AI Policy · Daily

The Senate's defense bill is set to carry three measures targeting China's AI chip access, as leaders folded the AI Overwatch, MATCH and Chip Security Acts into the NDAA manager's amendment. Commerce Under Secretary Jeffrey Kessler told the House Foreign Affairs Committee that few Nvidia H200 chips have reached China so far and that more semiconductor rules are coming, while ruling out any revival of the Biden-era global diffusion cap. The UAE won license-free access to advanced U.S. chips this week under a Commerce rule cast as a reward for Abu Dhabi's help against Iran, as G42 moves to reincorporate as a U.S. company. Twenty-six Meta employees sued in federal court in Oakland, alleging the company used internal AI tools to steer layoffs toward workers who had taken medical or parental leave. OpenAI employees donated more than $215,000 to a super PAC seeking tougher AI rules, positioning themselves against the pro-industry Leading the Future group backed by OpenAI president Greg Brockman.

I.Top Stories

Senate NDAA manager's amendment set to include three bills targeting China's AI chip access

The manager's amendment to the Senate National Defense Authorization Act will fold in the AI Overwatch Act, the MATCH Act and the Chip Security Act, Punchbowl News reported, citing multiple sources — a win for China hawks who have pushed the measures for months. The AI Overwatch Act would insert Congress into export-control decisions for AI chips bound for "countries of concern" like China, and the version in the amendment would shorten a ban on sales of advanced chips such as Nvidia's Blackwells to those countries from two years to 18 months. The MATCH Act aims to press U.S. allies to curb sales of chipmaking equipment to China, while the Chip Security Act would require companies to track their AI chips abroad to combat smuggling. Senators led by Jim Banks (R-Ind.) and Elizabeth Warren (D-Mass.) drove the push, though the broader NDAA remains uncertain after Democrats voted Tuesday against opening debate over the Iran war and the bill's cost.

Read at Punchbowl News ↗

House Foreign Affairs presses BIS chief Kessler on chip exports as H200 flow to China restarts

Under Secretary of Commerce for Industry and Security Jeffrey Kessler testified Tuesday before the House Foreign Affairs Committee and told lawmakers "very few" Nvidia H200 AI chips have been shipped to China, per The Hill and CNBC. Commerce authorized about 10 Chinese firms including Tencent and ByteDance to buy the H200 earlier this year while keeping Nvidia's more advanced Blackwell line barred from China, and Kessler's remarks confirm those shipments have begun. He also told the committee additional regulation of AI chips and semiconductors is coming, and said the administration will not replace the Biden era AI diffusion rule that capped global AI chip shipments, per Reuters.

Read at The Hill ↗ Read at CNBC ↗ Read at Reuters ↗

G42 plans to reincorporate in U.S. as Commerce's UAE chip rule takes effect

Commerce's rule removing license requirements for advanced computing exports to the UAE government and approved firms took formal effect this week. G42, the Abu Dhabi AI conglomerate at the center of the rule, plans to become a U.S. company, according to the Wall Street Journal, which framed the eased access as reward for Abu Dhabi's cooperation in the U.S. campaign against Iran. The rule moves the UAE into a country grouping that permits license free purchases of Nvidia AI chips, some commercial satellites and defense equipment, and dual-use civil nuclear items alongside NATO members, per Semafor. The UAE is the only country in the grouping that is not a member of a multilateral export control regime.

Read at WSJ ↗ Read at Semafor ↗

Twenty-six Meta workers sue in Oakland federal court alleging AI selected leave-takers for layoffs

Twenty-six Meta employees filed suit in federal court in Oakland, California late Monday, alleging Meta used internal AI systems, keystroke and activity monitoring data, AI token usage dashboards and algorithmically assisted performance rankings to select workers for layoffs, per AP. Each plaintiff took protected medical or parental leave or requested a disability accommodation, and all remain employed by Meta with separations scheduled to begin July 22. The complaint cites the Family and Medical Leave Act, the Americans with Disabilities Act, the Pregnancy Discrimination Act and Title VII disparate impact doctrine. The Trump administration has ordered federal agencies to deprioritize disparate impact enforcement, per the Washington Post. Meta said in a statement that the claims "lack merit and are not based on facts" and that workforce decisions "were and are made by people, not AI."

