Against the backdrop of US-China chip competition, Alibaba's T-Head has delivered something worth taking seriously.
Zhenwu M890—the latest AI accelerator chip from T-Head, announced yesterday. According to TrendForce and Wccftech specs, M890 delivers 3x the performance of its predecessor, with 144GB HBM3 memory, targeting both training and inference workloads.
The direct target: NVIDIA H20, the downgraded chip NVIDIA supplies to the Chinese market. Wccftech's headline was blunt: "Alibaba Targets NVIDIA's Hopper With Zhenwu M890."
Key specs:
- Performance: 3x previous generation
- Memory: 144GB HBM3
- Use case: Training + Inference
- Roadmap: New chips in Q3 2027 and Q3 2028
Important context: NVIDIA H20 is an export-controlled "China special"—performance well below H100/H200. If M890 can match or exceed H20, it gives Chinese AI companies an alternative to relying on NVIDIA's downgraded lineup.
But "3x previous gen" needs careful reading. Which previous gen? If it was already an entry-level product, 3x growth might just be catching up to a competitor's previous generation. Wccftech used "claiming 3x the H20 performance"—Alibaba's own claim, not a third-party benchmark.
T-Head also announced its own independent IPO plans. Per CRN Asia, T-Head is targeting the enterprise AI chip market and hopes to fundraise through a separate listing. This is a big move within Alibaba—spinning out the chip business to face capital markets independently.
Another notable detail: T-Head simultaneously announced the Qwen 3.7-Max large model. The chip + model combo mirrors NVIDIA's "CUDA + GPU" strategy. If M890 can run Qwen series training and inference, the ecosystem loop starts to take shape.
My take: M890 probably isn't competing with NVIDIA's high-end (H100/B200), but at the H20 tier, it's a credible alternative. For Chinese AI companies affected by export controls, this is one of the closest-to-production domestic AI training/inference chips available.
But chip competition isn't won on launch day—it's about ecosystem. CUDA's moat isn't hardware performance; it's a decade-plus of developer tools and software stack. T-Head needs to answer: what do developers code in? Are there mature framework supports? How high is the migration cost?
The launch event won't answer these. Only actual deployment will.
Key sources: