AI Compute Was A Verb Twenty Years Ago. Now It's A $500 Billion Asset.

A finance lead asks a vendor what they are buying for an AI project. The invoice says "AI compute." The line items underneath say chips, memory, networking, power, cooling, software, and a financing structure. None of those line items is "AI compute" by itself. And the more than $500 billion Nvidia and six of the largest Wall Street firms signed memorandums to mobilize this August is being written against the word at the top of the page, not the parts that age out underneath it.

"AI compute" used to be a verb. Computing meant to calculate, and the machines that did it were a back-office cost. In 2006, Amazon began selling processing capacity by the hour, and "compute" became a unit you could buy like electricity. A few years later, when a new generation of models needed banks of specialized chips to train, the word narrowed further. In the trade, "AI compute" stopped meaning any computer and started meaning a specific class of hardware, almost all of it supplied by one company that came to supply more than 90% of the world's data center GPUs.

Most of the bill is not the chip. A working AI stack is the chip plus the memory that holds the model's weights, the networking fabric that ties thousands of chips into a single training run, the power substation and the cooling system that keep the rack alive, the orchestration software that schedules work across the cluster, and the financing structure that puts the whole thing on a balance sheet. Each component has its own depreciation curve. The chips get refreshed on a roughly two-year cadence. The networking and software layers age on a different clock. The substation and cooling plant are infrastructure that lasts a decade or more, and they cannot be redeployed when the model architecture changes.

The August memorandums are priced against that gap. According to a Forbes contributor analysis, Nvidia and six of the largest firms on Wall Street signed agreements in August 2026 to mobilize more than $500 billion of outside capital around the category of AI compute. The memorandums are not the same as commitments; they are subject to execution, and the dollar target is the headline figure, not a deployed figure. What they show is that the category label, not the line item, has become the unit of finance.

Cloud bills used to meter compute by the hour. A buyer could read the invoice and know what they had consumed. The same word is now doing two jobs. On a vendor invoice, "AI compute" is a unit of consumption priced by the hour or by the token. On a financing term sheet, "AI compute" is a category of asset that a firm buys outright, depreciates over years, and carries on its books. The vendor invoice and the term sheet can be written against the same word and describe two different things.

The capital is being mobilized at the category level, against a label the people writing the checks may not be able to decompose. The source frames the bet as a bet on aging hardware: the chips are the pictured part, but the chipmaker will only backstop part of what the category turns out to be worth, and the surrounding stack ages out on a different clock. The people signing the checks need to know which parts of the stack they are actually buying, and the term sheet does not tell them.

Before evaluating one of these memorandums, ask the older question: when the document says "AI compute," is it naming a chip, a stack, or a financing category? The answer determines which depreciation curve the money is on, and which parts of the bill the chipmaker is actually on the hook for. The word has been doing too much work for too long, and the August memorandums are the first time the bill has been written at the scale where the ambiguity is the risk.