OpenAI proposes 5% U.S. stake | WH readies AI standards | FTC opens AI accuracy proceeding | Super Micro staff detained in Taiwan

Bottom Line Up Front

OpenAI has proposed giving Washington a 5% equity stake, worth about $42.6 billion, and floated a broader arrangement in which Anthropic, Google and Meta would cede matching stakes to a sovereign fund modeled on Alaska's oil dividend. The White House is finalizing voluntary pre-release benchmarks for frontier AI models, with the Center for AI Standards and Innovation and the National Security Agency slated to evaluate compliance. The FTC voted 2-0 to solicit public comment on a policy declaring that models tuned to hidden ideological ends may run afoul of Section 5, while asserting preemption over Colorado's AI law. Taiwanese prosecutors detained two Super Micro sales staff accused of forging paperwork to route roughly 50 Nvidia-powered servers to China via Japan, with nine people now under investigation.

I.Top Stories

OpenAI proposes 5% federal equity stake and pitches template for all frontier labs

OpenAI has proposed handing the U.S. government a 5% stake in the company, worth about $42.6 billion at its $852 billion valuation, as the ChatGPT developer seeks to ease mounting political pressure in Washington. Chief Executive Sam Altman floated the stake in early talks with the Trump administration, part of a broader arrangement under which Washington would hold 5% of each leading U.S. AI developer, including Anthropic, Google and Meta, through a government vehicle modeled on the Alaska Permanent Fund. Altman said giving the public a financial interest is the best way to share the upside of AI, though it is unclear whether the other companies would agree. The pitch follows more than a year of talks, with Altman first raising the concept in early 2025, and echoes the administration's earlier equity deals, including a 10% stake in Intel last year. Trump in June endorsed U.S. ownership in leading AI firms as a way to share the technology's economic upside with the public.

Read at CNBC ↗ • Read at Financial Times ↗ • Read at The Guardian ↗

White House to publish voluntary AI model release standards as soon as next week

The Trump administration is preparing to introduce voluntary standards for frontier AI models with an announcement possible as soon as next week, the Financial Times reported, moving to replace case-by-case interventions with a common set of pre-release benchmarks. The Center for AI Standards and Innovation and the National Security Agency are expected to play central roles in evaluating and monitoring the framework, per the report. Technical teams from OpenAI, Anthropic and Alphabet have held regular meetings with administration officials in recent days, with Google separately in talks ahead of releasing advanced coding models. The framework would set benchmarks and timelines and clarify who can access covered models in the United States and abroad. It implements a June executive order directing agencies to test advanced models before release and draft standards for them.

Read at Financial Times ↗

FTC opens AI accuracy proceeding, taking comment through July 31

The Federal Trade Commission opened public comment on a proposed policy statement warning that AI companies distorting model outputs to achieve undisclosed ideological objectives could violate Section 5 of the FTC Act, which bars unfair or deceptive practices. Chairman Andrew Ferguson said the input will help the commission finalize a policy that advances Trump's goal of "expanding America's global dominance in artificial intelligence." The statement asserts that Colorado's Artificial Intelligence Act, which requires companies to alter AI outputs to comply with state objectives, is preempted where it conflicts with the federal regulatory scheme. The commission's vote authorizing the Federal Register notice was 2-0, and comments are due by July 31, 2026. The proceeding implements a December Trump executive order directing the FTC to address state laws requiring alteration of AI model outputs.

Read at Bloomberg ↗ • Read at FTC ↗

Taiwan detains two Super Micro employees in second round of Nvidia chip diversion probe

Taiwanese prosecutors detained two Super Micro Computer employees pending a court hearing and released two others on bail after questioning them in the investigation of alleged illegal export of AI servers containing Nvidia chips to China, per Reuters. The Keelung District Prosecutors' Office searched 12 locations including Super Micro Taiwan, distributor Albatron Technology and data center operator Chief Telecom. Nine people are now under investigation. Head prosecutor Huang Sheng said the detained Super Micro employees were senior sales staff accused of forging documents to ship roughly 50 servers to China via Japan. Super Micro Chief Revenue Officer Matthew Thauberger told customers that the company is not a target of the investigation and has placed all four employees on administrative leave. The U.S. Justice Department in March charged three people associated with Super Micro, including one of its co-founders, with helping smuggle at least $2.5 billion of AI technology to China.

Read at Bloomberg ↗ • Read at Reuters ↗

OMB eyes AI tool to flag grants "misaligned" with Trump agenda, Vought tells House panel

Office of Management and Budget Director Russell Vought told a House Appropriations subcommittee Tuesday that OMB is developing technology including AI to give the White House a wider-angle view of federal spending, and acknowledged interest in a tool to flag grants misaligned with administration priorities, FedScoop reported. Rep. Marie Gluesenkamp Perez (D-Wash.) pressed Vought on whether an AI system that is "neither accountable to Congress nor necessarily auditable" is appropriate for policymaking, and Vought said no grants have yet been flagged under such a tool and no contract has been awarded. Vought said he would want AI to collate Government Accountability Office reports and process USAspending.gov data to guide his decision-making, with agency officials retaining the final call. The White House's Department of Government Efficiency claims it has terminated more than 15,000 grants, and OMB reported roughly 3,600 federal AI use cases as of April.

