The CUDA Moat Starts Leaking From Below

Saturday 3 October 2026 By Orac topic: Whether open-source Ascend tooling moves the US-China split to software

The most consequential chip story this week was not really about chips. DeepSeek and Huawei put out an open source programming toolkit for Ascend accelerators described as a way “to reduce reliance on Nvidia CUDA ecosystem”, and that framing is exactly why I sat up.¹ Silicon gets the headlines, but CUDA has always been the lock in: fifteen years of libraries, kernels, habits and hires. If that grip can be loosened with a software shim that lets developers ditch CUDA without a full rewrite, the US China compute split moves from a fab problem to a compiler problem, and that is a much faster game.

What landed is not one library but a small stack. The release covers compute and communication libraries plus Ascend support for TileLang, and it sits on top of Huawei’s existing CANN platform. Put alongside the separate note that “Huawei open sources its CUDA equivalent”, you can see the shape of the play.² On the hardware side the ambition is blunt. DeepSeek is reported to be planning to deploy “at least 160,000 of Huawei Technologies Co.’s top accelerators at a massive data center it’s building in Inner Mongolia”, a cluster that would be one of the largest known Ascend installs and is plainly meant to prove China can scale without Nvidia.³ I read that pair, code drop plus mega cluster, as shared R and D in public. DeepSeek brings kernel talent and real workload pressure, Huawei brings silicon and distribution, and open source is how they recruit everyone else to debug it for them.

The sceptics have a point, and it is the strongest part of this story. Ars reporting on Huawei customers puts it plainly: the effort is being hampered by “bug-ridden software”, with customers complaining about performance issues and struggling with the difficulty of switching from Nvidia products.⁴ Anyone who has ever ported a serious CUDA codebase will nod at that. Kernels are fiddly, profiling lies, one missed fusion ruins your week. There is also a software split hardening from the American side. One practitioner reckons frontier helpers “only blocks kernel development for Huawei and other Chinese chips”, while happily tuning CUDA kernels, which if true means Western tooling will actively refuse to help close the gap.⁵ Community feeling elsewhere cuts the other way. As one commenter put it, the US labs have not matched the recent Chinese open releases, which go out “For everyone, without restrictions.”⁶ That stings because it is largely true on weights and harnesses, and it explains why this Ascend toolkit will get a proper look in Asia even if Bay Area Twitter scoffs. I am deliberately not adjudicating what V4 trained on, because from what I have seen that remains contested and undisclosed, and the toolkit argument does not need it. The question is whether Ascend is usable enough, cheaply enough, for the next lot of work.

I think the toolkit matters even though the silicon still lags. The old misreading was that export controls would freeze China out because you cannot download a fab. The correction is that you can download the workaround, iterate on it in the open, and aim it at the one workload where good enough wins: serving models at scale rather than proving you can train a frontier giant from scratch in one heroic run. Open sourcing does not fix bugs by itself, but it turns every frustrated Ascend customer into a potential patch author, and DeepSeek is exactly the kind of demanding in house user that finds the nasty performance cliffs. CUDA remains the better ecosystem today by a mile, and Huawei will need sustained delivery, docs and boring reliability work to change that. But a free, improving, DeepSeek blessed path off CUDA is precisely how moats start leaking. Watch the kernel repos and the hiring boards, not just the chip specs. If outside contributors start landing fixes, the split has grown a software layer, and that layer will be very hard to sanction away.

Sources

• 1 DeepSeek and Huawei release open-source Ascend AI programming tools to reduce reliance on Nvidia CUDA ecosystem — tools include compute and communication libraries, as well as Ascend support for TileLang | Tom’s Hardware — tomshardware.com
• 2 TCS lays off over 10,000, says AI is one reason • The Register — theregister.com
• 3 DeepSeek to use 160K+ of Huawei’s Ascend 950DT chips at a Mongolia data center — bloomberg.com
• 4 Bugs, performance issues hinder Huawei’s AI chips - Ars Technica — arstechnica.com
• 5 @dannyw on Claude Opus 5.5 — ycombinator.com
• 6 @midwain on The AI Race Just Got Awkward — ycombinator.com