The Pingtouge AI chip Zhenwu M890 makes its debut, with performance improved up to 3 times.

Pingtouge just showed its hand at the 2026 Alibaba Cloud Summit, and the Zhenwu M890 is the most serious AI silicon China's hyperscalers have produced outside of Huawei's shadow. But let me be blunt: being "serious" isn't the same as being enough.

The raw numbers look respectable on paper. 144GB of memory crammed onto a single die, 800GB/s chip-to-chip bandwidth, and a claimed 3x performance leap over the 810E it replaces. Support for everything from FP32 down to FP4 means this thing is marketed as a Swiss Army knife — training on high precision, then slashing precision to FP8 or even FP4 when you just need raw inference throughput. Alibaba also shipped an in-house interconnect switch, the ICN Switch 1.0, enabling 64-card full-bandwidth clusters. On stage, they framed all of this as the backbone of a "芯-云-模型-推理" stack, their chip-to-cloud-to-model-to-inference pipeline for what they're calling the Agentic Era.

First, let's talk about the elephant in the room that Alibaba's PR carefully sidestepped: where does this actually land on the global performance curve? Nobody outside Alibaba has run independent benchmarks, so all we have is the internal claim of 3x over 810E. Three times a chip that was already a generation behind NVIDIA's H100 when it launched is not, mathematically, a frontier product. It's a chip designed to make Chinese cloud customers feel comfortable that they won't be stranded without CUDA access. That's a perfectly valid business goal — arguably the most important one in 2026 — but it's not the same as competing at the bleeding edge. The fact that Alibaba doesn't even bother to compare the M890 against NVIDIA or AMD in their press materials tells you everything you need to know about their confidence in a head-to-head race.

That said, dismissing this chip as merely "catching up" misses the real strategic picture. The interconnect story is where Pingtouge is actually making a clever bet. 64 cards at full bandwidth, enabled by a custom switch, is a meaningful architectural choice. The dirty secret of AI clusters is that FLOPS don't matter if your interconnect turns into a traffic jam the moment you scale past eight or sixteen nodes. NVIDIA's NVLink dominance isn't just about speed — it's about the entire ecosystem of switches, topology management, and software that makes large-scale training actually work. By building their own switch silicon in-house, Alibaba is attempting to own the full stack in a way that even most Western hyperscalers haven't bothered to. Google has TPUs and custom networking. Amazon has Trainium and EFA. But Alibaba is explicitly packaging the switch as a first-class citizen alongside the accelerator, which signals they've learned the painful lesson that you can't outsource your interconnect and still claim real independence.

The FP4 support is another angle worth dissecting. On the surface, training in FP32 and inferring in FP4 sounds like the kind of flexibility every cloud customer wants. In practice, FP4 is still a circus — the precision loss is real, the quantization techniques required to make it usable are finicky, and the number of models that actually benefit from sub-FP8 inference without catastrophic quality degradation is small. Alibaba is clearly betting that the economics will force the industry toward lower precision regardless. When you're serving millions of inference requests per second for agentic workloads — long-running, multi-step, tool-calling chains that consume orders of magnitude more compute than a simple chat response — every decimal point of efficiency matters. If FP4 gets you even 30% more throughput at acceptable quality, that translates to real margin at Alibaba Cloud scale. But this is a bet on the future of model architectures and quantization research, not a guaranteed win.

Now, the phrase "Agentic Era" deserves its own moment of scrutiny. Every major cloud vendor is currently tripping over themselves to brand their entire stack around AI agents — autonomous systems that plan, reason, call tools, and execute multi-step workflows. Alibaba is no exception. But branding your chip launch around agents is a narrative convenience, not a technical reality. The M890 doesn't have any agent-specific features. There's no hardware-accelerated planning module, no dedicated silicon for tool orchestration, or anything that would make it meaningfully better for agentic workloads than any other accelerator with comparable memory and bandwidth. What they're really saying is: "Our cloud, powered by our chips, will be the best place to run the agent frameworks we're also building." That's a platform play, pure and simple. And honestly? It's the right play. The winners in the agentic wave won't be whoever has the fastest single-chip FLOPS count — they'll be whoever can offer the most cost-effective, reliable, end-to-end system for deploying agents at scale.

Which brings me to the uncomfortable comparison nobody in Chinese tech wants to have: Huawei's Ascend series is still the eight-hundred-pound gorilla in the domestic AI chip space. Ascend 910C has real production volume, real customer adoption, and the implicit backing of national semiconductor policy. Pingtouge has been chasing respectability for years, and while the M890 is their most credible product yet, Alibaba's chip division still operates in Huawei's gravitational field. The difference is that Alibaba controls its own cloud — they don't need to sell M890s to third parties. They just need them to be good enough to run Alibaba's own models and serve Alibaba's own customers. That's a much lower bar than competing in the open market, and it's exactly the bar they should be aiming for.

Is the M890 a breakthrough? No. Is it a credible step toward reducing Chinese cloud's dependence on a supply chain that随时 could be severed by another round of export controls? Absolutely. The real question isn't whether this chip can beat NVIDIA — it can't, and everyone in the room knows it. The question is whether it can power a self-sufficient ecosystem where Chinese developers can build, train, and deploy AI at scale without ever needing to ask Washington for permission. On that metric, the M890 looks like a genuine piece of progress. Whether it's enough progress, fast enough, is a question that only the next two years of escalating geopolitical pressure will answer.