Huawei Open Sources a 505 Billion Parameter AI Model Built Without Nvidia Chips

Huawei Technologies Co. Ltd. open sourced openPangu-2.0-Pro, a 505 billion parameter mixture-of-experts AI model trained entirely on its own Ascend NPUs without any Nvidia hardware, on July 31. The model, with 18 billion active parameters per token and a 512K context window, was trained on roughly 34 trillion tokens, and its weights, inference code, and technical reports are available on GitCode's Ascend Tribe community. This release challenges the assumption underlying US export controls that frontier-scale AI training requires Nvidia chips, demonstrating that China can build such models with domestic silicon.

Huawei just released the full weights of a 505 billion parameter AI model trained entirely on its own Ascend chips, proving China can build frontier-scale AI without a single Nvidia GPU. On July 31, Huawei open sourced openPangu-2.0-Pro, a mixture of experts model with 505 billion total parameters and 18 billion active per token, trained on roughly 34 trillion tokens with a 512K context window. The company released the model weights, inference code, and technical reports on GitCode's Ascend Tribe community, and it did all of that training on its own Ascend NPUs. No Nvidia hardware touched this model at any point. That last detail is the whole story. For three years, the working assumption in Washington and in Silicon Valley has been that you simply can't train a model at this scale without Nvidia's H100s or their successors. Export controls were built on that premise. OpenPangu-2.0-Pro is a named, verifiable, open-weights model that breaks it. Huawei isn't hiding how it got here, either. The company has said it plans to open-source seven major training components in stages, including pre-training code, post-training code, and the training operators themselves. A smaller companion model, openPangu-2.0-Flash, with 92 billion total parameters and 6 billion active, already shipped on June 30 under the same plan. Anyone can now download the weights, read the technical report, and check the claims against reality. That's a different posture than a press release boasting about a chip roadmap. It's a model you can put on a GPU cluster, or an Ascend cluster, and run. The Chokepoint Story The timing isn't an accident. This is the same summer ASML's lithography monopoly and CXMT's IPO have both been reminding the market how much of the AI supply chain still runs through a handful of chokepoints Washington has spent years trying to control. OpenPangu-2.0-Pro adds a third data point: the chokepoint on advanced GPUs hasn't stopped China from fielding a frontier-scale model. It's forced Huawei to build its own stack instead, and that stack now appears to work. Nvidia's moat was never really about raw compute. It was about the argument that nobody else's silicon could get you to the frontier. Jensen Huang has said as much in public remarks about the difficulty of matching CUDA and Nvidia's networking stack. OpenPangu-2.0-Pro doesn't disprove that Nvidia hardware is faster per chip. Huawei itself claims its single-card throughput on Ascend is roughly double that of other mainstream open-source models running on the same silicon, which is itself an admission that efficiency, not just raw parameter count, is the harder problem. But throughput comparisons are beside the point here. The model exists, the weights are public, and it was built entirely on hardware the US government spent years trying to keep out of Huawei's hands. Frankly, the more interesting number isn't 505 billion. It's 34 trillion. Training tokens at that scale require sustained, reliable access to enough compute to run for months without falling over. Getting an Ascend cluster to hold together through a training run that long is an infrastructure achievement most outside observers assumed was still years away for Huawei, given how much of its chip production has been squeezed by sanctions on lithography equipment and advanced packaging. That's the harder problem, and Huawei apparently solved it. What Happens Next None of this means Nvidia's business is suddenly at risk in the markets that matter most to it. Not yet, anyway. Huawei's Ascend chips are still overwhelmingly a China-market story, largely because export rules keep them out of the US and most allied markets, and American cloud providers have no reason to switch. But the export control logic was never really about market share inside China. It was about denying China the ability to reach the frontier at all. OpenPangu-2.0-Pro, with its weights sitting on a public repository for anyone to inspect, is hard evidence that the denial strategy has a hole in it. What happens next matters more than the release itself. If Huawei follows through on open-sourcing the remaining training components, including the operators and post-training pipeline, other Chinese labs will be able to replicate the Ascend training recipe rather than reinvent it. That's how you turn one frontier model trained without Nvidia into a standard practice. Also read: Executive Order 14409 Just Set the Line for Dangerous AI Models https://startupfortune.com/executive-order-14409-just-set-the-line-for-dangerous-ai-models/ • Y Combinator Open-Sources QM, the AI Agent Harness It Uses to Run Itself https://startupfortune.com/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself/ • OpenAI's and Anthropic's AI Agents Escaped Testing and Hacked Real Firms https://startupfortune.com/openais-and-anthropics-ai-agents-escaped-testing-and-hacked-real-firms/