AI Policy · Daily

Trump pressed Hochul to lift New York's data center freeze, calling the one year pause on large facility permits a costly mistake that hands ground to China. Anthropic paired statehouse lobbying with new political money, backing Illinois' AI Safety Measures Act as FEC filings showed CEO Dario Amodei gave $1 million to the pro-regulation super PAC Public First. OpenAI published a "reverse federalism" framework, endorsing California, New York and Illinois frontier safety bills as building blocks toward a national standard rather than targets for preemption. Rep. Ted Lieu faulted House GOP leadership for passing no AI guardrails this term and cautioned that the Obernolte-Trahan draft's three year override of state model rules could prove risky. Threats against AI executives grew sevenfold between late February and May amid the public backlash, including an April attempt to firebomb Sam Altman's home.

I.Top Stories

Trump urges Hochul to reverse New York's data center moratorium

President Trump on Wednesday July 15 called Gov. Kathy Hochul's (D) new data center freeze a "terrible decision" and said New York should reverse the policy immediately, per CNBC. The clash comes one day after Hochul signed the executive order, covered in AIPD's July 14th edition. Trump warned the moratorium will cost the state money and jobs and cede ground to Chinese AI competition. Hochul's order blocks new environmental permits for data centers above 50 megawatts for up to a year to let state regulators write standards on water use, air quality and rate impact.

Read at CNBC ↗ • Read at Governor Kathy Hochul ↗

Anthropic presses state AI rules as Amodei makes his first $1 million political donation

Anthropic is running a state-by-state strategy to advance AI safety legislation, per Politico: state and local government lead Cesar Fernandez has endorsed Illinois' AI Safety Measures Act and is lobbying legislatures in California, New York and other states, contrasting with OpenAI's push for a federal framework or preemptionPreemptionWhen federal law overrides state law on the same subject. In AI policy it refers to proposals, in Congress or by executive action, to block or freeze the growing patchwork of state AI laws in favor of a single national standard.. The company's political spending is scaling up alongside the lobbying — CEO Dario Amodei gave $1 million in May to Public First, a super PAC backing candidates who support mandatory AI safeguards, in what appears to be his first seven-figure political donation, according to FEC filings released Wednesday. The filings also suggest a funding gap ahead: Leading the Future, a rival super PAC network backed by OpenAI President Greg Brockman and Andreessen Horowitz, ended June with $31 million on hand, against $1.8 million combined across Public First Action's three super PACs.

Read at Politico ↗ • Read at Politico ↗

OpenAI publishes 'reverse federalism' framework for state and federal AI rules

OpenAI Chief Global Affairs Officer Chris Lehane on Wednesday July 15 posted a corporate blog outlining a "reverse federalism" approach in which state AI laws help build a common national standard rather than being preemptedPreemptionWhen federal law overrides state law on the same subject. In AI policy it refers to proposals, in Congress or by executive action, to block or freeze the growing patchwork of state AI laws in favor of a single national standard.. The post endorses California, New York and Illinois frontier safety bills as directionally aligned on three elements: documented safety frameworks with public risk assessment disclosure, incident reporting and independent audits. It welcomes the Obernolte-Trahan federal discussion draft and calls for legislation to strengthen the Center for AI Standards and InnovationCAISIThe Center for AI Standards and Innovation, the US government's AI testing body inside NIST at the Commerce Department. Formerly the US AI Safety Institute, it evaluates frontier models with developers on a voluntary basis and works with the UK's AI Security Institute. (CAISI) as the federal testing home. OpenAI said it appreciates the Trump administration's goal of having a federal cyber evaluation framework in place by "early August."

Read at OpenAI ↗

Rep. Lieu says House GOP has passed no AI guardrails this term, warns against three year preemption ceiling

Rep. Ted Lieu (D-Calif.), vice chair of the House Democratic Caucus, said at The Hill's Hill Nation Summit that he remains "freaked out" by the pace of AI development and cannot name a single AI guardrail law passed this term by House Republican leadership, per The Hill. Lieu praised Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) for their draft national AI framework introduced last month. He warned against the draft's proposed three year override of state AI model development laws, calling a preemptionPreemptionWhen federal law overrides state law on the same subject. In AI policy it refers to proposals, in Congress or by executive action, to block or freeze the growing patchwork of state AI laws in favor of a single national standard. ceiling rather than a floor "a little dangerous" given how quickly the technology is moving. His remarks feed the House Democratic Caucus' Jeffries-convened commission effort to develop a consensus AI framework.

