US chip export controls: Deepseek and Huawei build a software moat around Ascend

Deepseek published open-source programming tools for Huawei's Ascend AI chips on 30 September, according to a post on the company's official WeChat channel. The centerpiece is TileLang, a programming language the company says is easier to work with than Nvidia's CUDA software platform. Reuters reported the release, and The Decoder covered it the same day.

This is the newest front in the US-China chip fight, and it is not about silicon. It is about the software layer that decides which chips developers can actually use.

What Deepseek actually shipped

The package includes libraries for computation and for moving data between chips, all released as open source. Deepseek said Huawei "fully supported" the work. The two companies also optimized a supernode, a cluster of 128 Ascend 950 chips. TileLang was originally developed by researchers at Peking University, and Deepseek has been using it for about a year, first testing it on older Nvidia chips. According to The New York Times, it is now the company's main tool for its work on artificial general intelligence.

Deepseek's argument is simple. Anyone trying to build an independent software ecosystem for AI chips first needs a universal language that is easy to program but still gets full performance out of the hardware. In the company's view, TileLang offers a simpler programming model than CUDA.

That matters because Nvidia's dominance does not come from chip design alone. It also rests on an estimated four million developers worldwide who build with CUDA. That ecosystem is the moat rivals like AMD have not been able to cross, even when their hardware looked just as strong on paper.

Huawei's gap, and the chips behind it

So far, Chinese model makers like Z.ai and Moonshot AI have moved faster than the country's chipmakers, according to the NYT. Huawei wants to close that gap. Two weeks before Deepseek's announcement, the company unveiled new AI processors and supernode systems and said they would be widely used for model training next year.

Huawei also admits it cannot keep up with demand at home, so it plans to sell fewer chips abroad. Huawei's current rotating chairman Eric Xu said the company cannot accept a future that hinges on whether others are willing to sell chips to China. He was referring to US export controls.

Inside China, the pressure is not only external. A week of reporting from Beijing and Shanghai by investor David Cheng, published on his Earned Intuition newsletter on 30 September, described youth unemployment at 17.9% as of July, July retail sales up 0.6% year over year against an expected 1.5%, and a real estate crash still weighing on consumption. Cheng wrote that AI is "the only engine of the Chinese economy that is working," propped up by state-subsidized electricity, fast-tracked IPOs and price wars the state tolerates.

如果不出海就白卷了 — if you don't go overseas, all that grinding was for nothing. Involution without export is wasted effort.

That line, which Cheng heard during his trip, captures the export logic behind tools like TileLang. If Chinese chips cannot be sold abroad in volume, Chinese software has to make them usable at home and attractive to buyers the US cannot reach.

How much of the CUDA moat is left

Research firm SemiAnalysis has been testing that question. After examining Jalapeño, OpenAI's inference chip, the analysts called the CUDA moat "potentially dead" because OpenAI gets new models running on its own hardware so quickly. Jalapeño beat Nvidia's Blackwell on performance per watt in most of the scenarios tested. According to SemiAnalysis, OpenAI models also helped design the chip, and those models run on Nvidia GPUs.

The analysts added their own caveats. They only tested scenarios that are relatively easy to optimize, with about 8,000 input tokens and 1,000 output tokens. They have not yet run AgentX, a benchmark that measures how AI agents handle multistep tasks. That is exactly where SemiAnalysis found Nvidia well ahead back in August. With AMD's current software stack, Nvidia would still come out cheaper per token even if AMD gave its hardware away. The authors do not see Nvidia's lasting advantage in the silicon itself. They see it in the software that links many chips into one system.

Huawei's chips were not part of the AgentX comparison. In an earlier analysis of DeepSeek V4, however, SemiAnalysis pointed out that Huawei's CANN software stack was the only one besides CUDA to support the model on day one.

The policy backdrop

All of this is happening while Washington debates how tightly to restrict the hardware. On 29 September, President Trump and the leaders of the largest US AI companies signed a set of voluntary safety standards after a lunch meeting at the White House, CBS News reported. "It's almost like a constitution, in a way," Trump said, flanked by SpaceX's Elon Musk, Meta's Mark Zuckerberg, Anthropic's Dario Amodei, Nvidia's Jensen Huang, OpenAI's Greg Brockman and Google's Sundar Pichai.

The one-page accord commits the companies to "four layers of controls and audits," including internal evaluations, audits by an external firm and reviews by each company's board. Asked whether the deal was binding, Trump said: "I think it's morally binding." He also said he plans to name an "AI czar" within three or four days, and signed an executive order directing the federal government to use the term "Super Intelligence" instead of artificial intelligence.

Democratic Senator Mark Warner of Virginia was not impressed. In a statement after the meeting, he called for Congress to pass mandatory testing, evaluation and incident reporting rules for the most advanced AI models. "The companies building the most powerful AI systems are warning us that the technology is advancing faster than our safeguards," Warner said. "The president's response? To rename it and tell the companies developing it to regulate themselves."

Trump has not backed stricter AI regulation, warning that a nationwide slowdown could dent the US economy and give China's tech industry a chance to catch up. That framing sits awkwardly next to the Deepseek-Huawei release, which is precisely an attempt to work around the export controls the US has already imposed.

Europe is moving in a different direction on a different file. The EU Council is pushing to allow police to order internet companies to scan private communications in "parts of a service" for a period of time, rather than requiring a warrant for a specific suspect, according to a leaked presidency note dated 18 September and reported by Reclaim The Net on 30 September. The sixth and potentially final trilogue on the Chat Control 2.0 legislation is set for Tuesday, 29 September. Former Pirate Party MEP Patrick Breyer called the proposal "mass surveillance by another name," and the Council's own legal service has previously said scanning an entire service or parts of it is "highly probable" to be found "general and indiscriminate" surveillance, and therefore unlawful.

What to watch

Two things decide whether the Deepseek-Huawei release changes the export-control picture. First, whether TileLang attracts developers outside China. An open-source language only becomes a moat if people build on it, and CUDA's four million developers will not switch without a reason that survives contact with production workloads.

Second, whether Washington responds with tighter software or services restrictions rather than more hardware rules. The current debate in Congress, per reporting from Lawfare and others, has focused on open-weight models and export controls on chips. A credible Chinese software stack for domestic chips is a different kind of problem, and it is not clear the existing toolkit addresses it.