DeepSeek and Huawei team up to build a real alternative to Nvidia

Why it matters: CUDA isn't just Nvidia's software, it's the reason switching away from Nvidia chips is so painful. Once a company builds its AI models and workflows around CUDA, moving to different hardware means months of rewriting code, debugging, and re-tuning performance from scratch.

That switching cost is Nvidia's real moat, arguably bigger than any raw speed advantage its chips have. TileLang is a direct attempt to lower that cost for anyone willing to try Huawei's hardware instead.

• TileLang: an open-source, higher-level programming language for AI chips.
• Goal: make coding for Huawei's Ascend chips simpler than coding for Nvidia's CUDA.
• Also open-sourced: "compute" and "communication" libraries, the building blocks for training AI across many chips at once.

Beyond the software, DeepSeek and Huawei built a system combining 128 of Huawei's Ascend 950 chips into what they're calling a supernode.

In real life think of a supernode less like 128 separate calculators and more like one giant brain made of 128 connected parts. If the parts don't talk to each other smoothly, the whole thing slows down no matter how fast any single part is.

By the numbers: the companies specifically focused on optimizing how well the chips communicate with each other, not just how fast each individual chip runs. For large AI training jobs, that connection speed between chips often matters as much as, or more than, raw per-chip performance.

DeepSeek isn't a random startup trying its luck. In January 2025, its R1 model shook up the entire AI industry by matching top US models' performance while reportedly costing around $5.6 million to train, a fraction of what US labs were assumed to need.

The big picture: that release triggered real policy consequences. It fueled debate over whether US export controls on advanced chips were actually working as intended, since DeepSeek had reportedly achieved frontier-level results using a mix of techniques and hardware that regulators hadn't fully anticipated.

The US tightened restrictions further in April 2025, requiring government approval before Nvidia could export certain chips, including its H20 model, to Chinese customers. That single policy shift reportedly cost Nvidia billions in expected China revenue almost overnight, and set off a scramble among Chinese AI companies to secure reliable domestic alternatives rather than depend on hardware that could be cut off again with little warning.

DeepSeek building tools specifically for Chinese-made hardware now looks less like a bold new direction and more like a direct, logical continuation of that same pressure, formalizing a relationship with Huawei that Chinese AI companies have been leaning toward for over a year.

Background: here's the part that makes this partnership less speculative than it might sound. Nvidia's own CEO, Jensen Huang, said in October 2025 that the company's China AI chip market share had already fallen from 95% to zero, a direct result of US export controls tightening throughout 2025.

By Nvidia's first-quarter 2026 results in May, Huang told CNBC he had "largely conceded" the Chinese market to Huawei, and the company reported zero data-center revenue from China that quarter. Nvidia had previously estimated it stood to lose as much as $22 billion in annual revenue tied to restricted China sales.

Why it matters: that context changes how this DeepSeek-Huawei partnership should be read. This isn't two companies trying to convince Chinese developers to abandon a market leader they're happily using today. Nvidia is already locked out of China's data-center AI chip market by regulation, not by choice. The real competition DeepSeek and Huawei are up against is the lingering technical debt of years of CUDA-based development, not Nvidia's current sales presence.

The catch: not immediately, and maybe not soon. Nvidia's advantage comes from years of accumulated developer familiarity, existing codebases, and a massive ecosystem of tools built specifically for CUDA. That kind of lock-in doesn't disappear because a competitor releases an alternative language.

What's next: but if TileLang genuinely delivers on being simpler to use, and Huawei's chips keep improving, the "porting tax," the cost and hassle of switching away from Nvidia, could shrink meaningfully over time. That's especially true for AI workloads specifically built for the Chinese market, where sourcing advanced Nvidia chips is already restricted or complicated.

Who's affected: Chinese AI companies building models specifically for domestic use are the most immediate beneficiaries here. Companies outside China with no export restrictions have far less reason to switch away from an ecosystem they already know works.

Who's affected: this partnership isn't happening in isolation. It fits a broader pattern of Chinese tech companies building complete, domestic alternatives to Western AI infrastructure, from chips to the software that runs on them, rather than trying to keep importing restricted hardware.

That pattern extends beyond just Huawei and DeepSeek. Multiple Chinese firms have been investing in domestic chip design, alternative software stacks, and homegrown cloud infrastructure since export restrictions tightened. A shared, open-source tool like TileLang could end up benefiting that entire ecosystem, not just the two companies that built it, if other Chinese AI firms adopt it for their own Ascend-based projects rather than each building something similar from scratch on their own.

Why it matters: open-sourcing TileLang, rather than keeping it proprietary to DeepSeek and Huawei, is a deliberate choice that maximizes how quickly it could spread across China's AI industry. A shared standard that many companies build on top of is far more likely to genuinely dent CUDA's dominance than a single company's internal tool ever could.

• This is a China-specific story for now. If you're not building AI systems constrained by US export rules, this partnership doesn't change your hardware choices today.
• Watch TileLang's actual adoption, not just its announcement. Open-sourcing a tool is easy; getting developers to actually rebuild their workflows around it is the hard part.
• Nvidia's dominance isn't going anywhere overnight, but efforts like this are exactly the kind of long-term pressure that could chip away at it over years, not months.
• DeepSeek keeps proving it's more than a one-time headline. This partnership shows it's building infrastructure, not just chasing another viral model release.