Read at The Hill ↗

AI backlash fuels surge in threats against tech executives

Mounting public opposition to AI has produced a wave of violent rhetoric and threats against industry leaders, the Wall Street Journal reported. Digital threats targeting AI chiefs and data centers grew sevenfold between late February and May before easing in June, per security firm Liferaft, and recent incidents include an April attempt to firebomb OpenAI CEO Sam Altman's home and a man who slipped into Anthropic's lobby warning that an executive was "going to be killed." Executive-protection spending has climbed across the sector — Oracle's rose 85.5% to $5.6 million in 2025 and Palantir's 150% to nearly $3 million — as public sentiment sours: 55% of Americans in a March Quinnipiac poll said AI is doing more harm than good.

Read at WSJ ↗

Democratic group launches tool to shape what AI chatbots tell voters

Run for Something, the progressive candidate recruitment organization, launched a tool in June called CampSight that lets candidates see how AI chatbots describe them and nudge those responses in their favor, Politico reported. The tool simulates voter personas — such as a homeowner or union member — to score how favorably models respond to policy questions, tracks the citations chatbots rely on (most commonly Ballotpedia and campaign websites) and recommends fixes like submitting a Ballotpedia profile or rewording an issues page. About two dozen campaigns are using CampSight, with roughly 350 candidates and advocacy organizations on the waitlist, as voters increasingly consult chatbots to make ballot decisions — a Cornell study found candidate-favoring chatbot conversations were about four times more effective at swaying voters than TV ads.

Read at Politico ↗

II.China Watch

China's memory champion CXMT opens the country's largest A-share IPO of 2026

CXMT, or ChangXin Memory Technologies, began online subscription for its Shanghai STAR Market initial public offering Thursday at a target of about 29.5 billion yuan ($4.3 billion), the largest A-share IPO of 2026 and the second largest in STAR Market history, per 36Kr. The Hefei-based company is now China's largest domestic memory chipmaker. The offering lands on a jittery note in the sector: South Korea's SK Hynix cleared a $26.5 billion New York listing on July 10, and a session later the memory sector suffered one of its most severe single-day drops in nearly two decades. The domestic listing gives Beijing a homegrown supplier for the DRAMDRAMThe working memory that phones, PCs, and AI servers use to hold data that a processor is actively handling. and high-bandwidth memory that AI accelerators depend on, as Washington weighs whether to tighten U.S. equipment restrictions on advanced memory production.

Read at 36Kr ↗ • Read at Reuters ↗

Huawei will publicly show its top-end AI training system at WAIC, aiming to fill the gap left by U.S. export controls on Nvidia

Huawei plans to demonstrate the Atlas 950 SuperPoD, its highest-end Ascend AI hardware, at the World Artificial Intelligence Conference in Shanghai for the first time, per Pandaily. The system links up to 8,192 Ascend NPU cards through Huawei's proprietary Lingqu interconnect, combining many chips to make up for lower single-chip performance. Volume shipments are expected in the fourth quarter of 2026, with ByteDance and Alibaba among the first customers. WAIC opens Friday with President Xi Jinping delivering the keynote for the first time (AIPD's July 14th edition), the venue Beijing is using to promote its Global AI Governance Initiative. The debut lands with Nvidia's B200 and next generation Rubin chips barred from China under U.S. export controls, and it hardens Huawei's vertically integrated NPU, MindSpore and cloud stack against further Commerce Department restrictions.

Read at Pandaily ↗

Xiaomi open-sources a 38 billion parameter embodied AI foundation model, joining a Chinese push to publish model weights

Xiaomi released Xiaomi-Robotics-U0, a multimodal autoregressive model, with code and weights posted to GitHub and HuggingFace, per TechNode. The model combines four capabilities in one framework: generating robot-ready scenes from text prompts, adapting recorded robot trajectories to new environments, producing task-conditioned interaction videos, and general image generation and editing. Xiaomi said the release addresses a data scarcity bottleneck in embodied AI development by letting researchers synthesize training data for scenarios physical robots cannot safely reach.

Read at TechNode ↗

Chinese humanoid startup LimX Dynamics posts autonomous-chore demo billed as Figure-level

LimX Dynamics released a video of its full-size Oli robot putting laundry in a hamper, tidying toys, stacking boxes and picking up floor clutter in a home, which QbitAI reported as a single three-minute take with no teleoperation or editing. The Shenzhen company credits its newly updated COSA 0.5 "brain system," which founder Zhang Wei describes as separate cognition, skill and motion-control modules rather than one large model. QbitAI called it China's first full-size humanoid to complete long-horizon household tasks autonomously and placed LimX alongside Silicon Valley's Figure — though the clip is a curated company demo that hasn't been independently verified.

