Deepseek ships open-source tools for Huawei Ascend chips
The TileLang language and new compute libraries aim to drive Ascend silicon without CUDA, tuned on a supernode of 128 Ascend 950 chips.
In short
Deepseek has released open-source tooling for Huawei's Ascend AI chips — the TileLang programming language plus libraries for computation and chip-to-chip data movement, tuned on a supernode of 128 Ascend 950 chips.
At a glance
- TileLang is an open-source language from Peking University researchers; Deepseek has been using it for about a year.
- The partners tuned a supernode of 128 Ascend 950 chips and shipped libraries for computation and chip-to-chip transfers.
- Nvidia's CUDA rests on an estimated four million developers worldwide, according to the report.
- SemiAnalysis noted Huawei's CANN was the only stack besides CUDA to support Deepseek V4 on day one.
- Rotating chairman Eric Xu says Huawei cannot accept a future decided by others' willingness to sell chips to China.
Deepseek has released open-source tooling for Huawei's Ascend AI chips — the TileLang programming language plus libraries for computation and chip-to-chip data movement, tuned on a supernode of 128 Ascend 950 chips.
What actually shipped
The tools are free and open source. Alongside the language itself, Deepseek points to libraries for compute operations and for moving data between many accelerators — the layer that decides how much of a cluster's theoretical throughput a large training run really gets. Deepseek says Huawei fully supported the work. The individual libraries are not named in the report.
TileLang as the CUDA alternative
TileLang is open source in its own right and came out of Peking University. Deepseek has been building with it for about a year, first on older Nvidia parts and now as the core tool for its AGI work. The pitch is a simpler programming model than Nvidia's CUDA. Deepseek's reasoning: anyone trying to stand up an independent software ecosystem for AI chips needs a common language first.
The bottleneck is software, not silicon
Nvidia's lead is not purely a chip-design story. It rests on an estimated four million developers worldwide who already know CUDA. Chinese model labs such as Z.ai and Moonshot AI have outpaced the country's chipmakers, and software is where that gap shows most clearly. SemiAnalysis pointed out that Huawei's CANN stack was the only one besides CUDA to support Deepseek V4 on day one.
Why Huawei needs the help
Pointing at US export controls, rotating chairman Eric Xu said the company cannot accept a fate decided by whether others are willing to sell chips to China. Huawei also concedes it cannot meet demand in its home market and plans to sell less abroad. An open language other Chinese labs can target widens the Ascend developer base without Huawei writing every kernel itself.
What is not verified
SemiAnalysis called the CUDA moat potentially dead after testing Jalapeño, OpenAI's inference chip — then hedged it, noting only relatively easy-to-optimize scenarios were measured and that Nvidia still looks ahead on agent workloads. None of that is an Ascend comparison. The report gives no performance figures for the 128-chip node, and there is no evidence yet on how many developers outside Deepseek use TileLang.
Source: the decoder
FAQ
What is TileLang?
An open-source programming language for AI chips, originally built by researchers at Peking University. Deepseek has used it for about a year and now treats it as its core tool, including on Huawei's Ascend hardware.
Do Huawei Ascend chips still need CUDA?
No — CUDA is Nvidia's platform. Ascend runs on Huawei's own CANN stack, and TileLang plus the new libraries are meant to make programming those chips simpler. The report offers no head-to-head performance numbers against CUDA.
How many chips are in the Huawei supernode Deepseek optimized?
128 Ascend 950 chips. Deepseek and Huawei tuned that node together, but no benchmark results for it have been published.
