in the company's history. DeepSeek had done something the consensus said was impossible: it trained a model competitive with OpenAI's best, on chips deliberately hobbled to comply with US export controls, and then it published the weights under an MIT license for anyone to download.
Eighteen months later, that is no longer a shock. It is the pattern. A cluster of Chinese labs — DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Zhipu's GLM, MiniMax — now sits at or near the open-weight frontier, and they got there while fighting the two constraints that were supposed to keep them a generation behind: no access to the best silicon, and far less capital than their American rivals. The interesting question is not whether they caught up. It is how, and why they insist on giving the result away.
Start with capability, because everything else follows from it. As recently as early 2025, the Center for Strategic and International Studies judged the US–China frontier gap at one-to-two years, with DeepSeek's models roughly comparable to what American labs had shipped in mid-2024. Through 2026 that lag compressed to a matter of months, and on the benchmarks that drive most real enterprise value it has effectively closed. We traced that convergence in detail in Closing the Gap; the short version is that four of the five open-weight families that matter are now Chinese.
The receipts are specific. DeepSeek's R1, released open, matched OpenAI's o1 on hard reasoning — 79.8% versus 79.2% on the AIME 2024 math competition, 97.3% versus 96.4% on MATH-500. Zhipu's GLM 5.2 lands within a single percentage point of Anthropic's Opus 4.8 on a closely watched agentic benchmark, at roughly a fifth of the cost, and its developer adoption on OpenRouter climbed faster than DeepSeek's own V4 launch. By June 2026, DeepSeek's V4 Flash had become the first open-weight model teams dropped straight into agentic pipelines as a plausible substitute for a frontier proprietary model — at $0.14 per million input tokens against the several-dollar rates the closed labs command, and as low as $0.029 on cached calls.
The honest caveat: the very top of the reasoning and reliable-agent tiers still belongs to the closed American labs, and OpenRouter's own read is that the open-weight gap is "real but narrow, and not widening." Narrow and not widening is exactly the condition that terrifies a business built on selling access to the frontier.
None of this happened on a level field. Since October 2022, Washington has banned the export to China of any AI accelerator at or above the capability of Nvidia's A100, then tightened the net repeatedly as Nvidia shipped deliberately degraded parts — the A800, the H800, the throttled H20 — to stay inside the line. DeepSeek trained V3 on 2,788,000 GPU-hours of those cut-down H800s for a reported final-run cost of about $5.6 million. Its parent hedge fund, High-Flyer, had stockpiled some 50,000 Hopper GPUs before the loopholes closed; before Washington shut it, more than $9 billion of A800 and H800 silicon flowed into China in a single year.
The domestic substitute is real but not close to sufficient. Huawei's Ascend line is the flagship, yet SMIC, its foundry, is stuck at a 7-nanometre process because the equipment it would need to go finer is itself export-controlled. China still lacks EUV lithography, enough advanced packaging, and unrestricted high-bandwidth memory — the three chokepoints that decide how good an AI chip can get. The arithmetic is stark: even on aggressive assumptions about Huawei's output, CSIS estimates it would supply only around 5% of Nvidia's aggregate AI compute in 2025, and that share falls toward 2% by 2027 as Nvidia scales faster. On raw compute, the gap is widening, not closing.
Money is tighter too, and the domestic market is a knife fight. In May 2026 DeepSeek triggered a price war with a permanent 75% cut to its V4 API rates; ByteDance and Tencent slashed prices to defend share. That race to zero is now colliding with physics. From mid-July 2026 DeepSeek is introducing time-of-day surge pricing — doubling peak-hour rates, with its V4 Pro model rising from 6 to 12 yuan per million output tokens — a move analysts read as a signal of tightening domestic compute supply, not just cooling competition. Goldman Sachs called it the price war "returning to rationality."
Here is the twist the export-control hawks did not price in: constraint is a powerful teacher. Denied the brute-force path of simply buying more and bigger chips, Chinese labs were forced to get efficient, and efficiency turned out to be transferable, cheap, and — once published — impossible to un-learn.
DeepSeek is the clearest case. Its architecture leaned on a fine-grained mixture-of-experts design and "multi-head latent attention," which cut the working-memory cost of inference by roughly 93%. That is how an open model could match a Western frontier system while costing about 94% less per input token. On the training side, DeepSeek-R1-Zero showed that large-scale reinforcement learning alone — with no supervised fine-tuning first — could coax strong reasoning out of a base model, a genuinely surprising result. Then the lab distilled that reasoning into six small dense models, from 1.5 to 70 billion parameters, built on Qwen and Llama; the 32B distillation reaches parity with OpenAI's o1-mini. Frontier reasoning, in other words, compressed to something you can run on a workstation.
Denied the best silicon, China's labs optimised the one thing export controls cannot ration: ideas. And ideas, once published, do not stay behind a border.
The state is now underwriting the effort as strategic infrastructure. When Beijing's "Big Fund" — the same vehicle it used to backstop chipmaker SMIC — moved into DeepSeek's financing, it signalled that frontier AI R&D now sits in the same national-priority bucket as semiconductors. The private capital followed: DeepSeek's first outside round reportedly ballooned from a $300 million target at a $10 billion valuation to something near $7 billion at $50 billion; Moonshot raised $2 billion at over $20 billion led by Meituan; Zhipu has pulled in more than $1.5 billion from Alibaba, Tencent and Qiming and is eyeing an IPO; StepFun is said to be closing on $2.5 billion.
Which brings us to the question that most puzzles Western observers. These are expensive models, built under duress, by companies that badly need revenue. Why publish the weights for anyone — including competitors — to download for free?
Because openness is the strategy, not a concession to it. Several logics stack on top of each other:
The mirror image explains the American position without needing to invoke virtue. OpenAI and Anthropic keep models closed because they genuinely hold the top of the frontier, because tens of billions in training and capex have to be earned back through metered access, and because a safety narrative gives a commercial instinct a principled frame. Notably, the strongest open-weight entrant from the US side is not a frontier lab at all — it is Nvidia, whose Nemotron 3 Ultra ranks second among open models, behind only GLM. The company that sells the shovels is happy to give away the maps; the companies selling the gold are not.
The real ceiling on China's ascent is not algorithms — it is atoms. Efficiency has bought the labs years, but every distillation trick eventually runs into the wall of how much advanced silicon actually exists inside the country, and that is a lithography-and-packaging problem measured in a decade, not a training run. It is the same chokepoint we mapped across the whole value chain in the AI stack's core sample.
Three things will tell you where this goes. Whether the domestic price war's shift "back to rationality" holds, or whether tightening compute forces consolidation among China's crowded field of labs. Whether the relentless downward pressure from open weights finally forces the closed American labs to cut prices or open something themselves. And whether Washington concludes that export controls, having pushed China toward exactly the efficient, open, hard-to-contain models that now undercut US firms globally, were a strategic own-goal. The models are open. The question of who benefits is still very much closed.
Three years ago the best downloadable model trailed the frontier by more than a year. In 2026 that lag is measured in months, and a wave of open-weight labs, most of them Chinese, is compressing the distance faster than almost anyone predicted.
AI models trained on public code are quietly turning Python and JavaScript into the default languages of software, and enterprise stacks risk being left behind.
The next leap in user experience is not prettier interfaces but delegation, where autonomous agents act on our intent so we touch less and offload more.
