Industry roundup, 2026-10-13: Reflection, research taste, and the capital wall
Sources: AI Breakfast, AI Weekly Espresso, Marcus on AI (Gmail); Import AI 476, The Information, The Decoder (RSS); X Following feed. Raw: raw/gmail/2026-10-13-newsletters.md, raw/rss/2026-10-1[23]-*.md, raw/twitter/feed/*-ranked.json.
Key items
- Nvidia may buy or acquihire Reflection AI (FT via Reuters/The Star). Nvidia already put in $800M; Reflection raised at a $25B pre-money valuation in April. An acquihire (hire staff, license tech) would skip antitrust review. Reflection's Beam is a 501B MoE (23B active) pitched on token efficiency, pretrained on 6,144 GB300 NVL72 GPUs for four weeks; Import AI notes its scores trail Chinese open models (GLM, Kimi, DeepSeek) (Import AI 476).
- Research taste is doubling every 3 months (Import AI on P-Zero's TasteVal). Frontier models' experimental research taste (designing experiments and interpreting results across 8 AI-development tasks) doubled every 3.0 months since Dec 2025 (95% CI 1.7-5.0). Opus 5.5 beats a 24-expert human baseline with a 2.30x compute multiplier at ~1/30 the cost per attempt. Jack Clark moves his full-RSI estimate from 2028 to 2027.
- Gary Marcus: "the frontier labs' WeWork moment?" (Marcus on AI). Tokenmaxxing ended, Chinese open weights forced price wars, OpenAI revised revenue forecasts down ~$20B, agent liability is rising ("Large Liability Modules"), and the Third Circuit held AI training on copyrighted material is not fair use. He questions trillion-dollar IPO valuations; Anthropic's S-1 is still not public.
- VCs securing compute for portfolio startups (The Information). a16z (20,000-GPU cluster since 2024) is in talks to expand via two backed data center companies; General Catalyst is building its own data center.
- Bain's Bridge Data Centres seeks to upsize a Malaysia loan from $2.8B to ~$6.6B, among the largest Asian data center debt deals (The Information/Bloomberg).
- Agent CPU and inference economics: UBS sees Cerebras revenue ~4x to $3.3B in 2027 and ~$23B by 2030 with ~$25B contracted backlog; Perplexity says revenue is growing 5x YoY on agent demand and sovereign funds want it run on their own hardware (X feed).
- Policy: the White House restored the UK AI Security Institute's early model access (The Decoder); CAISSI senior officials are being moved to report directly to the White House (The Information); Nadella's "assume a model is compromised" post (The Verge via AI Weekly).
- Infrastructure risk: Ukrainian drones shut Yandex's Vladimir data center, the third Yandex AI compute site hit in four days (Moscow Times). China plans $3.9T of grid buildout through 2049 to feed AI data centers (BloombergNEF).
- Science: Google DeepMind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems, with results peer reviewed in Science (The Decoder).
- Microsoft reorganized Azure leadership around AI (The Information), and its Office chief recounts the 2025 decision to pay for Anthropic models inside Copilot.
Relation to prior wiki pages
- 10-12's Looking Ahead flagged inference as a "permanent reverse auction". Marcus's essay and Agent Arena's cost-per-task collapse (89-96% cheaper to match July's leader) are two more data points for it; Microsoft's Maia 300 and Nvidia circling Reflection are the incumbents' responses (own the silicon, own an open lab).