At the ninth World AI Conference in Shanghai this month, Chinese President Xi Jinping addressed the gathering for the first time in nine years. It is indicative of the priority that AI occupies for the leadership. At the conference, Mr. Xi pitched Chinese AI models as a global public good. The contrast was clear as day. America is selling closed-source, proprietary models while China is offering open-weight models to the world. This distinction is not merely technical, but equally political.
Unlike Anthropic, OpenAI, and Google, which host their frontier models on their respective servers and offer services via an application programming interface (API), Chinese labs such as DeepSeek, Qwen, and Moonshot also release their weights for anyone to download. Unlike American models, these can be customised, modified, and even further trained. There is no subscription to cancel or terms of service. The maker cannot revoke or restrict access. So long as one can manage the requisite compute to host the model, even Papua New Guinea or Senegal can run it locally on their servers. And most importantly, the data never leaves the border.
AI diplomacy
However, one may wonder what explains China’s benevolence. Well, to begin with, the objective is to project China as a responsible AI player in the comity of nations. Unlike the Americans, who he alleged to have overstretched the concept of national security to AI, and thereby adopted an exclusionary approach, China is committed to sharing the benefits and fruits of AI. Mr. Xi’s framing was indicative in this regard. He described AI as “humanity’s collective wisdom,” a “historic opportunity” to “bridge the AI and digital divides,” and something that “should not be a solo performance by a single country but a symphony of international cooperation.” The underlying message, aimed at Washington, projected Beijing as the open and cooperative alternative to the U.S.’s AI models.
But there is a limiting factor at play too that explains Beijing’s benevolence — its compute constraint, both for training and inference.
Since 2022, the U.S. administration, in concert with its allies, has imposed a series of export restrictions on China, limiting its ability to both mass-import and mass-manufacture high-end chips.
The compute constraint
To understand this constraint, one must separate two very different tasks — training a model and serving it. Training is a one-time cost. A lab assembles a cluster of chips, runs it for a few months and comes up with a finished model. The fleet required is numerically small. For instance, GPT-4 was trained on an estimated 25,000 Nvidia chips, and Meta trained its Llama 3.1 on ~16,400 chips. DeepSeek, on the other hand, trained its acclaimed V3 model on ~2,000 export-compliant H800s. A frontier model, in other words, needs chips in the thousands, and it needs them only once.
Serving that model to the public, however, is an altogether different proposition. Inference scales with the user base and requires continuous expansion as it must answer hundreds of millions of user queries. OpenAI crossed a million GPUs in 2025, and Meta targeted the equivalent of 1.3 million by the end of 2025. Thus, for any model with a mass consumer base, the serving fleet dwarfs the training fleet, often by ten to a hundred times.
In other words, the distinction is that of a stockpile and a flow. To train a model, a lab can accumulate ten or twenty thousand chips through smuggling, stockpiling, and modest domestic production. This is precisely why the export controls could not prevent China from reaching parity at the model level. Inference, however, cannot be solved this way. The export controls mean that the most advanced chips remain out of bounds for Chinese firms. Additionally, the concessions the U.S. has offered exist largely on paper. Around 10 Chinese firms were cleared in principle to buy up to 75,000 H200 chips each, but not a single unit had been delivered by mid-2026, as sales remained stalled and Beijing itself directed firms towards domestic alternatives.
China can attempt to bridge the shortfall through domestic production, but even here, restrictions on export of EUVs continue to prove a handicap. SMIC’s 5nm line in China runs at roughly 20% yield without EUV tools, far short of commercial viability. This severely restricts the ability of Chinese labs to offer services via API alone. DeepSeek’s decision to freeze new registrations in January 2025 was influenced by a surge in traffic it could not serve, alongside a reported ‘large-scale malicious attack’.
Openness by necessity
Furthermore, given that Chinese models are nearly as good as the U.S.’s frontier models, their popularity is increasing. Data show that the Chinese models’ share of U.S. firms’ AI usage has hit a record 60% on OpenRouter — a marketplace for routing queries to various models based on requests.
American firms are preferring Chinese models as they provide comparable services at a fraction of the cost. But this success comes with a rider — the need for enormous compute. The solution thus is to simply outsource compute via open-weight models. Hence, China’s open-weight offering to the world is largely a product of its structural constraint. It can build models that rival the U.S. but cannot serve them at scale.
But Beijing cannot ignore the urge to securitise and restrain access for too long. Sooner or later, it will face that choice. Indeed, Beijing is already reported to be weighing restrictions on overseas access to its most capable models.
One can expect China’s frontier models to move towards proprietary models, while trailing models remaining open-weight. But so long as China cannot overcome the larger constraint, it is likely to continue fielding open-weight models. And China’s constraint is the world’s gain.
(Amit Kumar is a Staff Research Analyst at Takshashila Institution)
Published – July 30, 2026 08:30 am IST
