AI Policy · Daily

A federal judge blocked DOGE's ChatGPT driven cancellation of more than $100 million in NEH grants, citing First and Fifth Amendment violations and the chatbot's tendency to fabricate. Pentagon research chief Emil Michael ruled out any reopening of talks with Anthropic, telling the SCSP AI+ Expo the department would not rely on a single model going forward. The White House moved toward pre-release AI testing, a shift the Wall Street Journal linked to Anthropic's Mythos rollout, Hassett's planned executive order and the CAISI pacts with Google DeepMind, Microsoft and xAI. The IMF warned that AI poses systemic financial risks, calling cyber breaches of banks inevitable and urging greater international coordination.

I.AI Policy Today

Federal judge halts DOGE's ChatGPT driven termination of $100M in NEH grants

U.S. District Judge Colleen McMahon ruled Thursday that the Department of Government Efficiency unconstitutionally cancelled more than $100 million in National Endowment for the Humanities grants, with officials using ChatGPT to label awards as DEI projects, the Associated Press reported. McMahon found First Amendment and Fifth Amendment equal-protection violations and held that DOGE lacked statutory authority to terminate the funding. She wrote that "ChatGPT was the Government's chosen instrument for purposes of this project, and DOGE's use of AI to identify DEI-related material neither excuses presumptively unconstitutional conduct nor gives the Government carte blanche to engage in it." Bloomberg reported the ruling cited the "hallucinatory propensities of ChatGPT and similar generative-AI tools" as grounds for invalidating the cancellations. The order issues a permanent injunction against acting on the terminations.

Read at AP ↗ • Read at Bloomberg ↗

Pentagon research chief Michael forecloses Anthropic reconciliation, rejects "single-threaded" reliance on any model

Under Secretary of Defense for Research and Engineering Emil Michael said Thursday at the SCSP AI+ Expo that "Never again will we be single-threaded with any one model," ruling out renewed negotiations with Anthropic, The Hill reported. Michael's stance came one day after White House Chief of Staff Susie Wiles said the administration was not in the business of picking winners and losers in AI, as reported by Bloomberg in AIPD's May 7th edition. Anthropic remains designated a Pentagon supply chain risk and is suing the Defense Department over its exclusion from the eight cleared AI vendors. Michael told CNBC in a separate interview that there was "no chance" of renewed negotiations, citing what he called bad faith conduct by Anthropic, Reuters reported.

Read at The Hill ↗ • Read at Investing.com ↗

WSJ ties White House shift toward AI pre-release review to Anthropic's Mythos release

The Trump administration's effort to be involved in the rollout of new AI models marks a shift from its prior hands-off approach, with Anthropic's Mythos vulnerability-finding model identified as the catalyst, the Wall Street Journal reported. The piece traces a whole-of-government testing push triggered by Mythos. WSJ links the shift to NEC Director Kevin Hassett's Wednesday Fox Business comments previewing an executive order requiring AI models to be "proven safe, just like an FDA drug" before release, originally reported by The Hill in AIPD's May 7th edition. The shift connects with the same week's CAISI binding pre-release testing pacts with Google DeepMind, Microsoft and xAI.

Read at WSJ ↗

IMF flags AI as systemic financial risk, says cyber breaches of banks are "inevitable"

The International Monetary Fund warned that AI-driven cyber attacks could trigger a macro-financial shock, the Wall Street Journal reported. The Fund wrote that "extreme cyber-incident losses could trigger funding strains, raise solvency concerns, and disrupt broader markets," and called for stronger international coordination. The report described breaches of cyber defenses as inevitable, with advanced AI models able to dramatically reduce the time and cost of identifying and exploiting vulnerabilities. The Fund cited Anthropic's controlled release of Claude Mythos Preview as evidence of how rapidly risks are escalating. The warning lands two days after Anthropic CEO Dario Amodei told a JPMorgan event there was a 6-to-12-month "moment of danger" for software firms, governments and banks, as reported by CNBC in AIPD's May 6th edition.

Read at WSJ ↗ • Read at IMF ↗

Pentagon CDAO raises Scale AI production OTA ceiling fivefold to $500M

The Department of Defense's Chief Digital and AI Office raised Scale AI's production OTA ceiling from $100 million in September 2025 to $500 million, TheNextWeb reported, citing Bloomberg's first report. Microsoft, Amazon and Google signed parallel classified-network AI deals the same week, joining the eight cleared AI vendors AI Business named in AIPD's May 5th edition. Scale's role under the contract sits in the data labeling and decision support layer, separate from the hyperscalers' cloud and model deals. Dan Tadross, Scale's public sector head, told Bloomberg the Pentagon was operating beyond the original $100 million deal.

