Inside China's 800MW datacenter rival to OpenAI's Stargate

Google's 60-20-20 framework: how Gemini's product team balances shipping velocity with frontier innovation

Tulsee Doshi, Google's head of product for Gemini, outlines a methodology where 60% of effort goes to proven approaches, 20% to near-working ideas, and 20% to experimental bets. This framework offers concrete insight into how a major frontier lab structures product development and prioritization decisions across models like Gemini, Genie, Veo, and Lyria, balancing the need to ship competitive products with sustained innovation.

Meta's €53bn bond order book exceeds all European startup funding for 2025, exemplifying structural capital flight to US tech

Nicolas Colin argues Europe's "Europoor" phenomenon—where households save conservatively and institutional investors chase safe US borrowers—finances America's tech dominance while Europe captures only coupon returns. Meta's €53 billion euro bond order book alone exceeds all European startup funding for 2025, illustrating how this structural capital imbalance directly undermines European AI infrastructure competitiveness and reinforces US technological leadership.

Computer engineering graduates now have the second-highest unemployment rate among 73 majors, per New York Fed data

This signals potential labor market disruption from AI automation in technical roles, contrasting sharply with corporate productivity ambitions and suggesting uneven adoption benefits across the workforce.

We Visited China's Stargate in Inner Mongolia

SemiAnalysis published a field report on China's Galaxy datacenter campus in Inner Mongolia, a direct competitor to OpenAI's Stargate Abilene. Main takeaways:

• Galaxy's first 120MW phase was delivered within one year at one-fifth US construction costs, with full 800MW campus planned by end-2027, matching Abilene's scale
• Envision Group demonstrates extreme vertical integration: owns wind farms (420MW capacity, 60%+ power coverage), manufactures battery storage, designs 800V DC power systems, and operates blade factory adjacent to campus
• Galaxy deployed 5MW 800V solid-state transformer (SST) pilot converting 10kV AC directly to 800V DC, eliminating traditional UPS/rectifier stages; China has 25% of datacenters on DC power vs 3% in US
• Cooling design eliminates central chiller plant by running liquid and air-cooled servers on single water loop, a pattern operating at scale across northern China
• Inner Mongolia has 25GW of announced datacenter projects with 30,000 construction workers across 22 sites; region faces no anti-datacenter backlash unlike US, enabling acceleration

Why I'm becoming a China bear

Noah Smith in Noah Smith's Substack (noahpinion.blog) published an essay titled "Why I'm becoming a China bear". Main takeaways:

• China's manufacturing-dominant strategy is hitting saturation limits; manufacturing share of global output peaked during pandemic and has drifted down despite Xi's 4% GDP/year industrial subsidies
• China's domestic market cannot absorb the output of its subsidized manufacturing base, collapsing profits across auto, steel, solar sectors as oversupply meets weak real estate-driven demand
• AI represents a fundamental shift that favors adaptive, decentralized systems over China's centralized state-capacity model optimized for 20th-century industrial production
• Infrastructure overinvestment has left China with massive depreciation costs on underutilized assets, forcing continued spending on maintenance rather than growth
• Xi Jinping's political paranoia and economic policy exhaustion suggest China will remain powerful but unlikely to achieve 20th-century-style global hegemony

DeepFinance: Ultra Long-Horizon Benchmark for Private Equity and Investment Banking AI Workflows

Julius Research published a benchmark report on DeepFinance, an ultra long-horizon evaluation framework for AI agents on real-world Private Equity and Investment Banking tasks. Main takeaways:

• DeepFinance comprises 21 tasks averaging 27.1 hours of human analyst work, with 84 rubric criteria per task and realistic data rooms of ~19 files (emails, spreadsheets, PDFs, term sheets) from top PE/IB firms
• Frontier models (GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5) cluster at 54-58% rubric score; next-best (Muse Wonder) scores 44.2%; bottom tier (Nemotron, Inkling) scores 10-14%, showing extreme capability stratification
• Grading emphasizes trajectory and auditability over final numerical correctness: models scored on output accuracy, workflow quality, and judgment soundness across three categories, reflecting real deal review standards
• GPT-6.1 Sol dominates cost-performance frontier at $1.17/task (54.3% score); cheapest model above 30% threshold is GPT-6.1 Terra at $0.69/task (47.2% score)
• Benchmark uses model-native harnesses (Claude Code, Codex, OpenCode-AI, etc.) rather than universal constraints, testing agents in realistic professional environments with full tool access including web search