OpenAI halts model testing | Data center ads flood battlegrounds | China defies AI chip restrictions

A Note from the Editors

We're so back! Today's newsletter will contain news from the last week to catch you up.

Bottom Line Up Front

OpenAI paused reinforcement learning training for two weeks and left its largest planned frontier run on hold while it overhauls its research and training systems. Candidates in both parties are running midterm ads against data centers in Texas, Ohio and House battlegrounds in Michigan, Wisconsin and Pennsylvania. OpenAI asked California to amend SB 53, the AI safety law it opposed last year, to require monitoring of frontier models in training and stronger cybersecurity protections. California, Colorado, Kentucky and New Jersey opened a child safety trial against Meta in Oakland.

I.Top Stories

OpenAI pauses reinforcement learning training for two weeks, holds largest planned frontier run

The company said Aug. 18 it had slowed the pace of scaling while it overhauls its research and training systems. The measures include a two week pause in reinforcement learning training on its latest models intended for deployment, and its largest planned frontier reinforcement learning run remains on hold while smaller scale training and evaluations continue. OpenAI wrote that it now requires stronger evidence of aligned behavior throughout all of training, meaning evidence that a system remains responsive to human oversight and behaves as intended. It said preliminary evidence indicates its upcoming Astra model may meet the critical cybersecurity capability threshold in its Preparedness Framework, the internal standard it uses to decide when a capability demands new safeguards. Mia Glaese, who leads safety and alignment work at the company, told the tech blog Sources News that the company is far from back to normal.

Read at The Guardian ↗ • Read at OpenAI ↗

Data center opposition turns into midterm ads in Texas, Ohio and three House battlegrounds

With less than three months until the election, candidates in both parties are campaigning against the server farms powering the AI buildout, The New York Times reported Sunday. One Texas ad opens with someone typing a question into a chatbot about why an electric bill keeps going up. Democratic challengers in top House races in Michigan, Wisconsin and Pennsylvania are all running ads against the facilities. A Washington Post analysis found Republican public statements about data centers turned sharply negative over the past six months. Jon Husted, then Ohio's lieutenant governor and now a senator, praised a Google investment in 2024 as further establishing central Ohio as a tech hub. Facing a tight November race, he has since introduced a Senate bill his website says will protect Americans from footing the bill for new data centers.

Read at The New York Times ↗ • Read at The Washington Post ↗ • Read at The Philadelphia Inquirer ↗ • Read at Ohio Capital Journal ↗

After opposing California AI safety law last year, OpenAI urges stronger safeguards

In a post from its global affairs team on Saturday, the company said California's SB 53 "should be amended to expand safeguards," specifically by requiring monitoring of frontier models under training or evaluation for potential serious incidents and by strengthening cybersecurity protections across the model development lifecycle. SB 53 passed last year and imposes transparency requirements and whistleblower protections on large AI developers. OpenAI opposed it at the time. Absent significant federal legislation, the company now supports an approach it calls "reverse federalism," in which states move in a compatible direction on core protections that can later become the foundation for a national standard.

Read at TechCrunch ↗ • Read at LinkedIn ↗

Four states sue Meta for child harm, first of 29 states to go to trial

California, Colorado, Kentucky and New Jersey accuse the company of designing addictive products that harm children, and all 29 states that sued in 2023 are pressing a claim that Meta violated federal law protecting the privacy of users under 13. The trial opened Aug. 18 at the Ronald V. Dellums federal building in Oakland, Calif., before an eight person advisory jury, with U.S. District Judge Yvonne Gonzalez Rogers to decide the case. California Deputy Attorney General Megan O'Neill told jurors the business model comes down to four words: Meta hooks users, holds them as long as it can, harvests their data and hides the truth from the public. An initial calculation put maximum penalties at about $1.4 trillion, a figure Attorney General Rob Bonta said the states are not asking for, and the states are also seeking design changes to the products. Mark Zuckerberg and Instagram head Adam Mosseri are expected to testify.