Read at AP News ↗ Read at The Washington Post ↗

OpenAI employees donate more than $215,000 to super PAC pushing tougher AI rules

Seven current OpenAI employees and one former staffer have given more than $215,000 to Guardrails Alliance, a super PAC seeking stricter regulation of frontier AI labs, WIRED reported. The group launched last month with $5 million and casts itself as a populist counterweight to Leading the Future, the pro-AI industry PAC backed by more than $100 million from tech figures including OpenAI president and co-founder Greg Brockman. Research engineer Juan Felipe Cerón Uribe accounted for most of the total with a $200,000 donation, saying he wanted private companies held accountable for responsible AI development. The contributions are dwarfed by the $50 million Brockman and his wife, Anna, have pledged to Leading the Future, but they underscore growing internal tension at OpenAI over the company's ties to the group. Guardrails Alliance, which aims to raise $15 million this cycle, is expected to name more donors in a public campaign-finance filing Wednesday.

Read at WIRED ↗

White House launches 'Gold Eagle' clearinghouse to patch software flaws found by AI

The Trump administration launched Gold Eagle, an AI cybersecurity clearinghouse to help federal agencies, critical infrastructure operators and AI companies identify and patch software vulnerabilities discovered by advanced AI models, POLITICO reported. Unveiled Tuesday, it fulfills a requirement of Trump's June 2 executive order on AI security and will use leading models, including Anthropic's Mythos. National Cyber Director Sean Cairncross thanked open-source AI developers for their work on the effort.

Read at POLITICO ↗ Read at White House ↗

II.China Watch

Alibaba's Qwen becomes the model behind Apple Intelligence in China as regulator finalizes registration

On July 15, China's Cyberspace Administration filed seven on-device generative AI services for smartphones, including Apple Intelligence, Huawei's Xiaoyi and OPPO's AndesGPT, per TechNode. Alibaba said in response that its Qwen model will be integrated into Apple Intelligence as the underlying AI capability for Chinese users of iOS, iPadOS, macOS and visionOS, letting them access Qwen's text and image processing and content generation without leaving Apple apps. The filing brings Apple's China AI feature set inside the country's generative AI registration regime, which requires services to be approved by the CAC before public release under the 2023 Interim Measures for Generative AI Services.

Read at TechNode ↗

DeepSeek eyes a $71 billion valuation and 2027 mainland IPO as it races to lock in AI compute

DeepSeek is in talks with new investors on a fresh funding round at a pre-money valuation of about $71 billion, up 37% from a first outside round that closed just six weeks earlier, per QbitAI and the Financial Times. The Hangzhou-based model developer is also preparing paperwork for a 2027 initial public offering and may file its listing application as soon as this year, with accountants and bankers working to finalize financial reports by December 2026. Proceeds are earmarked for gigawatt-scale data centers, additional AI chip purchases and next generation agent research. Founder Liang Wenfeng's personal net worth has passed $36 billion on the strength of the new round.

Read at QbitAI ↗ Read at Financial Times ↗

Xiaomi's factory humanoid robot closes to within a point of human workers on core assembly

Xiaomi said its humanoid robot deployed on the SU7 electric vehicle line raised its success rate at a self-tapping nut loading station from 90.2% to 98% over four months of iteration, narrowing the gap with human worker qualification rates to a single percentage point, per TechNode. Two newer tasks, center console side panel sorting and parts bin folding, are already running at 90%, and the flexible workpiece station marks the first time a humanoid robot has run long duration continuous operations on soft materials in an automotive factory. The robot uses whole-body motion control, dual-arm coordination and force sensing, autonomously retrying a placement when one fails.

Read at TechNode ↗

JD.com breaks ground on its first embodied AI robot production hub in Guangzhou

JD.com broke ground on its first RoboBase facility in Guangzhou, a roughly 1 billion yuan ($140 million) complex due to complete construction by 2028 and reach full production by 2030 at a projected 1.75 billion yuan ($240 million) in annual output, per Pandaily. Founder Richard Liu said JD plans more than 80 RoboBase sites nationwide, with JD Logistics targeting the procurement of 3 million robots, 1 million autonomous vehicles and 100,000 drones over the next five years for warehousing and last-mile delivery. The company is positioning itself as a platform orchestrator rather than a robot maker, bundling supply chain access, open source models, real-world testing at JD stores and logistics parks, and capital for embodied AI startups.

Read at Pandaily ↗

Shenzhen humanoid startup LimX Dynamics hits a $2.3 billion valuation ahead of a Hong Kong listing

LimX Dynamics, founded by Southern University of Science and Technology professor Zhang Wei, raised $200 million in a pre-IPO round that values the Shenzhen humanoid maker at roughly $2.3 billion, with IDG Capital, Nio Capital and Abu Dhabi's Stone Venture participating, per Pandaily. The company said its strategy is to build fully autonomous humanoid robots for non-factory service settings rather than the auto plant use cases that dominate Chinese peers. Zhang said listing "is a must," and LimX is already in confidential review with Hong Kong as the likely venue.