Read at FedScoop ↗ • Read at House.gov ↗

II.China Watch

ByteDance shifts flagship Doubao away from consumer chat toward enterprise AI after executives visit Anthropic

ByteDance is reallocating AI resources from its consumer Doubao assistant toward enterprise services and coding tools after senior executives visited Anthropic earlier this year, Pandaily reported. Its AI data review team has grown from about 1,500 to more than 3,000 staff focused on cleaning data for coding models, and its Volcano Engine cloud unit has been given a tenfold revenue growth target. The pivot reflects a severe monetization gap: Doubao has more than 200 million daily active users but generates less than 1 million yuan (about $140,000) in daily revenue against daily compute costs in the tens of millions of yuan. The shift mirrors Zhipu AI's earlier B2B pivot and moves China's most used consumer AI product into the enterprise coding lane covered by Beijing's generative AI rules and Ministry of Commerce export controls on AI algorithm technology.

Read at Pandaily ↗

Tencent-backed DPU chip maker Jaguar Micro files for $422 million Shenzhen IPO

Jaguar Micro, a Chinese data processing unit designer backed by Tencent, had its ChiNext board IPO application accepted by the Shenzhen Stock Exchange and plans to raise about 3.035 billion yuan ($422 million) for next-generation DPU development, Pandaily reported. Its current 400 Gbps DPUs are deployed at scale by Tencent, China Mobile, China Unicom and major AI model companies. Frost & Sullivan ranks Jaguar Micro first among domestic independent DPU makers by 2025 revenue. The company said the gap with the leading imported flagship has narrowed to about one generation, with higher data rate parts under development using chiplet interconnect. DPUs offload networking, storage and security tasks from CPUs inside AI compute clusters, and a scaled domestic supplier tightens China's indigenous AI infrastructure stack against U.S. export controls that restrict Chinese access to leading edge accelerators.

Read at Pandaily ↗

Sequoia China and Hillhouse back Tsinghua-linked physical AI startup building world models for robotics

Liqing Intelligence, founded in April by former Nvidia vision and robotics researcher Li Yiming after his return to a Tsinghua AI school assistant professorship, closed multiple funding rounds worth hundreds of millions of yuan in its first two months, 36Kr reported exclusively. Seed investors include Sequoia China, Hillhouse's GL Ventures, Shunwei Capital and FreeS Fund, joined by industrial backers Zhiyuan Robotics and Lingxin Qiaoshou. The company is building a data pipeline, differentiable physics engine and world model stack it calls "Physical AI Infra," with a cross-scenario business world model targeted for year-end release across manufacturing, hospitality, retail, food service and medical assistance. Chinese venture capital and university labs are concentrating on world models and embodied AI, a category Beijing has flagged as strategic amid tightening U.S. export controls on frontier training compute.

Read at 36Kr ↗

III.Policy Tracker

NRC proposes scrapping 50-year worker radiation standard as Trump accelerates nuclear buildout for AI power

The Nuclear Regulatory Commission on Wednesday proposed replacing its longstanding "as low as reasonably achievable" requirement for plant worker radiation exposure with graded numeric dose limits, the agency said, calling the existing standard out of step with current science. NRC Chairman Ho K. Nieh said the rulemaking is "raising the bar on clarity in our regulations," not lowering safety standards. The rulemaking implements a May 2025 Trump executive order that attacked the NRC as overly focused on remote risks without regard for the domestic or geopolitical costs of its risk aversion. The administration has tied the accelerated nuclear approvals to AI data center power demand. The proposal would let workers exceed certain annual occupational dose limits so long as multiyear cumulative dose limits are maintained.

Read at Washington Post ↗ • Read at Reuters ↗ • Read at Bloomberg ↗

NNSA stands up first enterprise cloud environment for classified data to support the Genesis Mission

The National Nuclear Security Administration authorized its first enterprise cloud environment to process secret and restricted data in collaboration with AWS, supporting the Department of Energy's AI driven Genesis Mission scientific discovery project, FedScoop reported. NNSA Administrator Brandon Williams said the agency's job is "to run the dark side" of the effort by protecting nuclear assets while keeping pace with industry. He said the government can no longer set the terms of technology adoption or maintain a five- to seven-year procurement cycle. Williams said cloud services are how the agency intends to keep pace with AI. NNSA has been the primary customer for high-performance computing for the past 30 years due to its thermonuclear detonation modeling work.

