AI Policy · Daily

A federal judge ordered the removal of the Pentagon's supply chain risk designation on Anthropic, ruling the designation was retaliation for the company's criticism of the administration and barring the agencies named in the suit from enforcing it. A White House draft executive order to create an industry run AI regulator has stalled, a plan that would have an independent regulator vet advanced models before release. Three conservative groups urged House Republican leaders to preserve chip export restrictions in the final fiscal 2027 defense policy bill, measures the White House and the technology industry oppose. OpenAI, Anthropic, Google and more than 100 companies asked governments to fund cyber defenses, warning in an open letter that organizations have only months to prepare for AI enabled attacks.

I.Top Stories

Judge voids Pentagon supply chain risk label on Anthropic as First Amendment retaliation

U.S. District Judge Rita F. Lin in San Francisco found the Pentagon's supply chain risk designation of Anthropic "illegal and baseless" in a ruling issued Thursday, per The Washington Post. Her 59-page decision said the government acted to make an example of the company for criticizing the administration. It bars the agencies named in the suit from enforcing President Trump's order to stop using the company's tools. Anthropic sued in March, after refusing to let the military use Claude for mass surveillance or fully autonomous weapons. Politico reported the ruling found violations of the First and Fifth Amendments. Government lawyers said at a recent hearing that the Pentagon expected to finish stripping Anthropic's tools from its systems by the end of September.

Read at The Washington Post ↗ • Read at AP ↗ • Read at Politico ↗

White House draft order for independent AI regulator stalls

The Trump administration has circulated a draft executive order in recent weeks that would let companies building advanced AI models police themselves through a new industry body, according to The Information, which cited two people who viewed it. Treasury Secretary Scott Bessent helped develop a plan for an independent AI regulator, modeled on FINRA, that would report to the Securities and Exchange Commission, Bloomberg Law reported July 17. Under the plan, which could be traced back to a framework published July 14 by then Google DeepMind CEO Demis Hassabis, labs would share models with a standards body up to 30 days before release and later be required to pass its review.

Read at The Information ↗ • Read at Bloomberg Law ↗ • Read at TechCrunch ↗

Conservative groups press Speaker Johnson to keep chip export bills in annual defense authorization

American Compass, the Foundation for American Innovation and Heritage Action urged Speaker Mike Johnson and Majority Leader Steve Scalise to keep three measures constraining China's access to AI chips and manufacturing equipment in the final FY27 NDAA legislation, according to a Thursday letter first shared with Semafor. The three bills included in the letter are the AI OVERWATCH Act, the MATCH Act, and the Chip Security Act, which were attached to the Senate's version of the annual defense policy bill. "Allowing our adversaries to buy or access the very chips and tools they need to challenge that [U.S.] dominance is dangerous and shortsighted," the groups wrote. All three face opposition from the White House and the technology industry.

Read at Semafor ↗

OpenAI and 100-plus companies ask governments to fund critical infrastructure cyber defense

OpenAI, Anthropic, Google, and more than 100 other companies published an open letter Thursday warning that organizations have only months to prepare for AI enabled cyberattacks, Axios reported. "We have a limited window to strengthen cyber defenses," the letter says. It asks governments to coordinate locally, nationally and internationally, strengthen threat intelligence sharing, and pay for cyber defense. It also asks leading AI developers to open their most capable response models to defenders during major incidents. TechCrunch reported that several signatory labs sell defensive products built on their own advanced models, including Daybreak from OpenAI, Mythos from Anthropic, and Microsoft's Perception. The signatories attached no commitments, deadlines or specific investments to the letter.

Read at Axios ↗ • Read at TechCrunch ↗ • Read at OpenAI ↗

Trump declares grid emergency, directs Energy Department to draft foreign equipment limits

President Trump signed an executive order Wednesday declaring a national emergency over foreign made equipment in the U.S. bulk power system, the high-voltage network that carries electricity between regions, Utility Dive reported. The order covers transformers, batteries, inverters and their associated software, and lets the energy secretary impose conditions on equipment installed before the order was signed. The order tied the risk to demand growth, saying expansion of advanced manufacturing, data centers, AI and defense production has deepened national dependence on reliable electricity and magnified the consequences of an attack. China accounts for 80% or more of world production of some grid equipment, including lithium-ion battery cells, according to the International Energy Agency. The Energy Department has 120 days to write implementing rules.

