Two DeepSeek headlines landed within hours of each other on Monday, and read side by side they tell you pretty much everything about where the US-China AI race is heading: two separate stacks, built by two separate camps, with less and less overlap every month.

Let's start in Hangzhou. According to The Information, DeepSeek has gone well beyond being a Huawei customer and is now effectively a co-designer of Huawei's next-gen silicon. Huawei engineers have reportedly been stationed at DeepSeek's offices for an extended period, folding the AI lab's requests and suggestions directly into chip design decisions. In return, Huawei plans to hand DeepSeek most of the supply of its Ascend 960DT training chip when it launches next year. And on the software side, both companies have been using the open-source programming language TileLang as an alternative to Nvidia's CUDA, with DeepSeek's models now largely handled by TileLang running on Huawei chips.

None of this should come as a surprise. Back in September, when DeepSeek said it aimed to deploy over 160,000 Huawei AI chips, we wrote that DeepSeek's 160,000 Huawei Order Is A Real Threat To Nvidia In 2027. A few weeks later, on Sept 30, DeepSeek open-sourced an Ascend port of TileLang along with Ascend versions of its DeepGEMM and DeepEP libraries, explicitly crediting Huawei and touting a jointly optimized 128-chip Ascend 950 "supernode." Meanwhile Huawei pulled the launch of the Ascend 960DT forward to the first quarter of 2027, three quarters ahead of schedule.

What's new in Monday's report is the depth of the co-design. Nvidia spent two decades building its moat by tuning hardware and CUDA in lockstep with its biggest customers; China is now attempting the same trick with the lab and the chipmaker sitting, quite literally, in the same building. TileLang is still early and nobody serious thinks Ascend matches Blackwell (let alone Rubin) chip-for-chip, but the trajectory matters more than the snapshot. Which brings us back to a question we asked in August, and which nobody in Washington seems to have answered yet:

Meanwhile, In Washington...

The US response arrived the same morning, and it was, well, a procurement notice. The Department of Homeland Security published an order directing federal civilian agencies to remove DeepSeek products, and any third-party services built on the Chinese model, from their IT systems, and banned agencies from acquiring them going forward.

A few details worth flagging:

• The order (Exclusion and Removal Order #2026-01) was issued under the 2018 Federal Acquisition Supply Chain Security Act and was actually signed back on July 24; it only became public Monday via a Federal Register filing. So agencies have presumably been quietly scrubbing DeepSeek for three months already.
• It covers civilian executive branch agencies such as Treasury, State and Justice. The Pentagon and the intel community are carved out, mostly because the Pentagon already banned DeepSeek and High-Flyer AI via the FY2026 NDAA.
• DHS argues DeepSeek's models pose "multiple threat vectors" and that government use may result in the transfer of sensitive data to the PRC, pointing to China-based servers, PII collection at sign-up and possible user fingerprinting.
• The order doesn't lay out the underlying evidence and stops well short of an Entity List designation, so purely commercial use of DeepSeek is untouched... for now.

Step One Toward Banning All Chinese Open-Source AI?

Probably not step one... more like step four or five, and the important part is hiding in the fine print. The order reaches beyond DeepSeek's own app to third-party products "developed using DeepSeek implementations," which in plain English means any vendor tool built on a fine-tune or distillation of DeepSeek's open weights. And because federal acquisition rules bind prime contractors and subcontractors to these exclusion orders, the provenance of a model's weights has just become a procurement criterion for anyone selling software to Uncle Sam. That is a much bigger deal than banning a chatbot app nobody at the State Department was supposed to be using in the first place.

Look at the sequence so far: NASA, the Navy and roughly 17 states blocked DeepSeek on government devices in early 2025; the FY2026 NDAA booted it from the Pentagon and its contractors; now civilian agencies. And the pipeline behind it is already full:

• The Senate Armed Services Committee's version of the FY2027 defense bill would extend the Pentagon ban beyond DeepSeek to AI from ten other Chinese companies and their derivative models.
• In July, after Moonshot's Kimi K3 dropped (and sparked what we called another "DeepSeek moment" for US AI stocks), Axios reported the administration was reviving plans for de facto bans on Chinese open models, with Kimi and Alibaba's Qwen in the crosshairs via procurement rules, Entity List threats and security advisories. Scott Bessent added a sanctions threat over distillation a day later.
• And in August, The Information reported the US was mulling a self-regulatory organization for AI companies, which we flagged at the time for what it likely was:

If all of this feels familiar, it should: it's the TikTok playbook. TikTok was banned on federal devices in December 2022; the divest-or-ban law followed in April 2024. Government devices first, everyone else later.

That said, an outright ban on Chinese open weights would be a much harder lift than TikTok. The weights are already sitting on Hugging Face and on millions of hard drives, a weights-level ban runs into obvious First Amendment problems, and a lot of US startups (and, reportedly, at least one very large software company in Redmond) have built on Chinese open models precisely because they are cheap and nearly as good. As we laid out in Answering The "Trillion Dollar Question" back in June, the cost gap is enormous. So the more likely path is the de facto one: procurement bans that widen to every Chinese lab, then to contractors, then "advisories" that let enterprise compliance departments do the rest. No law required.

Decoupling, Courtesy Of Both Sides

The irony is that Beijing may get there first. Just last week China's commerce ministry floated export controls that would, for the first time, cover AI model weights and source code above certain performance thresholds (models trained with more than 10²⁶ FLOPs, over 2 trillion parameters, or ranking in the global top 10), even if it promises a fast-track license for open-source releases. Chinese AI commentators immediately took it to mean the next Kimi, GLM and DeepSeek flagships may never be open-sourced at all. So the free lunch of frontier-grade Chinese open weights could end from both directions at once.

And every step Washington takes to wall off DeepSeek pushes it deeper into Huawei's arms, which is exactly what Monday's Information report describes. The US is cutting DeepSeek out of its government while Beijing cuts DeepSeek off from Nvidia, and the result is two fully separate AI stacks: CUDA, Nvidia and closed US frontier models on one side; TileLang, Ascend and (maybe, for a while longer) open Chinese models on the other. Jensen Huang has been warning for years that export controls would hand Huawei the Chinese market on a silver platter. Turns out they also handed it a very motivated design partner.