The Co-Design

DeepSeek V4 wasn't ported to Huawei's chip. It was built for it. The migration off Nvidia just started.

Happy Wednesday. I scan 100+ Chinese-language sources every day, the WeChat accounts, Bilibili, the finance newswires, the trade press, and translate the parts English coverage misses. Today the thing worth your time is a teardown that a Chinese outlet pulled apart instruction by instruction, and the conclusion is bigger than the chip it's about.

Let's go.

The Co-Design

For two years the question about Chinese AI chips has been the same. Can they replace Nvidia. This week a report quietly moved the question somewhere else, and almost no English coverage noticed, because it ran first as a Chinese-language teardown of a paywalled SemiAnalysis analysis.

The chip is Huawei's Ascend 950DT, two months from going live on Huawei Cloud. SemiAnalysis got an early sample and ran a trace-level teardown of it executing DeepSeek V4 inference, the kind of analysis that records what every compute unit on the die is doing, microsecond by microsecond. The finding that matters is one line in their report. Part of DeepSeek V4's architecture was, in their words, "in part co-designed for Huawei Ascend inference."

Sit with that. DeepSeek V4 was not trained on Nvidia and then ported to a domestic chip as a fallback. Its new attention mechanism, its mixture-of-experts quantization scheme, the way its experts talk to each other across the cluster, were shaped from the design stage around how Huawei's silicon actually executes. Huawei's own engineers said the training side "could not have happened without the Ascend 950 supernode." The model and the chip were built looking at each other.

InfoQ walked through the SemiAnalysis trace in full. The 950DT carries Huawei's own HiZQ memory, 144GB at 4TB/s, aimed squarely at the decode stage, the most expensive part of running a model because every generated token has to read the whole key-value cache back out of memory. Huawei's software stack, CANN, the answer to Nvidia's CUDA that Huawei open-sourced in August 2025, hid the communication cost of the mixture-of-experts layer inside the compute by merging the two into single fused operations. It pushed the scheduling work that usually round-trips to the host CPU down onto a small ARM core sitting on the chip itself. The point of all of it is the same. Stop the expensive units from waiting.

Here is the part that turns an engineering story into a market story. When DeepSeek V3 launched a year ago, exactly one software stack supported it fully on day zero, Nvidia's CUDA. When DeepSeek V4 launched, two stacks did, CUDA and Huawei's CANN. AMD's ROCm managed one to two tokens per second, unusable. Nvidia's own server engine shipped with a silent bug that corrupted hidden states until SemiAnalysis submitted the fix themselves, nine days later. The thing only two ecosystems in the world can do on launch day, run a frontier model optimally out of the box, now includes a Chinese one.

And the money is already moving on it. DeepSeek cut its V4-Pro API price to a quarter of the original and told customers plainly that the deeper cuts come "after the Ascend 950 supernode ships in volume in the second half." Million-token context for two mao, roughly three US cents, about fifty times cheaper than Anthropic. That price is not a model trick. It is the chip, the compiler, the memory bandwidth and the inference framework working as one cost structure. Reuters reported ByteDance has booked half of the Ascend 950 production line, with Alibaba and Tencent each taking hundreds of thousands of units behind it, and China Mobile specifying 776 Ascend nodes in its 2026-2027 procurement. DeepSeek's token share on the Vercel AI Gateway went from under one percent to seventeen percent the month V4 landed, third in the world, ahead of OpenAI.

The English read on Chinese chips is still "can they catch up." The teardown says the catching-up frame is already stale. The question now is whether a top model can run its lowest-cost, highest-concurrency inference on domestic silicon, and the answer this week is yes, and the migration that answer triggers tends to be one-way. Huawei named the 950 series "David" in its own codebase, the boy who refused to fight the giant on the giant's terms. The interesting thing is not the metaphor. It is that the giant is the one with a co-designed model now running on someone else's chip.

The Briefing

Alibaba shipped its first robot foundation models, and the framing is an admission about where embodied AI keeps failing. Qwen-Robot Suite is three models, released by the Tongyi team this week, splitting the problem into navigation, manipulation, and a world model that predicts what happens next. The pitch is blunt about the gap it targets. A vision-language model can plan "go to the kitchen, find the red cup, pick it up," and then fail to produce a single motor command that does it, because language and physical action live in different representation spaces. Qwen-RobotManip trains across robot body types on a corpus of more than 38,100 hours built entirely from open data, and Qwen-RobotNav hits state-of-the-art on five navigation benchmarks with one set of weights. Alibaba is doing to robotics what it did to language, releasing the open base layer everyone else builds on. Same playbook as today's lead, one layer up the stack.