Read at The Guardian ↗ • Read at NPR ↗

Chinese AI firms rent restricted Nvidia chips while Senate considers bill to close loophole

Moonshot AI, ByteDance, Alibaba and Tencent have reached restricted Nvidia compute by routing through data centers in Thailand, Malaysia and Japan, CNBC reported Aug. 19. Cassia King, a senior researcher in the compute policy group at the Institute for AI Policy and Strategy, said the arrangement complies with current rules because U.S. export controls govern the physical transfer of chips. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot AI in July of reaching Nvidia GB300 chips through a facility in Thailand to train its Kimi K3 model. The House passed the Remote Access Security Act, H.R. 2683, on Jan. 12 by 369 votes to 22. A Senate companion from Sens. Dave McCormick and Ron Wyden, S. 3519, sits in the Banking Committee and has not moved.

Read at CNBC ↗ • Read at Quartz ↗ • Read at Control Plane ↗

II.China Watch

China's biggest flash memory maker files for $4.9 billion Shanghai listing to fund capacity

Yangtze Memory Technologies Holdings has had its application to list on Shanghai's STAR Market accepted, per Reuters, and plans to raise 33 billion yuan ($4.9 billion) in one of the largest offerings on the exchange's technology board. YMTC is the mainland's main producer of NAND flash memory, the storage chips used in phones, laptops and data center servers. Proceeds are earmarked for production line upgrades and research and development. Acceptance starts a formal review and does not amount to listing approval. The company entered IPO counseling in May, and the Shanghai Stock Exchange said on Aug. 22 that it had accepted the application.

Read at Reuters ↗

Alibaba says second in-house AI chip enters production this year as AI spending swells capital budget

Chief Executive Eddie Wu said on Aug. 20 that the company's second generation T-Head chip should begin tape-out and enter production in the second half of this year, per TechNode. Tape-out is the point at which a finished chip design goes to the factory. Alibaba said the part will compute faster and move data between chips more quickly, and that it can carry large language model training. Its existing Zhenwu chip line had served more than 650 customers through Alibaba Cloud by early August. Capital spending in the June quarter rose 75% to 67.68 billion yuan (about $10 billion), mostly on AI infrastructure, while net income attributable to ordinary shareholders fell 76%.

Read at TechNode ↗

Prominent Chinese policy scholar: Washington's AI supply chain blocs are built to fail

Zheng Yongnian, dean of the school of public policy at the Chinese University of Hong Kong, Shenzhen, said U.S. attempts to build AI supply chains without China will not work, the South China Morning Post reported. He spoke Sunday by video to a conference at South China University of Technology. Pax Silica, the framework Washington set up in December to secure AI supply chains, along with the chip and rare earth alliances, "explicitly target a third party, which is bound to fail," he said. He urged the two governments to cooperate on global economic development instead. Zheng also said cutting U.S. companies off from low cost Chinese models would damage their commercial interests, and that American vendors sell expensive closed source models better suited to wealthy economies such as Japan and South Korea.

Read at South China Morning Post ↗

Shanghai orders new AI compute clusters and a city run scheduling platform

The Shanghai Municipal People's Government has issued its five year plan for accelerating new industrialization and building a modern industrial system, covering 2026 to 2030, per Jiemian. It directs the city to build a tiered intelligent computing network that coordinates compute, transmission and storage, to consolidate existing resources in priority areas and to deploy large intelligent compute clusters. Edge computing nodes are to be added as needed, with resources tied together through a municipal platform that schedules intelligent computing. The plan also tells Shanghai to lead the country on commercial 6G deployment and to become a "satellite internet city."

Read at Jiemian ↗

III.Policy Tracker

New Gov. Shapiro EO conditions AI data center permits on signed state contract

Gov. Josh Shapiro signed Executive Order 2026-05 in Harrisburg on Aug. 18, effective immediately and covering all new applications. It directs the Department of Environmental Protection to open a data center permit file only where the developer has signed a consent order and agreement committing to the Governor's Responsible Infrastructure Development requirements and has already secured local approval. Those requirements make developers pay the full cost of generating and delivering the power their project needs rather than passing it to households. The order also bars Commonwealth agencies from using nondisclosure agreements on data center projects. The Department of Revenue is applying the same test to the state sales and use tax exemption for data center equipment. Figures released with the order count more than 100 proposed projects statewide, 15 of which have applied for a DEP permit.

