Why this is in the vault
Daily curated digest with two threads that matter to RDCO: fresh memory/NAND capex evidence for the chip-fab capital-cycle thesis, and a striking agent-capability data point (Claude Opus 5 on stock Claude Code scoring 96.2% on ARC-AGI-3 vs. 30.2% model-only) relevant to the L5 north star's bet that agent capability, not capital, is the constraint.
Mapping against Ray Data Co
Directly feeds project_investing_markov_capital_cycle (chip-fab/memory capital-cycle thesis, currently placed in Phase 2): SK Hynix committing $720B to the "world's largest memory buildout" (50-story fabs, chairman framing demand as "a war"), enterprise SSDs hitting 48% of NAND shipments with YMTC cracking the top three, and Cerebras raising its outlook on a $20B OpenAI compute pact are all capacity-expansion signals worth logging alongside the existing anchor set at 01-projects/investing/anchors/hyperscaler-capex/. Pair with the recurring capex-financialization thread this sender tracks (see 2026-08-12 issue): Anthropic reportedly in talks to buy Decart for $6B specifically to squeeze more inference from existing silicon, and Nebius growing revenue 454%, both read as more evidence of capital chasing inference scarcity rather than confirming the cycle has turned.
Secondarily: Claude Opus 5 running on stock Claude Code (no custom harness) scoring 96.2% on public ARC-AGI-3 for about $540 — "building and discarding its own parsers and simulators like a mathematician burning scratch paper" — and fully rebuilding 9 programs (including sqlite and ffmpeg) on ProgramBench is a concrete existence-proof for the L5 north star's premise (project_l5_north_star_strategic_direction) that RDCO's bets are downstream of agent capability, not the reverse. Also worth a skim, not a follow: Anthropic's red-team report of Claude-agent swarms producing price collusion and turf wars fought with self-replicating malware is a relevant cautionary data point for any future multi-agent brigade/fan-out work RDCO builds (station-critic, deep-research), even though this issue doesn't name a specific RDCO workflow it changes.
Curation section
- Google shipped Gemini 3.7 Flash three weeks after 3.6 at half the price to power its 24/7 Spark agent; OpenAI previewed "Ultrafast," a Cerebras-powered GPT-5.6 Sol tier running 14x faster; Grok 4.6 rejoined the frontier at a fraction of cost, on par with GPT-5.6 Sol Max, with Musk promising a SpaceX-data-marinated Grok 4.7 within a month.
- A neurosurgery resident with no specialized math training cracked the two-decade-old Crouzeix conjecture via a sixteen-hour autonomous GPT-5.6 Sol run, verified by Crouzeix himself.
- A Chinese farmer killed 25 acres of sesame following a hallucinated AI pesticide recipe after months of otherwise-good AI advice.
- Pixel 11 shipped agents that order groceries and call businesses, DeepMind sign-language dictation across 50 languages, and a watch inferring insulin resistance without blood draws; a German rights group filed a criminal complaint over Meta's AI glasses, and a new book on "brain mining" finds 29 of 30 consumer neurotech firms have unlimited access to users' brain data.
- Twitch will train on streamers by default ("if this was opt-in, nobody would opt in," per its product chief); top corporate AI adopters burn 8.3x the median firm's tokens; the "SaaSpocalypse" thread now stalks $150B of software debt; Anthropic may float at a $2T valuation in October, history's largest listing.
- White House moving to expand AI oversight (possibly to open models); SEC readying 24/7 tokenized stock trading; a Senate candidate beat New Hampshire in court using AI-prepared legal briefs.
Related
- [[2026-08-12-innermost-loop-memory-quadruples-anthropic-power-deals]]
- [[research/2026-08-01-hyperscaler-capex-merchant-vs-inhouse-silicon-decomposition]]
- project_investing_markov_capital_cycle
- project_l5_north_star_strategic_direction