The DeepSeek purge costs DeepSeek nothing: 2,255 federal AI awards now carry a model-provenance bill

The Department of Homeland Security published the Secretary's Exclusion and Removal Order #2026-01 on Oct 20, 2026 (91 FR 66346, Docket DHS-2026-0628), acting on a Federal Acquisition Security Council recommendation under the Federal Acquisition Supply Chain Security Act of 2018. It excludes information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence from federal civilian procurement, and removes DeepSeek and affiliate products from agency information technology. It also reaches third-party products developed using DeepSeek implementations, which is the clause that actually changes contractor behaviour.

DeepSeek's weights are free, so there is no invoice to cancel and no vendor to invoice. Enforcement runs through the acquisition rules instead. Under FAR 52.204-30, a contractor checks SAM.gov at least every three months while a contract runs, reports a covered article to the contracting officer within three business days, and follows up on mitigation within ten. The duty is limited to a reasonable inquiry into information already in the contractor's possession, and no internal or third-party audit is required — but the clause flows down to subcontractors, and the order reaches firms doing business with prime contractors.

CSIS's review of federal AI procurement finds $4.10B in net obligations across 2,255 AI-related awards from FY2019 through FY2025, with $1.16B obligated in FY2025 alone. Demand is dominated by the Pentagon, which takes 83% of obligations and about 64% of contract counts. Concentration is severe: 1,250 vendors won work, 66.2% of them won exactly one award, the median award runs about $150,000 a year, and the top three recipients take 30.4% of the dollars.

Governance is the gap. In CSIS's sample of 75 AI solicitation notices, more than half contain no clear governance language; explicit monitoring appears in 4 of 75, incident reporting in 3, and audit access or red teaming in fewer than 10% each — even as generative AI-tagged awards doubled from 56 in FY2024 to 118 in FY2025. GAO reached the same conclusion from the agency side: after reviewing 13 AI acquisitions at Defense, Homeland Security, GSA and Veterans Affairs, it issued four recommendations that every agency concurred to, requiring systematic collection of lessons learned, including contract clauses, because current policies do not require it.

No Chinese model vendor had meaningful federal revenue to lose. The exposure sits with the services firms whose revenue is federal. Booz Allen Hamilton generated roughly 98% of FY2026 revenue from U.S. government contracts, split $7,969M National Security (71%) and $3,248M Civil and Commercial (29%), on $11,217M of total revenue.

Accenture's fiscal 2026 10-K spells out why that is not free money: Defense Contract Audit Agency audits, unallowable costs that cannot be rebilled, False Claims Act exposure and suspension or debarment, alongside a Justice Department investigation into Accenture Federal Services' voluntary disclosure. In that frame a failed provenance attestation is not a paperwork problem — it carries a margin haircut.

NVIDIA's 10-Q for the quarter ended July 26, 2026 is the most consequential document in this chain, and not for the reason investors expect. Revenue was $96,221M, up 106% year over year, with data center revenue of $89,023M, up 117%. H200 shipments to China under U.S. licences were under 1% of data center revenue, the company booked a $0.4B charge on excess H200 inventory and purchase obligations in the first half, and it now calls itself foreclosed from competing in China's data center compute market. Fiscal 2026 revenue reached $215.9B, up 65%.

What matters more is a risk factor. The filing warns that restrictions on applications built on foundation models originating in China — DeepSeek, Qwen or Kimi — could have a material impact on the business, and notes that 38% of second-quarter revenue came from customers headquartered outside the United States. NVIDIA is exposed to inference volume, not to federal procurement, and the open-weight models it has spent 2026 tuning silicon for are the marginal demand.

Alphabet grew revenue 24% to $119,796M in the June 2026 quarter; Microsoft closed fiscal 2026 with $331,839M of revenue, up 18%, and $214.4B of Microsoft Cloud revenue, up 27%. Both answer a federal procurement question about where a model came from in one sentence. The open-weight Chinese stack cannot answer it at any price, which is exactly the feature now being bought.

Palantir is the purest read on the trade: June-quarter revenue of $1,935M, up 93%, with U.S. government revenue of $809M up 90% and a 47.1% operating margin, on a franchise that with Scale AI took about $597M — roughly 15% of all federal AI obligations — from fewer than 1% of contracts. Federal AI buying is already an oligopoly, and the order puts a compliance moat around it.

Alibaba's June-quarter revenue was CNY 267,809M, up 8.1%, and the Qwen family behind it took about 2,045M Hugging Face downloads in 2026 with 151,448 derivative repositories. Not a yuan of that is federal, and none of it is at risk. On Oct 20 China's commerce ministry said the order disregards the consensus reached by Xi and Trump, and officials warned of possible retaliation — the first sign that model provenance is now a bargaining chip.

The mechanism that matters is the list. H.R. 10670, introduced Oct 1, would bar Chinese AI models from government-issued devices and software systems and prohibit their procurement, and the Federal Acquisition Security Council already has the authority to recommend the exclusions DHS just executed. If Qwen, GLM or Kimi are named next, the friction moves out of one country's federal budget and into allied procurement standards and commercial enterprise rules.

The thesis fails if CISA grants broad exceptions, if a follow-up order narrows the derivative clause to exclude distillation and synthetic training data, or if agencies treat the purge as a one-time sweep with no attestation trail. It also fails in the opposite direction: if the rule becomes a template copied by allies and commercial buyers, the value at stake stops being a $4.10B federal budget line and becomes a global standard for what may ship into regulated systems.

The practical takeaway for the next two quarters: watch which contractors actually book model-provenance cost, and watch whether a second exclusion order names Qwen or GLM. The first tells you whether the cost line is real; the second tells you whether this is a tariff or a template.