Anthropic Committed 80 Billion Dollars to Compute in About a Week

Eighty billion dollars in about a week. Anthropic signed a 35 billion dollar cloud agreement with Lambda on 31 August, days after committing 45 billion to rent capacity from Nscale in West Virginia.

What the Anthropic Lambda data center deal covers

The Lambda capacity sits at a new site in Nueces County, Texas, drawing roughly 350 megawatts, and covers both training and inference. Anthropic does not own the building, the chips or the power. It is renting all three, on contracts long enough to be counted in tens of billions.

That is the part worth sitting with. A company valued on the quality of its models is making its largest financial commitments to property and electricity.

Nvidia is on three sides of the same table

The structure is the story. Hut 8, which pivoted out of crypto mining, is developing the data centre and leasing space to Nvidia. Lambda, in which Nvidia holds equity, installs Nvidia chips and sells the resulting capacity to Anthropic. On a campus planned for 700 megawatts, Nvidia is the chip supplier, the investor and the lease anchor at once.

Several outlets have started calling arrangements like this circular financing. That is a description rather than an accusation. It does mean the demand side and the supply side of the AI buildout are now tied together tightly enough that a wobble in one shows up in the other.

Why the power figure matters more than the dollar figure

Money can be raised quickly. Megawatts have to be permitted, built and connected, and that timeline does not compress. The reason these deals keep landing in Texas and West Virginia is that both have power available sooner than the alternatives.

For the hardware sitting inside these buildings, our pieces on AMD Helios against Nvidia Vera Rubin and what HBM4 memory actually does cover what is being installed and why it costs what it costs.

What to watch

Two things. Whether the next announcement names a power source rather than a dollar figure, and whether any of these leases begin before the capacity physically exists. Our semiconductor revenue forecast for 2026 covers the demand side of this, and the gap between committed spend and delivered megawatts is where the strain will show first.