China Isn’t Just Copying Nvidia’s Chips. It’s Rebuilding the Moat.

On September 30, DeepSeek released code for Huawei’s Ascend AI chips. One library handles the calculations models repeat millions of times. Another helps chips pass work to each other. DeepSeek and Huawei have also worked on a machine built around 128 Ascend 950 processors. The partnership puts a major model builder to work on the software that makes Chinese hardware usable. That matters more than another comparison of chip specifications.

The release has an awkward detail. Its best published communication results came from a special hardware configuration supplied to DeepSeek. The recommended commercial version was still due in October, and parts of the library remain unfinished. The installation notes are unusually frank about it. I like that. You learn more from an honest list of missing features than from another slide promising a million-chip cluster.

Nvidia’s defenders are too comfortable dismissing this attempt because a Huawei chip loses some one-on-one comparisons. Nvidia earns its position from an entire system of code, hardware, networks and customers who know how to use them. China is assembling its own version of that system. Washington has given its builders a powerful reason to keep going.

Nvidia’s moat lives in your code

Say your team has spent two years building an AI service on Nvidia GPUs. Your model runs in PyTorch. Your engineers know which kernels are fast, which ones crash and which version of a library broke last Tuesday’s deployment. Now a sales representative offers you another chip at a lower price.

How much lower? You have to retrain people, test every important operation and accept a fresh set of bugs. Meanwhile, customers expect the service to stay up. Hardware buyers can compare prices on a spreadsheet. The person responsible for getting the model running has a longer list.

CUDA is central to this advantage, but Nvidia sells much more than a programming tool. It supplies tuned libraries, networking, switches and complete systems. Its annual filing describes hundreds of software libraries and interfaces around its processors. Every developer who learns that stack makes it easier for the next employer to buy Nvidia. Every cloud provider that installs it makes the choice easier for the next customer.

Nvidia reported $89 billion in data-center revenue for the quarter ended July 26, 2026. Its companywide gross margin was 75%. Those numbers do not put a price on CUDA alone. They show the money available to a company that has made itself the familiar place to build and run AI.

Nvidia is also adapting. Its NVLink Fusion offering is designed to let large customers connect some of their custom processors to Nvidia’s platform. That is a clever defense: even when a buyer designs its own chip, Nvidia wants to supply the network and surrounding system. It has the cash to keep improving its software while smaller rivals spend theirs just getting a first version to work. A fast Chinese chip alone cannot undo that advantage. An engineer has to trust the whole route from code to a working service.

The U.S. made that route harder to buy in China

American export controls were meant to limit China’s access to advanced computing. They have done something else to the commercial market: they made the availability of Nvidia hardware uncertain. In January 2026, Washington began considering H200 export licenses case by case. That still leaves a Chinese buyer depending on approval, inspections and the next policy decision. Beijing has also restricted some sales. Nvidia says it shipped only a fraction of the H200 products licensed for specific Chinese customers.

In the quarter ended July 2026, those shipments amounted to less than 1% of Nvidia’s data-center revenue. The company described itself as effectively shut out of China’s data-center computing market, although it continued selling other products there. Then its filing spelled out the consequence: competitors were building larger developer and customer communities while Nvidia was absent.

That is the part of this policy I think Washington has mispriced. You can restrict chip deliveries and still strengthen the people building an alternative. A Chinese lab that spends a year fixing Ascend software has a stronger reason to stay on Ascend next year. The engineers it hires carry those skills to other companies. The cloud provider that learns to operate the machines can sell the experience to its customers.

The national-security argument for export controls involves more than a chip company’s revenue. The commercial cost is clear enough: China now has a large group of buyers who need domestic hardware to work, even when using it is difficult. That is a rare opening for an aspiring platform. It also changes what “competitive” means inside China. A domestic chip does not have to win every benchmark today if customers fear losing access to the imported one tomorrow. Give its developers enough time with real workloads, though, and convenience starts catching up with necessity.

