| AI Policy Institute | theaipi.org |
| Your day in AI policy starts here. | Washington, D.C. |
| Bottom Line Up Front |
The UK's AI minister criticized OpenAI for halting a major data center investment in Britain. OpenAI cited high UK energy costs and regulatory requirements for the pause. The UK had been positioned as a key international site in OpenAI's multibillion-dollar Stargate data center initiative before the project was suspended.
Anthropic released Claude Opus 4.7, which scored 64.3% on SWE-Bench Pro, a software engineering benchmark, more than 10 points above Opus 4.6. The company simultaneously launched a Cyber Verification Program to loosen cybersecurity guardrails for vetted security professionals. Anthropic said Opus 4.7 does not advance its capability frontier; Mythos Preview outperforms it on every evaluation. Findings from the deployment will inform the timeline for broader Mythos release.
European Union regulators plan to require Meta to modify how WhatsApp handles user data for AI purposes as part of an ongoing antitrust investigation, the Wall Street Journal reported. The order targets both privacy and competition concerns over using messaging data to train AI models. The EU is using existing competition law authority rather than new AI-specific legislation to regulate AI training data practices.
Nvidia, JPMorgan Chase, Google, Apple, and Microsoft described Anthropic's Mythos model as a "sharp improvement" over prior AI offerings for cybersecurity, Bloomberg reported. The five named companies were among the first outside organizations to receive access to the restricted model.
Senior Pentagon officials testified before Congress on the Golden Dome homeland missile defense system, which integrates AI for real-time threat assessment and automated coordination across multiple interceptor layers. The system would deploy AI-driven sensor fusion and automated targeting to protect the U.S. homeland against ballistic and hypersonic missile threats. Pentagon leaders characterized Golden Dome as the next generation of U.S. homeland defense infrastructure.
The UK government launched a £500 million ($675 million) sovereign AI fund to invest in British AI companies and reduce dependence on foreign AI technology providers. The fund targets early-stage AI startups building core capabilities within the UK. The launch came the same day as the UK AI minister's public criticism of OpenAI for pausing its British data center project.
Mao Shengyong, deputy head of the National Bureau of Statistics, told a State Council briefing that China's daily average AI token consumption exceeded 140 trillion in March, up more than 40% from the end of 2025, per TechNode. The figure, which tracks API calls across China's major model providers, represents a roughly 1,000-fold increase over two years. The bureau presented the data as part of a broader government assessment of China's commercial AI deployment.
Beijing-based embodied AI company TARS, founded in February 2025, closed a $455 million Pre-A funding round, the largest such round in China's embodied AI sector, per Pandaily and Caixin. The company has demonstrated a humanoid robot performing hand embroidery, a level of fine-motor dexterity uncommon in current humanoid platforms. The round is roughly ten times the typical Pre-A for Chinese AI startups.
Alibaba's Qwen team released Qwen3.6-35B-A3B under an Apache 2.0 license, a sparse mixture-of-experts model with 35 billion total parameters but only 3 billion active, per Pandaily. On agenticAI 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. coding benchmarks it outperforms Google's Gemma4-31B and competes with dense models ten times its active parameter count, while supporting a context window exceeding one million tokens. The model is the latest in a series of permissively licensed releases from Chinese AI labs this year.
A team from the Ningbo Institute of Industrial Technology at the Chinese Academy of Sciences developed a diamond-copper composite that improves AI data center cooling efficiency by up to 80%, per SCMP. The material has been deployed in an AI computing node in Zhengzhou, Henan province. Industry projections put the Chinese market for diamond-metal thermal materials at roughly 10 billion yuan ($1.4 billion).
AI systems deployed in combat are "black boxes" whose decision-making humans cannot interpret before the systems act, a neuroscience researcher argues in MIT Technology Review. The article cites AI systems in the Iran conflict that generate targets in real time, coordinate missile interceptions, and guide autonomous drone swarms. The analysis directly references the Anthropic-Pentagon dispute over military AI access and the current Pentagon guidelines requiring human operators to understand AI system decisions.
OpenAI released GPT-Rosalind, a model optimized for protein reasoning, genomics, and drug discovery, available as a research preview to qualified customers including Amgen and Moderna. The model is accessible through ChatGPT, Codex, and the API, with access restricted to organizations with demonstrated life sciences credentials. GPT-Rosalind puts OpenAI in direct competition with Google DeepMind in AI drug discovery and life sciences.
Physical Intelligence released its pi-0.7 model, which completed tasks it was never explicitly trained on by combining skills learned in different contexts, the company said. In demonstrations, the robot operated an air fryer with minimal training data; researchers reported capabilities are scaling "more than linearly with the amount of data." Step-by-step verbal coaching enabled the robot to perform novel tasks in laboratory conditions.
Zscaler joined OpenAI's Trusted Access for Cyber (TAC) program, gaining access to the restricted GPT-5.4-Cyber model for integration into its Zero Trust Exchange and SecOps platforms, per SiliconANGLE. OpenAI launched TAC last week alongside GPT-5.4-Cyber as a gated-access framework for vetted defenders, as reported by SiliconANGLE in AIPD's April 15th edition. Zscaler will integrate GPT-5.4-Cyber into its threat detection and response systems.
Anthropic signed a new London office lease with space to expand its UK staff from approximately 200 to more than 800, Wired reported. The expansion comes during the company's ongoing dispute over its Pentagon supply chain risk designation, as noted in AIPD's April 14th edition. The Pentagon supply chain risk designation coincides with Anthropic's largest international hiring push. The London office would become Anthropic's biggest location outside San Francisco.
Tesla and SpaceX's joint venture contacted chipmaking equipment suppliers including Applied Materials, Tokyo Electron, and Lam Research to solicit price quotes for the proposed Terafab facility, Bloomberg reported. Staff sought quotes for photomasks, substrates, etchers, depositors, and testers. A Musk-controlled chip fabrication facility would add a new variable to CHIPS Act subsidy allocation and semiconductor export control enforcement. Industry observers remain skeptical about the feasibility of a new entrant in advanced chip manufacturing.
Google's annual ads safety report shows the company blocked 8.3 billion ads globally in 2025, up from 5.1 billion in 2024, with AI systems catching more than 99% of policy-violating ads before users saw them. The company identified 602 million scam-related ads and 4 million advertiser accounts linked to scams. The shift toward content-level blocking without account suspension creates an enforcement model that platform regulators have not yet addressed. Despite the increase in blocked ads, Google suspended fewer advertiser accounts than in 2024.
AI-generated traffic to U.S. retail sites rose 393% year-over-year in Q1 2026, with AI-driven visits producing 37% higher revenue per visit than non-AI traffic, according to Adobe Analytics data covering more than one trillion site visits. AI visitors now convert at rates 42% above human shoppers, reversing a 38% deficit from a year ago. AI-intermediated commerce complicates FTC consumer protection and disclosure enforcement. Thirty-nine percent of surveyed consumers said they used AI tools for online shopping in Q1.