Anthropic Lost on Its Safety Policy. Alibaba Won on Lobbying. The Pentagon's AI Rule Now Sorts Vendors, Not Risks

The decisions are a month apart and rest on different statutes, which is why they point in opposite directions without cancelling each other out. Anthropic's challenge was reviewed under the Federal Acquisition Supply Chain Security Act, where the court tested the legality of the Defense Department's written findings. Alibaba's challenge was reviewed under the Title 10 bar on contracting with entities that retain lobbyists for designated Chinese military companies, where the court tested whether the government can disqualify a contractor over who that contractor hires to lobby. Anthropic lost the designation and stayed excluded; Alibaba kept the designation and won only the contract-disqualification rule.

The majority opinion is the part that will shape procurement for years. Anthropic can encode and enforce restrictions into Claude, and the record showed those restrictions had already prevented Claude from performing tasks government users asked for. The panel held that continued integration of Claude into Defense Department systems presented a statutorily covered national-security risk, warning of 'overly constrained AI models shutting down unexpectedly and thus causing important military operations to fail.' The rule that falls out: a model's built-in refusal now counts as evidence of unreliability, and the statute's less-intrusive-measures test is where a defense policy gets judged.

Judge Lee's order did not touch the designation. It barred the government from denying Pentagon contracts over who lobbies for Alibaba — companies cannot be denied Pentagon work because they, their parent companies or their subsidiaries do business with firms that lobby for Alibaba. Reuters reported that Alibaba had fired dozens of its Washington lobbyists after the June listing and can now keep lobbyists who also represent Pentagon contractors, removing a practical obstacle both to rebuilding a Washington presence and to fighting the designation in court. The statute at issue is 10 U.S.C. § 4663. The provision Alibaba is still fighting is Section 1260H itself.

On the test the D.C. Circuit just wrote, an auditable open-weight model is the safer procurement choice. Claude's problem was that it might decline a lawful mission; a downloadable model has no policy layer to decline anything and can be inspected before it runs. That logic points straight at Qwen, and it points somewhere the Pentagon cannot easily follow: Alibaba is barred from selling to the Defense Department at all, and from June 30, 2027 from having its technology bought on its behalf through intermediaries.

The demand is already real and awkward. The Federal Register's search tool ran on a small Alibaba Qwen open-weight model for at least a day in September before officials pulled it, after Senator Mark Warner asked whether US data had been processed by 'Alibaba-controlled systems' and Representative John Moolenaar argued that no federal entity should use a Chinese AI model. Meanwhile Alibaba's open-weight models passed 3 billion global downloads in six months, ahead of Meta and Google, and Chinese models now take 41.4% of generative-model downloads on Hugging Face — about five points more than US models.

The compute layer explains why the designation now matters commercially. In the quarter ended July 26, 2026, NVIDIA booked $96.2 billion of revenue and $89.0 billion of data-center revenue, up 117% year over year — and kept data-center Hopper shipments into China below 1% of data-center revenue after taking a $0.4 billion charge on H200 inventory and absorbing a 25% tariff it cannot pass to customers. When the export door is that narrow, the buyer builds the chip: Alibaba's T-Head unit unveiled the Zhenwu V900 in September, pitched it as China's most powerful AI accelerator, and tied it to a 20-gigawatt data-center build by 2032 after shipping more than 560,000 Zhenwu chips to outside customers.

The June quarter shows the AI strategy in the numbers, and it is expensive. Revenue rose 8.3% year over year to CNY 268.2 billion, but gross margin fell to 38.0% from 44.9%, operating income dropped 44.9% to CNY 19.3 billion, and net income fell 75.4% to CNY 10.6 billion, with diluted earnings per share of CNY 3.68 against CNY 18.00 a year earlier.

The business investors are buying is still small. AI-related product revenue reached RMB 12.4 billion with a twelfth consecutive quarter of triple-digit growth, and cloud external revenue growth accelerated to 45% — together under 5% of group revenue. Meanwhile capital spending over the trailing twelve months ran to CNY 155.5 billion against CNY 78.0 billion of operating cash flow, so free cash flow swung to negative CNY 77.5 billion from positive CNY 78.2 billion a year earlier. On the trailing twelve months through June 30, 2026 the equity trades at 1.7 times sales and 11.4 times EBITDA.