NextFin News - China's AI lab DeepSeek has joined forces with Huawei to release an open-source software toolkit for programming Huawei's Ascend AI chips, a move that targets the single biggest pillar of Nvidia's dominance: the CUDA programming platform that has locked in developers for nearly two decades. The toolkit, announced on DeepSeek's official WeChat account on Wednesday, includes a high-level language called TileLang and related compute and communication libraries, all free to download. The stakes go well beyond one product release. Nvidia's roughly 80% to 95% share of the AI accelerator market rests less on silicon than on the fact that an estimated 6 million developers write CUDA code and every major machine-learning framework treats CUDA as the first-class citizen. A language that promises to make non-Nvidia hardware "easy to program" is, in effect, an attempt to price that moat down to zero.
The announcement is the latest escalation in a structural shift that began well before this week. As U.S. export controls cut Chinese AI labs off from Nvidia's newest hardware, Beijing's technology companies stopped merely buying workarounds and started building the software layer that makes alternatives usable. The question is no longer whether China wants a CUDA-free stack, but whether it can make one that developers will actually adopt.
The Announcement: What DeepSeek Actually Released
DeepSeek said it is open-sourcing infrastructure for Huawei's Ascend platform, centered on TileLang — a high-level programming language the company positions as a simpler, more universal alternative to Nvidia's CUDA. The release also bundles the compute and communication libraries needed to run AI workloads across Ascend accelerators.
"To build a new generation of independent, self-controlled GPU software ecosystem, the first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware's full performance potential," DeepSeek said. "TileLang was created precisely to meet this need," the company added, saying it offers "a simpler programming model" than Nvidia's CUDA, improving development efficiency and simplifying code logic.
Two details in that statement matter. First, the goal is explicitly an "independent, self-controlled" ecosystem — the language of a supply chain under sanction, not a product chasing a feature gap. Second, DeepSeek frames TileLang as a high-level abstraction layer, which means its ambition is to sit above the hardware, not merely to patch Huawei's existing stack. If it works, Chinese developers write TileLang once and run it on whatever accelerator is available, rather than rewriting code each time the hardware changes.
The Trigger: How Export Controls Made a Rival Stack Inevitable
TileLang did not emerge from a vacuum. It is the product of a policy sequence that left Chinese AI labs with a narrowing set of options. The United States first restricted exports of Nvidia's China-specific A800 and H800 accelerators in October 2023. Nvidia responded by designing the H20, a chip scaled back to stay within the rules, and it sold hundreds of thousands of them into China — business that generated an estimated $12 billion to $15 billion for Nvidia in 2024 alone. Then, in April 2025, the Commerce Department reversed course and required a license for H20 exports, citing the risk the chips could be diverted into Chinese supercomputers. Nvidia disclosed a $4.5 billion charge tied to excess H20 inventory and purchase obligations, plus another $2.5 billion of H20 revenue it could not ship in that quarter.
That sequence taught Beijing a lesson that a single product announcement cannot undo: even a chip designed to comply with U.S. rules can be taken away. After April 2025, buying Nvidia was no longer a procurement decision; it was a bet on Washington's next policy move. For a lab like DeepSeek — whose flagship R1 model had been optimized to run on the very H20s that were suddenly restricted — the rational response was to stop optimizing for hardware it could not guarantee and start building for hardware it could reach.
That is the deeper significance of the DeepSeek-Huawei partnership. It is not simply a Chinese firm choosing a Chinese chip. It is a Chinese firm, having already proven it could extract extraordinary efficiency from sanctioned hardware, now committing to make the alternative stack usable for everyone else. Huawei brings the silicon and the CANN software stack; DeepSeek brings the systems-level credibility that CANN has lacked.
Why Software, Not Chips, Is the Real Battlefield
DeepSeek is an unusual actor for this fight, and that is what makes the announcement worth taking seriously. Its earlier breakthrough — the R1 reasoning model released in January 2025 — was built not by outspending rivals on hardware but by optimizing below the level where most developers work. R1, a 671-billion-parameter Mixture-of-Experts model, was trained on a cluster of 2,048 Nvidia H800 GPUs in roughly two months, achieving efficiency roughly ten times that of leading U.S. labs. It got there by rewriting performance-critical code in PTX, Nvidia's intermediate instruction set that sits between CUDA C++ and the chip's actual machine code, rather than relying on CUDA itself.
That history matters for two reasons. First, it proved DeepSeek has the low-level systems expertise that a CUDA replacement actually requires; most challengers never get past the driver layer. Second, and more important, it showed that the binding constraint on China's AI program is not ideas but access. When the best chip you can buy is restricted, the only lever left is software efficiency — and the team that squeezed 10x out of sanctioned H800s is precisely the team Huawei would want beside it.
This is the structural logic behind the DeepSeek-Huawei tie-up. Huawei's Ascend 910C accelerator is widely assessed to be competitive with Nvidia's H200-class hardware on paper, and U.S. officials have concluded that Huawei could produce "a few million" Ascend chips in 2026. But hardware parity means little without a software stack developers trust. As recently as 2025, analysts at the Center for Strategic and International Studies judged the Ascend-plus-CANN combination — CANN being Huawei's own Compute Architecture for Neural Networks, its answer to CUDA — "far from being competitive to CUDA," citing a weaker ecosystem and "many complex issues that will likely take years to sort out."
