DeepSeek and Huawei have open-sourced a set of programming tools built for Huawei’s Ascend AI chips, a release dated September 30, 2026 that is already being read across the developer world as the most direct software challenge yet to Nvidia’s CUDA platform. The toolkit includes a version of TileLang, a higher-level kernel-programming language DeepSeek has promoted as a simpler way to write optimized AI code than hand-tuned CUDA, according to reporting from The Next Web.
For engineers who build and deploy large language models, the news lands at a moment when access to Nvidia’s top-tier chips in China is effectively zero and when every major Chinese AI lab is hunting for a credible, production-grade alternative to the CUDA stack that has underpinned machine learning development for nearly two decades. The DeepSeek-Huawei release does not end Nvidia’s dominance overnight, but it is a concrete, downloadable answer to a question the industry has been asking since Washington tightened chip export rules: can China’s AI ecosystem actually run without Nvidia’s software layer?
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What DeepSeek and Huawei Actually Released
According to Tom’s Hardware, the release bundles six free, downloadable software modules targeting Huawei’s Ascend AI accelerators. The centerpiece is Ascend support for TileLang, which DeepSeek has pitched as a way to write AI kernels without wrestling with the low-level, hardware-specific detail that CUDA programming typically demands. Alongside it sits DeepGEMM-Ascend, a library tuned for the matrix-multiplication operations that dominate large language model training and inference.
The rest of the stack covers the plumbing that large-scale AI training actually depends on: libraries for general AI computation and for chip-to-chip communication, the kind of interconnect software needed to coordinate thousands of accelerators working on a single training run. Ascend deployments depend on Huawei’s CANN, or Compute Architecture for Neural Networks, as the underlying software toolkit. The September 30, 2026 release also added native support for Huawei’s newer Ascend 950 processor family, including code generation, automatic workload scheduling, and synchronization support, per Tom’s Hardware’s account of the release notes. DeepGEMM-Ascend is an Ascend port of DeepSeek’s optimized matrix-multiplication library that retains the same API as DeepGEMM’s other implementations, and DeepSeek also added Huawei support to TileKernels, with Nvidia CUDA and Huawei CANN backends exposed through corresponding Python APIs.
None of this is framed by either company as a finished, one-to-one CUDA replacement. It is, instead, an attempt to lower the barrier for developers writing AI kernels for Ascend hardware, so that porting a model off Nvidia GPUs becomes an engineering project rather than a research problem.
The Six-Module Ascend Toolkit, Mapped Out
The table below breaks down what each piece of the release is reported to do, based on the component descriptions published alongside the launch.
| Component | Category | Primary Function | Added / Updated |
|---|---|---|---|
| TileLang (Ascend port) | Kernel language | High-level language for writing AI kernels without raw CUDA-style code | September 30, 2026 release |
| DeepGEMM-Ascend | Math library | Optimized matrix multiplication for training and inference | September 30, 2026 release |
| AI computation library | Compute library | General-purpose AI compute kernels for Ascend hardware | September 30, 2026 release |
| Chip-to-chip communication library | Interconnect software | Coordinates data movement across multiple accelerators | September 30, 2026 release |
| Ascend 950 code generation | Compiler backend | Targets Huawei’s newer Ascend 950 chip family | September 30, 2026 update |
| Ascend 950 scheduling & sync | Runtime support | Automatic workload scheduling and cross-chip synchronization | September 30, 2026 update |
Why This Is Happening Now: China’s Chip Squeeze
This release did not happen in a vacuum. Nvidia’s H20 chip, the cut-down accelerator the company had designed specifically to stay inside U.S. export limits while still selling into China, was effectively banned from the Chinese market on April 9, 2025. The practical result, according to Nvidia’s own fiscal 2026 regulatory filing, was that restrictions had foreclosed Nvidia from China’s data-center compute market by the end of that fiscal year.
Nvidia CEO Jensen Huang has publicly acknowledged the scale of that collapse. At an industry event reported to have taken place in October 2025, Huang described Nvidia’s China market share falling from roughly 95% to effectively zero, a shift covered by outlets tracking the fallout, including Futu News. That filing also flagged a risk Nvidia’s own leadership has had to reckon with: locking Nvidia out of China does not remove demand for AI chips there, it just redirects developers toward whichever ecosystem can serve them, and Huawei is the most obvious domestic candidate.
