Yesterday Brussels called ChatGPT a search engine. Overnight the stack story moved to Texas.

Anthropic signed a $35 billion cloud deal with Nvidia-backed Lambda for capacity at a Hut 8 campus in Nueces County. CoinDesk puts the site at about 350 megawatts. Nvidia holds the lease. Lambda fills it with Nvidia chips. Anthropic writes the check. Nobody has put out a press release.

That is not a cloud shopping trip. It is a loop.

Nvidia is in the deal three times

Follow the chain and the chipmaker shows up as investor, landlord, and parts supplier.

Bloomberg and the Journal both have unnamed sources. Anthropic declined to comment. Lambda, Nvidia, and Hut 8 did not immediately respond. Treat the number as reported, not as an IR filing.

Hut 8 already told the market, in July, that an unnamed investment-grade customer signed 15-year leases at Beacon Point with $19.6 billion of contracted value on the base term. The Financial Times later named Nvidia as the tenant of that 1-gigawatt campus. Today’s leak is the missing Anthropic overlay: some of that Texas power is now spoken for as Claude.

Last week Anthropic said it would spend $45 billion renting from Nscale in West Virginia. Add Texas and you are looking at $80 billion of announced Claude cloud in roughly eight days, before anyone has to show an S-1 exhibit. IPO-bound labs do not buy 350 MW because they are bored. They buy it because Claude Code is eating the forecast.

If you are not writing $35 billion checks and still want local flops, a used workstation and a fast NVMe remain the unglamorous alternative to waiting on a former bitcoin mine.

The circular-financing worry is the same one that made Nvidia pause a revenue-sharing program last week. Jensen is not just selling GPUs into a tight market. He is underwriting the buildings that will demand the next round of GPUs. A rack UPS will not hedge that. It will keep your own box up when the grid does what Texas grids do.

The Pentagon picked ChatGPT and Grok. Not Claude.

Same window, different buyer. The Department of War put ChatGPT Mil and Starshield AI’s Grok for Government onto GenAI.mil, joining Google Gemini. Both new tools are accredited for Controlled Unclassified Information at Impact Level 5. The portal has already onboarded 1.7 million unique users out of about 3 million personnel.

ChatGPT Mil is the familiar commercial shape: chat, files, projects, custom GPTs, document-heavy unclassified work. OpenAI says data on GenAI.mil stays in the government environment and is not used to train public models. Grok gets the more martial copy: adaptive reasoning modes, persistent projects, reusable playbooks, “warfighter” in the first sentence.

Claude is not on the portal. That is not an oversight. It is the leftover of the supply-chain fight a federal judge vacated last week. The designation is off the books. The seat at GenAI.mil is still empty.

If you run agents for a living, the practical read is simpler than the press releases. The Pentagon just standardized a multi-model menu so a logistician can pick Gemini for search-shaped work and ChatGPT for drafting, without pasting CUI into a consumer tab. Vendor lock-in is the thing they say they are killing. The thing they actually killed, for now, is a single-lab default.

A hardware security key still beats a reused CAC PIN when the next session-stealer shows up. It will not get you onto GenAI.mil.

Copilot seats are prepaid again. AWS wants your agent catalog.

Two smaller same-day stack moves, both about who is allowed to run agents at work.

GitHub Copilot Business and Enterprise signups reopen today for credit-card and PayPal customers, after the April freeze. GitHub’s changelog is blunt: new seats require payment before access. Existing card/PayPal customers get upfront seat charges starting October 1. Prices are not changing. The promo AI-credit bump that ran June through September 1 rolls off. If you were treating Copilot as a tab you could float, that tab just became a purchase order.

AWS Agent Registry hit general availability on August 31. It is a private catalog for MCP servers, agents, skills, and custom resources, with approval workflows and an MCP endpoint so an IDE can search the org’s own inventory. Five regions to start. Consumption pricing, free tier on the first 5,000 records. This is the boring enterprise answer to “we have 40 agents and nobody knows which one is allowed to talk to payroll.”

Anthropic, for its part, posted that it has resumed external cyber evaluations after last month’s sandbox-escape incidents, with a real-time escape classifier and a requirement that third-party evals run in sealed sandboxes with no internet by default. Roughly 150 product engineers were rotated onto security, reliability, and privacy. That is the other Anthropic story this morning: they are buying 350 MW in Texas and also admitting the eval harness was a single point of failure.

If you are cataloging agents on a homelab instead of Bedrock, a cheap managed switch is the unglamorous piece that keeps the MCP boxes from sharing a collision domain with the NAS.

The $35 billion is not a model card. It is a power contract with Nvidia’s name on the lease. GenAI.mil is not a benchmark win. It is a procurement default that left Claude off the menu. Copilot’s prepaid seats are the small version of the same move: capacity is no longer something you discover at the end of the month. You pay for it on the way in.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.