"Welcome to August 6, 2026" — @theinnermostloop
Why this is in the vault
A paragraph-cluster digest with no single sustained argument; the two items worth keeping are SpaceX's Terafab memory-fab megaproject in Grimes County, TX (a $16.8B opening bid scaling toward 100M sq ft and 1TW/year of compute) as the latest data point on the memory-wall capex race, and OpenAI's "Agent Plugins" spec — a vendor-neutral bundle format for Skills + MCP servers backed by Amazon, Cursor, Microsoft, and Vercel — landing next to a wide scatter of model releases, biotech, and financing items.
Curation section
- Memory wall, megascale: SpaceX's Terafab (Grimes County, TX) opens at $16.8B / 3,000 jobs, scaling to 100M sq ft and 1TW/year of AI compute, split ~25% Optimus / 75% AI spacecraft; pitched as five times the size of today's largest building, justified by America having zero high-volume memory fabs. AMD bought Taalas (etches weights directly into silicon, ~17,000 tok/s serving Llama 3.1 8B on a 6nm chip, locked to one model until re-spin) and Nvidia is reportedly trimming planned memory in Rubin Ultra across three test builds — both are workarounds for the same bottleneck the fab is meant to solve at the supply side.
- Model releases: Meta's Muse Spark 1.2 hit 54 on the Intelligence Index (third release in four months), landing "6 points below Claude Opus 5 at ~1/6th the cost." OpenAI shipped GPT-5.6 Sol (tighter answers, effort slider, Luna as free default) and is reportedly training Astra, its largest pretrain since GPT-4.5. A $500M offshore labeling-data trade (same shops serving US labs also sell to Tencent/ByteDance/Alibaba/Ant) is drawing a "weights are policed, taste is not" framing.
- Agents standardizing: OpenAI published Agent Plugins, a vendor-neutral spec bundling Skills and MCP servers into portable client-loadable packages, backed by Amazon, Cursor, Microsoft, and Vercel. Suno issued content-provenance principles (watermarking, fingerprinting, downloads policy) in the same portability vein. OpenAI's rumored first Jony Ive device is a displayless, camera-and-mic "doughnut" puck, $300-400 in 2027.
- Financing strain: Alphabet sold $25B of bonds against $115B peak demand days after its first-ever negative free cash flow and $195-205B capex guidance; KC Fed's Jeff Schmid asked whether AI financing is becoming "too big to fail." US Saudi crude imports hit zero in July, first empty month since 1985. Anthropic is hiring an Insider Risk Investigator (up to $305k) to hunt exfiltration in its own logs.
- Biology as compile target: a Science paper (King et al.) used genome language models to design 16 functional phage genomes that beat bacteria already resistant to a natural phage. FDA approved Moderna's mFlusiva (first US mRNA flu shot, 27% better than standard dose, re-matchable in 2-3 months vs. 6). Penn engineered a lablab-bean-gum protein that cut HPV 93% in saliva.
- Infrastructure odds and ends: SpaceX will bolt femtocells onto Starlink dishes/Superchargers using EchoStar spectrum to reach phones without towers by late 2027; DeepMind's WeatherNext set new state of the art on cyclone tracking (open-sourced); a speculative paper price-tags megaengineering fixes to keep Earth habitable for ~9 million billion years.
No deep-fetches this issue. All citations route through substack.com/redirect wrappers with no clean destination resolvable in plaintext — consistent with prior Innermost Loop filings (see 2026-08-01 note).
Mapping against Ray Data Co
The load-bearing item is OpenAI's Agent Plugins spec, not Terafab: a vendor-neutral format for bundling Skills + MCP servers into portable, client-loadable packages — backed by Amazon, Cursor, Microsoft, and Vercel — is the industry standardizing on exactly the architecture RDCO already committed to internally (skills over commands per standing feedback, and the brigade-house v2 collapse to "markdown skills + the harness itself" documented in [[2026-07-18-agent-brigade-v2-simplification-design]]). RDCO's own house rebuild treated the library layer as the thing that broke and the skill/MCP concept layer as the thing that worked; Agent Plugins is external confirmation that layer is becoming a portable standard rather than an RDCO-specific convention — worth a check on whether brigade-house's skill/MCP packaging should target Agent Plugins compatibility once the spec stabilizes, rather than staying a bespoke format.
The Terafab / memory-wall cluster is a secondary, lower-confidence data point for the hyperscaler-capex anchor tracked in [[2026-06-23-stratechery-memory-chips-china-microsoft-deepseek]] and the investing memory-cycle thesis — one more concrete facility commitment on the supply side of the same bottleneck, not a standalone thesis input.
Related
- [[2026-07-18-agent-brigade-v2-simplification-design]]
- [[2026-05-11-innermostloop-harness-eats-the-model]]
- [[2026-06-23-stratechery-memory-chips-china-microsoft-deepseek]]
- [[2026-08-05-innermost-loop-tokenmaxxing-discovery-loop-buildout]]