In a moment that reflects the deeper logic of technological nationalism, China Telecom has trained two large language models — one exceeding a trillion parameters — entirely on domestically manufactured chips, signaling that Beijing's AI sector is no longer merely enduring American export restrictions but actively constructing an alternative to them. The effort, centered on Huawei's Ascend processors and Cambricon semiconductors, represents less a technical triumph than a strategic declaration: that necessity, applied with enough resources and resolve, can reshape the boundaries of what a nati
China Telecom trains advanced AI models on domestic Huawei chips amid US sanctions
China is building around sanctions, not waiting for them to lift.
Mark
So China Telecom trained AI models on Chinese chips instead of NVIDIA. Why does that matter beyond just being self-sufficient?
Mimi
Because it breaks the dependency. If you can't buy NVIDIA chips, you can't train large models—or you couldn't. Now China is showing that Huawei and Cambricon chips can do the job. That changes the calculus for every other Chinese AI company.
Luke
But we don't actually know how well these chips perform compared to NVIDIA. China Telecom published that they trained the models, not that they match NVIDIA's speed or efficiency.
Mimi
True. But the fact that they're publishing results at all suggests they're confident enough to stake their reputation on it. That's different from a year ago when the only option was to hide and buy black market chips.
Mark
What about the trillion-parameter model? Is that big?
Mimi
It's substantial. For comparison, some of the largest models in the world operate in that range. It shows China isn't just training small models—they're going after scale.
Luke
Though we don't know what that model does, how it performs, or whether it's actually useful. A trillion parameters is a number. Capability is something else.
Mark
So this is more about proving it's possible than proving it's better?
Mimi
Exactly. It's a proof of concept. It says to the rest of China's tech industry: you don't have to wait for American chips. You can build now.
Luke
And that matters geopolitically, which is probably why China Telecom announced it. It's not just a technical achievement—it's a political one.
Mimi
Right. It's saying China can compete in AI without relying on the US. Whether that's actually true at scale is still an open question.
Mark
What happens next?
Mimi
More companies will probably try the same thing. If Huawei and Cambricon can prove their chips work, demand will grow. The ecosystem gets stronger.
Luke
Or the chips hit a wall and companies realize they still need NVIDIA. We won't know for a while.
Le Pouls
- US export controls on NVIDIA's most powerful chips have forced China's AI industry into a corner, cutting off access to the hardware that once powered its most ambitious models.
- Rather than accept the constraint or rely on black-market workarounds, China Telecom has published working AI models trained entirely on Chinese silicon — turning a defensive posture into a public proof of concept.
- Huawei's Ascend chips and Cambricon processors are now being tested and deployed at scale, with China Telecom's State Cloud integrating both into live AI training infrastructure.
- The trillion-parameter model represents a significant leap in ambition, though domestic chips have not yet been proven to match NVIDIA's speed and scale across the broader industry.
- Each published result from China's domestic chip ecosystem makes the alternative supply chain more credible — and more attractive to other Chinese firms still weighing their options.
In a moment that reflects the deeper logic of technological nationalism, China Telecom has trained two large language models — one exceeding a trillion parameters — entirely on domestically manufactured chips, signaling that Beijing's AI sector is no longer merely enduring American export restrictions but actively constructing an alternative to them. The effort, centered on Huawei's Ascend processors and Cambricon semiconductors, represents less a technical triumph than a strategic declaration: that necessity, applied with enough resources and resolve, can reshape the boundaries of what a nation builds for itself. It is a reminder that geopolitical pressure does not simply constrain — it also creates.
China Telecom's Institute of AI announced this week that it has trained two large language models entirely on domestically manufactured chips — one called TeleChat2-115B, now publicly available, and a second, unnamed model built with a trillion parameters. Both were trained without a single American processor, a milestone that reflects how deliberately Beijing is moving to insulate its AI sector from US export controls.
For years, Chinese firms depended on NVIDIA's A100 and H100 chips to power their most capable systems. Washington's systematic restrictions on those exports left the industry with few good options: accept reduced performance, seek black-market alternatives, or build something new. China Telecom is demonstrating the third path — and publishing the results to prove it works.
Huawei is central to this effort. Its Ascend line of AI processors, including the recently tested Ascend 910C, has become a cornerstone of China Telecom's infrastructure. The company has also integrated chips from Cambricon, a domestic semiconductor startup, into its State Cloud AI training service. Neither company has released direct performance comparisons with NVIDIA, but the willingness to publish model results trained on their hardware suggests the gap may be closing.
The challenges ahead remain real. Domestic chips are not yet proven at the speed and scale NVIDIA delivers, and the broader ecosystem — supply chains, tooling, software — is still maturing. But China Telecom's announcement is less about solving the problem than about changing the terms of it. The country is no longer waiting for sanctions to ease. It is building around them, one published model at a time.
China Telecom's Institute of AI announced this week that it has successfully trained two large language models using tens of thousands of domestically manufactured chips—a milestone that signals how quickly Beijing is moving to insulate its artificial intelligence sector from American export controls. One model, called TeleChat2-115B, is publicly available. The other remains unnamed but carries substantially more computational weight: it was built with a trillion parameters, a measure of the model's complexity and capability. For context, TeleChat2-115B operates at just over 100 billion parameters. Both were trained entirely on Chinese silicon.
The announcement matters because it demonstrates that China's AI industry is no longer simply waiting for workarounds to US sanctions. Instead, it is building an alternative supply chain. For years, Chinese firms have relied on NVIDIA's A100 and H100 processors to train cutting-edge AI systems. Washington has systematically restricted exports of these chips, forcing Chinese companies to either go without or hunt for black-market alternatives. China Telecom's move suggests a third path: develop homegrown replacements and prove they work.
Huawei, the Shenzhen technology company that has itself been under US restrictions for years, is central to this strategy. The company has been developing its Ascend line of AI processors and recently began testing the Ascend 910C with potential customers. China Telecom has previously disclosed that it uses Huawei's Ascend chips in its AI infrastructure. The company has also integrated processors from Cambricon, a domestic semiconductor startup, into its State Cloud service for AI training. Neither Huawei nor Cambricon has disclosed exactly how their chips compare to NVIDIA's best offerings, but the fact that China Telecom is publishing results trained on these systems suggests the performance gap may be narrowing.
The broader picture is one of necessity driving innovation. US export restrictions have created a hard constraint: Chinese AI labs cannot easily access the hardware they once relied on. Some firms have responded by buying lower-powered alternatives and accepting reduced performance. Others have turned to the black market. But the most strategic response is the one China Telecom is demonstrating: invest in domestic chip makers, train models on their hardware, publish the results, and prove to other companies that the alternative ecosystem is viable.
This does not mean China has solved the problem. The country's AI industry still faces significant challenges in scaling training and deployment across the sector. Domestic chips may work, but they are not yet proven at the scale and speed that NVIDIA's processors deliver. The ecosystem is young. Supply chains are untested. But China Telecom's announcement is a signal that the country is no longer waiting for sanctions to be lifted. It is building around them. As geopolitical competition in AI intensifies, expect more announcements like this one—each one a small proof that the alternative is becoming real.
Citations marquantes
China Telecom's achievement demonstrates the country's strides towards chip independence amid increasing challenges in accessing foreign semiconductors— China Telecom Institute of AI (paraphrased)
Envie de l'histoire complète ?
Lire l'original sur Firstpost ↗
Technology