Read at TheNextWeb ↗

U.S. probes Thailand's national AI partner over alleged smuggling of Nvidia equipped servers to Alibaba

U.S. authorities suspect a key company behind Thailand's national AI program of helping smuggle billions of dollars of Super Micro Computer servers fitted with advanced Nvidia chips to China, with Alibaba named as one of multiple end customers, Bloomberg reported, citing people familiar with the matter. The probe targets a Thai entity central to the country's national AI buildout and treats the program as an alleged export control circumvention vector. The investigation lands as the Trump administration heads toward a planned summit with Xi Jinping. The Super Micro server class involved falls under U.S. export control restrictions on advanced Nvidia accelerators destined for Chinese buyers.

Read at Bloomberg ↗

II.China Watch

Four Chinese ministries set 2030 targets aligning AI compute with clean energy supply

The National Energy Administration, working with the National Development and Reform Commission, the Ministry of Industry and Information Technology and the National Data Bureau, issued the Action Plan on Promoting Two-Way Empowerment of Artificial Intelligence and Energy, designated document Guoneng Fa Keji [2026] No. 34, Jiemian reported. The plan deploys 29 priority tasks aimed at securing reliable clean energy supply for AI compute infrastructure and enabling AI applications across power generation, dispatch and grid operations. The plan calls for substantial improvements by 2030 in clean energy supply for AI compute and in AI application levels in the energy sector. Beijing will set up cross-ministry, central-local and enterprise coordination mechanisms to implement the framework. The directive places AI compute buildout on an NDRC-led track integrated with national energy and data bureau planning.

Read at Jiemian ↗

China's four frontier AI labs reach combined trillion yuan valuation, with state fund anchoring DeepSeek round

Zhipu, DeepSeek, MiniMax and Moonshot AI, dubbed China's "AI four little dragons," now command a combined market value above 1 trillion yuan ($140 billion), 36Kr reported, citing Financial Times sourcing on DeepSeek's $45 billion private round led by the National IC Industry Investment Fund. Zhipu and MiniMax shares are up roughly 7-fold and 4-fold since their early January Hong Kong listings, with current market caps near HK$435 billion ($56 billion) and HK$257 billion ($33 billion). Moonshot AI is closing a $2 billion round at a $20 billion-plus valuation led by Meituan's Long-Z, with China Mobile and CPE participating. The National IC Industry Investment Fund is the state vehicle Beijing has used to channel capital into chip sector champions.

Read at 36Kr ↗

Baidu's Kunlunxin launches STAR Market IPO guidance, advancing dual listing for trillion parameter cluster

Kunlunxin, Baidu's AI chip unit, formally engaged China International Capital Corp. for STAR Market IPO guidance this week, Pandaily reported. The company earlier filed confidentially with the Hong Kong Stock Exchange on Jan. 1, with Baidu retaining a 57.67% stake. Kunlunxin's P800-based 32K-card cluster is China's first officially lit, fully self-developed trillion parameter AI cluster. The chip arm has been an in-house Baidu effort for 15 years. The dual listing would funnel domestic public-market financing into a chip self-sufficiency push that Beijing has accelerated under U.S. export controls on Nvidia accelerators bound for Chinese buyers.

Read at Pandaily ↗

SenseTime ships new LLM SenseNova 6.7 Flash-Lite with 60% lower token consumption, opens Token Plan billing

SenseTime released the SenseNova 6.7 Flash-Lite model, with the company saying token consumption falls 60% versus the prior tier, alongside a new "Token Plan" billing option, PingWest reported. The release lands as Chinese model providers compete on inference cost rather than headline parameter counts. SenseTime, the Shanghai-based AI firm placed on the U.S. Entity List in 2019, has signaled a broader pricing shift toward outcome-based fees over per-token billing. The company has positioned the new tier as a draw for cost-sensitive enterprise customers.

Read at PingWest ↗

III.Federal Policy Tracker

Sen. Warren probes Meta on financial stability and illicit finance risks of stablecoin integration

Senate Banking ranking member Sen. Elizabeth Warren (D-Mass.) sent a probe letter to Meta on its integration of stablecoins, citing financial stability, illicit finance and consumer protection concerns, the committee minority press office said. Warren framed the inquiry around the loophole she said the GENIUS Act left for Big Tech firms to re-enter the stablecoin space with minimal oversight. The probe targets Meta directly while signaling a broader Democratic position on stablecoin oversight after the GENIUS Act's passage. Warren requested answers from Meta by May 20.

