Why China Is Going All-In on Open-Source AI

If you have a structural lead in hardware and supply chains, the rational move is to give software away. Make it free, make it ubiquitous, and let it erase the one place your competitor is ahead. That is the logic behind China’s open-weight AI strategy, and it is worth taking seriously even if you build for the other side.

The framing comes from a conversation on Naval’s The AI Industrial Revolution, which seeded this piece: the observation that “all the open source heft is coming from China,” and that China is “going all in on it because they have hardware superiority.” The rest of this post is an attempt to check that thesis against the public record — what holds up, what is softer than the headlines suggest, and where the United States still has the floor.

The diffusion play #

Start with the thing that is no longer arguable: Chinese open-weight models have caught up on adoption. For the first time, Chinese models overtook US models in global downloads, capturing 17.1% of downloads on Hugging Face in the year ending August 2025 versus 15.86% for US models, per a joint MIT/Hugging Face study (MIT Technology Review). The engines here are DeepSeek and Alibaba’s Qwen. By another cut of the data, Chinese models reached roughly 30% of global usage share by the end of 2025, up from about 13% at the start of the year (ISEAS).

That did not happen by accident; it happened by design. Lacking the GPU stockpiles and venture capital of US frontier labs, Chinese labs picked open distribution as asymmetric competition — maximize diffusion, developer goodwill, and downstream data instead of guarding a proprietary moat. One Vertex Ventures analyst put the contrast cleanly: US companies “build for perfection” while Chinese AI companies “build for diffusion” (Fortune). When your goal is reach rather than margin, free weights are not a concession — they are the weapon.

The ecosystem effect compounds. Qwen now anchors the largest derivative family in open AI: the US-China Economic and Security Review Commission’s March 2026 report documents more than 100,000 Qwen derivative models, more than Google and Meta combined (USCC). Every fine-tune is a developer who now reaches for a Chinese base model by default.

Why open source compounds a hardware lead #

The deeper argument in the episode is that open source does not just win developers — it feeds the physical economy that China already dominates. The USCC report frames this as “two reinforcing loops.” A digital loop: open adoption drives model iteration. And a physical loop: models deployed across factories and robotics generate real-world data that flows back into the next model (USCC). The Commission’s blunt conclusion is that US chip export controls were not designed to counter this — you can throttle training compute without touching the deployment flywheel.

This is the builder-relevant insight. If software generation sits upstream of every hardware pipeline, then on-demand software is a force multiplier for whoever already owns the factories. China installs more industrial robots than any other country and runs the deepest electronics-manufacturing cluster on earth; pour cheap, ubiquitous, locally-runnable models into that base and the software disadvantage that Silicon Valley counted on starts to evaporate. The open-weight models are the solvent.

(A note on sourcing: precise figures often quoted here — exact robot-install percentages, actuator and rare-earth shares, Shenzhen prototype-cycle times — trace back to a TechCrunch piece that, on inspection, covers humanoid-robot shipments and does not contain those industrial-robot statistics. I have left them out rather than cite numbers a source does not support. The directional claim that China leads in industrial deployment is carried by the USCC physical-loop framing above, not by those orphaned stats.)

The sovereignty pull #

The clearest evidence the thesis is operative is who is buying. Malaysia announced its national AI ecosystem would run on DeepSeek’s open weights — not OpenAI, Anthropic, or Google — citing cost, data sovereignty, and the ability to run models on domestic infrastructure without a US API dependency. In parallel, Singapore’s government-backed AI Singapore program chose Qwen over Meta’s Llama as the foundation for its regional model (MIT Technology Review).

Neither is a fringe actor. Both are sophisticated tech hubs with deep Western institutional ties, and both looked at the menu and picked the open Chinese option. That is exactly the mechanism the thesis predicts: open weights neutralize US software lock-in, because a sovereign buyer can take the model home, run it on its own metal, and owe no one a per-token rent.

The honest counter-section #

None of this means the race is over, and the brief is full of reasons to keep the enthusiasm bounded.

The frontier is still US-held. Chinese open models still trail the top proprietary US models by roughly 8% on key software-development benchmarks as of late 2025 (Fortune). DeepSeek shipped a preview of its long-awaited V4 in April 2026 (CNBC); it is strong on math and coding, but the very top of the world-knowledge frontier remains closed and American. Diffusion is winning the middle of the market while the ceiling stays elsewhere.

Compute is a real tax. DeepSeek’s founder, Liang Wenfeng, has acknowledged that Chinese firms “have to consume twice the computing power to achieve the same results,” and potentially four times more once you account for data gaps. DeepSeek only dodged a crippling shortage because Liang’s hedge fund, High-Flyer, stockpiled an estimated 10,000–50,000 Nvidia A100s before export controls tightened — and even then it built its breakout reasoning model in roughly two months for under $6M on export-capped chips (MIT Technology Review). Impressive, but it is efficiency born of scarcity, not abundance.

Hardware self-sufficiency is incomplete. Domestically, roughly 40% of China’s AI-chip market is now served by Chinese silicon, with Huawei accounting for about half of that (The Substrate). Huawei’s own stated target for 2026 is on the order of 600,000 Ascend accelerators — a serious ramp, though widely-circulated higher figures should be treated with caution. And the ceiling is structural: SMIC and domestic fabs are pushing deep-ultraviolet lithography to its limits, yielding frontier chips at higher cost and lower yield than TSMC or ASML EUV processes (The Substrate). The flywheel runs; it just runs at a premium.

The substrate is still American. Even with Chinese models on top of the download charts, the infrastructure underneath — the major clouds, and Hugging Face itself — is US-controlled. The framing that “China is winning the open-source AI race, but a US company still controls everything underneath” comes from an analyst quote I was unable to independently verify (the source did not render for retrieval), so treat the exact wording as unconfirmed; the underlying point about American platform control is, however, straightforwardly true of where these models are hosted and distributed.

Open source cuts both ways. The US camp is itself split on whether open weights are a threat or an accelerant. Former Google CEO Eric Schmidt has been cited arguing the US should cultivate a vibrant open-source ecosystem to move faster, even as frontier labs treat proprietary weights as a moat — and China’s open bets could inadvertently lift US capability if American developers fine-tune Chinese base models. I could not verify that Schmidt attribution directly (the source returned a 403), so flag it as unconfirmed; the strategic tension it describes is real regardless of who said what.

What a builder should take from this #

The thesis survives contact with the evidence in its strong form and loses some of its swagger in the details. China is genuinely going all-in on open source, the adoption numbers are real, and the deployment flywheel into a hardware-heavy economy is a coherent reason export controls feel less decisive than they were meant to be. What is not true is that the gap has closed: the frontier, the leading-edge fabs, and the platform layer all still sit on the American side, and Chinese efficiency is partly a workaround for compute it cannot yet get.

If you build, the practical reading is simple. Treat open Chinese weights as a first-class option, not a novelty — for sovereignty-sensitive, cost-sensitive, or air-gapped deployments they are already the default for serious buyers. But do not confuse “most downloaded” with “best,” and do not assume the infrastructure underneath you is neutral. The interesting contest over the next few years is not models versus models. It is diffusion versus the frontier — and right now both sides are winning a different half of it.