DeepSeek is doing the work a chipmaker needs

DeepSeek’s new DeepGEMM-Ascend library handles matrix multiplication, a calculation used throughout AI models. It keeps an interface compatible with DeepSeek’s existing library, which reduces the amount of familiar code an engineer has to throw away. It still requires Huawei’s CANN software and TorchNPU, the bridge to PyTorch. That is a useful improvement, with installation work attached.

The other release, DeepEP-Ascend, deals with communication. Many current AI models have a large collection of specialized parts, often called experts. For each piece of text you send, the model calls on a few of them. If those experts sit on different processors, the system has to move data across the network and bring the result back. Slow communication means your expensive chips spend too much time waiting.

DeepSeek measured its library on groups of eight to 128 Ascend 950DT processors. The results are its own benchmarks on a specified setup. They do not tell us whether a normal customer gets the same performance, and they do not establish a cost advantage over Nvidia. Some communication features are still being built. The validated setup uses precise versions of CANN, Python, PyTorch and TorchNPU, plus the special pre-release hardware configuration I mentioned earlier. A team copying the commands onto a different machine has more testing ahead of it.

Putting the code in public lets other engineers inspect it, report problems and adapt it. DeepSeek also points to TileLang, a higher-level language for writing fast AI operations, and there is an Ascend version available. If these tools make the next model easier to run, Huawei benefits without having to employ everyone who writes its software.

Huawei has been working on the wider developer experience as well. Its TorchNPU project connects Ascend to PyTorch, the framework many model builders already know. PyTorch’s own documentation includes the Ascend extension as a device-backend example. Huawei says Ascend supports more than 90 outside open-source projects. I would treat that count as a starting point. A developer cares whether her specific model runs, whether the documentation matches the installed version and whether someone fixes a bug before her deadline. Huawei needs to win those mundane tests repeatedly.

DeepSeek gets something too: another source of computing power in a market where access to Nvidia is political. I do not think the two companies have to agree on everything for this to work. Their immediate incentives line up. Huawei needs demanding models on its chips; DeepSeek needs chips it can obtain and program. The code turns that shared problem into something other teams can use.

Over time, the feedback runs deeper. A model builder who knows the available memory and network can change how a model distributes its work. Hardware engineers then see which operations slow that model down and improve the next chip or its software. Nvidia has benefited from this kind of close contact with model builders for years. Chinese companies now have enough domestic demand to keep the conversation going across product generations. That matters more to me than a promise to match Nvidia’s latest processor on a single benchmark.

Huawei is building a larger machine around weaker chips

A processor does not work alone when you train or serve a large model. It needs memory and fast connections to other processors. Once enough chips are involved, moving information becomes a serious expense in both time and electricity. Huawei has spent years building networking equipment, and it is applying that experience to AI systems.

Its researchers described a CloudMatrix384 machine connecting 384 Ascend 910C chips and 192 Kunpeng CPUs. They also designed software to split the work of answering a prompt, generating new words and managing stored context. They are trying to make the machine as a whole do useful work, even when each processor has limitations. The paper reports performance under its test conditions. It does not give buyers a full comparison of purchase price, energy use and reliability against an equivalent Nvidia installation.

Huawei’s latest Atlas 960E design scales to 4,096 processors and uses optical connections near the chips. That addresses a problem created by scale itself. Thousands of processors need thousands of reliable ways to talk. Huawei says its optical units reduce the number of conventional modules needed in that configuration and cut power consumption in the interconnect. Those are company claims about its design, and they deserve testing in customers’ data centers.

I am skeptical of the giant cluster numbers both sides like to announce. Four thousand chips in an architectural specification tells you little about the cost of serving one million customers. I want to see a company install the machine, keep it running and get a lower cost for the same quality of answer. Huawei has a credible strategy for reaching that point. We still need the operating results.

Alibaba wants a piece of the same market

China’s alternative to Nvidia may have several owners. Alibaba designs Zhenwu chips through T-Head, develops Qwen models and rents computing through Alibaba Cloud. It can test its hardware on its own workloads and offer it to customers through a service they already buy.