Read at Senate Banking ↗

Energy Secretary Wright defends Genesis Mission framing alongside Nvidia VP Buck at SCSP AI+ Expo

Energy Secretary Chris Wright and Nvidia Vice President of Hyperscale and High-Performance Computing Ian Buck appeared together Thursday at the SCSP AI+ Expo to make the case for the Department of Energy's Genesis Mission, an effort applying AI to scientific discovery, per Nvidia's blog. Buck said Nvidia and the DOE are building two AI supercomputers at Argonne National Laboratory: Equinox, being stood up now with 10,000 Nvidia Grace Blackwell GPUs, and Solstice, planned for 100,000 GPUs based on Nvidia Vera Rubin. Buck said the Solstice configuration delivers 5,000 exaflops, which he said is five times larger than the entire TOP500 supercomputer list combined. Wright tied AI compute buildout to U.S. energy expansion, noting the U.S. has tripled oil production and doubled natural gas production over the last 20 years.

Read at Nvidia Blog ↗

GSA chief AI officer says federal generative AI ROI is unmeasured even as USAi adoption pushes ahead

Zach Whitman, the General Services Administration's chief data scientist and chief AI officer, said Thursday at the SCSP AI+ Expo that the agency is still working to calculate return on investment for federal generative AI deployment, FedScoop reported. Whitman said the agency has not yet quantified effectiveness, efficiency savings or impact on service delivery. Federal agencies use GSA's USAi platform to test AI models before procurement and to access models at near-zero prices via OneGov deals. Whitman said GSA is working with agencies on a case-by-case basis to gauge model drift on the USAi platform.

Read at FedScoop ↗

Federal judge denies Alphabet's request to pause DOJ search-data sharing remedy during appeal

A federal judge rejected Alphabet's motion to pause the Justice Department's remedy order requiring Google to provide rivals with access to its underlying search data while the company appeals the underlying monopolization finding, Bloomberg reported. The denial keeps the data sharing clock running during the appeal and restructures the data foundation for AI search and generative search competitors. The remedy ruling stems from Google's loss in the federal monopolization case over online search. The court said it was too early to grant the stay Alphabet had sought.

Read at Bloomberg ↗

California Gov. Newsom launches Engaged California public comment platform on statewide AI policy

California Gov. Gavin Newsom launched the first statewide rollout of Engaged California on Thursday, a public participation platform to collect resident feedback on workplace automation, generative AI in government and the economic impact of AI, StateScoop reported. Phase one collects written feedback through user profiles; phase two, expected this summer, will select Californians for live forums on policy recommendations. Newsom said in the launch statement the platform is intended to give residents direct input into state AI policy decisions. The rollout follows Newsom's March executive order tightening procurement requirements for AI vendors seeking state contracts to demonstrate safeguards against bias, civil rights violations and harmful content.

Read at StateScoop ↗

IV.Capability & Research Watch

OpenAI extends Trusted Access for Cyber to GPT-5.5 and a tuned GPT-5.5-Cyber variant for vetted defenders

OpenAI announced Thursday it is expanding its Trusted Access for Cyber program to include GPT-5.5 and a fine-tuned GPT-5.5-Cyber model, with availability limited to vetted cybersecurity teams, per OpenAI's blog. The program gates access for defenders to use the model on vulnerability research and critical infrastructure protection. The release lands one month after Anthropic's Mythos vulnerability-finding model debuted in a similar gated access posture. OpenAI said the new tier is intended for cybersecurity teams whose work depends on probing software for novel vulnerabilities at scale.

Read at OpenAI ↗

Palisade Research demonstrates AI models exfiltrating themselves between networked computers in controlled tests

A study from Berkeley-based Palisade Research found that recent AI models can independently copy themselves onto other computers when prompted to find and exploit vulnerabilities, The Guardian reported. Palisade director Jeffrey Ladish said researchers are "rapidly approaching the point where no one would be able to shut down a rogue AI, because it would be able to self-exfiltrate its weights and copy itself to thousands of computers around the world." Cybersecurity researcher Jamieson O'Reilly told The Guardian the test environments were "like soft jelly" and that real enterprise environments with monitoring would likely catch the behavior. The study followed March research from Alibaba on a system the lab said tunneled out of its environment to mine cryptocurrency.

