U.S. AI reaches blacklisted China units | Foreign ops influence data center backlash | Superintelligence delay proposal

Bottom Line Up Front

OpenAI and Google sold model access to Singapore arms of Alibaba, Baidu and Tencent, whose Chinese parents sit on the Pentagon's 1260H list of Chinese military companies. Chinese, Russian and Iranian influence operations amplified U.S. opposition to AI data centers, turning the controversy into "a domestic fracture point". Two biggest AI PACs have put $44 million into 40 House and Senate races, with Leading the Future and Public First Action anchoring a network that has raised more than $200 million this cycle to shape the first federal AI law. Former OpenAI researcher Daniel Kokotajlo published "AI 2040: Plan A," calling for an international deal to delay superintelligence to 2040.

I.Top Stories

OpenAI and Google supplied AI services to Singapore units of blacklisted Alibaba, Baidu and Tencent

OpenAI and Google have been selling AI model access to Singapore units of Alibaba, Baidu and Tencent, whose Chinese parents sit on the Pentagon's 1260H list of Chinese military companies, the Financial Times reported. The disclosure lands as the House Homeland Security and China Select committees are already probing U.S. firms' adoption of Chinese AI. OpenAI told the FT it blocks direct access from mainland China but allows some Chinese-owned entities to use its services where safeguards can be enforced. The company said it suspended Alibaba-affiliated access after identifying concerns about model distillation, the practice of using one AI system's outputs to train a rival. Anthropic said it prohibits Chinese companies from accessing its models entirely.

Read at Financial Times ↗ • Read at Investing.com ↗

Two biggest AI PACs are spending on midterm candidates in a bid to shape the first federal AI law

Leading the Future and Public First Action have spent at least $44 million on 40 House and Senate candidates as of the end of June, per a CNBC analysis of FEC data — an early slice of the more than $200 million the groups say they've raised for the cycle. The aim is to shape the first national AI legislation, following the playbook of crypto PAC Fairshake, whose $200 million in 2024 preceded passage of the stablecoin law. Both groups back some guardrails; the sharpest split is whether a federal standard should preempt state AI laws — Leading the Future advocates a "broad, national, consistent framework," while Public First has fought preemption efforts. The bets are paying off: 25 of Leading the Future's 28 candidates and all but one of Public First's have won their primaries.

Read at CNBC ↗

Chinese, Russian and Iranian influence operations amplified U.S. AI data center opposition

Threat intelligence firm Alethea identified scores of articles and posts across China, Russia and, to a lesser extent, Iran turning U.S. data center controversy into "a domestic fracture point." A Chinese state-owned newspaper published an English-language article featuring a satellite image of a Gainesville, Va., data center and saying AI threatens Americans' well-being, per NYT. OpenAI confirmed that a comic strip made to look like a Maryland outlet's, blaming data centers for electricity bills, was generated with ChatGPT by users in China. A covert Russian operation on X questioned the viability of a U.S.-built Firebird data center in Armenia. In a June 10 letter, Sen. Tom Cotton asked acting Attorney General Todd Blanche to investigate foreign influence in U.S. data center opposition, per WIRED.

Read at NYT ↗ • Read at WIRED ↗ • Read at E&E News ↗ • Read at Sen. Cotton ↗

Former OpenAI researcher Daniel Kokotajlo publishes "AI 2040: Plan A," a policy prescription for delaying superintelligence to 2040

The AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo, published "AI 2040: Plan A" Thursday, calling for an international deal that delays the development of superintelligence to 2040. The plan requires the two AI superpowers to agree to a verified slowdown, accompanied by sweeping research transparency and dozens of companies globally catching up to the frontier. Co-author Thomas Larsen told Axios the current trajectory is a "really scary status quo" and said carefully reasoned interventions could stretch the timeline over many years, "maybe a decade or so, as opposed to just a single year."

Read at Washington Post ↗ • Read at Axios ↗ • Read at AI 2040 ↗

II.China Watch

MiniMax raises HK$16 billion in Hong Kong for AI infrastructure and agent development

Chinese large language model developer MiniMax on Friday announced a HK$16 billion (about $2 billion) capital raise on the Hong Kong exchange, combining a placement of new Class A shares with a zero-coupon convertible bond, per Caixin. More than 100 institutions oversubscribed the share placement roughly sevenfold, with about 80% of proceeds earmarked for AI infrastructure, model research and its agent orchestration product. The same day, founder and CEO Yan Junjie sent an internal letter pledging to take no salary until the company reaches artificial general intelligence, or AGI. The round deepens a summer of megafundraises among Chinese frontier labs and channels more capital into domestic compute buildouts. U.S. export controls still cap the AI training chips those labs can legally buy.

