DeepSeek and Huawei Technologies are open-sourcing core software tools for Huawei Ascend chips, directly challenging Nvidia’s CUDA developer ecosystem. The partnership seeks to lower software barriers for Chinese artificial intelligence firms working around strict U.S. export controls and searching for domestic computing alternatives.
The shared infrastructure includes compute and communication libraries alongside TileLang, a high-level programming language designed to give developers a simpler programming model
than Nvidia’s dominant CUDA platform.
The collaboration centers on a joint supernode
system built on 128 Ascend 950 processors, with Huawei providing full software support to optimize both computation and cross-device communication. By open-sourcing these foundational programming tools, the companies are aiming at the half of Nvidia’s market dominance that relies on software lock-in rather than silicon performance alone. Software collaboration is now underway between the Shenzhen telecom titan and the artificial intelligence model developer based in Hangzhou, tackling a key yet frequently overlooked foundation of Nvidia’s enduring global market power.
The Long Shadow of Nvidia’s CUDA Developer Platform
For more than a decade, Nvidia’s Compute Unified Device Architecture has served as the default language for building neural networks on graphical processing units. Released in 2006, CUDA became the AI industry’s go-to platform for parallel computing. Industry analysts compare CUDA’s sticky ecosystem to Apple’s iOS, because mature software tools and developer familiarity create massive switching costs.
Alternative software stacks have historically struggled to gain traction. Rival frameworks such as AMD’s ROCm, Intel’s oneAPI, and the multi-company OpenCL failed to unseat CUDA after years of development. In Bloomberg’s assessment, Nvidia’s superior software ecosystem has consistently given developers a compelling reason to stick with its hardware, even as Huawei aggressively pushed domestic chip shipments.
U.S. Export Controls and the Push for Domestic Chip Self-Reliance
The timing of the DeepSeek and Huawei partnership intersects with tightened U.S. semiconductor export restrictions that have largely locked Nvidia out of mainland China. Sales of H200 chips to Chinese customers accounted for less than 1% of Nvidia’s data center revenue in the quarter ended July 26, 2026, when total revenue reached $96.2 billion, up 106% year-over-year. Following a series of regulatory updates, Nvidia initially developed a modified H20 processor specifically for the Chinese market, watched subsequent policy changes block those transactions, and ultimately secured authorization in December to supply the more capable H200 on the condition that the American authorities retain a 25% share of every transaction. Even so, Nvidia’s finance chief told analysts in February that the company had yet to generate any revenue from the approved China chips and did not know whether any imports would actually be allowed, meaning near-term dollars at risk are small.
At the same time, market activity in China continues through gray-market channels and secondary supply networks. In a keynote speech last month, Huawei rotating chairman Eric شو told a keynote audience that he had surpassed Nvidia in terms of local market share. As Beijing steps up its drive to minimize reliance on foreign semiconductor technology, domestic alternatives like CXMT and Moore Threads have both launched public stock offerings over the course of the past year. Nvidia shares finished Tuesday at $227.21, marking a 0.72% decline, and advanced 0.44% to $228.22 during Wednesday premarket trading around 6:07 a.m. ET.
China Trails US Models Amid Hardware Shortages
Despite the strategic significance of releasing TileLang, market observers caution that software releases alone cannot instantly close capability gaps. Independent analysts note that China’s frontier models still trail U.S. state-of-the-art systems by roughly three to six months. Furthermore, U.S. officials assert that DeepSeek’s V4 model was trained using smuggled Nvidia Blackwell hardware, and DeepSeek has been described as extraordinarily dependent
on American technology while facing internal chip shortages.
Hardware manufacturing limits present another friction point. While DeepSeek has planned to install more than 160,000 Huawei Ascend chips at a new data center facility in Inner Mongolia, industry analysts point out that Huawei faces ongoing challenges in scaling physical silicon production to meet domestic AI demand. DeepSeek itself stated that it cannot serve its V4 Pro model to most customers because it lacks enough chips, which undercuts the open-source software effort until Huawei hardware and manufacturing facilities can fill the gap.
