This article draws on Business Outreach Magazine's report on DeepSeek and Huawei.
The case for treating the software layer as secondary is stronger than the optimists admit. A rival to a mature compute platform takes years of compilers, libraries, debuggers and documentation, plus the population of engineers who know which of them lie to you. A team priced a move off Nvidia and gave up in the second week, when the estimate for re-writing their attention kernels passed what they expected to save on hardware across three years. They concluded that the moat is a country, and that no one captures a country with one open-source release.
That conclusion interprets the problem correctly and then assigns the cost to the incorrect area. Switching cost lands in two spots: the first afternoon an engineer spends with unfamiliar hardware, and the handful of operations their model cannot run without. DeepSeek and Huawei have aimed at both, which is why their release matters more than its size suggests.
What a chip monopoly actually is
Start with what a monopoly in this market consists of. It is the point at which a developer's choice stops feeling like a choice. EE Times reported that Nvidia's early investment in CUDA turned a graphics processor into the standard platform for AI work, and that moving to another vendor is now cost-inefficient, with proprietary standards raising the barrier for anyone trying to enter. Read that as a statement about behaviour. The chip is fast; the reason people stay is that the code they already have runs on it.
Business Outreach Magazine reports that DeepSeek and Huawei are releasing open-source programming tools for Huawei's Ascend chips, including TileLang, a high-level language, alongside compute and communication libraries. The stated aim is to make Ascend "considerably easier" for developers, which is a modest phrase for the thing that decides whether anyone starts.
Why the tooling cannot be attacked alone
Habits are cheap to keep and expensive to break, which is why a programming model cannot be changed on its own. HFS Research set out how the AI supply chain has concentrated: a small group of firms own compute, models and the fabrication underneath them, invest in one another, and keep demand circulating inside a closed loop. The same research details an enterprise rebuilding its pipelines after a vendor changed its pricing mid-contract, and draws the conclusion that portability has stopped being an IT convenience and become a survival strategy.
That is the correct mental model. The chain that EE Times described runs through chip design, foundry capacity and high-bandwidth memory, with Nvidia's tooling as one link. You do not break that chain by hitting its strongest link. You break it where a buyer has an alternative and the switch is affordable, and for most teams that link is the programming model: the thing an engineer touches before any benchmark result exists.
TileLang is the layer in the stack
DeepSeek asserts that it is useful: code becomes simpler without giving up access to the performance underneath, so the developer stops managing every hardware detail by hand and the reasoning in the source stays legible. Less code logic means fewer places for errors.
The table below sets the release against the layers a working platform has to cover.
| Layer of the stack | What a developer needs from it | What this release puts in place | What is still open |
|---|---|---|---|
| Writing the inner loop | A way to express a kernel and tune it without hand-managing every hardware detail | TileLang, a higher-level language above Ascend C with access to the hardware below | Published tuning practice and worked examples a new team can copy |
| Repeated operations | Optimised building blocks for the arithmetic a model performs over and over | Compute libraries for Ascend | Coverage of every operator a real model uses, and an owner when one of them is slow |
| Work across many chips | Fast exchange of gradients and activations between accelerators | Distributed communication libraries | Evidence at the scale of the reader's own cluster, not a demonstration cluster |
| Continuity | Confidence that code written now still runs in a few years | An open-source repository anyone can inspect and change | A version policy, a support commitment, and a record of keeping both |
The language is the visible part of the release. The libraries are the load-bearing part.
What only appears at cluster scale
A single chip tells you very little. Training or serving a large model can occupy hundreds or thousands of accelerators, and at that size arithmetic speed matters less than the speed of the conversation between chips. Communication libraries are what make that conversation efficient, and they are the reason this arrived as a set of components that includes the language.
This is also where a challenger is most exposed. Distributed performance is the hardest thing to show in a repository and the easiest thing for an incumbent to stay ahead on, because it depends on interconnects, topology and firmware as much as on the software layer. A team that benchmarks one Ascend card and extrapolates has gained no knowledge about the configuration it will actually run on.
What open source does to adoption, and what it does not
Releasing the tools under an open licence is a decision about who does the work. Developers can read the code, try it on their own models and send back fixes, without waiting for one company to ship every feature. DeepSeek has run this play before with its models, and the approach has done more for its adoption than any benchmark table.
It is not a warranty. A repository is not a platform. Platforms need releases that do not break, documentation that matches the code, and people who answer when an operator is missing. Business Outreach Magazine makes the same caution explicit: A simpler programming model is an assertion about the form of the work. It does not state that the software has caught up with a platform that has been accumulating for far longer.
Who the first demanding customer is
Toolchains improve because somebody with a production deadline is using them, and here the collaboration has more weight than the language release. DeepSeek says the operators used in its own training environment now have high-performance implementations on Ascend. Business Outreach Magazine reported that Huawei said its Ascend supernode systems would support DeepSeek's V4 models, and that DeepSeek has since been linked with plans to put a very large number of Huawei accelerators into new infrastructure. A model developer with real workloads is a different kind of first user from a benchmarking team.
The same principle shows up in policy. The UK AI Hardware Plan sets out an end-to-end pipeline running from development through demonstration and deployment to scale, and it uses public procurement, including advance commitments to buy novel inference chips, to give hardware firms a customer before they have a product history. WIRED reported that British suppliers get priority in that procurement and that the government is deliberately targeting specialist inference silicon. The plan stops short of trying to match the largest training systems.
Two countries, one lesson. A stack becomes a real alternative when a demanding user is already running production workloads on it. The software layer is where that user starts, and it is where the monopoly is weakest.
The Ascend release lowers the first afternoon's cost and covers the operations a real model needs. CUDA replacement is a later question. An open-source release has a smaller job: the first afternoon, and the handful of operations that determine whether the afternoon ends in a working kernel or a flight home. DeepSeek and Huawei have put something in both places. That is why the release matters more than its size suggests.
Sources
- UK AI Hardware Plan - GOV.UK
- How are Monopolies formed in the Generative AI Hardware Ecosystem? - EE Times
- Before AI breaks the market, let's break the monopoly - HFS Research
- The UK Is Betting on a Billion-Dollar AI Supercomputer to Kick Its Addiction to US Tech - WIRED
Source: DeepSeek Huawei AI Chip Software: Nvidia CUDA Rival Explained, Business Outreach Magazine


