The Labs Gave Up the Building and Moved Into Yours

In Abilene, Texas, the country's flagship AI campus got smaller while everyone watched. Oracle and OpenAI scrapped the planned expansion from 1.2 to 2 gigawatts after the financing talks collapsed, capping construction where it already stood; Nvidia paid a $150 million deposit just to keep its chips in the portion that survives, and Meta began negotiating to lease the abandoned space ^[1]. It is one plot of dirt, and it runs the whole year's ledger. The shed column came first. In January, OpenAI told eleven governments to build the data centers themselves — the "OpenAI for Countries" pitch was installation without ownership, the company handling the software while the host pours the concrete ^[2]. The compute OpenAI still needs, it now rents back — from AWS and Oracle — after the financing difficulties ^[3]. Its CFO warned the company may struggle to fund the compute contracts it has already signed ^[4]. The hinge is the UK, where the whole swap got compressed into one announcement: when OpenAI paused the Stargate UK project over energy costs, the same statement insisted the public-services integration would continue under an earlier memorandum of understanding ^[3]. The campus was shelved; the installation into the British state was not.

AI compute is foundational to that goal — we continue to explore Stargate UK and will move forward when the right conditions, such as regulation and the cost of energy, enable long-term infrastructure investment. — OpenAI

So while the dirt changed hands, the labs were moving into the other side. Both opened implementation subsidiaries — OpenAI's Deployment Company and Anthropic's Ode — framed explicitly as moving value capture from model training toward the integration of AI into the services sector, more than three-quarters of U.S. GDP ^[5]. Claude sits inside Excel, PowerPoint and Slack ^[6]. ChatGPT sits in Estonia's secondary-school classrooms ^[2]. A leaked finance section would put it inside bank accounts ^[7]. Anthropic opened a wet lab where Claude directs robotic units through experiments ^[8]. The engine pushing all of it: raw intelligence got cheap. DeepSeek's permanent 75 percent cut made equivalent work 12 to 19 times cheaper than the American frontier models ^[9], and the world's token usage followed the cheap models ^[10]. The labs spent the year cutting prices while chasing profitability ^[11], and analysts read the price war as pushing the money out of the model and into the workflow ^[9][5]. Anthropic's product chief put the denial on the record, in case anyone mistook this for a land grab.

It’s not a product that’s trying to own every workflow. — Scott White

The base the labs shed did not vanish. It landed on parties built to carry it. The UAE is redesigning its Stargate with blast-resistant concrete and air defenses after an Iranian armed-forces video named the first-phase cluster as a target — a state bolting armor onto what began as a company's showcase campus ^[12]. Globe Telecom will spend at least $1 billion on data centers under Manila's keep-it-domestic mandate ^[13]. Bell Canada and Cisco are pairing up to sell sovereign AI on Canadian soil ^[14]. Nvidia is turning chips into an insurable, financeable asset class so smaller firms can borrow against them ^[15]. And Oracle, the strained landlord, now holds a record backlog with nearly half its remaining obligations tied to a single tenant — OpenAI — whose own miss already slumped the stock ^[16]. The risk migrated with the watts, and carrying it shows. The institutions being installed into are already drafting the uninstall. India's central bank proposed a kill switch to instantly override, suspend or deactivate AI models, citing supply-chain risk from dominant providers ^[17]. The Financial Stability Board proposed human-approval gates for high-risk transactions, because agents can bypass human intent ^[18]. The UK's competition regulator held businesses responsible for consumer-law compliance no matter who built the agent ^[19]. None of this is hypothetical practice: KPMG's AI chief requires a final manual approval before any agent executes a transaction, and puts it plainly.

You need to build agents to check your agents. You need to check the outputs. — Todd Lohr

Millennium's digital twins, rolling out across its 7,000 employees, are barred from making decisions on their behalf ^[20]. The Indian and stability-board rules are drafts, not yet in force; the UK ruling is the one already on the books. The labs say they aren't trying to own the workflows. What the regulators' drafts are quietly pricing now is whether the installed layer is detachable. The year left one demonstrated precedent: the physical layer came out fast when the money failed, and gigawatts proved removable. Software threaded through a ministry, a classroom, a bank account is a different material.