| AI Policy Institute | theaipi.org |
| Your day in AI policy starts here. | Washington, D.C. |
| Bottom Line Up Front |
| Were you forwarded this email? Subscribe → |
The bill calls for a federal pause on new AI data center construction pending environmental review, oversight rules, and worker protections. The proposal covers energy consumption, environmental effects, and labor standards tied to AI infrastructure buildout. No Republican co-sponsors have been announced.
The White House budget justification designates $202.2 million and 730 full-time positions for EPA to advance AI capabilities in FY2027. EPA's overall budget would drop from $8.8 billion to $4.2 billion under the proposal. Data center permitting and AI infrastructure support are listed among the agency's five stated goals.
Vice President Vance and Treasury Secretary Bessent personally questioned technology companies about AI security risks before Anthropic deployed its Mythos model, per CNBC. Bessent and Fed Chair Powell separately convened heads of the largest US banks to address Mythos cyber threats; the model generated working exploits for zero-day vulnerabilities 72.4% of the time during testing. Vance's participation had not been previously reported.
The Federal Trade Commission banned Air AI for deceptive practices related to AI business opportunity marketing. The enforcement action uses the agency's existing Section 5 consumer protection authority rather than AI-specific legislation. The settlement was announced March 24, 2026.
Environmental organizations are challenging the permit for a power plant supplying xAI's Mississippi data center, per CNBC. xAI, now a wholly owned subsidiary of SpaceX following their February 2026 merger, received the permit over objections citing air quality and energy grid impact. The legal challenge was filed after state regulators approved the permit.
Researchers published a wafer-scale two-dimensional semiconductor growth technique they say achieves over 1,000 times faster production than standard methods, per SCMP. The work, published in Advanced Materials, targets next-generation chips for AI and low-power computing by addressing physical limits of silicon scaling. The 2D approach offers a potential path to advanced semiconductors that does not require EUV lithography equipment.
Huawei's Kirin 9030 chip, fabricated by SMIC on its N+3 process node, uses DUV-based quadruple patterning to approach transistor densities typically requiring restricted EUV equipment, per analysis from TechInsights and Semiwiki cited by Huxiu. The chip is deployed in Huawei's Mate 80 Pro and RS series smartphones. DUV lithography equipment remains outside the scope of current US export controls targeting Chinese chipmakers.
Chinese banks are moving cautiously on AI model risk as Beijing prioritizes financial stability, per SCMP. Regulators have increased monitoring of AI-related threats and now require cybersecurity assessments tied to AI model deployments. The requirements apply to AI systems used in both customer-facing and back-office banking operations.
ByteDance released Coze 2.5, an AI agent development platform equipping autonomous agents with persistent memory, cloud-based virtual devices, and specialized skills across video creation and programming, per Pandaily. The platform enables agents that operate continuously beyond chat interfaces. The AI agent infrastructure layer is a market segment not yet addressed by existing US export controls.
Cybersecurity experts argue that Anthropic's Mythos model is a wake-up call about systemic software vulnerabilities accumulated over decades, not a novel AI weapon, per Wired. The appropriate policy target is software infrastructure security rather than AI model regulation alone. The vulnerabilities Mythos finds at scale were built by developers who treated security as an afterthought for years.
Google developed TurboQuant, a quantization algorithm that reduces the amount of memory required to run AI models, per the Financial Times. Semiconductor experts cited by FT cautioned that more efficient AI could drive even greater chip demand by enabling new applications at lower cost. Samsung and SK Hynix, which supply the majority of high-bandwidth memory for AI workloads, both posted record quarterly revenue driven by AI demand.
Elon Musk has lost a series of legal cases in recent months, ranging from shareholder fraud claims to allegations that rivals stole AI trade secrets, per the Financial Times. The Musk v. OpenAI trial begins April 27 and will decide whether Musk can block OpenAI's for-profit conversion. The restructuring at issue is valued at over $100 billion.
Anthropic is closing the competitive gap with OpenAI in US enterprise AI adoption, driven by strong demand for Claude Code, per the Financial Times. The shift follows Anthropic's reported $30 billion annual revenue run rate earlier this month. Claude Code, a coding-focused AI product, has been the primary driver of new enterprise account growth at Anthropic.
Japan approved an additional ¥631.5 billion (approximately $4 billion) in subsidies for Rapidus, bringing total government support to approximately $16 billion, per Bloomberg. Rapidus aims to manufacture advanced semiconductors for AI applications, though analysts widely regard the venture as a long shot. The additional funding tranche was approved April 11.