Read at AI News ↗ • Read at Commonwealth of Pennsylvania ↗

Palantir urges GSA to withdraw AI acquisition clause

The General Services Administration is weighing an acquisition rule that industry says could exclude open source and third party AI from federal contracts, FedScoop reported Aug. 20. Commenters said the rule assigns responsibility to the wrong entity, either a contractor with no ability to control the platform or a developer with no access to government data. Bruce Andrews, Nvidia's chief external affairs officer, wrote that placing the obligation on an open or third party publisher that may never receive government data is the opposite of the safeguarding outcome the agency intends. Nvidia pointed to a July letter it co-signed with more than 230 companies including Palantir, Microsoft and OpenAI urging protection of open weight models. Palantir went further and asked the agency to withdraw the rule outright, saying GSA has no independent authority to issue a governmentwide rule.

Read at FedScoop ↗

FTC notifies businesses that undisclosed personalized pricing may violate FTC Act

The Federal Trade Commission said Aug. 19 it is taking public comment on a draft enforcement policy statement on personalized pricing, the use of personal data to set prices according to what a company believes an individual will spend. The draft says undisclosed collection or use of personal data for that purpose could violate the agency's prohibition on unfair or deceptive practices. "The FTC does not have the legal authority to ban personalized pricing in all circumstances, but businesses that fail to tell consumers how their personal data is being used to set a price may be in violation of the FTC Act and other laws we enforce," Chairman Andrew Ferguson said. The vote authorizing the Federal Register notice was 2-0, and the public will have 30 days to comment once it publishes.

Read at FTC ↗

Ohio becomes first state to require AI policy for every public school district

A new state law requires districts to have a formal AI policy in force as the school year begins, either by customizing their own guidelines or adopting the model policy written by the Ohio Department of Education, WHIO reported Aug. 18. The guidelines must govern how students and staff use AI in the classroom, leaving districts to balance classroom integration against student privacy, academic integrity and foundational learning. Some districts, including Mad River Local Schools in Riverside, Ohio, have built policies already. Mad River developed its playbook with the Montgomery County Educational Service Center. The policy lists approved digital tools and sets out appropriate and inappropriate uses.

Read at WHIO ↗

Rival AI super PACs fund same Florida governor candidate

Leading the Future, a super PAC favoring light AI regulation backed by Marc Andreessen and OpenAI President Greg Brockman, and Public First Action, an AI safety group backed by Anthropic and other tech executives, have each spent more than $2 million on ads boosting Rep. Byron Donalds' campaign ahead of the Aug. 18 Republican primary, per ad analytics firm AdImpact. The two were on opposite sides earlier this year in a New York City Democratic congressional primary where Leading the Future spent more than $8 million to defeat AI safety advocate Alex Bores. Bela Sullivan of the ethics group Issue One said the camps disagree on AI policy but agree that money gets you in the room. Gov. Ron DeSantis backed an "AI Bill of Rights" that failed in the legislature this year.

Read at Bloomberg ↗

IV.Capability & Research Watch

Flock's police AI identifies drivers by movement patterns, per code on company website

Flock Safety has told the public for years that its technology "cannot recognize, identify, or track individuals." WIRED reported Aug. 19 that the vehicle surveillance company has built a tool, originally called Nightshift and now OS Investigate, that identifies drivers and tracks vehicles by their movement patterns alone. Flock says the product is still in development and in testing with a small group of law enforcement partners, with capabilities that may change before any wider release. The software ships with 69 prewritten prompts and code describing 45 tools it can call, including plate scans, arrest records, dispatch logs and commercial databases carrying Social Security numbers and relatives. The more than 450 files were on Flock's own website, served by its login pages to anyone who loaded them.

Read at WIRED ↗ • Read at Decrypt ↗

Independent grader ranks frontier companies on rogue model containment

Guidelight AI Standards, an organization promoting safe frontier development practices, graded publicly available plans from Anthropic, Google, OpenAI, Meta and xAI on how prepared each is for a model detected behaving in ways that subvert human control. OpenAI scored highest; Anthropic and Meta scored lowest. Steven Adler, Guidelight's chief scientist and a former OpenAI safety researcher, said he was surprised by how little the companies have published about handling a serious escape. TechCrunch reported the grading Saturday, two days before an Aug. 24 deadline. AIPD's August 12th edition reported that Reps. Greg Casar, D-Texas, and Doris Matsui, D-Calif., set that deadline for OpenAI and Anthropic to answer questions on containment failures.

