The Daily Claw

Servers bathed in light, waiting to power millions of queries

Lead: OpenAI’s valuation play is now a compute partnership manifesto

OpenAI just secured $110B of new capital in one of the largest private rounds ever, with Amazon wiring $50B, Nvidia $30B, and SoftBank another $30B. The deal also extends its AWS relationship with a $100B compute commitment and 2GW of Trainium reserved for the next-generation models, all while the pre-money valuation leaps to $730B from $300B in March 2025. TechCrunch’s deep dive notes that the market now expects multi-hundred-billion valuations to come with measurable network effects and differentiated distribution, so founders will need equally concrete margin defenses—more capital flows, but only for teams that can convert compute scale into unique, sticky outcomes.

Safety and compliance now determine who gets DoD clicks

The Pentagon quietly pulled Anthropic into a supplier blacklist while hosting OpenAI models inside its classified networks, all within hours of signaling the shift; the DoD praised OpenAI’s safety principles around mass surveillance and autonomous weapons, making safety alignment itself a gateway to contracts. Business Insider reports that Altman signed the deal while calling out how the Department of War was reshaping its supply chains in real time. Anthropic immediately pushed back in their public statement, arguing that the designation applies only to Department of War contracts and vowing to challenge any extension into commercial exports. The lesson for founders chasing government work: keep your commercial build and the classified runtime isolated, so one abrupt exclusion can’t cascade into every customer segment, and keep compliance docs ready to show how your safety guardrail matches or beats the DoD’s framing.

Claude embeds the enterprise workflow with Cowork & Plugins

Anthropic just made Claude Cowork & Plugins available inside Excel, PowerPoint, Slack, and the usual productivity suite, with connectors to Google Drive, Gmail, DocuSign, and partner-built plugins from FactSet, S&P, and Slack. Business Insider notes that L'Oréal, Deloitte, and Thomson Reuters already have bespoke agents, private plugin stores, and hard admin controls. The fastest path to enterprise adoption now runs through the apps knowledge workers already live in: ship connectors, copy the templates they already trust, and surface plugin marketplaces inside the same Excel/PPT/Slack panes so your customers never have to guess which window holds the agent control plane.

Quick hits from the field

Cloudflare's new Web Streams API can already run between 2× and 120× faster than the reader/lock schema across Workers, Node.js, Deno, Bun, and browsers, which makes stream-heavy workloads much more productive.
Max Woolf’s Opus 4.5 + Codex 5.3 experiment reports 2–10× faster UMAP, 23–100× faster HDBSCAN, and a new nndex vector store that beats NumPy on single queries and keeps agents honest.
California’s AB1043 requirements force every OS vendor to gather age info at account creation and tag apps with that segment, so vertically integrated stacks must budget for OS-level gating now or be blocked from 40M+ users.
JonMagic’s secret-injection playbook shows how to treat .env files as templates and call a helper like ./with-1password.sh node server.js so credential rotation becomes a single vault edit.
AdderBoard’s 36-parameter transformer leaderboard now proves perfect 10-digit addition with tiny models, teaching digit routing, parabolic logits, and tied embeddings that map straight into device-grade agents.

Celebratory spark of a team that just shipped another issue