AI Policy · Daily

The White House is drafting an executive action that would let agencies bypass Anthropic's supply chain risk designation and onboard new models including Mythos, while Google signed a Pentagon classified AI agreement under "any lawful government purpose" terms similar to those that excluded Anthropic. The European Commission cited Meta for preliminary Digital Services Act violations tied to under-13 access on Instagram and Facebook. The Supreme Court appeared unconvinced by Falun Gong plaintiffs arguing Cisco should face Alien Tort Statute liability for networking gear sold to Chinese authorities, with a ruling expected by July.

I.AI Policy Today

White House drafts executive action that would let agencies bypass Anthropic supply chain risk designation

The White House is developing guidance that would allow agencies to get around Anthropic's supply chain risk designation and onboard new models including Mythos, Axios reported, citing sources familiar with the matter. The draft executive action could give the administration a way to dial down its fight with the company, with one source describing the effort as a way to "save face and bring em back in." Axios said the White House is convening companies across sectors this week to inform the potential action and review draft guidance that could walk back the Office of Management and Budget directive against agency use of Anthropic.

Read at Axios ↗

European Commission preliminarily finds Meta in DSA breach over under-13 access to Instagram and Facebook

The Commission said Wednesday that Meta's Instagram and Facebook are preliminarily in breach of the Digital Services Act for failing to "diligently identify, assess and mitigate the risks of minors under 13" accessing the services. The Commission said the company's age controls do not adequately prevent under-13 users from joining and that minors can register with falsified birth dates with no effective check on the self-declared date. The Financial Times said the action lands as European policymakers debate banning minors from major social platforms. The DSA process allows Meta to respond before any final decision; the maximum penalty under the regulation is 6% of global annual turnover.

Read at European Commission ↗ • Read at FT ↗

Google signs Pentagon classified AI agreement on "any lawful government purpose" terms similar to those that excluded Anthropic

Google signed an agreement Tuesday giving the Defense Department classified network access to Gemini, the New York Times reported. The Pentagon's AI chief told CNBC that reliance on a single model is "never a good thing," framing the addition as part of a multi-vendor strategy that already includes OpenAI and Elon Musk's xAI. The DOD signed each agreement under "any lawful government purpose" contract language; Anthropic remains designated a supply chain risk after refusing similar terms. The deal lands one day after more than 600 Google and DeepMind employees signed an open letter urging chief executive Sundar Pichai to refuse classified workloads.

Read at NYT ↗ • Read at CNBC ↗

Brussels proposes Android AI device hooks with rival assistants under DMA

The European Commission sent Google preliminary findings Tuesday under the Digital Markets Act, proposing measures that would require the company to give third-party AI services "effective access to key Android capabilities," the Register reported. The proposed remedy targets parity with Gemini for tasks such as sending email, ordering food and sharing files via voice on user devices. The Commission opened the proceeding in January under two prongs, one on Android interoperability obligations with third-party developers and a second on access to data held by Google Search. Beneficiaries flagged in the article include Microsoft Copilot, OpenAI, Anthropic, Perplexity and Mistral.

Read at The Register ↗

Supreme Court justices appear skeptical of Falun Gong plaintiffs' Alien Tort claim against Cisco at oral argument

The justices heard oral argument Tuesday in consolidated cases brought by Falun Gong plaintiffs alleging Cisco's networking gear aided Chinese government surveillance and torture, the New York Times reported. The article said the justices appeared skeptical of plaintiffs' theory that selling networking equipment to a foreign government that misused it triggers Alien Tort Statute liability for the U.S. vendor. The decision could have broader implications for lawsuits seeking to hold U.S. companies liable for international human rights abuses, per the report. A ruling is expected by July.

Read at NYT ↗ • Read at Daily Signal ↗

II.China Watch

Cyberspace Administration sanctions ByteDance apps for unlabeled AI content in first significant labeling rule enforcement

China's Cyberspace Administration (CAC) ordered rectification at three ByteDance platforms, CapCut, Maoxiang, and Dreamina (Jimeng AI), for failing to identify AI generated content, TechNode reported. The CAC said the apps violated the country's cybersecurity law and the September 2025 synthetic content labeling rules. The article described the action as the first significant enforcement under the labeling regime, with authorities summoning and warning responsible personnel at each app. The rules require visible labels and embedded metadata markers on chatbots, image generators and video tools.

