On October 20, 2026, the Department of Homeland Security published a notice in the Federal Register under the 2018 Federal Acquisition Supply Chain Security Act, ordering (1) the exclusion of information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence and its affiliates from federal civilian agency procurement, and (2) the removal from civilian agency IT systems of DeepSeek and affiliate products, as well as third-party products and services developed using DeepSeek implementations.
The order covers federal civilian agencies such as Treasury, State and Justice, and excludes the Department of Defense. It was originally issued around July 2026 and made public around October 20. Agencies may request exceptions or delays in writing within 60 days of issuance, but only with a "compelling justification."
On the same day, DeepSeek open-sourced DeepInfer together with the deepseek-recipe tokenization recipe on GitHub under the Apache 2.0 license, aimed at production deployment of DeepSeek-V4.1-Flash. The repository passed ten thousand stars within a week of creation.
DeepInfer supports NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 accelerators at launch, and is built mainly on vLLM while drawing on SGLang and TensorRT-LLM. It ships model-optimized kernels, cross-chip partitioning, and speculative decoding.
DeepSeek's own published figures claim more than 25,000 decode tokens per second, a single cluster processing 155 trillion tokens per day, and roughly 38 dollars of revenue per GPU per hour under off-peak power pricing. These are the company's own measurements and have not been independently reproduced.
The current release does not include a full cache system, prompt prefill/decode disaggregation, or multi-machine load balancing — an honest, scoped first drop rather than a finished product.
Chinese Foreign Ministry spokesperson Mao Ning said China consistently opposes over-generalizing the national security concept and politicizing science, technology and trade issues, and will firmly safeguard the legitimate rights and interests of Chinese enterprises.
Why This Is the Most Fun — and Most Surprising — Story of the Day
Let us start with the part that made a lot of engineers laugh out loud: the timing was not a coincidence. The Federal Register notice went up, and a few hours later a new GitHub repository went up. When a government publishes an exclusion order and a company answers with an open-source release on the same afternoon, that is not a press strategy. That is a face-to-face response, delivered in the only language that does not require a press office.
The genuinely surprising thing here is what was given away. The AI industry has spent two years arguing about open weights — releasing model parameters so researchers can poke at them. That is generous, but it is not the crown jewels. An inference engine is the crown jewels. It is the thing that decides how many users a given piece of hardware can serve, how much each token costs, and whether a lab's margins are 30 percent or 60 percent. Inference engines have historically been either closed or dominated by NVIDIA's ecosystem. DeepInfer hands over the craft of "how we run models cheap." A lab does not do that by accident.
And then there is the detail that matters more than the drama: dual-chip support. NVIDIA SM100 and Huawei Ascend 950, on day one. This is the quiet demolition of the entire logic of a compute blockade. The blockade's premise is that model and chip can be locked together — cut off the silicon and the model cannot run. A chip-agnostic inference engine simply unbundles that relationship. Use NVIDIA today, switch to Ascend tomorrow, and the application layer barely notices. The wall was built in hardware; the answer was written in software.
There is also a rich seam of dark comedy in the ban's wording. The removal scope extends to "third-party products and services developed using DeepSeek implementations." Read that twice. It sweeps in thousands of downstream fine-tunes, wrappers, and internal tools built by American contractors on an openly published model. Every federal vendor now has to audit its stack for descendants of a model, and every one of those audits costs real money. Meanwhile the notice governs procurement — it does not govern GitHub. Open-source code, once released, cannot be recalled. The order can close a procurement door; it cannot close a repository that has already been cloned tens of thousands of times.
The second-order effects are where this gets genuinely interesting for everyone else. American cloud vendors and hyperscalers can now study DeepInfer and potentially run the world's most-discussed open model family more cheaply on hardware they already own. A ban intended to keep a model out of the federal government may, in practice, accelerate the adoption of its engineering everywhere else. That is not a prediction, it is a pattern that has repeated every time a technology has been declared contraband.
Two honest caveats. First, the headline numbers — 155 trillion tokens per day, 38 dollars per GPU-hour — are vendor self-reports with no independent benchmark, and the repository itself is explicit about what is missing. Second, the ban is real and it bites: it is now a government-wide civilian rule under a supply-chain security framework rather than a scattering of agency-level hesitations. Both things are true at once, and that simultaneity is exactly why this story is the day's most fun and most surprising at the same time.
Source
Federal Register / DHS notice / GitHub deepseek-ai/DeepInfer / Reuters / CNBC / The Next Web / Tencent News / ZAKER / Science and Technology Daily / threerouter
Top comments (1)
😀