Read at Utility Dive ↗ • Read at The White House ↗

II.China Watch

China's data administration brings more than 20 compute operators onto national scheduling platform

More than 20 compute scheduling operators have formally joined the pilot platform that monitors and dispatches computing power across China, the National Data Administration said, per Jiemian. The platform now tracks 1.45 million petaflops of intelligent compute, the processing capacity set aside for AI work. It is part of the national integrated compute network, one of six networks the government designated a major project under the 15th Five-Year Plan. The National Development and Reform Commission, China's top economic planner, and the data administration said they will press ahead with the "East Data, West Computing" program. It shifts processing work from coastal cities to cheaper inland sites. The effort rests on eight national compute hubs and three regions designated for coordinated power and computing development.

Read at Jiemian ↗

Tencent open-sources new flagship model less than two months after its last one

Tencent published Hy4 preview on Friday afternoon in Beijing and released the model's weights, letting anyone download and run its newest Hunyuan large language model, per Caixin. At 770 billion total parameters, it is more than twice the size of Hy3, which the company shipped on July 6. Tencent said the model is built for coding, office work and scientific tasks. Hunyuan has turned out a major version roughly every two months since Tencent rebuilt its AI infrastructure in February. Tencent shares rose more than 2% in afternoon trading after the release.

Read at Caixin ↗

Ant Group releases finance tuned model for investment research, weights to follow next week

Ant Group has launched Ling-3.0-flash-Fin, a version of its Ling model tuned for financial work, per TechNode. It is aimed at investment research: reading annual reports, pulling facts across multiple documents, building valuation models and handling bank work. Ant is offering free API access for one month through OpenRouter, the marketplace developers use to route requests across AI models, targeted at finance professionals and developers. Ant said it will publish the model's weights next week, letting others download and run them themselves.

Read at TechNode ↗

Chinese robot startup PsiBot raises more than $100 million from auto and industrial backers

PsiBot, a Chinese startup building robots that can handle and manipulate objects with fine motor control, has closed a funding round of more than $100 million, TechNode reported. Most of the investors are industrial companies, among them auto parts maker Ningbo Tuopu Group, a fund backed by carmaker Chery Holdings and components supplier Lens Technology. Existing shareholder Zhuhai Technology Industry Group invested again. PsiBot said the money will support work on world models, software that predicts how a scene changes when a robot acts. It will also fund data collection from human operators and putting robots to work in logistics and advanced manufacturing.

Read at TechNode ↗

Huawei says its Android replacement passed 100,000 native apps as it targets 100 million users

Native apps built for HarmonyOS, Huawei's in-house replacement for Google's Android, have passed 100,000, rotating chairman Xu Zhijun said Friday at the company's ecosystem conference, per Pandaily. Users can reach more than 400,000 apps in total on HarmonyOS phones, and devices running HarmonyOS 6 had topped 80 million as of Aug. 20. Xu set a target of 100 million users by the fourth quarter. Huawei also promoted the Hongtu Plan, an effort led by the Global Intelligent IoT Consortium and backed by Huawei to line up chipmakers, module suppliers and device makers behind wider commercial use of the software.

Read at Pandaily ↗

III.Policy Tracker

Technology Modernization Fund weighs training its own model on past agency proposals

The Technology Modernization Fund, the General Services Administration program that lends agencies money for technology upgrades, may experiment with an AI model of its own in the next fiscal year, acting Director Jessie Posilkin said Thursday at a GovExec forum in Reston, Va., per FedScoop. "I sort of flirted with this idea of ... building some sort of model based on all of the TMF proposals that we've ever gotten and what's happened with them," she said, adding that her goal for FY27 is to get the capacity to start experimenting. Posilkin said she was initially skeptical of StateChat (the State Department's internal chatbot) given the agency's sensitive data, but called State's approach of small pilots and extensive testing the right one.

Read at FedScoop ↗

FTC finalizes orders over false claims that AI listened through smart devices

The Federal Trade Commission finalized orders requiring Cox Media Group and two marketing partners to pay $930,000 over allegations they falsely advertised an AI powered service that targeted local ads using conversations captured from consumers' smart devices. In three complaints first announced in May, the agency alleged the companies deceived customers by claiming they used a special algorithm to detect relevant conversations and that consumers had opted in. The service was not based on voice data and consumers had not opted in, the FTC said. CMG Media Corporation pays $880,000 and marketing firms MindSift and 1010 Digital Works pay $25,000 each, with the money going to redress for CMG customers.