In July, Alibaba opened its T-Head SAIL software stack and said it had shipped 560,000 Zhenwu chips cumulatively by April. That company figure covers a range of use; it tells us little about how many top-tier models run on them. In September, Alibaba unveiled the Zhenwu V900. Commercial release is planned for early 2027.

Alibaba and Huawei have different advantages. Huawei can sell hardware and software to many operators. Alibaba owns a cloud where it can place its chips, tune Qwen for them and sell customers the finished service. They also compete for programmers. If their software stacks demand separate versions of every important operation, developers will have to do the same work twice. Shared tools such as PyTorch help, provided the features people need run properly on each backend.

I would rather see those companies competing than one state-favored winner collecting a captive market. Rivalry gives buyers a chance to demand better tools and lower prices. China’s push for independence becomes much more convincing when a customer has a choice among domestic products.

Outside China, the sales pitch changes. A cloud operator with reliable access to Nvidia has less reason to absorb the switching cost I described earlier. Huawei or Alibaba has to offer a better price, a more useful service or an application Nvidia’s partners have neglected. Open-source software may travel first because developers can try it before committing to a new data center. Winning China gives these firms experience and scale; it does not automatically hand them the rest of the world. Nvidia’s global customer base remains a formidable defense.

Production is the part nobody can patch with code

Huawei’s rotating chairman Eric Xu admitted in September that production was falling short of Chinese demand for its AI computing equipment. The shortage limits overseas sales. Huawei plans to release its next 960DT and 960PR processors in 2027. Alibaba’s V900 also has a 2027 commercial date. You can believe in the direction of these products and still refuse to count them as machines already delivered.

Large AI systems need memory, packaging, switches, cooling, electricity and technicians who can repair them. Using more processors to make up for slower individual chips increases demands on several of those inputs. A benchmark that ignores power and downtime tells only part of the cost. The right comparison is useful work per dollar and per watt, measured on the models customers actually run.

Nvidia depends on a concentrated supply chain of its own. Its filing names TSMC and Samsung among its wafer suppliers, and SK Hynix, Micron and Samsung for memory. Yet it has spent years building relationships across that chain. Chinese chipmakers have to secure enough of the same kinds of inputs while navigating restrictions on advanced manufacturing equipment. Better software gives them a reason to solve that production problem; it does not solve it for them.

Huawei has machines and customers. DeepSeek has released practical code. Alibaba has existing chip shipments and a cloud business. The newest hardware and software still face capacity limits, incomplete features and questions about operating cost. Those details decide whether this becomes a profitable industry or an extremely expensive campaign for self-sufficiency.

The second order matters more than the first

Investors often treat this as a simple transfer of revenue from Nvidia to a Chinese rival. Nvidia is still growing rapidly elsewhere. Huawei’s 2025 annual results show CNY880.9 billion in companywide sales and CNY192.3 billion in research and development spending, but they do not tell us the profit on an Ascend system. Alibaba might earn its return through cloud subscriptions and lower model-serving costs rather than a chip margin. China can build substantial AI capacity without producing a company that earns Nvidia’s margins.

I care more about the next order than a headline benchmark. If a customer finishes a model on Ascend, will it use Ascend for the next one? Can a new engineering team install the tools without Huawei or DeepSeek walking it through the process? When Alibaba offers more Zhenwu capacity, will its customer take it for the price and service, or because another chip is unavailable?

If customers keep buying these systems once they have other options, Huawei and Alibaba have earned their place. Until then, some of their sales tell us as much about restricted supply as they do about product quality.

DeepSeek’s September release still needs special configuration, and other programmers have plenty of work left. I think dismissing it is a mistake. Each completed tool gives Chinese labs one less reason to return to CUDA, and export controls keep supplying a reason to finish the tools. Nvidia can remain dominant globally while losing its place as the default for an entire generation of developers in China. That is a serious change, even if no Chinese chip wins the one-on-one race people keep drawing.