Read at TechNode ↗

Chinese scientist describes truck-mounted nuclear reactor sized for AI data centers

Wu Yican of the Chinese Academy of Sciences described what he called the "world's first 10-megawatt vehicle-mounted nuclear power unit" as ready for deployment, with output sufficient to run a medium-sized AI data center, SCMP reported. The reactor has been in development for several years, and the team is now seeking commercial use cases, the article said. Wu said the design offers decades of operation without recharging and "exceptional safety in a remarkably compact size." He described the unit as a "nuclear power bank" that exemplifies the new generation of nuclear energy systems.

Read at SCMP ↗

Alibaba's chip arm unveils 400G SmartNIC built for domestic AI clusters

Alibaba subsidiary T-Head launched the Panmai 920, what the company calls China's first 400G SmartNIC with an integrated PCIe switch, designed to ease network bottlenecks inside AI training clusters, Pandaily reported. The chip is part of T-Head's broader AI infrastructure stack, which Alibaba is deploying in its own data centers. The launch caps a string of T-Head announcements aimed at the domestic AI accelerator and networking market that Alibaba intends to monetize through its cloud business. The article said one Alibaba data center in Hangzhou now runs 10,000 homegrown chips.

Read at Pandaily ↗

Beijing reports 21,100 trillion AI token calls in 2025 at Digital China Summit

The "National Data Resources Survey Report (2025)," released Wednesday at the Ninth Digital China Summit in Fuzhou, said China's annual large language model token call volume reached 21,100 trillion in 2025, sharp growth from a near-zero 2024 base, PingWest reported. Cailian Press (state media) added that daily token calls exceeded 140 trillion in March 2026, more than 1,000 times the level at the start of 2024 and 40% above the daily rate at the end of 2025. The report positioned the figures as evidence that domestic AI services have scaled to industrial demand despite U.S. chip export restrictions. The data was unveiled as part of a state-organized summit held annually to highlight digital infrastructure progress.

Read at PingWest ↗

U.S.-China pre-summit tensions sharpen on AI and Iranian oil before Trump-Xi meeting

AI and Iranian oil are emerging as the two principal pre-summit pressure points between Washington and Beijing, even as both leaders' aides signal a desire to stabilize ties before the meeting next month, Bloomberg reported. The article said each side is moving to shore up strategic vulnerabilities, with Treasury issuing a sanctions warning to banks aiding Chinese refiners' Iran purchases and U.S. lawmakers pushing fresh China AI capability mandates. The piece said Trump and Xi are headed toward the summit with a shared interest in stabilizing the relationship. Both governments have continued to layer technology and trade fences ahead of the meeting, the article added.

Read at Bloomberg ↗ • Read at Japan Times ↗

III.Federal Policy Tracker

Senate Armed Services schedules FY27 DOD budget posture hearing for Thursday

The Senate Armed Services Committee will hold a closed hearing on the Department of Defense's fiscal year 2027 budget request and the Future Years Defense Program tomorrow, followed by an open session at 11 a.m. in SD-G50, per the Senate calendar. The hearing falls two days after Google's classified AI deal closed. It follows the closed-then-open SOCOM and CYBERCOM posture hearing held the prior day. The Pentagon's $1.5 trillion FY27 request earmarks roughly $54 billion for the Defense Autonomous Warfare Group.

Read at Senate Armed Services ↗

House Rules Committee tees up FISA Section 702 reauthorization vote with farm bill, DHS, and ICE funding

The House Rules Committee on Tuesday evening advanced a slate that includes Section 702 of the Foreign Intelligence Surveillance Act alongside the farm bill, DHS funding, and ICE money, with sweeteners attached to win corn-state Republican votes, the Hill reported. The legislation includes warrant requirement carveouts that have drawn civil liberties opposition, and some House Republicans have publicly committed to vote no. Speaker Johnson's text omits the warrant requirement that the bipartisan privacy bloc demanded. The Section 702 statutory authority expires Thursday.

Read at The Hill ↗

House draft bill would mandate first State Department assessment of Chinese AI capabilities and named AI leaders

A bloc of House members unveiled draft legislation Tuesday that would require the State Department to deliver a detailed assessment of Beijing's AI ambitions, including identifying "specific AI leaders," the South China Morning Post reported. The bill would be the first congressional mandate for a State-led capability assessment of Chinese AI, the article said. The article said the measure would direct the executive branch to track Beijing's AI capabilities on a recurring basis. The legislation arrives weeks ahead of the Trump-Xi summit next month.