The Software Gap Behind the Hardware Gap
Export controls target hardware, but the harder problem for any Nvidia alternative has always been software. Huawei’s Ascend chips are not CUDA-compatible, which means any team that wants to move a model off Nvidia GPUs has historically had to rewrite low-level kernels by hand. That is the exact gap DeepSeek and Huawei are now trying to close with a free, open-source toolchain rather than a proprietary one, a strategy that mirrors how DeepSeek built its reputation in the first place: by publishing capable AI software at no cost and letting adoption do the marketing.
Inside Nvidia’s CUDA Moat
To understand why an open-source kernel language is treated as significant news, it helps to understand what CUDA actually is and how long Nvidia has had to build around it. Nvidia launched CUDA in 2007, giving developers a way to write general-purpose code for GPUs nearly two decades before the current AI boom made that capability indispensable, a head start detailed in a retrospective on Jensen Huang’s long CUDA bet. That head start compounded. An OECD policy paper published in 2025 estimated that Nvidia holds more than 80% of the market for GPUs used in AI workloads, a figure that reflects not just chip shipments but the accumulated weight of software, libraries, and institutional habit built on top of CUDA.
Nvidia’s own disclosures back that up from the inside. The company’s ecosystem materials describe more than 4 million developers working with Nvidia’s AI tools, alongside more than 40,000 companies and roughly 15,000 startups enrolled in its Inception program, a scale that no rival GPU software stack currently approaches. That is the installed base DeepSeek and Huawei are trying to chip away at, not by out-competing CUDA on raw performance, but by making it cheap and fast for a developer to try the alternative.
DeepSeek’s Track Record of Shaking Up the AI Hardware Narrative
This is not the first time DeepSeek has forced a rethink of assumptions about Nvidia’s grip on AI infrastructure. DeepSeek-V3 launched on January 10, 2025, with the company’s researchers saying it was trained using Nvidia H800 chips at a reported cost of less than $6 million, a figure that stunned an industry accustomed to training budgets an order of magnitude higher, according to Reuters. DeepSeek followed with its R1 reasoning model later that same month.
The January 2025 Market Shock
The market reaction was immediate and severe. On January 27, 2025, Nvidia shares fell just under 17%, according to Reuters citing LSEG data, wiping out roughly $593 billion in market capitalization in a single session, which Reuters described at the time as the largest one-day market-cap loss for any company in Wall Street history. TechCrunch calculated the close-to-close decline at 16.9% and put the loss at nearly $600 billion. The trigger was investor concern that DeepSeek’s cheaper training approach undercut the assumption that advanced AI required ever-larger fleets of Nvidia chips.
That history matters for reading the Ascend toolkit release correctly. DeepSeek has already shown it can move markets with a software story, not just a model release. Pairing that reputation with Huawei’s hardware and distribution inside China gives this latest open-source push a credibility that a smaller or newer lab’s CUDA-alternative project would not carry.
CUDA vs ROCm vs the Ascend-TileLang Stack
Nvidia is not the only incumbent with a software moat to defend, and CUDA is not the only alternative developers have tried before. AMD has spent years building ROCm as its own CUDA alternative for Instinct GPUs, with mixed adoption results that remain hard to quantify independently; no reliably sourced developer-count or market-share figure for ROCm exists in the same way Nvidia publishes for CUDA, according to comparison coverage from AIMultiple. The table below lines up the three stacks as they stand in October 2026.
| Platform | Backer(s) | Primary Hardware | Kernel Language | Developer Ecosystem | China Access |
|---|---|---|---|---|---|
| CUDA | Nvidia | Nvidia GPUs (H100, H200, RTX Pro, etc.) | CUDA C/C++ | 4M+ developers, 40,000+ companies (Nvidia ecosystem figures) | H20 banned since April 9, 2025; top-tier access effectively zero |
| ROCm | AMD | AMD Instinct accelerators | HIP / ROCm | No independently verified adoption figure available | Available, subject to separate export rules |
| Ascend + TileLang / DeepGEMM-Ascend | Huawei & DeepSeek | Huawei Ascend chips (910B, 950 family) | TileLang (Ascend implementation expanded September 30, 2026, with Ascend 950 support) | Early stage; no verified developer count yet |
The honest reading of that table is that Nvidia’s advantage is not just technical, it is a matter of scale and time. ROCm has had years to build adoption and still lacks CUDA’s reach. Ascend and TileLang are starting from an even earlier point, but with one advantage ROCm never had: a captive, chip-starved market in China that has strong incentive to make the new stack work.