Read at Pandaily ↗

III.Policy Tracker

NSF awards 12 second-cohort Regional Innovation Engines across 20 states

The U.S. National Science Foundation announced 12 new Regional Innovation Engines awards spanning 20 states, the second cohort of a program that funds multi-state research and workforce clusters in emerging technologies including AI. Brian Stone, performing the duties of NSF director, said the Engines will be "transformational for America's innovation infrastructure."

Read at NSF ↗

Army awards Neros $500 million contract for Ukraine-style FPV drones

The U.S. Army awarded drone startup Neros a $500 million contract to mass-produce first-person-view attack drones cost competitive with Chinese models, per the Wall Street Journal. Neros is on course to be the first winner of the Pentagon's 18-month Drone Dominance competition to fulfill its order, and founder Soren Monroe-Anderson, a former drone racing world champion, moved production into a new factory in Torrance, California, per the Washington Post. The Drone Dominance program, overseen by the Defense Innovation Unit, pits roughly 26 companies against one another for a share of $1.1 billion in defense contracts aimed at 300,000 attritable drones, cheap enough to be lost in combat. Next year's defense budget calls for $54.6 billion to fund an expanded drone warfare unit.

Read at WSJ ↗ Read at The Washington Post ↗

IV.Capability & Research Watch

arXiv paper proposes framework for standardizing frontier model CBRN uplift evaluation

A new arXiv preprint introduces a "Threshold Exceedance Criteria" framework for evaluating frontier language models' potential to uplift chemical, biological, radiological and nuclear threats. The framework decomposes evaluation into non-expert participant eligibility, defined threat scope and statistical estimation of exceedance. The authors position the methodology explicitly for "policymakers and model developers."

Read at arXiv ↗

PJM capacity auction adds $6.3 billion in data center charges across 13 states

PJM Interconnection's latest capacity auction procured $16.4 billion in power supply for the year starting June 2028, with data centers accounting for roughly $6.3 billion of the total across the grid's 13 states and Washington, D.C., per Bloomberg. The auction tied a record set in late 2025, and PJM's independent market monitor said cumulative data center burden across four auctions now approaches $30 billion for ratepayers. The auction fell 6.8 gigawatts short of PJM's future supply target, the third straight time the market has failed to secure enough capacity to meet reliability standards, per the New York Times. Prices hit the market's $325 per-megawatt-day cap. Without the cap, the auction would have cleared at $554.72 for a total closer to $30 billion, and Chicago-zone prices would have exceeded $775.

Read at Bloomberg ↗ Read at NYT ↗

V.Industry & Market Watch

IBM sheds $69 billion in market value after AI linked profit warning

IBM shares plunged more than 25% Tuesday after the company issued a preliminary second-quarter profit warning, wiping roughly $69 billion in market value in a single session on track to exceed IBM's 1987 Black Monday decline, per WSJ. CEO Arvind Krishna said in a letter to investors that clients in late June shifted spending toward servers, storage and memory to secure supply ahead of expected AI linked price increases, pulling budget away from IBM's higher-margin mainframe and software products, per the Guardian. Preliminary second-quarter revenue is expected at $17.2 billion versus $17.86 billion consensus. Microsoft, ServiceNow and Salesforce fell between 3% and 5% on the news.

Read at WSJ ↗ Read at The Guardian ↗

VI.Global & Geopolitics

Albanese unveils single-window Australian AI office, fast-track data center approvals and early-2027 legislation on water and copyright

Prime Minister Anthony Albanese, in a University of Sydney speech Wednesday, announced the creation of an Office of Artificial Intelligence inside his department to oversee economic, social, security and environmental AI issues, per The Guardian. The government will introduce faster approval processes for AI projects including data centers. It will legislate national AI standards it aims to pass early next year, per ABC News. Under the standards, new data centers will be forced to underwrite or supply their own power and minimize water use. Albanese also committed to legislative protection for Australian creators against use of their work to train AI without consent.

Read at The Guardian ↗ Read at ABC News ↗ Read at Prime Minister of Australia ↗

Japan enacts election law mandating AI labels on social media

Japan's parliament passed cross-party legislation requiring AI generated images and videos to be labeled when shared on social media during election campaigns, per The News. The internal affairs minister announced the rules Tuesday, July 14. The law takes effect in March 2027 and does not include penalties for noncompliance.

Read at The News ↗