Read at QbitAI ↗

China's tech giants race to build a "national-level AI app," each betting on a different flagship

ByteDance is bundling short video, e-commerce, workplace collaboration and music services into its Doubao chatbot; Tencent has accelerated a WeChat Agent expected to roll out broadly in the second half of 2026; Alibaba is layering its Qwen assistant across Taobao, DingTalk and Quark; and Ant Group has staked its Alipay AI agentAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. Abao alongside a separate health-focused agent, A-Fu, per 36Kr. Ant this month took a stake of more than 28% in Boohee, a Chinese diet-tracking service, becoming its largest external shareholder. The bet frames weight management as a "national-level scenario" pairing AI dietary analysis with home body fat measurement. Mind share and ecosystem lock-in, not raw user counts, will decide the race, since AI apps lack the social graph switching costs that anchored WeChat and Alipay a decade earlier.

Read at 36Kr ↗

III.Policy Tracker

Whitmer launches Michigan data center pledge as a conditioning alternative to New York freeze

Gov. Gretchen Whitmer (D-Mich.) on Wednesday July 15 unveiled a 10-commitment Michigan Affordability and Responsible Growth Pledge covering minimum billing demand, contract termination fees, credit and collateral requirements, and minimum contract terms for hyperscale data center projects, per Politico. Whitmer called on the state legislature to pass a companion package from state Sen. Kevin Hertel (D) capping water use, requiring wastewater and pollution reporting, and mandating community benefit agreements. Google's regional policy lead Grace Walovich said Google would sign the pledge, tying it to the company's Van Buren Township project with DTE Energy. The commitments track the nonbinding White House Ratepayer Protection Pledge originally adopted in March and expanded Tuesday July 14.

Read at POLITICO ↗ • Read at Michigan.gov ↗

GSA AI acquisition rule listening session draws Palantir warning agencies could route around GSA

GSA held a listening session Tuesday July 14 on its June 17 proposed AI acquisition rule; the public comment window closes August 3, per FedScoop. Palantir federal policy lead Menaka Kalaskar said the material change notification requirements of seven or 30 days are unworkable for SaaS companies and warned that if the rule proceeds as written, "government agencies will have to turn to non-GSA vehicles for the most advanced LLM-powered solutions." George Washington University's Jessica Tillipman flagged concerns about how the rule defines protected data, while Nvidia's Shane Shaneman said data handling responsibility should shift to system integrators. GSA acting Federal Acquisition Service commissioner Laura Stanton defended the rule as balancing agency adoption with data protection.

Read at FedScoop ↗

Palm Beach County commissioners reject Project Tango AI data center after 12-hour hearing

Palm Beach County commissioners on Wednesday evening July 15 voted to deny without prejudice the Project Tango hyperscale AI data center near Arden, Florida, after more than 12 hours of testimony from residents, teachers and community leaders, per WFLX. All commissioners except Maria G. Marino voted to reject the application. Developers said the facility's closed-loop cooling system would use about 600,000 gallons of water without daily replacement, while the Western Palm Beach County Alliance said a hyperscale AI data center did not fit the site's 2016 light-industrial designation. The property retains its existing entitlements for more than 2 million square feet of warehouse and data center development.

Read at Bloomberg ↗ • Read at WFLX ↗

IV.Capability & Research Watch

OpenAI unveils GPT-Red automated red-teamer, claims 6x resilience gains for GPT-5.6 Sol

OpenAI on Wednesday July 15 unveiled GPT-Red, an internal automated red-teaming system that uses AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. in adversarial self-play to probe frontier models for prompt injection and cyberattack vulnerabilities, per MIT Technology Review. The company said training against GPT-Red delivered a 6x improvement in resilience for its GPT-5.6 Sol model. The tool was released alongside OpenAI's GPT-5.6 Sol release notes.

Read at MIT Technology Review ↗ • Read at OpenAI ↗

Microsoft Patch Tuesday resolves record 570 vulnerabilities, cites AI assisted discovery

Microsoft's Patch Tuesday resolved 570 security vulnerabilities across Windows, Office, Azure, Xbox and other products this week, per TechCrunch. The company attributed the record volume to its use of AI in vulnerability discovery. The patch cycle followed the White House's July 14 launch of the Gold Eagle AI cybersecurity clearinghouse, an initiative to coordinate patching of software flaws found by AI models.

Read at TechCrunch ↗

V.Industry & Market Watch

Thinking Machines releases Inkling, its first foundation model, as open weights

Mira Murati's Thinking Machines on Wednesday July 15 released Inkling, its first foundation model, with full open weightsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. posted to Hugging Face and fine-tuning available on the company's Tinker platform, per Axios. Inkling activates about 41 billion of its 975 billion parameters per typical prompt to keep costs down, and was trained on roughly 45 trillion tokens of text, image, audio and video, per SiliconANGLE. Thinking Machines said the model matched Nvidia's Nemotron 3 Ultra on coding tasks despite using two-thirds fewer tokens in early tests. The company raised a $2 billion seed at a $12 billion valuation in 2025 and plans to generate revenue through its Tinker fine-tuning platform rather than metered API access.

