Huawei names its 2027 Nvidia challengers: the 960DT and Ascend 960PR
Huawei put dates and part numbers on its AI chip roadmap for the first time on Thursday. In the same hour, two labs published work that shrinks the machine a 35B model needs and shows what coding agents can build when nobody hands them a strategy.
Huawei's rotating chairman David Wang told a company event in Shanghai that Huawei will ship the 960DT in the first quarter of 2027 and the Ascend 960PR in the third quarter, naming two specific parts for the first time rather than describing a roadmap in general terms. Wang framed the plan around interconnect, not silicon: Huawei's UnifiedBus technology is the "key link" in its next-generation AI systems, he said, and the company has now developed 11 semiconductors built on it. Its largest linked systems — Huawei calls them superclusters — can tie together up to 1 million AI processors, and it has already shipped more than 1,000 smaller linked systems, or supernodes, to more than 370 customers. Wang did not say how many chips fit in a single supernode or name any of those customers.
That framing is a direct answer to the constraint Washington imposed. Export controls keep China from freely buying the most advanced Nvidia parts, so the way to more compute is to wire more locally available chips together — which only pays off if those chips can exchange data fast enough. It also continues an argument Huawei's supervisory board chairman Guo Ping made internally last week, when Huawei's chairman says its computing goal is to become Nvidia and conceded the company's real advantage is running other labs' models rather than owning a frontier one. Huawei's own numbers show where the gap sits: Wang said its AI chip ecosystem has 5,270 monthly active developers. Nvidia's moat was never the die; it is CUDA, and five thousand developers is a rounding error against it. Naming the chips is the easy half of the announcement — the developer count is the half that decides whether the 960 series matters outside China.
A streaming inference engine called Edge0 serves a 35B-parameter mixture-of-experts model at 20 tokens per second inside 3 GiB of peak active memory on a single 24GB machine — with the weights living on an SSD instead of in RAM. The paper behind it, from Yu Lin and co-authors at the Chinese Academy of Sciences' Institute of Computing Technology, targets the half of the memory wall that sparsity does not solve: a 35B model is 19.5GB at 4-bit, and offloading to SSD normally fails because layer N+1's experts must be chosen before layer N's output exists, so the reads cannot start early enough. Edge0's answer is a per-layer "prerouter" that predicts the next layer's routing one token ahead and consumes the prediction as the routing itself, so nothing is dropped, plus a recovery LoRA that pays back the quality lost to int4 quantization. Across five public benchmarks the released 35B tier lands within a few points of its fp16 teacher, and the framework, checkpoints and adapters are open, with MLX support for Apple Silicon today. Total parameter count stops dictating resident memory — that is the part worth keeping.
Frontier coding agents can now build winning game players from scratch: given a game description, a raw observation-and-action interface and an empty policy file, a general-purpose coding agent beating every fair built-in StarCraft II AI and two cheating variants in a single autonomous session. The paper's Gauntlet framework freezes the agent's finished controller and scores it on held-out instances with zero model calls during play, which is what makes the result more than a demo. On an unpublished procedural roguelike, held-out success ranged from 0% to 86% and exposed a sharp generational threshold — every observed session of the newest-generation system beat the best session of its predecessor. Single-session programs also won complete games of Freeciv by total conquest on held-out seeds. No perception stack, no skill library, no planner supplied by the researchers; the agent wrote the controller. That is a different claim from "agents can play games," and it is the one with engineering consequences.
What to watch: whether Huawei's 2027 dates survive contact with fabrication capacity, and whether anyone reproduces Edge0's 3 GiB figure on hardware that isn't an Apple Silicon Mac.
Would you trust a 35B model streaming off your laptop's SSD for work you'd normally send to an API? Tell us in the comments.
Sources: Edge0 (GitHub)