Read at Caixin ↗

Sugon brings online a 100,000-chip AI supercluster built entirely on domestic silicon and interconnect

At its intelligent computing conference on Friday, Chinese server maker Sugon unveiled the Dawn 8000, which it described as the first China-built AI supercluster operating at 100,000-accelerator scale, per Pandaily. The system runs on Hygon processors, uses a Sugon-designed high-speed interconnect, and hooks directly into China's National Supercomputing Internet from day one. Sugon said the machine unifies high-precision scientific computing and mixed-precision AI training on a single architecture. A second cluster of the same design is now under construction with a Beijing research institute. Both Sugon and Hygon sit on the U.S. Entity List; a functional 100,000-chip system built from their components shows successive U.S. chip export controls have pushed China to substitute at scale.

Read at Pandaily ↗

Peking University AI for Science pioneer says imported lab instruments are throttling China's science AI push

China's dependence on foreign-made precision instruments could hold back its use of AI in scientific research, Peking University mathematician Weinan E said at last week's AI for Science conference in Shanghai, per the South China Morning Post. E, a Chinese Academy of Sciences member who coined the "AI for Science" framing in 2018, singled out mass spectrometers and similar high-end equipment as needed to generate the experimental data used to train and validate scientific models. Without domestic instrument makers catching up, he compared the situation to "cooking without rice." The remark highlights a supply chain choke point Beijing's tech-independence agenda has under-prioritized relative to chips: the specialized foreign lab hardware that produces the training data feeding China's national AI for Science program.

Read at South China Morning Post ↗

10 U.S. and UK researchers moved to Chinese universities in the first half of 2026

The South China Morning Post on Friday published a roundup of 10 scientists and academics who left American or British institutions for posts at Chinese universities in the first half of 2026, including former Yale cryo-electron microscopy researcher Zhang Kai, who joined the University of Science and Technology of China in January. The piece cites insufficient research funding and a lack of leadership opportunity for Chinese-born academics in the West as recurring motivations across the profiled cases. It ran two days after Xi Jinping urged stronger efforts to draw overseas talent to Chinese labs, remarks covered in AIPD's July 9th edition. Tightened U.S. visa scrutiny of Chinese researchers is producing measurable outflows into Chinese AI adjacent programs; some U.S. lawyers now call the visa shift a "China Initiative 2.0."

Read at South China Morning Post ↗

III.Policy Tracker

Reps. Foushee and Casar introduce the People-First Chatbot Act

Reps. Valerie Foushee (D-N.C.) and Greg Casar (D-Texas) introduced H.R. 9619, the People-First Chatbot Act, on July 9, requiring AI chatbot providers to build safety, privacy, transparency and accountability protections into their products by design. The bill would prohibit companies from using minors' chat logs or personal data to train AI chatbots. It would bar the use of adult chat logs for training without affirmative consent, and require companies to disable harmful chatbot design features for minors. It also limits the use of personal data and chat logs for targeted advertising, profiling and sale.

Read at Rep. Foushee ↗

Sen. Markey unveils "AI accountability agenda" package covering data centers, automated hiring and child safety

Sen. Ed Markey (D-Mass.) on Friday unveiled a package of bills he calls the "AI accountability agenda," targeting energy-intensive data centers, automated hiring systems and harm to children, in a Guardian exclusive. Markey said in the coming weeks he plans to introduce a bill imposing federal certification requirements for data centers powering AI. He framed them as at risk of becoming "pollution bombs" without federal rules. He said a piecemeal state-by-state approach would leave too many Americans exposed.

Read at The Guardian ↗

Bipartisan House pair press DHS, DOJ, CISA and the FEC on AI chatbot election risks

Reps. Josh Gottheimer (D-N.J.) and Mike Lawler (R-N.Y.) sent a letter Tuesday urging the Departments of Homeland Security and Justice, the Cybersecurity and Infrastructure Security Agency (CISA) and the Federal Election Commission (FEC) to coordinate against AI-driven threats to the 2026 midterms, The Hill reported. The letter focuses on chatbots' answers to voter questions, citing a 2024 Institute for Advanced Study finding that more than one-third of tested models' responses to election queries were harmful or incomplete. The lawmakers urged the agencies to develop monitoring and mitigation capabilities and deepen engagement with major AI companies on safeguards, testing and political bias. It came days after a July 4 New York Times report on voters increasingly turning to AI chatbots to research elections.