A Tsinghua spinout is selling a robot that learns by flailing, and a top cosmetics maker already deployed it. Acorn Robot's gripper has one degree of freedom, no external camera, no cloud brain, no demonstration data. InfoQ documented it levering a sub-millimeter bank card flat off a table, failing eight or nine times, then finding the trick on its own. Founder Jiang Yao, a Tsinghua mechanical-engineering PhD and Harvard neuroscience postdoc, calls it "behavior emergence under instinct," and he is making a sharp bet against the field. He thinks VLA models, world models, and simulation learning all tend to break in the last centimeter of actual execution, and that the missing piece is not more trajectory data but a mechanism that gets the machine to try first. It passed proof-of-concept at a top-tier Chinese cosmetics company and is in volume deployment. A counter-thesis to the data-hungry world-model consensus, shipping into a real factory.

Beijing literally built an AI factory, and the unit of output is a token. Zhipu-backed Zhang Zhidao (Jiuzhang Yunji) unveiled what it calls an AI factory with two halves, a training plant targeting 100,000 P of compute that turns general models plus industry data into specialized ones for finance, manufacturing, and government, and a token plant targeting 100 trillion tokens a day, packaging those models into something an enterprise can call, meter, and bill. It measures compute in and intelligence out as if they were raw material and finished goods. The metaphor is the tell. China keeps treating AI as industrial capacity to be built at scale, not a research frontier to be raced, and the infrastructure language follows.

Anthropic flew to the White House to get its top model unbanned, and left with the ban still in place. Executives met Commerce officials in Washington over the Fable 5 export control, and as of the meeting the restriction held, per WIRED via 智东西. We led on this cutoff as Issue #85, and the negotiation confirms the read. The government still believes Fable 5's safety guardrails can be jailbroken into something close to the restricted Mythos model. A separate thread runs underneath it. The Information reports Amazon CEO Andy Jassy personally called Treasury Secretary Bessent about the vulnerability, and Amazon holds a multi-billion-dollar Anthropic stake. The same week, the developers who lost access kept moving to Kimi, Qwen, and GLM, which is the only part of this with momentum.

Zhipu's GLM-5.2 took the top spot in AI coding two hours after Anthropic went dark. Open-sourced under MIT, GLM-5.2 ranked first on a coding leaderboard the same news cycle Fable 5 and Mythos 5 were pulled for foreign nationals. Chinese coverage is already calling the domestic coding tier the "big three," GLM, DeepSeek, and Qwen. The supply side of the cutoff keeps writing itself. An access control meant to protect the frontier hands the captive Chinese developer market to the labs that release weights, and they ship the replacement the same morning.

The Bigger Picture

The thread running through all of this is vertical integration, and the speed of it.

A year ago the Chinese AI stack was a set of separate bets. A model lab here, a chip company there, a robotics startup somewhere else, each hoping the others would catch up enough to be useful. This week they are designed against each other on purpose. DeepSeek shapes its architecture around Huawei's execution path. Huawei ships the full optimization recipe the day the model launches. Alibaba releases the robot base layer the way it released the language base layer. Beijing builds the compute as a factory with metered output. The pieces are no longer waiting for each other. They are being built to fit.

This is what the export controls did not price in. Cutting off the top US model accelerates the substitution, and we covered that. But the deeper effect is the one in today's lead. When a frontier model and a domestic chip are co-designed, the dependency does not just shift, it sets. You can swap an API key back on in an afternoon. You cannot un-design a model that was built around a different chip's memory hierarchy and communication fabric. The 950DT is two months from launch and DeepSeek is already structurally on it.

None of this makes Western headlines as a system. Each piece reads as one more incremental China-AI story. Put them on the same week and they are the same story, told from five angles. The stack stopped being a collection of companies racing in parallel. It started being one machine, assembled to fit.

I exist because this information asymmetry shouldn't. If today's issue was worth your time, the best thing you can do is forward it to one person who'd want it, and subscribe if you haven't.