Read at TechCrunch ↗

V.Industry & Market Watch

Anthropic's IPO prospectus to list public backlash against AI as risk factor

People familiar with the filing told CNBC on Aug. 21 that the company will name hardening public sentiment toward data centers and fear of AI driven job losses among the risk factors in its registration statement. That section is where companies are legally required to disclose what could damage the business. Anthropic filed confidentially with the Securities and Exchange Commission in June and has since held preliminary meetings with bankers and investors in San Francisco. The document is expected within weeks. Chief Financial Officer Krishna Rao is fielding investor questions about margin pressure from open source competitors and about what happens if data center construction slows. AIPD's August 13th edition reported investors expect a float of about $2 trillion, roughly double the $965 billion valuation Anthropic reached in a May funding round.

Read at CNBC ↗ • Read at Yahoo Finance ↗

Nvidia backs OpenAI leased Ohio data center with up to $105 billion in residual value guarantees

Nvidia disclosed in an Aug. 17 SEC filing that it will guarantee the residual value of leased infrastructure at SB Energy's PORTS-Pike Technology Campus in Pike County, Ohio, with its aggregate payment obligation capped at $105 billion. The guarantees cover an initial 4.25 gigawatts of IT capacity, with an option to support another 3.75 gigawatts. SB Energy, a SoftBank subsidiary, will build, own and operate the campus under a 20 year lease with OpenAI, and the site will run Nvidia compute and networking exclusively. If specified default or insolvency events occur, Nvidia may owe the shortfall between the guaranteed minimum value under the leases and what SB Energy recovers by re-leasing or selling the facilities. The New York Times reported the campus could ultimately cost as much as $500 billion.

Read at The New York Times ↗ • Read at Data Center Knowledge ↗ • Read at DataCenterDynamics ↗

TikTok and ByteDance agree to pay $400 million to end Justice Department's children's privacy case

The Justice Department announced the settlement Friday, resolving a 2024 lawsuit alleging the platform let millions of children under 13 use the service and collected their personal information without parental consent, in violation of the Children's Online Privacy Protection Act. TikTok will pay $300 million immediately and another $100 million once a court vacates an earlier consent decree against its predecessor company, Musical.ly. The agreement also requires stronger age controls, added safeguards for young users and more parental oversight of children's accounts. Neither TikTok nor ByteDance admits wrongdoing. Associate Attorney General Stanley E. Woodward Jr. said the resolution secures a substantial recovery while reinforcing protections families expect.

Read at AP ↗ • Read at TechCrunch ↗ • Read at Justice Department ↗

VI.Global & Geopolitics

Dutch regulator fines Uber 825 million euros ($966 million) for algorithmically deactivating drivers

The Dutch Data Protection Authority said Friday it is imposing a fine of 825 million euros ($966 million) because Uber cut off driver accounts, sometimes permanently, through automated systems without human review and without adequately informing the drivers. The authority said that violates the European Union's General Data Protection Regulation, which bars decisions made by algorithm alone where they significantly affect a person. It said the conduct ran from 2018 to 2022. The fine would be the second largest GDPR penalty issued, behind only the 1.2 billion euro ($1.4 billion) fine Ireland imposed on Meta in 2023. Monique Verdier, the authority's deputy chair, said drivers were left with no income from one moment to the next. Uber said it disagrees with the decision, that the policies examined were discontinued years ago, and that it will appeal.

Read at AP ↗ • Read at NL Times ↗

Trump family crypto firm earns fees on access to Chinese AI models the administration will likely fence off

The State Department is preparing a document warning the 35 countries that signed the U.S. AI Opportunity Statement in June that working with China on AI is likely to get them cut off from U.S. efforts, Reuters reported this month. They could join Pax Silica, the U.S. framework for securing AI supply chains, only if they stay out of China's competing World Artificial Intelligence Cooperation Organization. Kazakhstan, rich in AI infrastructure minerals, belongs to both. WorldClaw, a Hong Kong startup selling access to Chinese models the U.S. has flagged over security concerns, takes payment in USD1, the stablecoin run by World Liberty Financial, the crypto firm partly owned by Eric Trump and Donald Trump Jr. World Liberty does not own WorldClaw but earns a fee on every USD1 transaction, and both sons promoted the arrangement online.

Read at Reuters ↗