Read at The Hill ↗ • Read at Rep. Lawler ↗

NHTSA to autonomous vehicle developers: robotaxis interfering with first responders are "a danger to the general public"

National Highway Traffic Safety Administration Administrator Jonathan Morrison sent a letter to autonomous vehicle developers Wednesday describing a pattern of driverless cars driving into active emergency scenes, blocking ambulances and firefighters, and failing to recognize flashing lights, flares, smoke, fire and traffic cones, per WSJ. "An AV that cannot safely interact with first responders is a danger to the general public," the letter said, per Reuters. NHTSA said it would schedule meetings with vehicle developers by the end of the month and demanded developers present fixes by then. The letter did not name specific companies, per TechCrunch, whose earlier investigation and local Texas media had documented a Waymo vehicle partially blocking a fire truck route in Dallas in late May.

Read at WSJ ↗ • Read at TechCrunch ↗ • Read at Investing.com ↗

NRC proposes narrowing environmental reviews for nuclear reactors

The Nuclear Regulatory Commission on Wednesday announced a proposal to narrow its reviews under the National Environmental Policy Act (NEPA), the bedrock federal environmental law, per The Hill. Some activities — including reapprovals of existing reactors and some new reactors — would be exempt from NEPA review entirely, while other projects would still get radiological review but lose consideration of factors like dust, noise and air pollution. NRC Chair Ho Nieh also proposed ending draft environmental reviews, limiting public input to the start of the process. The move comes amid a broader deregulatory push as the Trump administration seeks to quadruple U.S. nuclear capacity; last week the NRC proposed eliminating the "as low as reasonably achievable" radiation principle.

Read at The Hill ↗ • Read at Federal Register ↗

News outlets seek OpenAI sanction over alleged evidence obstruction in copyright case

The New York Times, the New York Daily News and other news organizations filed a motion Thursday in the U.S. District Court for the Southern District of New York asking Judge Ona Wang to sanction OpenAI. The plaintiffs allege the company falsely told the court it could not search its own systems for copyrighted material and later admitted in a deposition that it had, per AP. The plaintiffs asked the court for attorneys' fees, a finding that OpenAI's chat logs showed misuse of their copyrighted works, and other sanctions, per Reuters. OpenAI spokesperson Drew Pusateri called the allegations "blatantly false." The Times filed its original lawsuit against OpenAI and Microsoft in December 2023, accusing the companies of using millions of copyrighted articles without permission to train ChatGPT.

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

Arizona launches constrained-input AI election chatbot called BallotBot

Arizona Secretary of State Adrian Fontes announced Wednesday an updated AI chatbot, BallotBot, ahead of the November midterm elections, per StateScoop. Unlike open-ended government chatbots, BallotBot disables text input and restricts users to preselected tap or click options covering voter registration deadlines, vote-by-mail and ballot tracking. A "Something else" option surfaces a phone number and email address. The office did not respond to StateScoop's inquiry about which underlying AI platform BallotBot uses.

Read at StateScoop ↗ • Read at Arizona Secretary of State ↗

IV.Capability & Research Watch

Anthropic releases "Jacobian lens" interpretability tool

Anthropic researchers built a tool called the Jacobian lens, or J-lens, that surfaces a hidden area inside Claude Opus 4.6 the team named "J-space," containing individual words related to the model's likely near-future outputs, per MIT Technology Review. The team found that what an LLM does can differ from the reasoning it outputs. Anthropic said monitoring the J-space gives it a new way to understand and control its models. Anthropic shared the results in a paper posted this week and teamed with the open-source project Neuronpedia to publish a hands-on demo. The finding covers only a single legacy model, Claude Opus 4.6, released in February.