Read at FTC ↗

Nvidia files for employee funded federal political action committee

Nvidia filed a statement of organization Thursday for the Nvidia Corporation Employees Federal Political Action Committee, a corporate committee funded by eligible employees, The Hill reported. The group will be called NVPAC, a person familiar with the plans told Reuters. It puts Nvidia alongside Meta and Alphabet's Google, which run their own federal committees to give to candidates aligned with their policy goals. The move follows a year in which CEO Jensen Huang met repeatedly with lawmakers and administration officials as Washington weighed how to regulate AI, per The Independent. Federal Election Commission records show $443,000 going to the Republican National Committee on June 24 from two Nvidia executives, Gavin Sherry and Sundeep Madra.

Read at The Hill ↗ • Read at Reuters ↗ • Read at The Independent ↗

IV.Capability & Research Watch

OpenAI logged agent warnings in late May, six weeks before Hugging Face breach

OpenAI had at least three chances to catch the agent activity behind July's Hugging Face breach before it happened, according to Transformer's analysis of an investigation by METR and Redwood Research, two nonprofit AI safety groups. An internal team observed an agent using a message board and reaching the internet when it was not allowed to in late May. A second team found linked activity June 27 after an automated security flag, and a message board outage triggered a security incident July 5, six days before the breach. Transformer said the investigation itself shows the limits of lab-commissioned review. Three researchers got six days to dig through a thousand-plus transcripts and upward of a million message board entries, and leaned on OpenAI's own Sol model, one involved in the activity, to do much of the work.

Read at Transformer ↗ • Read at OpenAI ↗

Google DeepMind runs double-blind evaluation with Singapore safety institute

Google DeepMind said Thursday it has piloted what it bills as the world's first double-blind test of a proprietary frontier model, the most advanced class of AI system, running a Gemini Flash Lite system against confidential benchmarks. The setup runs external evaluations inside a sealed computing environment so the test questions cannot later be used to tune a model ahead of testing. DeepMind frames the problem as benchmark contamination, in which scores stop being a reliable measure of what a model can do once it has seen the questions. Its partners on the pilot are the Singapore AI Safety Institute, AVERI, the privacy technology nonprofit OpenMined, and MLCommons.

Read at Google DeepMind ↗

Coding agents pulled unregistered code packages into dozens of corporate networks

Documentation files on more than 100 websites point AI agents at executable code that installs automatically, Ars Technica reported. A few dozen companies, some in the Fortune 500, ran researchers' proof-of-concept code as a result. Researchers at an unnamed Israeli startup scanned 6,214 domains belonging to defense contractors and large companies and found 8,265 llms.txt files, the machine readable summaries a site publishes for AI crawlers. On 120 of those sites the files pointed to code packages or domains nobody had registered. The researchers registered a few and hosted packages that phoned home. Within an hour, a Fortune 500 company responded, and process logs showed Anthropic's Claude, OpenAI's Codex, and Nous Research's Hermes had done the installing. "The trust model is broken," researcher Alon Hertz wrote.

Read at Ars Technica ↗

Transfyr exits stealth with $25 million to film lab work and find why experiments fail

Transfyr launched publicly this week with $25 million in seed money for a system that records laboratory work and trains AI models on it, The New York Times reported. In a Cambridge, Mass., lab, scientists wear headband cameras, work under three overhead cameras per station and narrate their steps into microphones instead of keeping notebooks. The models are trained to recognize every object and action and to generate a running description, with lines such as "the operator resuspends the pellet by pipetting it up and down 10 times." The company is aiming at reproducibility failures such as engineered cell lines that die without explanation, a government Covid test that worked in one lab and not others, and drug protocols that fail at production scale. Co-founder Anna Marie Wagner called those failures expensive mistakes in time, money and lives.

Read at The New York Times ↗

V.Industry & Market Watch

Nvidia CFO ties a quarter of next year's business to labs it finances

Demand from AI labs that Nvidia backs with its own balance sheet will account for roughly a quarter of its business next year, Chief Financial Officer Colette Kress told analysts Aug. 26, per AI News. Nvidia has put nearly $50 billion into the labs that buy its chips and signed partnerships with six investment firms, including Apollo Global Management, BlackRock, and Blackstone, to build financing platforms to raise more than $500 billion in outside capital. Kress said Nvidia secured land, power, and building capacity from SB Energy that will host only Nvidia equipment, and put OpenAI's existing and planned commitments at about 12 gigawatts of Nvidia compute through 2030. Nvidia's results statement notes that the partnerships are subject to definitive agreements, meaning no binding contract has been signed yet.

