China Is Winning the Open-Weight AI Race, Says Hugging Face CEO

A Blunt Assessment from the Open-Source Hub’s Leader

When the CEO of the world’s largest open-source AI platform tells you who is winning, you listen. On August 3, 2026, Hugging Face chief executive Clément Delangue appeared on CNBC’s Squawk on the Street and delivered a message that sent ripples through Silicon Valley boardrooms: “China is winning the AI race and dominating on open models.”

Delangue did not hedge. He pointed to the explosion of high-quality open-weight models flowing out of Chinese labs — DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi, Z.ai’s GLM, and MiniMax — and predicted that Chinese tools could match or surpass the capabilities of US frontier labs (OpenAI, Anthropic, Google) as soon as the end of 2026, or by 2027 at the latest.

The statement came just days before an August 7 CNBC deep-dive that confirmed the trend with hard data: China’s AI firms are not only closing the performance gap with US frontier labs, but Chinese models are seeing rapidly rising adoption globally — with one critical caveat that keeps the United States in the game.

By the Numbers: How China Closed the Gap

The data backing Delangue’s claim is striking. Consider the following metrics, drawn from Hugging Face’s own platform telemetry, OpenRouter’s routing logs, and Stanford HAI’s 2026 AI Index:

• 41% of all Hugging Face model downloads now originate from Chinese developers, up from roughly 1% of global open-model usage in late 2024. Qwen and DeepSeek alone account for the lion’s share.
• 61% of tokens consumed on OpenRouter — the largest neutral LLM routing service — were served by Chinese open-weight models as of May 2026, according to analysis by DataGravity.
• Nearly 30% of global AI usage runs on Chinese open-source models, per Stanford HAI’s December 2025 assessment, with Qwen and DeepSeek leading the pack.
• 70% of newly created derivative open models worldwide are built on Chinese foundation models — predominantly Qwen — according to GenAI Assembling research tracking the period from late 2023 through March 2026.
• The US-China benchmark performance gap narrowed from 103 points in January 2024 to just 23 points by February 2025, and Stanford HAI’s 2026 AI Index declared the race effectively “neck and neck.”

These numbers represent a structural shift, not a temporary blip. China made a deliberate strategic bet on open-weight releases — publishing model weights under permissive MIT and Apache 2.0 licenses — and that bet is paying off in adoption, derivative innovation, and developer mindshare.

The Open-Weight Strategy: Why It Works

China’s dominance in open-weight AI is not accidental. It is the product of a coordinated industrial strategy that the US-China Economic and Security Review Commission detailed in a March 2026 report titled Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance. The report’s thesis is that China created a self-reinforcing cycle: by releasing models freely, Chinese labs attract global developers; those developers build derivative models and tools that improve the ecosystem; the improved ecosystem draws more users back to Chinese platforms.

The key Chinese labs executing this strategy include:

• Alibaba (Qwen): The broadest product line in Chinese AI, ranked as the second-largest provider globally with 13.9% market share and 2.77 trillion weekly tokens. Roughly 70% of new global derivative models trace their lineage to Qwen.
• DeepSeek: The original disruptor whose V3 and V4 series proved that frontier-level reasoning could be achieved at a fraction of Western training costs, forcing US labs to rethink their pricing.
• Moonshot AI (Kimi): Released Kimi K3 in July 2026, the world’s largest open-weight model, which rivals leading American systems and dominates major coding benchmarks.
• Z.ai (GLM): The GLM 5.2 series ships under Apache 2.0 and has become a staple for developers building agentic applications.
• MiniMax: A rising player in multimodal open models, competing directly with Western alternatives on cost-per-token.

Together, these labs have turned China into the world’s supplier of affordable intelligence at the open tier. As DataGravity’s Chris Zeoli put it in June 2026: “China supplies the majority of the world’s open-model tokens, sets the price of intelligence at the open tier, and leads on derivative innovation.”

Washington Pushes Back — But Adoption Keeps Climbing

The growing reliance on Chinese AI models has not gone unnoticed in Washington. Throughout 2026, US lawmakers have ramped up pressure on companies adopting Chinese models. In July, members of Congress sent information requests to DoorDash and other firms demanding details about their use of Chinese AI infrastructure. The core concern is twofold: national security (data could flow through Chinese-controlled infrastructure) and industrial competitiveness (US labs lose revenue and feedback loops when developers choose cheaper Chinese alternatives).

But the political pressure is running headlong into economic reality. Chinese models are dramatically cheaper to run — often an order of magnitude less expensive per million tokens than US frontier APIs. For cost-sensitive applications, the choice is not ideological; it is arithmetic. A startup building a document-processing pipeline will choose Qwen at $0.20 per million tokens over a US equivalent at $2.00, regardless of congressional letters.

This tension — between security hawks in Washington and pragmatic engineering teams across corporate America — is one of the defining dynamics of the 2026 AI landscape.

The One Advantage America Still Holds

Despite the open-weight surge, the August 7 CNBC analysis identified one area where the United States maintains a decisive edge: advanced AI chips. While Chinese firms have reported record semiconductor revenue driven by AI demand, Chinese-made accelerators still lag cutting-edge American hardware from NVIDIA, AMD, and others in raw performance.

The Brookings Institution noted in April 2026 that “Chinese AI chips are not expected to close the performance gap with cutting-edge American AI chips for the foreseeable future.” Export restrictions on advanced NVIDIA GPUs — the H200, B200, and Rubin series — remain a significant bottleneck for Chinese labs attempting to train next-generation frontier models.

This hardware advantage translates into a training advantage: US labs can still build the largest, most compute-intensive models. But the open-weight revolution means that once a model is released, the advantage of having trained it erodes quickly as the global community fine-tunes, distills, and builds upon it — often using Chinese base models because they are free.

What This Means for the AI Industry

The implications of China’s open-weight dominance extend far beyond a single quarter’s download statistics:

1. Developer mindshare is shifting. When 70% of derivative models trace back to Qwen, Alibaba’s ecosystem becomes the default substrate for open-source AI innovation. This is a moat that compounds over time.
2. The price floor for intelligence is now set in China. US labs must either compete on price (difficult given their capital-intensive training regimes) or differentiate on proprietary capabilities that open weights cannot replicate.
3. Policy responses are fragmenting. The EU is exploring its own AI sovereignty initiatives, while the US oscillates between export controls and domestic investment incentives. Meanwhile, Chinese models continue to flow freely to developers worldwide.
4. The open-weight vs. closed-weight debate is being settled by adoption. When the best free models are Chinese, developers vote with their download buttons.

Looking Ahead

Delangue’s prediction — that Chinese open-weight models catch up to US frontier capabilities by late 2026 or 2027 — is not a fringe view. It reflects the consensus forming among practitioners who watch the benchmark curves converge month over month. The question is no longer whether China will close the gap, but what the global AI industry looks like when it does.

For now, the open-weight revolution continues to accelerate. Every new Qwen release, every DeepSeek breakthrough, every Kimi milestone pushes the frontier of what is freely available further out — and widens the gap between what governments want developers to use and what developers actually choose. As 2026 unfolds, the most consequential AI competition may not be between GPT and Claude, but between Washington’s policy levers and the irresistible economics of free, frontier-grade models from China.