THE SWELL #17
SAIL Weekly Digest May 11–15, 2026 | Issue #17
Ten days inside China’s AI labs produced one counterintuitive finding: the export controls meant to slow them down actually made them sharper. SAIL writers filed dispatches this week on an ecosystem that turned a massive compute deficit into a strategic discipline. By extracting more intelligence per unit of hardware, these labs have transformed scarcity into a capability-generating advantage. China is now proving that engineering constraints can accelerate progress rather than just hinder it.
The Week in AI
The Trump-Xi AI Safety Accord: During his historic visit to Beijing on May 14, President Trump met with Xi Jinping to discuss "red lines" for AI. While export controls remain a stalemate, the two leaders reportedly reached a baseline agreement on non-binding safety standards for frontier models and the establishment of an AI-crisis hotline.
Anthropic’s 80x Growth & "Colossus" Expansion: CEO Dario Amodei revealed that Anthropic achieved an unprecedented 80-fold increase in revenue and usage in Q1 2026. To meet this demand, the company disclosed massive new compute deals, including a pivotal partnership with SpaceX to utilize "Colossus" capacity alongside gigawatt-scale commitments from Amazon and Google.
DeepSeek V4 Redefines Efficiency: Following its late-April launch, DeepSeek V4 dominated technical discussions this week as researchers confirmed its "capability-generating" efficiency. Despite China’s hardware constraints, the 1.6T parameter model is matching U.S. frontier benchmarks by extracting 4–7x more intelligence per unit of compute, signaling a shift from brute-force scaling to high-efficiency architecture.
This Week from SAIL Authors
Inside China’s AI Labs: Field Notes
Inside the Chinese AI labs where America’s AI controls created its toughest competition — Across 14 labs in Beijing, Hangzhou, and Shanghai, the team found export controls have become “capability-generating”: Chinese labs extract 4–7x more intelligence per unit of compute than naive scaling predicts, worth roughly two to three years of efficiency gains. — Exponential View
A big lesson of my China visit: compute shortages are holding back Chinese AI — The entire Chinese AI industry holds about as much compute (1.86GW) as a single US company, and that scarcity is the dominant structural force shaping how its labs build. — Understanding AI
ChinAI Mood: April 26 – May 4, 2026 — Field notes from roughly 85% of China’s leading labs: “cracked interns” in their mid-20s, no “philosopher kings,” and a 3:1:1 compute split that guards GPUs for research over costly one-shot training runs. — Interconnected
The Open-Weight Advantage
How open model ecosystems compound — With roughly 80% of a frontier model’s cost going to R&D rather than the final training run, China’s open, knowledge-sharing ecosystem de-risks research across labs — a structural cost advantage that could raise its long-term ceiling. — Interconnects
The Breaking Compute Bargain
The broken bargain of Moore’s Law — TSMC’s refusal to adopt ASML’s High-NA EUV through 2029 may mark the first sustained reversal in transistors-per-wafer-dollar since 2019 — the economic bargain behind the digital economy is bending. — Exponential View
Exponential View #573: Are the AI labs building for an intelligence explosion? — Jack Clark puts 60% odds on a frontier model training its successor by 2028; Azeem reads the labs’ compute land grab as ambiguous — positioning for an intelligence explosion, or just keeping the lights on. — Exponential View
The Trump-Xi Summit
Macartney to Mar-a-Lago — Julian Gewirtz and Matt Sheehan frame the visit as a “stalemate summit”: de-escalation that Beijing still approaches competitively, with echoes of the 1793 Macartney mission and Nixon’s 1972 opening. — ChinaTalk
Xi-Trump to talk AI Safety, Huh? — Part two: the Mythos moment has pushed AI safety back onto the bilateral agenda, though the panel sets low expectations for deliverables beyond technical testing-and-evaluation talks. — ChinaTalk
Full Library
Access the complete, searchable archive of SAIL Media in our Sitemap.
Xi-Trump to talk AI Safety, Huh? — Jordan Schneider | ChinaTalk
Macartney to Mar-a-Lago — Jordan Schneider | ChinaTalk
Inside the Chinese AI labs where America’s AI controls created its toughest competition — Azeem Azhar | Exponential View
A big lesson of my China visit: compute shortages are holding back Chinese AI — Timothy B. Lee | Understanding AI
ChinAI Mood: April 26 – May 4, 2026 — Kevin Xu | Interconnected
How open model ecosystems compound — Nathan Lambert | Interconnects
The broken bargain of Moore’s Law — Azeem Azhar | Exponential View
Exponential View #573: Are the AI labs building for an intelligence explosion? — Azeem Azhar | Exponential View
Join us next Thursday for our Weekly Substack LIVE!