Chinese AI developer DeepSeek has released a free, open-source programming toolkit optimised for Huawei's Ascend processors, in a move that targets the software moat Nvidia has built around its AI hardware business.

The toolkit, announced on DeepSeek's WeChat account, includes the TileLang platform — a framework the company positions as a rival to Nvidia's CUDA, the programming environment that has become the de facto global standard for developing AI models. The release bundles compute and communication libraries alongside the language itself.

DeepSeek worked directly with Huawei on the Ascend 950 chips, and praised the vendor for what it called "unwavering and strong support" during the research and development process.

The company did not shy away from the strategic intent behind the release. "To do a good job, one must first have the right tools," DeepSeek stated, arguing that the first step in building an independent GPU software ecosystem is an easy-to-programme high-level language that "can simultaneously reach the hardware performance limits."

On the practical benefits for developers, DeepSeek claimed TileLang is "simpler to programme, significantly improving development efficiency and simplifying code logic." That claim comes from the company itself rather than independent benchmarks, and CUDA's two-decade head start in tooling, documentation and developer mindshare remains a formidable barrier for any challenger — the comparison with the Nvidia platform in this announcement is DeepSeek's own framing.

Hardware order underpins software push

The software release follows reporting by Bloomberg earlier this month that DeepSeek plans to deploy at least 160,000 Huawei AI accelerators at a new data centre in Inner Mongolia. An order of that scale would rank among the largest single deployments of domestically produced AI compute in China, and it gives the TileLang toolkit an immediate, high-volume user base. Software ecosystems typically gain traction when a major operator of hardware adopts them internally before pushing them outward, and DeepSeek's dual role as both Ascend customer and toolkit developer fits that pattern.

Why software matters in the chip war

The scarcity of CUDA-compatible alternatives has been a persistent structural problem for Chinese AI companies. US export controls have restricted access to Nvidia's most advanced GPUs, pushing developers toward domestic accelerators — but hardware alone has never been enough. AI teams build models on software stacks, and the cost of migrating code away from an established ecosystem is measured in engineer-months and lost performance.

This is the gap Huawei and DeepSeek are attempting to close. By open-sourcing the toolkit rather than keeping it proprietary, the partners lower the barrier for third-party developers to write for Ascend silicon, replicating the community-driven dynamics that made CUDA dominant. The compute and communication libraries included in the package matter here: distributed training across large accelerator clusters depends heavily on optimised communication primitives, and shipping these alongside the language suggests the toolkit is aimed at production-scale workloads rather than academic experimentation.

The road ahead

For Huawei, the Ascend 950 sits at the centre of its push to make domestic AI infrastructure credible at scale, and DeepSeek — whose low-cost model training runs have already disrupted assumptions about compute requirements — is arguably its most valuable showcase customer.

The immediate test will be adoption. A free download and enthusiastic vendor quotes do not constitute an ecosystem; that requires developers outside the two companies' orbit to build, debug and ship models on Ascend hardware. With at least 160,000 accelerators reportedly heading to Inner Mongolia, DeepSeek has given itself the incentive and the scale to make TileLang work in practice.