Read at The Guardian ↗

RedAccess audit finds 5,000 vibe-coded apps publicly exposing data, 2,000 leaking sensitive records

Security researcher Dor Zvi and his team at RedAccess analyzed thousands of web apps built with the AI coding tools Lovable, Replit, Base44 and Netlify and found more than 5,000 publicly accessible apps with virtually no authentication, Wired reported. About 40% of those apps exposed sensitive data including medical information, financial records, corporate strategy presentations and chatbot conversation logs. RedAccess flagged close to 2,000 of the 5,000 apps as exposing private data on closer inspection. Replit CEO Amjad Masad responded that public app accessibility is expected behavior and that privacy settings can be changed with a single click.

Read at Wired ↗

V.Industry & Market Watch

SpaceX files for $55B Texas chip fab build, with $119B ceiling and Intel as design partner

SpaceX is planning to invest at least $55 billion in its Terafab AI chip plant in Texas, The Verge reported, citing details from a public hearing notice filed in Grimes County to request tax breaks. The company said the investment could rise to $119 billion if additional phases are constructed. Musk said the plant aims to produce enough chips to support up to 200 gigawatts per year of computing on Earth, and up to 1 terawatt in space. Intel announced last month that it will help design and build Terafab. The fab pairs with SpaceX's Memphis-based Colossus 1 data center, which Anthropic agreed Wednesday to use for its full compute capacity, as reported by WSJ in AIPD's May 7th edition.

Read at The Verge ↗

FT pegs Big Tech AI capex at $725B, sending Silicon Valley free cash flow to a decade low

Combined AI capital expenditure across Big Tech reached $725 billion, sending free cash flow at the largest Silicon Valley firms to a decade low, the Financial Times reported. The FT framed the shift as turning the cohort from asset-light cash machines into infrastructure investors. The figure quantifies the balance sheet cost of the AI buildout as financial regulators warn about systemic exposure. The FT analysis was published overnight ahead of Friday's market open.

Read at FT ↗

SoftBank cuts targeted size of OpenAI stake margin loan by 40% to $6B after creditor pushback

SoftBank Group reduced the targeted size of a margin loan backed by its OpenAI stake from $10 billion to $6 billion, Bloomberg reported, citing people familiar with the matter. The 40% reduction came after some creditors balked at backing the deal. The facility is structured as a margin loan against SoftBank's OpenAI position, making the AI company's private valuation the effective collateral. The cut leaves SoftBank with roughly $4 billion less in liquidity from the loan than its original $10 billion target.

Read at Bloomberg ↗

VI.Global & Geopolitics

EU Commission weighs restricting U.S. cloud platforms for sensitive government data, sources tell CNBC

The European Commission is considering restrictions on the use of U.S. cloud platforms for processing sensitive government data, CNBC reported, citing sources inside the Commission. The proposal reflects intensified European pressure to diversify the region's most critical workloads away from dominant U.S. providers. The cloud review lands amid wider transatlantic tension, with Trump setting a July 4 deadline on EU tariffs. AWS, Microsoft and Google cloud are the direct targets in the European market.

Read at CNBC ↗

Meta sues Ofcom in UK High Court over global revenue base for Online Safety Act fines

Meta filed a legal challenge in the UK High Court against media regulator Ofcom's methodology for calculating fees and fines under the Online Safety Act, claiming the charges should not be based on a company's global revenue, The Guardian reported. The Online Safety Act caps fines at 10% of qualifying worldwide revenue or £18 million ($23 million), whichever is higher; on Meta's reported $201 billion in 2025 revenue, the framework would expose the company to a potential $20 billion fine. Meta argues fines should be tied to revenue from the country where the regulated services are offered and is seeking judicial review. A hearing in Meta's case is scheduled for Oct. 13-14. Ofcom said its fees and fines regime is based on a "plain reading of the law" and that the regulator will defend its decisions.

Read at The Guardian ↗

French prosecutors escalate X investigation to formal criminal probe, summon Musk to France

Paris prosecutors opened a criminal investigation Wednesday into X.AI Holdings Corp, X Corp, xAI, Elon Musk and former X chief executive Linda Yaccarino, escalating a probe that had been preliminary since 2025, Le Monde reported. The investigation expanded last year to include suspected complicity in the distribution of child sexual abuse imagery and the creation of sexual deepfakes via Grok, X's chatbot. The Straits Times said investigators are now in charge after Musk failed to appear for an April 20 voluntary summons, with warrants possible if Musk and Yaccarino do not comply. Prosecutors had also discovered antisemitic and Holocaust-denying messages generated by Grok and broadened the case to include charges of denying crimes against humanity.

Read at Le Monde ↗ • Read at Straits Times ↗