China's AI Chip Gamble: Can Huawei, Alibaba, and Baidu Really Dethrone Nvidia?
Nvidia's dominance in China's AI sector has been undeniable for years. Their GPUs have been the workhorse for everything from search engines to generative AI. But with tightening export rules from the US, China's been forced to look inward, betting on domestic alternatives like Huawei, Alibaba, Baidu, and Cambricon to fill the void. The question isn't whether they can build chips – it's whether they can build competitive chips, and fast enough to avoid falling behind.
The shift away from Nvidia wasn't purely about technological capability; it was a strategic decision. Reports of state media labeling Nvidia's H20 as potentially compromised, coupled with alleged directives to cancel orders from companies like Alibaba and ByteDance, paint a picture of Beijing losing patience. The risk of relying on a foreign supplier, even with "China-only" versions of chips, apparently outweighed the immediate performance benefits.
Huawei is arguably the frontrunner. Their Ascend line of AI chips has been maturing under sanctions, and their roadmap extends to 2027 with the promise of petaflop-level performance. The Ascend 910B, while trailing Nvidia's H20 in memory capacity and interconnect speed, has become the go-to option for those cut off from Nvidia. Huawei even claimed it outperformed the A100 (Nvidia's top chip from 2020) in some training tasks back in 2024. But these claims, without rigorous, third-party benchmarks, should be taken with a grain of salt.
Alibaba, through its T-Head unit, is also making a play with its PPU chip, positioned as a direct rival to the H20. The company is using over 16,000 PPUs in a China Unicom data center, suggesting a serious commitment. The motivation here is clear: Alibaba Cloud's business depends on reliable access to AI chips, and they can't afford to be at the mercy of US export controls.
Baidu, not to be left out, unveiled a 30,000-chip cluster powered by its P800 processors. These chips, according to Guosen Securities, reach roughly 345 Tflops at FP16, putting them in the same ballpark as Huawei's 910B and Nvidia's A100. Baidu's stock has jumped 64 percent this year, with the Kunlun reveal playing a central role. But the future isn’t all sunshine. Rumors out of Seoul suggest Samsung has paused production of Baidu’s 4-nm designs.
Cambricon, a publicly traded chip company, has seen its share price jump nearly 500 percent over the past year. Their MLU 590 chip, delivering 345 Tflops at FP16, has been a game-changer, restoring investor confidence. The MLU 690, currently in development, is rumored to potentially rival Nvidia's H100 in some metrics.
But raw hardware performance is only half the battle. Nvidia's CUDA platform has a massive head start, and Chinese developers are accustomed to it. Forcing them to adapt to proprietary software stacks from Huawei, Alibaba, or Baidu could introduce friction and slow down AI development. DeepSeek's reported delays in developing its next AI model, attributed to switching to Huawei's chips, illustrate this challenge. Is the juice worth the squeeze?
It’s like switching from a manual transmission to an automatic. Sure, the automatic might be newer, but you lose a level of control, and there's a learning curve involved. How many developers will be willing to make that switch?
And this is the part of the report that I find genuinely puzzling: Why not invest in building an open-source alternative to CUDA, rather than fragmenting the ecosystem with competing, proprietary platforms? It seems like a missed opportunity for collaboration.
The Chinese government is essentially betting that strategic independence is worth the potential short-term performance hit and the cost of re-tooling the software ecosystem. Whether that bet pays off remains to be seen. As China's AI Chip Race: Tech Giants Challenge Nvidia highlights, this is a complex and evolving landscape.
The question isn't whether China can build AI chips. They're already doing it. The real question is whether they can achieve parity with Nvidia in terms of performance, software support, and end-user trust. The jury's still out, but one thing's clear: China's no longer content playing second fiddle in the AI race.
