The US Banned DeepSeek From Federal Agencies — and DeepSeek Answered by Open-Sourcing Its Production Engine | CloudEmythos

Illustration: CloudEmythos. Washington ordered DeepSeek technology out of federal agencies; DeepSeek answered by open-sourcing the engine that runs its service.

TL;DR — Around October 19, 2026, the U.S. Department of Homeland Security ordered federal civilian agencies to remove all DeepSeek technology from their systems — including third-party tools built on it. The very next day, DeepSeek open-sourced DeepInfer, the inference engine behind its production services, under Apache-2.0 — with first-class Huawei Ascend 950 support alongside Nvidia's flagship GPUs. The two events together are the clearest picture yet of a splitting global AI stack. All performance figures below are DeepSeek's own self-reported claims, not independently verified.

The Ban: What DHS Actually Ordered

The Department of Homeland Security's removal order is broader than a simple "don't use the chatbot" directive. Per reporting verified this week, the order covers:

• All federal civilian agencies must remove DeepSeek technology from their systems.
• Not just DeepSeek's own models and apps — the order extends to third-party tools built on DeepSeek technology. If a vendor product embeds a DeepSeek model or API, it has to go too.
• This is a supply-chain purge, not a warning about a single product. Washington is treating DeepSeek as a tainted input to the entire software stack.

We covered the full order in detail in our DHS DeepSeek ban article — including the Pentagon carve-out, the Federal Register mechanics, and why the order landed when it did. If you want the policy mechanics, start there. This article is about what happened next.

Context: The "dhs" Google Trends breakout this week traces directly to this order — it surged as news of the federal purge spread. The story traveled faster than the explanation. Most coverage summarized the ban; almost none connected it to what DeepSeek did the morning after.

The Answer: DeepInfer, One Day Later

On October 20, 2026 — one day after the DHS order — DeepSeek published DeepInfer on GitHub under the permissive Apache-2.0 license. This is not a model release. It is the inference engine: the software layer that takes an already-trained model and serves it to users at scale.

• What it runs: DeepSeek-V4.1-Flash, in offline inference mode only for this first release. Not a complete serving stack — KV cache, prefill/decode disaggregation, and load balancing across replicas are not included yet.
• What chips it targets: Nvidia SM100 (Blackwell) GPUs on CUDA 13 and Huawei Ascend 950 NPUs on CANN 9.2 — both as first-class targets. Per multiple outlets, this is the first time a frontier lab has given Huawei's silicon the same framework-level support as Nvidia's flagship data-center GPUs.
• The performance claims (DeepSeek's own, unverified): 155 trillion tokens per day per cluster, 96% input cache hit rate, ~400ms time-to-first-token, 230 tokens per second. Reported by outlets covering the release; not independently tested.
• The reversal: in April 2025, DeepSeek said it had considered releasing its full inference engine but decided against it. DeepInfer reverses that position — the whole engine, not just pieces.

We went deep on the technical stack — vLLM lineage, DSpark speculative decoding, the custom kernel stack (FlashMLA, DeepGEMM, DeepEP) — in our DeepInfer technical breakdown. The repo crossed ~8,000 GitHub stars within hours of launch, one of the biggest developer threads of the week.

What DeepInfer Actually Is (Plain English)

If you are not an infrastructure engineer, here is why an "inference framework" matters more than it sounds.

A trained AI model is just a very large file of numbers. To answer your question, a system has to load those numbers onto chips, split them across multiple processors, batch your request with thousands of others, and stream the answer back token by token. That orchestration layer is the inference engine — and it is a huge share of what every AI company pays per answer. Cheaper, faster inference means cheaper API prices and lower serving costs.

Why open-sourcing the engine is the deeper move: Open-sourcing a model gives developers weights they can run. Open-sourcing the serving engine gives developers the recipe for running models at frontier scale — the scheduling, the kernels, the multi-chip orchestration. It is the difference between handing someone a cake recipe and handing them the blueprints to the bakery. And this bakery's blueprints say, right there in the supported-hardware table: Huawei Ascend 950, first-class.

That hardware line is the real story. For years, the global AI ecosystem has been a Nvidia monoculture at the framework layer — CUDA, cuDNN, NCCL, and everything built on top. A frontier lab shipping first-class support for a Chinese NPU is a software vote for a post-Nvidia serving world. Anyone who can run DeepInfer on Ascend hardware now has a path to frontier-class inference that never touches an American GPU.

The Timing: Coincidence or Message?

Let's be honest about what we know and don't know.

• What we know: the DHS removal order landed around October 19; DeepInfer was published October 20. That sequence is factual and verified.
• What we don't know: whether the release date was chosen because of the order. Software release schedules are set weeks in advance; a federal order on Monday does not rewrite a GitHub release pipeline by Tuesday morning. DeepSeek had been signaling an engine release since its April 2025 reversal.
• What the symbolism is regardless: Washington tried to wall DeepSeek out of government systems. DeepSeek responded — deliberately or not — by giving the world its serving engine, with Huawei silicon as a first-class citizen. The message writes itself either way: you can ban the company from your agencies; you cannot ban an Apache-2.0 codebase from the internet.