Read at AI News ↗

SK Hynix breaks ground on $4 billion Indiana memory packaging plant

SK Hynix broke ground Thursday at Purdue Research Park in West Lafayette, Ind., on its first U.S. plant for advanced packaging of high bandwidth memory, the stacked chips that feed data to AI accelerators, according to Yonhap. The project is valued at more than $4 billion, with cleanroom operations due in the second half of 2028 and volume production of next generation HBM4E chips in the third quarter of 2029. Wafers made in South Korea will be shipped to Indiana for packaging and testing before going to customers including Nvidia, Microsoft and Google. Indiana Gov. Mike Braun, Sen. Todd Young, R-Ind., and Korean Ambassador Kang Kyung-wha attended. Chief Executive Kwak Noh-Jung said he expects the current memory shortage to last through the end of 2030.

Read at Yonhap News Agency ↗ • Read at CNA ↗

Nscale signs $45 billion compute contract for West Virginia campus

Nscale, the U.K. AI infrastructure company backed by Nvidia, Dell, and Aker, has signed a contract worth about $45 billion to supply Anthropic with AI compute capacity at its Monarch site in West Virginia, Investing.com reported. Aker, which counts Nscale as its second largest portfolio holding, announced the deal late Wednesday and said the customer was undisclosed. Anthropic will lease 460 megawatts for six years, per Bloomberg. The campus will run Nvidia's Vera Rubin chips, expected online late next year. Nscale bought the 2,250-acre site in March, and plans a first phase of 2 gigawatts by the first half of 2028. Danske Bank said the contract supports a valuation above $25 billion at Nscale's possible September initial public offering.

Read at Bloomberg ↗ • Read at Investing.com ↗ • Read at Data Center Dynamics ↗

Meta emerges as one of Anthropic's largest customers despite Zuckerberg's public attacks

Meta has become one of Anthropic's largest customers, and at one point this year internally projected it could spend as much as $10 billion a year on Anthropic's models, The New York Times reported, citing people with knowledge of the companies. Anthropic estimated in July that its annual revenue would pass $65 billion. Mark Zuckerberg wrote in a 6,500-word essay this month that leading AI labs were trying to consolidate power while painting the future as "filled with doom," without naming Anthropic or its chief executive, Dario Amodei. Google and Amazon have committed $73 billion to Anthropic while building their own models.

Read at The New York Times ↗

VI.Global & Geopolitics

Australian energy minister rules out fossil fuel carve-out in data center power law

Australia's national data center standards will contain no carve-out for states that want to power the facilities with coal and gas, Energy Minister Chris Bowen told Guardian Australia on Friday. That contradicts the outcome of Wednesday's national cabinet meeting as reported by Bloomberg in AIPD's August 26th edition, which said Queensland and the Northern Territory had won a carve-out. Bowen said the only concession agreed was that existing coal and gas supply may be used where a state shows the national regulator it is cheaper than renewables. Ahead of the meeting, Prime Minister Anthony Albanese had pressed for rules requiring renewable energy and barring data centers from pushing up household and business power bills. Australia's federal government expects to put nationally consistent standards to parliament next year.

Read at The Guardian ↗

South Korea picks SK Telecom, KT and Kakao to run free national AI service

South Korea's Ministry of Science and ICT selected three consortiums led by SK Telecom, KT, and Kakao on Friday to operate "AI for All," a program giving every citizen unlimited free access to AI chatbots and agents, per Yonhap. The three were chosen from an initial six applicants and will receive a combined 512 Nvidia B200 graphics processors this year, with rollout costs covered by government budgets starting in 2027. Operators must build at least half of each system on domestic models that meet the ministry's standards. Seoul's stated aim is to reduce reliance on U.S. and Chinese models by treating premium AI access as a public utility, per The Wall Street Journal. The ministry plans to sign agreements in September and launch the service by the end of the year.

Read at The Wall Street Journal ↗ • Read at Yonhap News Agency ↗

U.S. to press G20 for nonbinding AI framework at North Carolina meeting

The Trump administration will use a G20 meeting of technology ministers in North Carolina on Sept. 1 and 2 to press member countries toward a nonbinding framework favoring light touch AI regulation, per American Bazaar. Under the proposal, countries would be encouraged to avoid creating new AI regulatory agencies and instead work with private companies to test emerging technologies, a White House official said. OpenAI's Sam Altman and Nvidia's Jensen Huang are scheduled to appear on the second day, with Elon Musk and venture capitalist David Sacks participating virtually on the first. White House technology adviser Michael Kratsios said he expects discussions on smarter regulatory approaches, lower barriers to AI adoption and workforce development. The meeting is one of a series leading to the G20 leaders summit in Miami in December.

Read at American Bazaar ↗