Daily Newsletter Digest

White House prioritizes open AI models, proposes GPU compute markets

June 15th 2026

Today's News

U.S. government forces Anthropic to suspend Claude 5 models globally over national security concerns

Interconnects by Nathan Lambert + Techmeme

The White House compelled Anthropic to shut down access to both Fable 5 and Mythos 5 models for all customers following an export control order, marking an unprecedented use of export controls targeting a specific AI company's models. Reports cite a suspected China-linked group accessing Mythos and a jailbreak vulnerability reported by Amazon researchers as triggers. This signals a dramatic shift from the ChatGPT-era regulatory approach to direct executive intervention on model releases based on political and security judgments, with the Trump administration demonstrating willingness to unilaterally restrict U.S. AI capabilities on national security grounds.

SemiAnalysis STEEL lab reveals SMIC N+3 achieves TSMC N6-class density but at higher cost and complexity

SemiAnalysis

SemiAnalysis's new Teardown Engineering & Evaluation Lab published detailed reverse engineering of Huawei's Kirin 9030 SoC on SMIC's N+3 process, showing China's most advanced node reaches a minimum metal pitch of 32.5nm—about 10% tighter than Intel's 18A at 36nm. However, N+3 achieves TSMC N6-class logic density only through aggressive DUV multi-patterning and design-technology co-optimization, requiring far more complexity, higher cost, and worse power efficiency than leading-edge nodes. The Kirin 9030 Pro performs similarly to three-year-old Android flagships and trails current flagship SoCs from Apple, Qualcomm, MediaTek, and Samsung by multiple generations, illustrating both China's progress under export controls and the significant constraints those controls impose.

Sequent launches with $100–150M to pursue independent alignment research outside frontier labs

Jack Clark from Import AI

Researchers from the UK AI Security Institute and Timaeus have formed a nonprofit focused on theoretical alignment approaches that could provide confidence before superintelligent AI systems are deployed, targeting 40–80 employees. The organization explicitly states "alignment is not on track" to match the timeline of ASI development, signaling that serious researchers believe current frontier lab approaches are insufficient and that independent safety research organizations may be necessary to maintain credibility and raise alarms if needed.

FrontierCode benchmark shows Claude Opus 4.8 at only 13.4% on hardest tier, suggesting durable eval

Jack Clark from Import AI

Cognition released a hand-curated coding benchmark with 150 tasks graded for production-ready code quality (correctness, test coverage, style adherence, codebase compatibility), with top models struggling significantly on the Diamond difficulty tier. Hard evals that resist saturation are rare and valuable for tracking progress; the author predicts 70%+ Diamond performance by June 2027.

Hyperliquid governance approves routing ~90% of treasury yield to automated HYPE buybacks

Today in DeFi

Hyperliquid governance passed with 69% validator approval to route approximately 90% of Hyperliquid treasury yield to automated HYPE buybacks starting August 26, with the mechanism expected to add $135–197M annually in buyback volume. This represents a material structural change to token economics and treasury management for a major DEX.

China's financial and educational systems show signs of structural strain amid technological pivot

Noahpinion + Brian Roemmele from The Brian Roemmele Report.

China's banks are evergreening bad loans at scale, mirroring Japan's 1990s zombie company trap: official non-performing loan ratios have fallen even as loss-making enterprises surged since 2021, with below-market lending rising sharply and many firms unable to service debt interest. This suggests widespread unrecognized bad debt being rolled over to keep unprofitable companies afloat, consuming scarce labor, materials, and capital that productive companies need and potentially explaining China's current "involution" and price-slashing across manufacturing sectors. Simultaneously, the government is cutting 12,200 university degree programs in a five-year overhaul, prioritizing AI, robotics, and semiconductors while eliminating arts and humanities—a shift reflecting strategy to align education with technological self-reliance and address youth unemployment, with nine universities adding embodied intelligence programs to support humanoid robot initiatives.

Worth Reading

The White House's plan for open models & AI research in the U.S.

Interconnects

Nathan Lambert in Interconnects published an essay titled "The White House's plan for open models & AI research in the U.S.". Main takeaways:

  • White House explicitly prioritizes investment in open-weight models as strategic priority, reversing Biden-era policy that grouped open models with closed models under compute thresholds
  • Federal government proposes financial markets for GPU compute access (spot/forward markets) to reduce long-term contract requirements blocking academic and startup access to large-scale compute
  • NAIRR pilot expansion and NSF AI funding remain positive outliers in federal R&D budget, but document does not address immigration barriers preventing leading AI researchers from attending US conferences or joining institutions
  • Lambert notes ~18-month window to develop 'American DeepSeek' before academic community left with compute but no frontier-competitive open models as Chinese labs dominate open model releases
Read more

The EDA Primer: From RTL to Silicon

SemiAnalysis

SemiAnalysis published a comprehensive technical primer on Electronic Design Automation (EDA) and the chip design flow. Main takeaways:

  • Verification now consumes up to 70% of total chip design project effort, with verification engineers the fastest-growing job category in semiconductor development
  • Design productivity improves only ~20% annually while chip complexity grows ~50% yearly, creating an exponential engineering bottleneck that EDA tools must address
  • The modern EDA landscape is dominated by three companies: Synopsys (Design Compiler, logic synthesis), Cadence (IC layout and place-and-route), and Siemens EDA (acquired Mentor Graphics for $4.5B in 2017)
  • Advanced SoCs like AMD MI455X pack 320 billion transistors across 12 logic dies with 2nm/3nm processes, hybrid bonding, HBM4, and 224G SerDes—designs that test companies' tooling, methodology, and human capital organization
  • Single advanced mask sets cost tens of millions; every respin adds months to schedule and represents a major balance sheet impact when A0 rarely goes into production
Read more

China's Financial and Fiscal Decay

Rhodium Group

Rhodium Group published a research report on China's deteriorating financial and fiscal systems as tools for economic management. Main takeaways:

  • Credit growth hit record lows in early 2026 (6.1% y/y), down from 9-18% historical averages, with 58% of loans at or below the 3% loan prime rate, indicating banks increasingly finance unproductive state enterprises rather than private investment.
  • Fiscal revenues declined 1.7% outright in 2025 despite government projections of 0.1% growth; tax revenues rose only 0.8% while non-tax revenues fell 11.3%, pushing fiscal deficit to 9.0% of GDP with further widening expected in 2026.
  • Land sales revenues collapsed nearly 15% in 2025 to 4.15 trillion yuan (down from 8.7 trillion yuan peak in 2021), reducing local government capacity to service debt and manage fiscal obligations.
  • Banking system profitability deteriorated despite 520 billion yuan in capital injections; net interest margins halved from 270 basis points (2014) to 142 basis points (Q4 2025), with returns on assets at only 0.6% and annual profits flat since 2023.
  • The financial system has become 'zombified'—unable to fail but increasingly inefficient at generating growth, with declining payoff from lending and fiscal spending while private sector investment remains weak.
Read more

Articles Read