Why this is in the vault
Daily curated digest with two threads that matter to RDCO: the memory-price/capex-financialization signal feeding the chip-fab capital-cycle thesis, and a striking example of extreme sub-agent fan-out (60 subagents, 31M tokens) producing a formally-verified math result — a real-world data point on the ceiling of the fan-out pattern RDCO's own workflows are built around.
Mapping against Ray Data Co
Directly feeds project_investing_markov_capital_cycle (chip-fab/memory capital-cycle thesis, currently placed in Phase 2): this issue reports memory prices have "roughly quadrupled in a year" — pushing Apple toward blacklisted Chinese chipmakers and Microsoft to ramp its own Maia chips hoping Anthropic will adopt them — which is fresh anchor-quality evidence of the capacity-constraint phase the thesis is tracking. Pair with the capex-financialization thread: Jensen Huang unveiled $500B+ in "AI factory" commitments and called chips "the first time technology chips have become an investable asset class," while skeptics drew an explicit 2008-subprime comparison to sliced GPU revenue streams — a framing RDCO should track as a bear-case input, not just bull confirmation, for the next /investing:build-thesis or /investing:edgar-watch pass. Anthropic's own infrastructure moves ($9.1B/20-year Riot Platforms power deal, the new Theseus Infrastructure JV with Macquarie and GIC, pledging to cover consumer electricity-hike spillover) are a first-party data point that the labs themselves are now underwriting power-market risk, which is the same site-constraint dynamic flagged in the 2026-08-10 issue's data-center-ban item.
Secondarily, the unreleased research Claude pushing the Riemann hypothesis zeta-zero bound from 41.6% to 67.2% using 60 subagents and a formally verified Lean proof is a concrete existence-proof for extreme sub-agent fan-out on a single hard problem — worth keeping in mind alongside RDCO's own fan-out patterns (process-newsletter, deep-research, brigade stations) as a reference point for how far the pattern scales when the task has a checkable formal answer.
Curation section
- ZeroBench "impossible" visual benchmark: GPT-5.6 Sol first to reach the 30% human baseline (Opus 5 at 26%, Fable 5 at 24%).
- Hidden reasoning-trace extraction from Claude/GPT/Gemini surfaced signs of Chinese models trained on rivals' outputs; Nvidia responded by funding an in-house open-model family (Nemotron 3.5 Lightning, 30B MoE) plus a "Switchyard" router claiming frontier accuracy at a third of Opus 4.8's cost.
- Zuckerberg's "The Future is for Everyone" essay argues for safety via power balance among many personal superintelligences rather than one aligned system.
- OpenAI split Daybreak trusted-access into tiers and shipped GPT-5.6-Cyber (95% of advanced cyber tasks vs. 1.5% for the civilian model), already crediting it with finding two chained Chrome V8 zero-days.
- Anthropic pledged imperceptible watermarks + signed C2PA metadata on future Claude outputs under the EU AI Act.
- Foxconn's AI hardware crossed half of revenue for the first time; CoreWeave grew revenue 112% YoY to $2.58B with a $104B backlog including new Anthropic and Meta business.
- Dyna Robotics' Dyna-2 showed the first human-to-robot transfer scaling law on unseen robot bodies, pretrained on a million hours of human video.
- Anthropic's IPO chatter puts the company at a $965B valuation ahead of a potential largest-ever IPO.
Related
- [[2026-08-10-innermost-loop-machine-traffic-majority-auto-mode-default]]
- [[2026-07-30-innermost-loop-ai-capex-fragility-digest]]
- [[research/2026-08-01-hyperscaler-capex-merchant-vs-inhouse-silicon-decomposition]]
- project_investing_markov_capital_cycle