CloudEmythos take: The smartest reading is not "DeepSeek is trolling Washington" — it is that open source is DeepSeek's asymmetric answer to every restriction. Export controls can throttle chips. Federal orders can purge software. But an MIT/Apache-licensed codebase, once published, replicates faster than policy can chase it. DeepInfer is the clearest demonstration yet that in the AI stack war, openness is the weapon of the side with fewer chips.

The Bigger Picture: The AI Stack Is Splitting in Two

Zoom out, and the DHS-DeepSeek week fits a pattern that has been accelerating all year: the global AI stack is dividing into two ecosystems.

The American stack: hyperscalers are building software moats around their own silicon. Google's TorchTPU lets developers run PyTorch workloads natively on TPUs; AWS's TorchNeuron does the same for Trainium and Inferentia. The message: the model code is portable, the hardware underneath is ours.

The Chinese stack: DeepSeek is building the software case for Huawei's Ascend ecosystem — not by locking developers in, but by open-sourcing the framework layer so anyone can run frontier-class serving on non-Nvidia chips. The message: the hardware is ours too, and the software is free.

Notice the asymmetry. The American play is proprietary frameworks around proprietary chips — you get the portability, Amazon and Google get the lock-in. The Chinese play, at least from DeepSeek, is open frameworks around domestic chips — you get the freedom, Huawei gets the ecosystem. Which strategy wins the next decade of AI infrastructure is genuinely an open question, and it will not be decided by federal removal orders.

One more data point for the trend file: Japan just deployed AI agents to 180,000 government workers — with published safety standards — the same week Washington was purging DeepSeek from its agencies. Different governments, opposite directions. The global AI map is being redrawn in real time.

What to watch next: (1) whether any U.S. lab matches DeepInfer's openness on the serving layer — the pressure is on; (2) whether Ascend 950 performance claims hold up under independent testing once researchers run the released code; (3) whether the DHS order's "third-party tools" clause triggers a wave of vendor audits — that is where the real compliance pain will land.

Related reading

• DHS Bans DeepSeek From Federal Agencies — What the Order Actually Says
• DeepSeek Open-Sourced DeepInfer — and Gave Huawei's Ascend First-Class Support
• Japan Just Gave 180,000 Government Workers Their Own AI Agents

Frequently Asked Questions

What did the DHS order against DeepSeek actually require?

The Department of Homeland Security ordered all federal civilian agencies to remove DeepSeek technology from their systems — not just DeepSeek's own models and apps, but also third-party tools built on DeepSeek technology. It is a full purge of the supply chain, not just a ban on direct use.

What is DeepInfer?

DeepInfer is the high-performance inference framework DeepSeek says powers its own large-scale production AI services. On October 20, 2026, DeepSeek published it on GitHub under the permissive Apache-2.0 license. An inference engine is the software that serves an already-trained model to users — splitting the model across chips, batching requests, and deciding how fast answers come back. It is a large part of what an AI company pays per answer.

Which chips does DeepInfer support?

The first release supports two chip families as first-class targets: Nvidia SM100 (Blackwell) GPUs running CUDA 13, and Huawei Ascend 950 NPUs running CANN 9.2. Per multiple outlets, this is the first time a frontier AI lab has given Huawei's Ascend silicon the same first-class framework support as Nvidia's flagship data-center GPUs.

Why does the timing of these two events matter?

The DHS removal order landed around October 19, 2026, and DeepSeek open-sourced DeepInfer the very next day, October 20. The timing may be coincidence — release schedules are set weeks in advance — but the symbolism is hard to miss: Washington tried to wall DeepSeek out of government systems, and DeepSeek responded by giving the world its serving engine, with Huawei silicon as a first-class citizen.

Are DeepSeek's DeepInfer performance numbers independently verified?

No. DeepSeek's production figures — 155 trillion tokens per day per cluster, a 96% input cache hit rate, around 400 milliseconds time-to-first-token, and 230 tokens per second — are DeepSeek's own claims reported by outlets covering the release. They have not been independently tested. Treat them as the lab's self-reported numbers, not audited benchmarks.

What does this mean for the US-China AI technology split?

The two events together deepen the picture of a splitting AI stack. On the American side, hyperscalers are building software around their own silicon: Google's TorchTPU, AWS's TorchNeuron. On the Chinese side, DeepSeek is building the software case for Huawei's Ascend ecosystem — open-sourcing the framework layer so anyone can run frontier-class serving on non-Nvidia chips. The hardware war now has an open-source software front.

Published October 21, 2026. All performance figures are DeepSeek's self-reported production numbers, not independently verified. DHS order details per reporting verified October 21, 2026. The timing-connection framing is editorial analysis, not a claim about DeepSeek's release planning.