A Simplified Look at the Kernel-Writing Gap
The pitch behind TileLang is abstraction: write less hardware-specific boilerplate and let the compiler handle more of the optimization work that a CUDA kernel normally requires by hand. The simplified, conceptual comparison below illustrates the difference in approach rather than reproducing any specific vendor code.
// Conceptual CUDA-style kernel: manual thread/block indexing,
// manual memory layout, manual tiling
__global__ void matmul_kernel(float* A, float* B, float* C, int N) {
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
// developer manages tiling, shared memory, and sync explicitly
}
# Conceptual TileLang-style kernel: higher-level tile abstraction,
# scheduling and memory handled by the compiler
def matmul(A, B, C, N):
with T.Tile(A, B, C, size=N) as tile:
tile.compute() # compiler handles indexing, tiling, syncThat abstraction is exactly why Nvidia’s own CUDA ecosystem also offers higher-level libraries on top of raw CUDA C, so the real test for TileLang on Ascend will not be whether it is easier to write than bare-metal CUDA, but whether it performs close enough to hand-tuned kernels to be worth the switch.
Market Impact: Nvidia, AMD, and the Open-Source Scramble
For Nvidia, the near-term financial exposure from this specific release is limited, since the company has already written China’s top-tier data-center market down to effectively zero in its own filings. The longer-term risk is different: every month that Huawei and DeepSeek’s stack gets easier to use is a month that Chinese developers, and potentially developers in other markets facing their own export restrictions, build fluency in a non-CUDA toolchain. That is the dynamic Nvidia’s own filing reportedly warned about, the possibility that exclusion from China accelerates a rival ecosystem’s growth rather than merely shrinking Nvidia’s addressable market.
The timing also lands awkwardly next to Nvidia’s broader hardware strategy. The company has reportedly been winding down RTX 5090 production to redirect capacity toward RTX Pro chips, and separately pushed its own $20 billion open-source bet aimed at competing with China’s open models. Both moves read as Nvidia trying to defend share on two fronts at once: consumer and professional GPU supply at home, and open-source credibility abroad. AMD, meanwhile, faces a market where its own ROCm push has to compete not just with CUDA but now with a second, China-backed open-source alternative gaining attention at the same time.
China’s Broader Open-Source AI Push
The Ascend toolkit is one piece of a wider pattern of Chinese AI labs choosing open-source releases as a competitive strategy rather than a defensive one. DeepSeek’s own models continue to be benchmarked directly against Western rivals, including in recent comparisons of DeepSeek’s flash-tier models against Claude Haiku 5.5 and Gemini 3.8 Flash, and DeepSeek released DeepSeek-V4.1-Flash-1017, an updated checkpoint of DeepSeek-V4.1-Flash, on October 17, 2026, after launching its V4 preview on April 25, 2026 with Huawei Ascend Super Node support, while other labs have made similar open-weight plays, such as Reflection AI’s open-source push aimed at matching China’s frontier models. The strategic logic is consistent across all of these: open distribution builds developer habit faster than a closed, paid platform can, especially among developers who cannot access the market leader’s hardware in the first place.
That same climate has also drawn regulatory and enforcement attention on the hardware side. U.S. authorities have pursued cases tied to the illegal movement of restricted chips, including a recent case in which a California executive was charged in connection with an alleged $300 million Nvidia GPU smuggling scheme, a reminder that the export-control regime this toolkit is responding to is being actively tested on multiple fronts at once.
What Developers Actually Need to See Before Switching
Open-sourcing a toolchain is the easy part. The harder test, and the one that will determine whether this release matters beyond a single news cycle, is whether Ascend plus TileLang can hold up under three things developers actually check before they migrate production workloads: performance parity on real training and inference workloads, stability of the compiler and runtime across driver updates, and a community large enough to answer questions when something breaks. None of those three have independent, verified benchmarks yet. The release is still young, and the claims about ease-of-use are, so far, coming from the two organizations that built the toolkit. A complete, production-equivalent replacement for Nvidia CUDA has not been confirmed.
That is not a knock on the release, it is simply the normal early life cycle of any new developer platform. CUDA itself took years after its 2007 launch to become the default choice it is today. The difference this time is the speed at which Chinese developers have incentive to adopt whatever works, given that the alternative is simply not having access to competitive Nvidia hardware at all. It also arrives alongside a broader wave of specialized coding-assistant models, such as JetBrains’ Mellum 2.1 model aimed at coding agents, underscoring how much of the current AI infrastructure race is now being fought at the developer-tooling layer rather than purely on chip specs.
