• Nvidia CUDA’s hold on AI computing rests on scarce kernel engineering skills as much as on its chips.
• ByteDance’s CUDA Agent and DeepSeek’s new Ascend tools both chip away at the cost of leaving Nvidia.

Chinese AI developer DeepSeek on September 30 released a free set of programming tools for Huawei’s Ascend chips, built around a language it says is simpler to work with than Nvidia CUDA. CUDA is the software layer that has kept most of the world’s AI developers on Nvidia hardware for close to two decades, and it is the main reason buying a rival chip is harder than it sounds.

In a post on its WeChat account, DeepSeek said Huawei supported the work throughout, and that the two companies jointly worked on a computing cluster built from 128 Ascend 950 chips, Reuters reported. Each tool it released matches one DeepSeek had already published for Nvidia GPUs, according to the company’s announcement.

Read alongside a second piece of Chinese research, the release points at the same target. In late February, ByteDance’s Seed team and Tsinghua University’s Institute for AI Industry Research published CUDA Agent, a model trained to write and speed up CUDA code without a human engineer. One project makes it easier to program chips that are not Nvidia’s. The other suggests the expertise that makes Nvidia hard to leave can be automated.

Why Nvidia CUDA is really a people problem

A GPU does not run ordinary software. Code has to be rewritten into small programs called kernels that split a job across thousands of processing cores at once. Writing one that works is manageable. Writing one that runs fast takes an engineer who understands how the chip shuffles data around, and those engineers are scarce and expensive.

The scale of that work shows in Nvidia’s own numbers. Ian Buck, the company’s vice-president of hyperscale and high-performance computing, told Asia-Pacific media at GTC 2026 in March that 400 Nvidia engineers spent four months tuning the DeepSeek-R1 model for its GB200 systems, Computer Weekly reported. The result was four times the performance on the same hardware, with no new chips involved. Nvidia said it marked CUDA’s 20th year at the same event with more than 6 million developers on the platform.

Every rival chip needs that tuning redone from scratch for its own design. That is the bill a buyer faces when they consider switching, and it is the bill both Chinese projects are trying to cut.

A model that tunes its own code

CUDA Agent works the way a human engineer does. It writes a kernel, compiles it, checks the answer is right, times how fast it runs and then tries again, for up to 200 rounds on a single task. The researchers scored every attempt, penalising kernels that failed the correctness check and paying out most for those that beat both of PyTorch’s existing modes.

On KernelBench, a 250-task test built at Stanford, the system produced correct code 98.8% of the time and beat torch. compile, PyTorch’s built-in optimiser, on 96.8% of tasks, at 2.11 times the optimiser’s speed on average, the paper said. ByteDance’s starting model, before this training, had beaten that same optimiser on only 27.2% of tasks.

Those numbers come with caveats. The researchers never tested the system against human engineers, despite how the paper has been read online. KernelBench tasks are small, self-contained problems rather than real workloads, and the headline wins often came from the model noticing that the original code was doing unnecessary work. 

The team also said it did not test against more advanced compiler frameworks such as TVM, and that its reliance on a large GPU pool could limit access for the wider research community. The trained model has not been released, though the team published its training data and setup on GitHub.

What it means for China’s exit from Nvidia

China’s problem has never been a shortage of chips to build. Huawei said at Huawei Connect 2026 in Shanghai last month that more than 1,000 Ascend supernodes have been deployed, and Huawei Cloud chief executive Zhou Yuefeng said its Ascend 950 cluster service opened to Chinese customers on September 30 and reaches other markets on November 30, TechNode reported. The problem is that moving a model onto those chips means redoing years of software work that Nvidia’s ecosystem already did.

Both Chinese projects chip away at that cost from different directions. TileLang, the language at the centre of DeepSeek’s release, lets a developer write a kernel once and have a compiler translate it for whichever chip it is pointed at. It was not DeepSeek’s invention. It came from a Peking University team led by Professor Zhi Yang and was open-sourced in January 2025, according to its GitHub page, which added Ascend 950 support on September 30.

CUDA Agent, meanwhile, was trained and tested only on Nvidia hardware, and its authors claim nothing about other chips. But the method itself is not tied to Nvidia. It needs practice problems and a way to automatically check whether code is correct and time how fast it runs, which any chipmaker can supply for its own hardware. If the same approach can be pointed at TileLang or Ascend, the tuning work that makes leaving Nvidia expensive becomes something a model does rather than a team a company has to hire.

Washington’s export controls were designed on the assumption that withholding the best chips slows China down. The pressure has been working in both directions. Nvidia said in its latest quarterly filing that Chinese restrictions left it unable to sell its H20 inventory even after Washington began issuing licences, and that sales of the more powerful H200 ran into the same problem. H200 sales to Chinese buyers came to less than 1% of Nvidia’s data centre revenue in the quarter ended July 26, Bloomberg reported. ByteDance ran CUDA Agent’s kernel testing on a pool of 128 Nvidia H20s, the cut-down chip built for the China market.

Nvidia reads it the other way

Asked at GTC whether AI writing code would weaken CUDA’s position, Buck said agents are already being used to write and tune CUDA kernels, including for models such as DeepSeek. “It’s actually accelerating CUDA adoption,” he said, telling Computer Weekly that Nvidia’s own researchers had seen their productivity climb using Claude alongside the company’s Warp framework.

The CUDA Agent paper partly backs him up. Every kernel it wrote was CUDA code running on Nvidia chips, using Nvidia’s own libraries where the rules allowed. Cheaper CUDA work makes Nvidia hardware more attractive, not less, and Nvidia chief executive Jensen Huang told the GTC audience that hundreds of millions of CUDA-capable GPUs are now installed worldwide.

The field also has a credibility problem that argues for patience. The CUDA Agent authors point to an earlier Sakana AI system where a claimed 150-times speed-up turned out to be three times slower once an outside tester ran it. They locked their own test scripts behind file permissions and checked every kernel against five random inputs to stop their model gaming the score. TWA found no independent replication of CUDA Agent’s own results.

Neither ByteDance nor Tsinghua has said publicly whether they intend to point the method at chips other than Nvidia’s. Perhaps that is the answer worth waiting for.