CapEx, OpEx and the depreciation argument

Two true numbers

In the first quarter of 2026, xAI reported an operating loss of about $2.5 billion on revenue of roughly $800 million. Part of its hardware was Colossus 1, a cluster of computers in Memphis. A few months later, a tenant reserved that cluster, plus capacity in a second one, for fees at an annualized rate of $15 billion. Colossus 1’s cash costs are a few hundred million. On paper, if that cluster carries the whole fee, the margin is near 98 percent.

The two numbers do not cover the same scope. The loss is the whole company, and the margin is one rented cluster. But the cluster’s hardware did not change between them. A signature changed. These two true numbers teach how AI infrastructure economics actually work. They make you keep two ideas separate that most discussions mix: cash is a fact, but profit is partly a choice.

This essay examines that difference with real, reported numbers. It ends with the one calculation that is worth remembering.

A machine that costs forty years of itself

First, the shape of the thing. Epoch AI made a bottom-up model of a one-gigawatt facility. It costs about $38 billion to build and about $0.9 billion a year to operate. Build cost is roughly forty times a year of running cost.

That ratio is very unusual:

• An airline’s aircraft cost a single-digit multiple of their annual operating bill. Fuel and crew are the largest costs.
• A semiconductor fab is the classic capital-intensive business. Even there, the materials and labor bills are a meaningful fraction of its depreciation.
• An AI data center is closer to a bridge than to a factory. Nearly all of the cost occurs on the day that the owner commits the capital. The costs in the decades after are very small in comparison.

A check from the flow side gives the same ratio, and this type of cross-check is a good habit. Epoch’s annualized cost for the same facility is $8.5 billion a year. The $0.9 billion of cash operations inside it is about eleven percent. All the rest is capital recovery. The stock view and the flow view are two different accounting frames, and they agree. This business is ninety percent about capital.

One consequence follows immediately. Electricity is the operating cost that gets most public discussion. In both preset scenarios of the model below, it is one to three percent of revenue. The power bill decides almost nothing about whether these businesses work. The capital decides it, and above all, the speed at which the owner charges the capital against income.

The declared number

That speed has a name, depreciation. The important point is that nobody measures it. Management declares it.

Look at the declarations over time:

• Amazon carried servers at three years and moved to four in 2020.
• It moved to five in 2022.
• It moved to six in 2024.
• Effective the start of 2025, it moved a subset back to five years. It gave the pace of AI development as the reason.
• Microsoft moved from four to six.

The physical silicon did not change on any of those dates. Billions of dollars of reported profit did. Critics argue that the honest number for a frontier training GPU is two or three years. That is roughly how long a GPU stays at the frontier.

The defense is the value cascade. When newer parts replace a GPU in training, the GPU does not become scrap. It moves to inference, a cheaper market, and continues to earn for years. Yesterday’s flagship phone becomes today’s midrange phone in the same way.

Both positions are reasonable, and nobody outside the company can check either one. The difference between them is worth more than all the operating lines together. You can check that claim yourself. In the model below, move the book life between two and seven years. The business changes from a comfortable book profit to a heavy book loss. Its cash flow does not change by one dollar.

Memphis, worked honestly

Now the clearest public example, because here the sources report both sides of the ledger.

The machine is Colossus 1, the first phase of xAI’s Memphis site. xAI built it with 100,000 H100s in 122 days and later expanded it. The full Colossus complex reached 555,000 GPUs. Introl reports that the GPUs cost about $18 billion, an average near $32,400 per accelerator. Outside estimates put Colossus 1’s own build near $7 billion.

The contract: in May 2026, Anthropic reserved Colossus 1 for $1.25 billion per month. Colossus 1 is more than 220,000 GPUs and 300 megawatts. The SpaceX IPO filing gives more detail on the terms:

• The fee covers capacity across Colossus and Colossus II.
• Payments continue through May 2029.
• Capacity ramped in May and June 2026 at a reduced fee.
• Either party may end the agreement on 90 days’ notice.

The payment buys reserved capacity, not usage. While the agreement runs, Anthropic pays even if no job runs.

