How China’s AI Push Caught Up Despite US Chip Restrictions
China's AI firms are releasing competitive open-weight models despite GPU export bans. Here's how they did it and what it means for the global AI race.
Despite US chip export restrictions that limit access to high-end GPUs, Chinese AI firms are shipping models that benchmark closely against leading Western alternatives. Moonshot AI's Kimi K3, a 2.8 trillion parameter model, matched Anthropic's Fable on most metrics within weeks of that model's launch. The key factor: Chinese labs are releasing models as open-weight or open-source, letting any developer run them locally, which is building global adoption fast and creating a strategic long game that deserves a closer look from any business relying on AI infrastructure.
What happened
| Detail | Fact |
|---|---|
| Event | Shanghai World AI Conference, July 17, 2026 |
| Notable model | Kimi K3 by Moonshot AI, 2.8 trillion parameters |
| Benchmark result | Matched Anthropic’s Fable on most parameters |
| Anthropic credit offer | $100 in credits to paying users for Fable this month |
| India’s position | Skipped Shanghai conference despite hosting AI Impact Summit in February 2026 |
At the Shanghai World AI Conference on July 17, Chinese president Xi Jinping called on nations to embrace “openness and win-win” principles in AI development. Chinese firms are currently acting on exactly that framing. Moonshot AI’s Kimi K3 arrived weeks after Anthropic launched its Fable model to wide attention, and independent evaluations put K3 on par with Fable across most metrics.
K3 is open-weight, meaning the model’s trained parameters are publicly available. Anyone with enough compute can download and run it without paying a licensing fee or routing data through a third-party API.
How did China close the capability gap?
The standard explanation for why China should fall behind in AI is simple: US export controls restrict access to advanced GPUs (the chips that train large models and run inference on them). But Chinese labs have worked around this in two ways. First, they source larger quantities of lower-specification chips. Second, they benefit from cheap and abundant electricity, which offsets the efficiency gap of running more modest hardware at scale.
The result is a competitive ecosystem. Alibaba has released several lightweight models aimed at developers who want to run inference on their own hardware rather than cloud APIs. DeepSeek, the startup that first rattled US AI confidence earlier this year, offers its models both as a hosted service and open-source. Tencent and Baidu have also been steered toward AI by government encouragement. Across all these firms, Chinese models have been downloaded millions of times from HuggingFace, often picked up by developers building narrow, task-specific applications.
If you want broader context on how Chinese AI models are being perceived internationally, our earlier coverage of Kimi, DeepSeek, and the US reaction to Chinese AI is a good starting point.
Open-source vs captive: why it matters for your business
Western frontier models from OpenAI, Anthropic, and most of Google sit behind APIs. You pay per token, you accept their terms, and your data passes through their infrastructure. These are “captive” models. Chinese alternatives like Kimi K3 and DeepSeek are open-weight or open-source, which changes the economics in a few concrete ways:
- No per-token costs once you have the compute to run them.
- Third-party inference providers can host them and compete on price, which drives down costs further.
- No data leaves your infrastructure, which matters for regulated industries or privacy-sensitive workflows.
That said, Western labs are currently subsidising their own models heavily. Anthropic is offering $100 in credits to paying users just to try Fable. Once promotional pricing normalises, the cost comparison shifts and open-source becomes more attractive to infrastructure providers.
For businesses thinking about integrating AI into their operations, the choice of model provider is increasingly a strategic infrastructure decision, not just a feature comparison.
The trust problem and geopolitical friction
Capability parity does not automatically mean adoption. A persistent trust deficit surrounds Chinese models, particularly in India. Despite that, open-source Chinese models are being used by Indian inference aggregators like Neysa and have been reviewed by CERT-In (India’s Computer Emergency Response Team) because open-weight models eliminate data leakage concerns: if you run the model locally, no data goes anywhere.
India notably skipped the Shanghai conference, even after hosting the AI Impact Summit in February 2026. The government is backing domestic alternatives like Sarvam AI, though resources are limited. China, for its part, blocked Meta’s acquisition of AI firm Manus and restricted international travel for key AI personnel to prevent talent and IP from being absorbed into the US-led ecosystem.
How long will the open-source wave last?
This is the real strategic question. Open-weight models are expensive to train. The companies with the compute to build them at scale may eventually decide a closed, monetised model is a better business. Alibaba has already started building captive models that run on its own infrastructure and are not open-source. That is a data point worth watching.
The theory being tested: once the world has used Chinese open models long enough to trust their quality, those same firms may introduce closed, hosted versions and find an audience willing to accept the geopolitical baggage. Whether that plays out depends on whether current open models remain freely available and whether researchers studying the published model weights and accompanying literature develop effective domestic alternatives.
Our take
The open-source angle is not altruism. It is a distribution strategy. Chinese firms cannot easily sell API access to Western businesses with data residency requirements or political concerns, but they can make a model so useful and so free that developers build on it anyway. By the time a captive version arrives, the switching cost is high.
For businesses actually making AI decisions right now: the practical implication is that open-weight Chinese models are worth evaluating on technical merit for internal tooling, especially where data privacy matters. The geopolitical risk is real but it cuts both ways. Dependency on a single US hyperscaler has its own concentration risk.
Xi Jinping’s closing note at Shanghai pointed to AI “moving from the digital world” into physical applications. That is the next phase to watch. If you are thinking through how these shifts affect your own AI integration roadmap, the model landscape is going to look very different in 12 months.
The smartest move right now is to architect your AI stack so you can swap model providers without rebuilding everything around them.
Frequently asked questions
How is China building competitive AI models despite chip export restrictions?
Chinese labs use larger quantities of lower-specification chips combined with cheap and abundant electricity to compensate for the lack of advanced GPUs restricted by US export controls. Models like Kimi K3 (2.8 trillion parameters) have matched Western frontier models on most benchmarks using this approach.
What is the difference between open-source and captive AI models?
Open-source or open-weight models make their trained parameters publicly available, so anyone can download and run them locally without paying per-token fees or sending data to a third party. Captive models from firms like OpenAI and Anthropic are accessed via paid APIs and the source weights are not shared publicly.
Is it safe to use Chinese AI models like DeepSeek or Kimi?
Open-weight Chinese models can be run entirely on your own infrastructure, which eliminates data leakage concerns since no data is transmitted externally. India's CERT-In and inference aggregators like Neysa have used them on this basis. Political and supply-chain trust concerns remain, particularly for enterprise and government use cases.
Will Chinese AI models stay open-source?
Not necessarily. Open-source model training is expensive, and Alibaba has already begun building captive (closed) models hosted on its own infrastructure. The concern is that firms may shift to closed, monetised models once open-source has built sufficient global developer adoption.