Alibaba has made an unusually candid admission for a technology company. Its in-house AI chips are not as good as those made by Nvidia or AMD. And, the company concedes, they may never be.
Alibaba's semiconductor unit T-Head unveiled the Zhenwu 810E in January 2026. The chips have since achieved scaled mass production, with more than 470,000 units cumulatively delivered. CEO Yongming Wu revealed the scale of output during the company's earnings call this week.
The volume is impressive in isolation. The Zhenwu 810E is equipped with 96GB of HBM2e memory and offers 700GB/s of inter-chip bandwidth, with performance described as comparable to Nvidia's H20. To put that in context, Nvidia has publicly stated it manufactures around six million GPUs annually. If Alibaba has delivered 470,000 units since January, it is currently operating at a comparable production rate.
But the underlying problem remains unresolved. "Given that our chips still lag behind foreign counterparts and performance in various respects, we aspire to engage in more profound co-design with Alibaba's cloud infrastructure and the Qwen model," Wu said during the call. In other words, Alibaba cannot beat Nvidia at raw speed. So it is betting on a different strategy entirely.
Full-Stack Integration as Competitive Answer
The company's playbook here is not novel, but it is disciplined. Rather than try to match Nvidia's architectural prowess or manufacturing scale, Alibaba is optimising its entire ecosystem around its own silicon. The Zhenwu 810E has been deployed in large clusters on Alibaba Cloud, serving more than 400 customers, including State Grid, the Chinese Academy of Sciences, and XPeng Motors, and has been deeply optimised for the training and inference of its Qwen large language models.
This approach addresses a real constraint. Against the backdrop of an intense trade stand-off with the United States, Beijing views the move as a major power play to significantly boost the development of its domestic semiconductor sector. Beijing's confidence reflects its assessment that domestic alternatives now provide comparable performance to American offerings.
For Alibaba, the geopolitical pressure is also commercial opportunity. Wu emphasised that developing its own chips gives Alibaba "guaranteed supply of AI computing power" in a context of US export controls. When supply is constrained and competition is fierce, controlling your own foundational layer becomes valuable.
Racing the Cloud Market
Alibaba's cloud business revenue jumped 36% in the latest quarter to 43.3 billion yuan from a year ago. That growth is real. But it is also expensive. Profit for the quarter was down 67% year-on-year, in part due to growing marketing and sales expenses.
Wu believes the unit economics improve if Alibaba can optimise inferencing costs by controlling more of the stack. The company's AI strategic goal is to exceed $100 billion in cloud and AI commercial revenue, including MaaS, within five years. That target sounds audacious on its surface. But given current trajectory and if the full-stack model actually delivers the efficiency gains Alibaba is banking on, it is plausible.
The deeper question is whether the world really needs another mainstream AI chip supplier. Nvidia's dominance rests partly on technical superiority and partly on switching costs and developer lock-in through CUDA. Alibaba's solution sidesteps that lock-in by accepting modest performance trade-offs in exchange for lower inferencing costs and tighter integration. That works for captive workloads inside Alibaba's own cloud and for customers optimised for Qwen. It is less clear whether it can sustain long-term competition in an open market.
For now, Alibaba has telegraphed its ambition clearly: it will compete not by beating Nvidia at the spec sheet, but by offering a complete package at lower cost for customers willing to standardise on its platform. That is a legitimate strategy. Whether it succeeds depends far less on chip performance than on whether customers value the promise of full integration enough to accept a trade-off in raw horsepower.