Read at WSJ ↗ • Read at Axios ↗ • Read at SiliconANGLE ↗

Hyundai workers strike over humanoid robots as automaker takes full ownership of Boston Dynamics

Workers at Hyundai Motor's Ulsan plant in South Korea launched a three-day partial strike this week over the planned introduction of the Atlas humanoid robot, in what WSJ called the first known auto factory shutdown tied to humanoid robot deployment. The four-hour daily stoppages are expected to cost about 5,000 vehicles in lost production and roughly 200 billion won ($134 million) in sales, per MK, with union leaders demanding wage protections, a retirement age of 65 and formal income security guarantees before any Atlas deployment. Hyundai plans to deploy Atlas at its Metaplant America facility in Georgia beginning in 2028 for repetitive tasks such as parts sequencing, expanding to assembly work by 2030. The labor fight comes as Hyundai Motor Group moves to acquire SoftBank's remaining 9.65% stake in Boston Dynamics, taking its ownership to 100%, per Yonhap — Hyundai bought an initial 80% from SoftBank for 1 trillion won ($674 million) in 2021 and had built its holding to 90.35%. The company said full ownership will support an "end-to-end AI robotics value chain" pairing Boston Dynamics' physical AI expertise with Hyundai's manufacturing capabilities.

Read at WSJ ↗ • Read at MK ↗ • Read at Yonhap ↗

TSMC pledges another $100 billion for U.S. chip factories, bringing total U.S. commitment to $265 billion

Taiwan Semiconductor Manufacturing Co. announced an additional $100 billion in U.S. chipmaking capacity, taking its total pledged American investment to $265 billion as part of a bilateral deal with the Trump administration, per Bloomberg. TSMC also reported Q2 profit up 77% year-over-year to a record 706.6 billion New Taiwan dollars (about $22 billion). It raised its 2026 revenue growth forecast to slightly above 40% year-on-year, up from a prior above-30% outlook, per AP. CEO C.C. Wei said the new commitment will fund four additional Arizona fabrication plants focused on 2-nanometer and below chips, on top of the six fabs already committed under the prior $165 billion pledge. The company also lifted its 2026 capital expenditure budget to $60 billion to $64 billion from an earlier $52 billion to $56 billion range, per Business Standard.

Read at Bloomberg ↗ • Read at AP ↗ • Read at Business Standard ↗

VI.Global & Geopolitics

Japan launches physical AI robotics push with Nvidia and plans 27,500 Rubin chip buy

Japan is planning to purchase 27,500 next generation Nvidia Rubin chips to build a domestic foundation model for robotics, per Bloomberg. The buy tracks a broader physical AI initiative unveiled in Tokyo Thursday July 16 by Nvidia CEO Jensen Huang and Fujitsu CEO Takahito Tokita alongside the CEOs of Fanuc, Yaskawa Electric and Kawasaki Heavy Industries, per AP. The consortium aims to build smart robots that can work in factories, homes and hospitals to address Japan's acute labor shortage. Prime Minister Sanae Takaichi's government recently announced a plan for more than 370 trillion yen (about $2.3 trillion) in public and private investment in technology fields including physical AI, semiconductors and data centers by 2040.

Read at Bloomberg ↗ • Read at AP ↗

Australia's national AI and data center rules will exempt projects under construction

Prime Minister Anthony Albanese on Thursday July 16 said Australia's national AI and data center regulatory framework, unveiled Wednesday (AIPD's July 15th edition), will not apply to projects already under construction, saying "you can't retrofit," per The Standard. The rules will require data center operators to cover the cost of new energy generation for their projects. Climate Council CEO Amanda McKenzie said Australia's roughly 20 gigawatt data center pipeline is largely not yet under construction. Albanese also pledged separate compensation rules for writers, musicians and creators whose work is used in AI training, calling unpaid use "a form of theft."

Read at The Standard ↗

UK opens formal investigation into TikTok's child safety protections

UK regulator Ofcom on Thursday July 16 opened a formal investigation into whether TikTok fails to protect children from harmful content, per The Guardian. The probe operates under the Online Safety Act, the UK's platform liability statute giving Ofcom powers to levy fines and impose behavioral orders. It follows the Labour government's proposal for a voluntary midnight-to-6-a.m. teen social media curfew, as flagged in AIPD's July 15th edition.

Read at The Guardian ↗