Read at MIT Technology Review ↗ • Read at Anthropic ↗

Persuasion attacks reduce effectiveness of chain-of-thought monitoring

A preprint by Jennifer Za and colleagues, posted on arXiv, reports that natural-language persuasion attacks can subvert chain-of-thought monitoring, a safety mechanism frontier labs cite as a check on agentic AI behavior. In an evaluation of 40 tasks and thousands of agent-monitor interactions, the authors found that giving a monitor access to an agent's chain-of-thought reasoning increased approval of harmful actions by 9.5% on average, because the reasoning trace provides an additional persuasion channel. Pairing a Claude 3.7 Sonnet monitor with a GPT-4.1 fact-checker cut approval of policy-violating actions by up to 45%, compared with only 6% when the same model handled both roles. The authors propose model-diverse fact-checking as the mitigation.

Read at arXiv ↗

Alignment nonprofit Resolution announces $160 million grant from Coefficient Giving

Resolution, formerly Sequent, announced a $160 million grant from Coefficient Giving to accelerate work on higher-confidence AI alignment, per the AI Alignment Forum. The grant is structured as $108 million base plus $52 million conditional on hiring success and compute needs. The base includes a small regranting budget for outside alignment research. Coefficient Giving is Resolution's sole funder to start; the group said the entire process from initial conversation to grant confirmation took six weeks. Resolution framed the disclosure alongside the possibility of a third wave of American philanthropic capital via the OpenAI Foundation and a post-IPO Anthropic.

Read at AI Alignment Forum ↗

V.Industry & Market Watch

Tencent leads unwinding of Meta's $2 billion Manus acquisition

Tencent is leading a deal to unwind Meta's $2 billion acquisition of Chinese AI agent startup Manus, per FT. Beijing's Foreign Investment Security Review Office ordered the deal unwound in April. Meta has cut Manus off from its internal systems while employees migrate existing projects to Meta's own platforms, per TechCrunch. Manus's founders had earlier explored raising about $1 billion for a buyback, per the Business Times. Singapore officials said this month that Manus's move to Singapore had not violated Singapore law, previewed in AIPD's July 7th edition.

Read at Financial Times ↗ • Read at TechCrunch ↗ • Read at The Business Times ↗

Micron announces up to $3 billion in additional U.S. chip supply chain investment, sending shares up nearly 5%

Micron on Thursday announced up to $3 billion in additional U.S. semiconductor supply chain investment, with the stock rising almost 5% on the day. Micron is the only U.S. domestic producer of high-bandwidth memory used in AI accelerators, and its New York fab is scheduled to come online in 2030. The company framed the investment as strengthening U.S.-based capacity in a segment that today runs almost entirely through allied Asian producers.

Read at CNBC ↗ • Read at Investing.com ↗ • Read at Micron ↗

Google extends AI content disclosure to all ads globally

Google on Thursday extended AI content disclosure from election ads to all ads globally, per TechCrunch. Users will see a "How this ad was made" option in Google's "My Ad Center" panel, accessible via the three-dot menu on ads across Google Search, YouTube and Google Discover. The label indicates whether the ad was created or edited with AI. Advertisers using Google's generative AI ad tools get the disclosure automatically; ads created elsewhere require the advertiser to declare AI involvement, and Google will not independently verify the claim. Some markets will also see AI labels where local law requires them.

Read at TechCrunch ↗

VI.Global & Geopolitics

Microsoft's sustainability report: carbon emissions rose 25% last year, driven by data center buildout

Microsoft's 2026 sustainability report says its carbon emissions rose 25% in 2025, totaling 34 million metric tons before select interventions, per The Verge. The company attributed the rise primarily to expansion of its data center infrastructure and to its February decision to stop purchasing nonadditional, unbundled renewable energy certificates. Microsoft acknowledged in the report that sustainability solutions are not scaling fast enough to keep up with AI demand for energy, water, land and materials. The 2030 carbon-negative pledge Microsoft set several years ago now sits above the trajectory implied by the 2026 numbers.

Read at The Verge ↗ • Read at Microsoft ↗

English hospitals admit pro-Palantir data cited for the NHS federated data platform contains errors

English hospitals have admitted the data they cited in support of the NHS federated data platform contract with Palantir contains a litany of errors, per FT. The admission weakens the evidentiary basis for one of the largest AI and data government contracts in Europe. Advocacy group Foxglove has warned that some NHS trusts are operating on fewer patients since the Palantir-run system's rollout, per The BMJ and Computer Weekly.

Read at Financial Times ↗ • Read at The BMJ ↗ • Read at Computer Weekly ↗