In August 2026, NVIDIA quietly rolled out a feature on its Build platform (build.nvidia.com) that deserves closer attention: any developer could apply for a free API key—no credit card required—and gain access to five top Chinese AI models, including DeepSeek V4 Flash, MiniMax M3, Qwen 3.5-397B, Kimi K2.6, and GLM-5.1, all built on an interface standard fully compatible with OpenAI. According to CNBC, NVIDIA characterized the move in public statements as an effort to "support global developers in building applications on the American technology stack."
That framing itself warrants scrutiny: what is actually being promoted is the model weights of Chinese companies, while what is called the "American technology stack" is the inference infrastructure NVIDIA provides. This choice of packaging precisely reveals the deeper commercial calculus behind the move.
When Hardware Bans Meet the Free Flow of Software
Over the past two years, the US government has steadily escalated chip export controls against China: A100 and H100 were successively placed on restricted lists, and in April 2025, H20 was also subjected to export licensing requirements over concerns it could be used in Chinese supercomputers. NVIDIA consequently lost billions of dollars in potential revenue.
Yet in that very same period, NVIDIA has been actively helping spread the inference capabilities of Chinese models to developers around the world. This is not a policy loophole but a deliberate strategic choice: Washington can control the physical flow of silicon, but it cannot prevent model weights from being replicated across global servers as bits. NVIDIA has chosen to put down roots in that gap.
That gap is already producing tangible effects. According to MindStudio's analysis, the US blockade on GPU exports has inadvertently accelerated efficiency innovation among Chinese models such as DeepSeek—forced to train on constrained compute, they instead achieved lower inference costs. The net result: the ban did not weaken the competitiveness of Chinese models; rather, it gave them a price advantage in the global market.
61%: A Number That Makes Washington Uneasy
Market data presents a more direct reality than policy debates. According to OpenRouter platform data cited by TechBriefly and multiple other outlets, Chinese AI models accounted for approximately 61% of the platform's token consumption by mid-2026, up from less than 1.2% at the end of 2024. Chinese models took all five top spots on OpenRouter's global usage rankings, with code generation tokens from Qwen and DeepSeek combined accounting for close to half of the platform's total.
The driving force behind this explosive growth is not politics but a dual edge in price and capability. The cost-performance choices developers make in specific projects, once aggregated, form structural data that policymakers can hardly ignore. NVIDIA Build's free APIs are, in effect, another tap on the accelerator for this trend.
NVIDIA's Business Logic: Whosever Model Runs, It Uses My Chips
To understand this move, one must first recognize NVIDIA's profit position. Whether the model ultimately running is GPT-4o, Claude, DeepSeek, or Qwen, the underlying inference hardware can hardly bypass NVIDIA's GPUs. The free inference offered on the Build platform runs on NVIDIA's own DGX cloud infrastructure.
In other words, by distributing free APIs for Chinese models, NVIDIA is running a "platform tax" business that locks global developers into its own infrastructure. Once developers write their code in an OpenAI-compatible format and connect to NVIDIA's endpoints, the future switching costs become prohibitively high. This is the same playbook Intel used to bind developers with free compiler toolchains—the only difference is that today's target is AI inference.
Under this logic, Chinese models are not competitors to NVIDIA but content assets that increase platform stickiness. NVIDIA's real rivals are AWS Bedrock, Azure AI Studio, and Google Vertex AI—all of which are vying for the same developer inference traffic.
Fractured Signals from Washington
The contradictions at the policy level are equally clear. On July 24, 2026, 25 companies—including NVIDIA, Microsoft, Meta, IBM, and Palantir—jointly signed an open letter titled "Open Weights and American AI Leadership," urging Washington to avoid "premature restrictions" on downloadable AI models. Notably, OpenAI, Anthropic, and Google were all absent from the list of signatories.
The structure of that list is itself a message: companies with closed commercial models chose to remain silent, while those dependent on underlying infrastructure and open ecosystems chose to speak up. Each party's commercial interests determined its policy stance—not the other way around.
Meanwhile, the Trump administration is reportedly reviewing whether to ban the use of Chinese AI models in the United States, and multiple congressional committees are holding hearings on the matter. NVIDIA's intensified platform support for Chinese models at this moment is, in essence, an attempt to preemptively secure a fait accompli before the policy window closes.
The Real Boundaries of the Free API
One point needs to be clarified at the technical level: the current free API on NVIDIA's Build platform is subject to a rate limit of roughly 40 requests per minute, and its terms of use restrict it to development and testing purposes, not production deployment. This means it can cultivate developer habits but cannot replace enterprise-grade inference procurement. True commercial-scale usage still requires paid access to NVIDIA's NIM microservices or self-procured hardware.
This boundary delineates NVIDIA's real intent: not to provide production capability for free, but to lock developers' toolchain switching costs into the NVIDIA ecosystem through a zero-barrier trial experience. What is free is the entry; what is paid is the scale.
A Framework That Should Not Be Easily Accepted
Around this event, two opposing simplified narratives currently prevail: one is "NVIDIA stabbing US national security in the back," the other is "technology knows no borders, cooperation brings win-win." Both replace analysis with emotion.
A more accurate description is this: NVIDIA is executing a commercial strategy that maximizes its own platform value within the gaps of export controls, and the objective result is accelerated penetration of Chinese models into the global developer community. This is neither treason nor noble technological openness—it is the rational profit-maximizing behavior of a multi-trillion-dollar company operating under geopolitical constraints.
The question that truly needs to be answered is not whether NVIDIA "should" be doing this, but rather: when chip controls cannot stop the spread of models, where exactly does the actual reach of US AI policy toward China lie? From the current evidence, that boundary is far blurrier than policy documents suggest.
Sources: - [Nvidia is bolstering support for Chinese open AI models as it warns of White House crackdown](https://www.cnbc.com/2026/08/27/nvidia-chinese-ai-models.html) - [Chinese AI Models Top OpenRouter; Claude at 13.3%](https://tech-insider.org/au/chinese-ai-models-openrouter-2026/) - [Chinese LLMs Take Top Five Spots On OpenRouter](https://dataconomy.com/2026/07/29/chinese-ai-models-openrouter-top-five/) - [Why US Export Controls on GPUs Accidentally Made DeepSeek V4 Cheaper Than Any American Model](https://www.mindstudio.ai/blog/us-export-controls-deepseek-v4-cheaper-training) - [Nvidia and 24 other companies sign open-weights letter](https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-and-24-other-companies-sign-open-weights-letter-as-washington-weighs-chinese-ai-model-ban) - [Open-Weight AI Debate 2026: Why Big Tech Fights Regulation](https://www.edenai.co/post/the-open-weight-ai-debate-nvidia-microsoft-meta-push-back-on-regulation) - [NVIDIA Is Offering 80+ AI Models for Free via APIs](https://medium.com/coding-nexus/nvidia-is-offering-80-ai-models-for-free-via-apis-fc64b38276b8)© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接