Anthropic's gigawatt: when compute is measured in power plants

When Anthropic and Google announced their expanded deal, the number that travelled wasn't a count of chips. It was a unit of power. Anthropic was getting access to more than a gigawatt of capacity. Read that twice. A software company described its growth in the same units a utility uses to describe a power station. That shift in vocabulary is the whole story.

The deal, in numbers

On 23 October 2025, Anthropic confirmed an expanded agreement to use up to one million Google TPUs, bringing more than a gigawatt of compute capacity online in 2026. The arrangement is reported to be worth tens of billions of dollars (Anthropic, CNBC).

For scale: industry estimates put the all-in cost of building a single gigawatt of AI data center near $50 billion, of which roughly $35 billion is the chips themselves (Data Center Dynamics). A gigawatt is not a procurement line item. It is a piece of national infrastructure that happens to be privately financed.

Why "gigawatt" replaced "how many chips"

For years, the AI arms race was counted in accelerators: how many H100s, how many GPUs in the cluster. That framing has quietly died. The leaders now talk in watts, because power, not silicon, is the thing they cannot conjure on demand. You can place a larger chip order. You cannot place an order for a gigawatt of dependable electricity and have it show up next quarter.

So the unit of ambition migrated to the genuinely scarce resource. When Anthropic says "a gigawatt," it is not being poetic. It is naming the actual ceiling on how big its models can get. Chips are a check you write. Power is a place you have to find, build, and feed for a decade.

The cost that compounds quietly

Here is what the gigawatt framing exposes. A chip is bought once. The power to run it is bought every hour, for years. Over a GPU's useful life, the electricity bill can rival or exceed the price of the hardware. That means the single number with the most leverage over the economics of AI is not the sticker price of the accelerator. It is the price of a kilowatt-hour, multiplied by every hour the machine runs.

Anthropic's deal makes this concrete at the largest scale. But the same arithmetic governs a single rack. If power is the denominator of AI, then the cheapest reliable power wins, not by a little, but compounded across every hour of every year the cluster is alive.