Put the two together at a five-year book life. Assume that Colossus 1 carries the whole fee at the full rate, and that neither party ends the agreement. The arithmetic is almost too good:

• Revenue of $15 billion a year.
• Electricity of about $160 million, at 300 MW and 90 percent utilization.
• Other cash costs of perhaps $140 million.
• Cash payback of the estimated build in under six months.
• A book operating profit above $13 billion a year.

Numbers that good need their caveats stated plainly:

• The $7 billion is an outside estimate. It is suspiciously close to the cost of 220,000 accelerators alone at xAI’s own average. Thus it may cover silicon and not much site. For that reason, the build cost is a slider in the model, not a constant.
• The GPU count, the power and the build estimate are for Colossus 1 only. Part of the fee pays for Colossus II capacity. Thus the true revenue per Colossus 1 GPU is lower than shown, by an amount that is not public.
• Nothing here includes financing, and at this scale interest is not a small item.
• A contract is a counterparty, not a certainty. With 90 days’ notice on either side, this one is less certain than its end date suggests.

But the most important lesson comes from the before-and-after. In the quarter before the signature, its owner reported the $2.5 billion operating loss from the start of this essay. That loss covers all of xAI, not only this cluster. Cost discipline did not close that gap and could not have. Against the capital, every cash cost on the books is almost zero. A tenant closed the gap. In this business, utilization and contracts do not only change the outcome. They are the outcome.

The number that travels

You can reduce any compute deal to one comparable figure. Make this a habit. Companies select headline numbers for effect, but they cannot select this one.

Divide the money by the GPUs and the hours. For the Memphis reservation, assume that Colossus 1 carries the whole fee: $15 billion a year, across 220,000 GPUs, at 8,760 hours per year, is $7.78 per GPU-hour. Part of the fee pays for Colossus II capacity, so this is an upper bound.

Now you can compare the deal with the market:

• Reported one-year H100 contracts moved from about $1.70 to $2.35 per hour into early 2026.
• GB200 reservations start near $2.63.
• A plausible blend for this fleet is near $3.50 per GPU-hour.

Thus the reservation price is at most roughly twice the going contract rate. On-demand prices for the identical H100 are from $0.80 to $11.10 across providers. That fourteen-fold spread shows that nobody should quote “the market price” without context.

Is a rate up to twice the market an overpay? That is the wrong question, and the reason is the point of the exercise. An exclusive reservation of a full named cluster is available on a signature, for a fixed fee. It is a different product from an hourly instance. A chartered aircraft and a seat on an aircraft are different products in the same way. The division does not tell you that the deal is expensive. It tells you what the buyer actually purchased: not compute, but assured access. In a market where the queue for power is years long, assured access carries a premium. Now you can see an upper bound on its size.

Declared numbers, real consequences

Look at the full pattern. The AI build-out depends to a remarkable degree on numbers that people declare and do not measure:

• the useful life of a GPU
• the basis of a gigawatt
• the scope of a cost estimate.

Each declaration is defensible, and each moves billions. A quarter later, people often treat each one as a fact of nature.

The defense is not cynicism. It is the habit that this essay used. Find the declared number inside any figure that you see, and ask what happens when it moves. Sometimes the answer is nothing, and the figure is sturdy. Sometimes, as with book life, the full conclusion is inside the declaration. In this industry, most of financial literacy is to know which type of number you have. One division and one slider show you.

What to take away

1. Build cost is about forty times annual running cost. These businesses are about capital allocation. The operating costs, electricity included, are almost decorative.
2. Management declares depreciation and does not measure it. One company moved it from three to four, five and six years, then back to five. When you read AI profits, first ask what book life they use.
3. Cash and book separate early and by a large amount. xAI reported a $2.5 billion quarterly loss. Months later, its Memphis cluster showed a margin near 98 percent on paper. The difference was a tenant, which is to say utilization.
4. Reduce every deal to dollars per GPU-hour. The Memphis reservation is at most $7.78 against market contracts near $3.50. The premium is the price of reserved capacity.
5. The consequential numbers in this industry are declarations. Find them before you trust anything that uses them.