DeepSeek Open-Sources TileLang: Filling the Last Gap in China's AI Ecosystem

A Quiet But Significant Open-Source Release

Just before China's National Day holiday, DeepSeek^Ad quietly released its low-level toolchain TileLang to the open-source community. Unlike splashy product launches, this move came with almost no fanfare — but to Bilibili creator Mingyan (明远说), the significance of this release was hard to overstate. In his view, it signals that China's AI ecosystem is inching closer to a complete, self-sustaining loop.

The timing was notable. The day before the release, China's Ministry of Science and Technology held a press conference explicitly calling for full-stack breakthroughs in integrated circuits, industrial machine tools, and precision instruments. The very next day, DeepSeek open-sourced the low-level toolchain it uses to train trillion-parameter models — and immediately added support for Huawei's Ascend chips. For most end users, this kind of infrastructure tooling is invisible. But it touches exactly the weakest link in China's domestic AI stack.

The Bottleneck Isn't the Chip — It's the Ecosystem

To understand why TileLang matters, you need to look at how the AI stack is structured. At the bottom sits the chip. NVIDIA has long dominated here, but Huawei has been steadily closing the hardware gap with its large-scale Ascend clusters. The real bottleneck lies in the software layer in between.

Think of it like HarmonyOS: the operating system exists, but it lacks the rich app ecosystem that attracts developers. NVIDIA's moat lies in its CUDA ecosystem — built up over nearly two decades, it includes not just low-level drivers but a vast library of matrix operations, operator libraries, and tooling that developers can use off the shelf. Writing training code or building MoE models on top of this mature stack is relatively straightforward, because so much foundational work has already been contributed by countless vendors and developers.

Huawei's Ascend has its own software stack, CANN, but it's nowhere near as mature. The common workaround has been to "translate" CUDA code into something CANN can understand — but this is not a true low-level replacement. It's expensive to maintain, and performance is often compromised.

A note on CUDA's moat: CUDA (Compute Unified Device Architecture) was introduced by NVIDIA in 2006. Over nearly two decades, it has accumulated a massive upper-layer toolchain — cuDNN (deep learning operator library), cuBLAS (linear algebra library), NCCL (multi-GPU communication library) — along with millions of developers and a vast ocean of open-source projects. Most AI developers simply assume CUDA is the underlying infrastructure when writing training scripts. PyTorch and JAX are both deeply optimized for it. This means that even if a competing chip matches NVIDIA on raw compute, developers still face enormous porting and rewriting costs without an equivalent software ecosystem — which is precisely why the ecosystem is harder to catch up to than the chip itself.

What TileLang Fills In

DeepSeek has long been doing bespoke development work for Huawei Ascend — technically demanding work that few others have invested in. TileLang, now open-sourced, fills in exactly that missing middle layer of the ecosystem, and makes it freely available.

The core value is this: developers can now write a single codebase that compiles and runs on both NVIDIA servers and Ascend servers, without having to do the complex Ascend-specific porting work separately. This benefits not just DeepSeek itself, but any model company — Doubao, Yuanbao, Alibaba's model lineup — that wants to be compatible with both hardware platforms with minimal code changes.

In other words, TileLang flattens the development divide between NVIDIA and Ascend at the toolchain level, lowering the barrier for domestic chips to actually get used.

TileLang's technical positioning can be understood as a domain-specific language (DSL) or compiler intermediate layer for AI computation — similar to the role LLVM plays in general-purpose compilation. Developers describe computation logic (matrix multiplication, attention operators, etc.) using a unified abstract syntax; TileLang's compiler backends then translate that into native instructions for different hardware — whether NVIDIA's PTX instruction set or Huawei Ascend's Da Vinci architecture. This "write once, run anywhere" philosophy echoes WebAssembly in the web world or Flutter in mobile development. The challenge, of course, is that delivering near-peak performance on each hardware target requires careful tuning around each chip's memory hierarchy and compute unit layout — and that is precisely the technical asset DeepSeek has accumulated through years of deep work on Ascend.

Closing the Loop: From Chip to Application

With this piece in place, the full AI stack is starting to form a closed loop: chips can be domestically produced, open-source models are no longer behind the curve (and are narrowing the gap with closed-source ones), the software ecosystem is being filled in, and work continues down the stack toward manufacturing and lithography equipment.

Mingyan offers a provocative take in his video: OpenAI's recent use of "safety" as a reason to pause may reflect diminishing returns from simply scaling compute — it's getting harder to produce a dramatically more capable model just by throwing more hardware at the problem. He argues the center of gravity in AI competition is shifting from model capability to application deployment: who can actually roll out large models across different companies and industries. And that, he contends, is where China has a structural advantage. From applications to models, to the software ecosystem, to chips and manufacturing — an entire chain is taking shape.

It's worth noting that the claim about "OpenAI using safety as cover for a training plateau" is Mingyan's personal speculation, with no direct supporting evidence. Readers should treat it as such.

MoE (Mixture of Experts) is key context for understanding why DeepSeek has such a strong need for low-level tooling. MoE architectures divide model parameters across multiple "expert" sub-networks, activating only a subset per inference pass — enabling massive parameter counts without proportional increases in compute. DeepSeek-V3 and similar models use this architecture to reach trillion-parameter scale. But MoE places extreme demands on multi-GPU communication, memory bandwidth, and operator scheduling. Off-the-shelf CUDA operator libraries often can't fully utilize Ascend hardware for these workloads — purpose-built low-level operators are required. That is exactly the pain point TileLang is designed to address.

An Industrial Philosophy: Strength Through Self-Reliance

In the second half of the video, Mingyan takes a clear stance on industrial policy. He says he has opposed purchasing the castrated, export-controlled NVIDIA chips sold specifically to China — overpriced, underperforming, and potentially carrying backdoor risks. His view: rather than accepting dependence, it's better to block those sales entirely and force domestic alternatives to develop.

He goes further, predicting that when technology restrictions genuinely fail to contain China's progress, NVIDIA may actually be willing to sell its advanced chips to China on normal terms — because commercial incentives will override containment logic. But even then, he argues, any procurement must hold firm to non-negotiable conditions: no backdoors, parity pricing with the U.S. market, no discriminatory sales terms, and oversight by a coordinating body to prevent companies from scrambling individually and creating new vulnerabilities.

Underlying this logic is his consistent stance as a political-economy commentator: China's AI trajectory is more promising, rooted in the system's capacity for organized coordination. He acknowledges that when he made similar arguments two years ago, he was widely dismissed.

A Balanced Take: Real Progress and Grand Narratives

Objectively, the open-sourcing of TileLang is a concrete and meaningful step in strengthening China's domestic AI software stack. Its impact in lowering the barrier to using Ascend in production is real.

That said, several caveats are in order. The analysis in this article draws primarily from a single Bilibili source, and contains a heavy dose of subjective political-economic framing. An open-source toolchain release is not the same as a mature ecosystem — CUDA's moat, built over nearly two decades, remains formidable. The United States retains a commanding lead in the overall AI closed-loop for now. Whether TileLang achieves wide adoption, and how it performs in practice, will take time and independent benchmarking to determine.

For readers, the right posture is to appreciate the genuine technical significance of this step — while keeping a healthy skepticism toward the grander narratives built around it.