• DeepSeek and Huawei released open-source programming tools for Ascend processors on September 30, aiming to build a Chinese alternative to Nvidia (NVDA)’s AI computing ecosystem.
  • The collaboration focuses on software and systems-level optimization, including a 128-chip Ascend 950 supernode, rather than a jointly designed chip.
  • Nvidia’s dominance faces a growing challenge in China, but manufacturing constraints and Nvidia’s global lead temper the threat.

A Software-First Offensive

DeepSeek and Huawei are intensifying efforts to create a domestic AI computing stack that could reduce China’s reliance on Nvidia, according to people familiar with the matter. The two companies on September 30 released open-source computing and communication libraries for Huawei’s Ascend processors, along with a programming language called TileLang designed to simplify AI-chip programming. They also optimized computation and communication across a 128-chip Ascend 950 “supernode,” marking a significant step in scaling domestic hardware for large AI workloads.

The move underscores a strategic shift: rather than merely designing a competing chip, DeepSeek and Huawei are targeting Nvidia’s moat—its CUDA software ecosystem and developer familiarity. “The real battle is at the systems level,” said an industry analyst who declined to be named. “Better software can make Ascend more usable without matching Nvidia’s top-end chips.”

Training, Inference, and a Separate Chip Project

The collaboration extends beyond software. On September 21, The Information reported that DeepSeek is increasingly using Huawei processors for model training—a more demanding task than inference. This suggests a push toward domestic training infrastructure, though it does not prove Nvidia has been replaced across DeepSeek’s operations. Separately, Reuters (TRI) reported in July that DeepSeek was developing its own inference chip, recruiting chip-design engineers and discussing partnerships with manufacturers. That project, described as early-stage, could eventually reduce reliance on both Nvidia and Huawei.

Huawei’s financials reveal the cost of this drive for self-sufficiency. First-half 2026 revenue rose 9.6% to CNY467.82 billion, but net profit fell 36% to CNY23.81 billion as R&D spending jumped 25.2%. Operating cash flow turned negative, and inventories swelled 42% from year-end 2025. “The economics are challenging,” said a person close to the company. “Building an alternative ecosystem requires heavy investment.”

Nvidia’s Grip and the Road Ahead

Nvidia remains the incumbent to beat. Its latest quarter, ended July 26, 2026, saw revenue surge 106% year over year to $96.2 billion, with data-center revenue up 117%. Notably, its outlook assumed no China data-center compute revenue, highlighting the impact of US export controls. Those restrictions cut both ways: they push Chinese firms like DeepSeek toward Huawei, but also limit access to advanced manufacturing and high-bandwidth memory, constraining domestic chip production.

Investors are watching closely. Nvidia shares slipped about 1.6% in premarket trading after Reuters’ July report on DeepSeek’s chip project, though the dip was tied to a related development, not the September collaboration. Analyst Richard Windsor of Radio Free Mobile remains skeptical: “Without access to leading-edge manufacturing, DeepSeek would have little prospect of selling competitive silicon outside China.”

For now, the DeepSeek-Huawei alliance represents a growing Chinese AI ecosystem that could erode Nvidia’s influence in the world’s second-largest economy. But the technological and manufacturing hurdles are formidable, and Nvidia’s global lead remains intact. As one Beijing-based fund manager put it, “This is a marathon, not a sprint.”

Correction: An earlier version of this article misstated the date of the open-source tools release. It was September 30, not September 21.