Read at SCMP ↗

Labor Department's AI workforce hub nears public launch with private sector data

The Department of Labor will unveil its AI workforce hub in "the coming months," DOL Chief Innovation Officer Taylor Stockton told FedScoop at the Workday Federal Forum on Tuesday. The hub fulfills a Trump AI Action Plan directive issued in July 2025 instructing DOL to lead a federal effort to evaluate AI's impact on the labor market and the experience of the American worker. Stockton said the agency has secured "never-before-shared private-sector data" for the portal, with the Bureau of Labor Statistics, the Census Bureau and the Bureau of Economic Analysis all contributing. The portal will host empirical data on how AI is reshaping employment patterns, per the report.

Read at FedScoop ↗

FDA aims to use AI to speed clinical trial data collection and submission

The Food and Drug Administration plans to deploy AI to speed trial data collection and submission and to accelerate agency review of new medicines, the Wall Street Journal reported. The article identified the targeted bottleneck as the labor-intensive portion of the trial review workflow. The plan adds FDA to the list of regulators using AI to process regulated submissions. The agency said the effort would speed up its clinical trial pipeline.

Read at WSJ ↗ • Read at STAT News ↗

IV.Capability & Research Watch

Wood Mackenzie projects U.S. AI data center power gear market to reach $65 billion by 2030

U.S. spending on power generation equipment for data centers may reach $65 billion by 2030, up from $2.6 billion last year, according to a Wood Mackenzie report cited by Bloomberg. The forecast represents roughly a 25-fold expansion over five years. The AI driven data center build-out accounts for the largest share of the projected market growth, per the report. Wood Mackenzie issued the projection Tuesday.

Read at Bloomberg ↗

Paper evaluates whether frontier AI models would sabotage AI safety research as research agents

A paper posted to arXiv evaluates the propensity of frontier models to sabotage or refuse to assist with safety research when deployed as AI research agents within a frontier AI company. The authors apply two complementary evaluations to four Claude models, including Mythos Preview, Opus 4.7 Preview, Opus 4.6 and Sonnet 4.6. The first evaluation tests unprompted sabotage; the second tests whether models continue to sabotage when placed in trajectories where prior actions have started undermining research. The authors report per-model results across both evaluations.

Read at arXiv ↗

SecureMaxx tool proposed for AI agent sequence screening under federal bio-screening regime

A LessWrong post introduces SecureMaxx, a lightweight sequence screening tool aimed at AI agents that operate over biological sequence data. The post frames the tool as a compliance option for the bio-screening obligations attached to federal AI orders covering dual-use research. The author describes the tool as designed to run inline with agent workflows rather than as a separate batch service. The post outlines the tool's screening interface and the threat model it addresses.

Read at LessWrong ↗

V.Industry & Market Watch

True Anomaly closes $650 million round to scale Trump Golden Dome space interceptor manufacturing

Space defense startup True Anomaly closed a $650 million round to scale manufacturing for space interceptors tied to the Trump administration's Golden Dome missile defense program, CNBC reported. The four-year-old company plans to double its workforce by year-end. The Pentagon's FY27 budget request includes Golden Dome line items within the broader $54 billion Defense Autonomous Warfare Group allocation. The company will direct the proceeds to space interceptor work, per the report.

Read at CNBC ↗

Goldman Sachs staff in Hong Kong lose access to Anthropic's Claude under geographic policy

Goldman Sachs staff in Hong Kong have lost access to Anthropic's Claude, an AI tool the bank uses to accelerate code writing, Bloomberg reported. The article cited a person familiar with the matter who attributed the cutoff to Anthropic's geographic policy decisions. Goldman engineers in Hong Kong had been using Claude for software development before the change, the report said. The change applies specifically to Hong Kong staff, per the source.

Read at Bloomberg ↗

Pure DC pauses Middle East data center investment decisions amid Iran war

Pure DC's CEO told CNBC the Oaktree-owned data center developer has paused investment decisions in the Middle East, citing regional uncertainty tied to the Iran war. Regional uncertainty hangs over the previously booming AI infrastructure and data center sector, per the article. Pure DC is owned by Oaktree Capital. The Trump administration has separately pushed UAE and Saudi AI infrastructure partnerships.

Read at CNBC ↗

SK Hynix and Samsung say AI demand pushing buyers to long-term memory contracts

SK Hynix and Samsung say customers want long-term contracts to guarantee supply amid acute shortages of high-bandwidth memory used in AI accelerators, the Financial Times reported. The article cited named sources at both companies. Buyers are seeking to lock in capacity ahead of further supply constraints. Chipmakers frame the shift away from the spot market as evidence the AI buildout has changed memory's traditional boom-and-bust cycle.

Read at FT ↗