Historical Parallels: Platform Wars Rarely End Quickly
Software platform shifts in computing tend to move far slower than the headlines around them suggest. CUDA’s own 2007 launch did not immediately dethrone general-purpose CPU computing for parallel workloads, it took the better part of a decade of library building, academic adoption, and the eventual deep learning boom for CUDA to become as entrenched as it is now. x86 versus ARM in servers, and Windows versus Linux in enterprise computing, both followed similarly slow arcs where the incumbent’s ecosystem advantage outlasted several waves of “this changes everything” coverage for the challenger.
The lesson for the Ascend-TileLang release is that its significance should be measured in a multi-year trajectory of developer adoption and benchmark credibility, not in whether it immediately dents Nvidia’s quarterly numbers. The January 2025 DeepSeek stock shock is the closest recent comparison, and even that, a nearly $600 billion single-day market cap swing, did not meaningfully reduce Nvidia’s data-center GPU shipments over the following year.
Five Predictions for the Next 12 Months
- Expect other Chinese AI labs, such as those building models compared in recent open-source frontier rankings, to publish their own ports to the Ascend-TileLang stack rather than build competing kernel languages from scratch.
- Expect AMD to lean harder into ROCm marketing and documentation in response, since a credible third CUDA alternative narrows the argument that ROCm is the only viable non-Nvidia option.
- Expect independent, third-party benchmarks comparing Ascend-TileLang training throughput against Nvidia H100/H200-class hardware to surface within two to three quarters, and expect the numbers to be contested by both sides.
- Expect continued U.S. policy debate over whether tightening export controls further actually accelerates China’s push toward self-sufficient AI software, given Nvidia’s own filing already flags that risk.
- Expect Nvidia to respond primarily through hardware and pricing moves rather than open-sourcing core CUDA internals, continuing the pattern seen in its RTX Pro channel strategy and its own open-source model investments.
What This Means for Developers Outside China
For developers in the U.S., Europe, and other markets with uninterrupted Nvidia access, the near-term practical impact is small. Most teams outside China have no immediate reason to port workloads to Ascend hardware they likely cannot buy. The more durable relevance is architectural: TileLang itself is a general-purpose kernel language, not an Ascend-exclusive tool, and any gains it makes in simplifying kernel development could eventually influence how other hardware vendors, or even Nvidia itself, think about lowering the barrier to writing optimized AI code. A useful tool rarely stays confined to the market that produced it.
Frequently Asked Questions
What did DeepSeek and Huawei release in September 2026?
On September 30, 2026, a set of six free, open-source software modules for Huawei’s Ascend AI chips, including an Ascend version of the TileLang kernel language and a matrix-multiplication library called DeepGEMM-Ascend, according to The Next Web.
What is TileLang?
TileLang is a higher-level programming language for writing AI kernels that DeepSeek has positioned as simpler to use than writing low-level CUDA code directly.
What is DeepGEMM-Ascend?
A library optimized for matrix multiplication, one of the core mathematical operations behind large language model training and inference, adapted for Huawei’s Ascend hardware.
Why are DeepSeek and Huawei building a CUDA alternative?
Because Nvidia’s top-tier chips have effectively been locked out of China’s data-center market since the H20 ban took hold in 2025, leaving Chinese AI developers with strong incentive to build a usable software ecosystem around domestic hardware like Huawei’s Ascend chips.
Is Huawei Ascend hardware compatible with CUDA code?
No. Ascend chips are not CUDA-compatible, which is exactly the gap this new open-source toolkit is trying to close by giving developers an easier path to writing Ascend-native kernels.
How dominant is Nvidia’s CUDA ecosystem right now?
Nvidia’s ecosystem materials describe more than 4 million developers and over 40,000 companies using its AI tools, and an OECD paper published in 2025 estimated Nvidia holds more than 80% of the market for AI-focused GPUs.
Did this news affect Nvidia’s stock price?
There is no confirmed, immediate stock-price reaction tied specifically to the September 30, 2026 toolkit release. The much larger market reaction tied to DeepSeek happened earlier, on January 27, 2025, when Nvidia shares fell just under 17% following the DeepSeek-V3 and R1 model releases, according to Reuters.
Where can developers get the new Ascend tools?
The modules were released as free, open-source downloads, according to reporting on the release; developers interested in the toolchain should consult DeepSeek’s and Huawei’s official open-source repositories directly for installation and documentation.
