On Tuesday morning DeepSeek pushed DeepInfer to GitHub under the Apache 2.0 license: the in-house serving stack behind DeepSeek-V4.1-Flash, released as open source, with 2,015 points and 785 comments on Hacker News before most of Europe had finished lunch. The repo was created on October 16 and had 4,320 stars four days later; by the time I looked it was at 10.8k stars and 950 forks.

Everyone read the license. Almost nobody read the README.

Under “Supported hardware” there are exactly two entries: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

That is the whole story, and it is a much more interesting one than “China’s best model lab gives away its crown jewels.”

Two Sockets, One Message

An Apache license is a promise not to sue you. A hardware support list is a roadmap.

DeepSeek did not release a general-purpose inference engine. It released the kernels, parallelism layout, speculative decoding and scheduler for one model family, tuned for two pieces of silicon. Treating Huawei’s Ascend 950 as a first-class target in the same commit as Nvidia’s SM100 is not a courtesy to the open-source community. It is the single most valuable thing anyone could have done this year for Ascend adoption, and it did not require a marketing budget.

Porting a production inference stack is miserable, unglamorous, months-long work — the kind of thing an ops team defers until someone hands them a working reference and a reason. DeepSeek just handed them both. The reason is that there are now two vendors who can sell you the chip, and DeepSeek’s economics improve in direct proportion to how many of them exist.

What Isn’t in the Box

The repo is candid about its own gaps. A complete inference service requires a KV cache system — not included. Prefill/decode disaggregation is mostly implemented but not finished. Load balancing across data-parallel replicas is missing. The public release does offline inference for one model, V4.1-Flash, and nothing else.

Compare that to vLLM or SGLang, which serve dozens of model families and ship an OpenAI-compatible server out of the box. DeepInfer is not competing with them. It is a reference implementation: a build sheet for someone standing up a production service, addressed to a cloud operator rather than a hobbyist with a single GPU and a weekend.

An inference engineer at a US neocloud, mid-deploy on Tuesday, put it to me this way in a Slack DM: “I cloned it, ran the smoke test, and went looking for the KV cache. It’s a beautifully organized half of a serving stack.” Then he added that he’d be watching the commit log.

So the 10.8k stars are measuring something other than use. They are measuring aspiration — an enormous crowd of developers admiring a thing most of them cannot deploy on hardware they cannot buy. That is not nothing. It is just not adoption.

The Reversal Is the Tell

In April 2025, DeepSeek posted a short GitHub note explaining that it had considered open-sourcing its inference engine and decided against it. Three reasons: the code was an old fork, it was welded to internal infrastructure, and a small research team had no time to maintain a large open-source project.

Every one of those reasons was true. And none of them changed between April 2025 and October 2026. The team did not triple. The code did not stop being welded to a cluster.

What changed is that DeepSeek now has a silicon partner whose success is worth absorbing a maintenance burden for. The 2025 note was an engineering answer to an engineering question. Tuesday’s release is a business answer to a business question, wearing the same clothes.

Who Has to Update Their Priors

Start with the open-source triumphalists, who will file this under “another moat destroyed.” The moat was not destroyed. It moved — from the model layer down to the fab. DeepSeek did not commoditize inference in the abstract; it commoditized Nvidia’s position in Chinese inference specifically. Those are different acts with different beneficiaries, and only one of them helps a Shenzhen data center operator’s procurement department.

Then the export-control optimists, who have spent years treating advanced silicon as the chokepoint that settles everything. A chokepoint on hardware is a chokepoint on hardware. Software crosses a border as a tarball. If Ascend 950 now ships with a credible production engine from the country’s most respected model lab, the gap being closed is not the model gap — it is the operations gap, and the operations gap is the one that decides whether a data center is worth building at all.

And finally the open-weights-is-the-danger caucus, who will reflexively log this as another leak. The weights were already open. What is new is the plumbing. Plumbing is what converts a benchmark into an invoice, and invoices are what build capacity.

DeepSeek’s own announcement says the point is to lower compute costs, cut energy use, and make AI more affordable. Fine. Also true: every dollar of that savings is denominated in a chip somebody has to buy, and there are now two names that can appear on the purchase order instead of one.

That is not charity. That is a supply chain being assembled in public, with a permissive license as the packaging material.

The 785 comments were about the license. The two lines in the README were about the next five years.

Sources