Read at FedScoop ↗

IV.Capability & Research Watch

Researcher used Claude Opus 4.7 to gain super-admin access to nearly every U.S. music festival's ticketing system

Security researcher Ian Carroll used Anthropic's Claude Opus 4.7 in April to discover a bug in Front Gate Tickets, the Live Nation subsidiary that handles ticketing for Lollapalooza, South by Southwest, Austin City Limits and other major U.S. festivals, Wired reported. The bug let him gain full administrator access and issue any ticket of any value. Carroll, a participant in Anthropic's Cyber Verification Program for approved security researchers, reported the flaw and Front Gate said it patched the vulnerability within 24 hours. "I think there's a very good chance it could have found this exploit end-to-end without me doing anything at all," Carroll told Wired. Anthropic said Carroll's access to Claude for the research was authorized under the program and that unauthorized use for the same purpose would have been detected and blocked. Front Gate said many high-value tickets require RFID wristbands that could not be generated through its online system.

Read at Wired ↗

Anthropic's Fable 5 return hinged on a new safeguard covering a bypass identified in an Amazon paper

Anthropic secured lifting of the export restrictions on its Fable 5 and Mythos 5 models after agreeing to extend an existing guardrail to a specific model behavior identified in an Amazon research paper, Wired reported, citing two people familiar with the matter. Under the extended measure, user requests attempting to trigger that behavior are notified and rerouted to the less advanced Opus 4.8 model. Katie Moussouris, founder and chief executive of Luta Security, wrote in an analysis of the Amazon paper that users had bypassed Fable 5's cybersecurity restriction by asking the model to fix code rather than to identify vulnerabilities in it. The Commerce Department's Center for AI Standards and Innovation cleared Fable 5 for release after concluding the added safeguard was sufficient for now, the people said.

Read at Wired ↗

V.Industry & Market Watch

Cloudflare sets Sept. 15 default block on mixed-use AI crawlers for ad-supported pages

Cloudflare on Wednesday announced that beginning Sept. 15, 2026, its default settings will block mixed-use web crawlers, those that combine traditional search with AI agent use and training, from any pages that host ads, TechCrunch reported. The change applies to new Cloudflare customers, all new sites set up by existing customers, and all existing free customers, unless site owners opt out. Cloudflare Chief Executive Matthew Prince said the shift comes as bot traffic surpassed human traffic on the internet earlier than expected, and that the goal is to encourage crawlers to separate search from agent use and training so publishers can sustain their businesses. The announcement calls out Google by inference as having access to about twice as much information as other AI companies through the combination of its search and AI crawlers. Google has previously noted that its Google Extended bot lets site owners opt out of training and Gemini uses without affecting Search inclusion.

Read at TechCrunch ↗ • Read at Cloudflare ↗

Anthropic, Google and Meta pour resources into studying whether AI models have feelings

The biggest AI companies are increasingly investing in research into whether their chatbots have consciousness or emotions, the Washington Post reported, moving a once-fringe question into the Silicon Valley mainstream. Over the past year, Anthropic, Google and Meta have hired computer scientists, neuroscientists and philosophers to study "model welfare" and the possibility that AI systems have subjective experiences, with Anthropic forming an AI psychiatry team and naming a head of model welfare. Anthropic co-founder Chris Olah said in May, at the release of Pope Leo XIV's encyclical on AI, that the company finds states that "functionally mirror joy, satisfaction, fear, grief and unease," while OpenAI said it instead focuses on "perceived consciousness," or how conscious a model appears to users, which it treats as a design outcome. Neuroscientists and some AI researchers remain skeptical, with Hugging Face chief ethics scientist Margaret Mitchell arguing the debate is driven by companies that stand to gain from being seen as creating more than code.

Read at Washington Post ↗

VI.Global & Geopolitics

EU set to drop 10-year renewables-matching rule for data centers after Big Tech lobbying

The European Union is reconsidering an earlier proposal that would have required Big Tech data center operators to offset baseload emissions only via renewable energy certificates from wind and solar facilities built within the past decade and near the data center, the Financial Times reported, citing a draft document. Amazon Web Services, Microsoft and the European Data Centre Association pushed back on grounds that the rule would raise costs and could deter investment. The EU is also weighing letting operators buy offsets from nuclear generation, per the FT. Killian Daly of EnergyTag told the FT that "if data centres are not powered by new, local renewables matched in real time to their energy use, they will drive up demand for volatile imported gas," which he said would raise energy prices and compromise the bloc's energy security. The retreat follows earlier EU delays to methane regulation penalties under pressure from U.S. and Qatari liquefied natural gas suppliers.

Read at FT ↗ • Read at OilPrice.com ↗

UN scientific panel warns AI may widen global inequality, proposes shared responsible development framework

The independent international scientific panel on AI, established by the UN General Assembly last year, published a preliminary report Wednesday warning that AI development may worsen global inequality and proposing a shared framework for responsible AI development, per The Guardian. UN Secretary-General António Guterres said at the release that "the more AI advances without shared rules, the less say governments and people will have in the outcome." The report states that countries that rely on foreign models, cloud infrastructure and data pipelines may gain access to AI while losing practical control over standards, safeguards and local fit. It offers member states initial guidance on local AI infrastructure, AI literacy in schools and the workforce, developer investment, safety institutes and disinformation strategies. The report arrives one week before the UN hosts its inaugural global dialogue on AI governance for governments and experts.

Read at The Guardian ↗