TileLang is the attempt to shortcut that timeline. The mechanism is straightforward: if developers stop writing CUDA and start writing TileLang, the code is no longer bound to Nvidia's hardware. Over time, that turns a chip shortage into a permanent loss of share for Nvidia — not because the chips are worse, but because the code is no longer written for them. This is why the software layer, not the transistor count, is where the war is fought.
The Counter-Case: CUDA's Moat Is a Network Effect, Not a Feature List
The strongest argument against reading this as an imminent threat to Nvidia is that CUDA's dominance is a network effect, not a checklist of capabilities. Nearly 20 years of development, 6 million developers, a library of hundreds of optimized kernels, and every major framework treating CUDA as the default create a gravity that a new language does not escape simply by being "simpler." History is littered with technically competent CUDA challengers — AMD's ROCm, Intel's oneAPI — that failed to dislodge the incumbent because developers do not rewrite working code for abstraction's sake. They rewrite it when the alternative is unavailable or unaffordable. That is exactly the condition U.S. export controls created inside China — which is why this move is most credible as a China-contained parallel stack, not a global CUDA killer.
There is also a second-order tension inside the partnership itself. TileLang is pitched as a universal, hardware-agnostic language. If it succeeds, it reduces lock-in not only to CUDA but also to Huawei's own CANN. Huawei gains a broader developer base today but cedes some control over the stack tomorrow. For DeepSeek, that is the point: it wants a programming layer it controls, not a dependency swap from one vendor to another. But that split incentive could slow adoption among Huawei loyalists who would rather deepen CANN than empower a neutral abstraction.
The hardware gap has not fully closed either. Analysts assessing the Ascend 910C note that it offers about 80% of the memory bandwidth of Nvidia's China-specific H20 chip while still using the older HBM2E memory standard, two generations behind the most advanced AI chips. That gap matters most for training frontier models, where bandwidth is the bottleneck; it matters less for inference, which is cheaper, more forgiving, and the segment where China's domestic stack is most competitive. Many Chinese firms still rely on Nvidia for training workloads, and that reliance will not end with a language release.
What This Means for Nvidia: Share Erosion, Not Revenue Shock
The near-term financial impact is limited. Nvidia's data-center revenue is driven by U.S. hyperscalers and allied markets that remain closed to Huawei, and TileLang does not change that overnight. The real exposure is the long-term share trajectory, and there the arithmetic is already moving against the incumbent. Industry analysts project Nvidia's accelerator share declining from roughly 87% in 2024 to about 75% by 2026 as AMD and custom silicon scale, while absolute revenue still grows from around $100 billion to more than $150 billion because the total market is expanding faster than share is eroding. Custom ASICs alone could capture 10% to 20% of AI training and inference by the end of 2026, up from under 5%.
TileLang accelerates that erosion in one specific way: it lowers the cost of the "second source" decision. A Chinese cloud provider that once tolerated Huawei's weaker software because it had no choice can now tolerate it more cheaply, because part of the programming burden shifts from the hardware to the language. That is how a geopolitical workaround becomes a commercial competitor — quietly, workload by workload, until the share numbers reflect a new reality.
There is also a valuation dimension. Nvidia trades on the assumption that CUDA's switching costs are effectively permanent. Every credible data point that those costs are falling — DeepSeek's PTX work in 2025, this open-source release in 2026 — chips at the multiple, not just the revenue line. Investors price this risk suddenly, not gradually: the January 2025 DeepSeek shock erased more than $600 billion from Nvidia's market value in a single session, a reminder that the market reprices the moat the moment it looks porous. With the stock trading around 20 times forward earnings and Wall Street's median 12-month price target near $315 to $325, the risk is not that Nvidia's earnings collapse — it is that the premium investors pay for CUDA's durability quietly compresses.
Scenarios: What Could Break the Thesis
Three paths lie ahead, and they point in different directions depending on the time horizon.
In the base case, TileLang becomes the default programming layer for Huawei Ascend deployments inside China, lifting Huawei's share of the Chinese AI accelerator market — particularly in inference — while CUDA retains its global franchise and its hold on Chinese training workloads. Nvidia's share erodes at the margin, but revenue keeps growing with the market. This is the outcome the current share-price trajectory already implies.
In the upside case for the challengers, TileLang proves genuinely portable and begins attracting developers outside China, perhaps among cost-sensitive startups or emerging markets unwilling to bet their stack on U.S. policy. That would transform the project from a sanctioned-workaround into a real second ecosystem — the scenario under which Nvidia's multiple, not just its China revenue, comes under sustained pressure.
In the downside case, adoption stalls. Developers resist rewriting production code; Huawei's CANN politics resist a neutral abstraction layer; and the memory-bandwidth gap keeps frontier training on Nvidia. Twelve months from now, TileLang has no meaningful production deployment, and Huawei's share of China's AI accelerator market remains confined to inference-only workloads. In that world, CUDA's moat has held, and this week's announcement is a regional workaround, not a regime change.
The falsifying signal for the structural-threat thesis is concrete and observable: meaningful production deployment of TileLang outside China, or Huawei crossing into a dominant share of China's domestic AI accelerator market across both training and inference. Without either, the network effect has won again.
The bottom line: DeepSeek and Huawei have not replaced CUDA this week — but they have drawn the blueprint for how it could be replaced, and the motivation to follow it is only growing. Nvidia's moat was never the chip; it was two decades of code written for the chip. The moment that code becomes portable, the moat becomes a memory.
Explore more exclusive insights at nextfin.ai.
