AI Demand Starts Paying for Data Centers

AI revenue has reached a level that can cover data-center depreciation for the second straight quarter.

Artificial intelligence is finally producing enough revenue to cover a big chunk of the hardware bill. In the first quarter of 2026, global AI sales outside China hit $25 billion, while estimated depreciation tied to data centers and chips came to $21 billion.

That matters because the biggest US tech companies are still on track to spend as much as $725 billion this year on capital expenditures, most of it tied to AI infrastructure. The money is flowing into chips, server racks, power, and the buildings that keep the whole system running.

AI spending is finally meeting a real revenue stream

The core question hanging over the AI boom is simple: can demand catch up with spending? The latest data from Exponential View says yes, at least for now. Its report says AI sales have now cleared depreciation costs for two quarters in a row, which is a much better sign than the usual hype cycle where revenue lags far behind infrastructure buildout.

That does not mean the business is flush with profit. Depreciation still eats more than two thirds of revenue, so there is not much room left for power bills, staff, financing costs, and the inevitable mistakes that come with building too much too fast. Still, the direction is moving the right way.

The report is based on a dataset tracking spending across more than 1,000 companies, using filings, executive comments, press coverage, and cloud disclosures. It also tries to avoid double-counting across the AI supply chain, which matters because the same dollar can show up multiple times if you are not careful.

• AI sales outside China: $25 billion in Q1 2026
• Estimated depreciation: $21 billion
• Generative AI revenue over the last 12 months: $110 billion
• Big US tech capex plan: up to $725 billion this year

The hardware bill is still enormous

The spending numbers are hard to ignore. Meta, Alphabet, Microsoft, and Amazon are collectively planning to spend up to $725 billion on capital expenditures this year, according to the Yahoo Finance report based on Bloomberg’s coverage. A huge share of that goes into AI infrastructure, especially data centers and the chips inside them.

Azeem Azhar, founder of Exponential View, told Bloomberg News that the economics are just about clearing the depreciation hurdle and improving over time. His point is important: if the companies were already printing huge margins on the first wave of AI infrastructure, they would probably be underinvesting instead of overinvesting.

“It just about clears the depreciation hurdle, and roughly speaking, it’s improving over time.” — Azeem Azhar, founder of Exponential View

The problem is that the margin for error is still thin. If demand slows, if power becomes more expensive, or if chip replacement cycles shorten faster than expected, the economics can turn quickly. That is why the financing structure matters just as much as raw revenue.

Exponential View says more of the risk is shifting into capital markets through leases, debt, and equity, especially among the so-called neoclouds. In plain English: some companies are funding the AI buildout with borrowed money and outside capital, not just operating cash flow.

Old chips are not dying as fast as skeptics hoped

One of the strongest arguments against the AI spending spree has been depreciation. If GPUs lose value quickly, then the whole investment case gets weaker. Michael Burry, the investor famous for betting against the housing market before the 2008 crisis, has called understated depreciation “one of the most common frauds of the modern era.”

But the report says older hardware is holding value better than many critics expected. The rental price for an hour of access to Nvidia’s H100 chip remains close to 80% of its launch level. That suggests demand for compute is still strong enough to keep older accelerators useful longer than a doomsday model would predict.

Nvidia also matters here because its Blackwell chips are the new hot item, yet demand has been strong enough that older A100 servers have not been retired quickly at Amazon Web Services. Matt Garman, AWS chief executive officer, said in February that the company had not retired six-year-old Nvidia A100 servers because customers still wanted them.

• H100 rental price: nearly 80% of launch level
• A100 servers: still in use after about six years at AWS
• Blackwell supply: tight enough to keep older chips valuable
• Assumed depreciation life: six years for IT equipment

Cheaper models are changing who gets paid

The other big shift is on the demand side. OpenRouter, a platform that gives developers access to multiple models, shows the share of tokens requested from Google, OpenAI, and Anthropic fell to 33% in June 2026 from 72% a year earlier. That is a huge change in a very short time.

The obvious takeaway is that power users are chasing cheaper and faster models for simpler tasks. You do not need a top-tier reasoning model to pull a number from a receipt and drop it into an expense spreadsheet. That kind of work can move to lighter, cheaper systems without hurting the user experience much.

Azhar’s argument is that this does not kill the market for major foundation-model companies, but it does raise the bar for pricing. If the most common tasks get commoditized, the premium has to come from better tools, deeper integration, and stronger lock-in around workflows that are expensive to replace.

That creates a split market. One side is infrastructure-heavy and capital intensive. The other side is software-heavy and price competitive. The winners will probably be the companies that can sit in both camps without burning through cash too quickly.

The next test is margin, not demand

The headline number here is not just that AI revenue is growing. It is that revenue is finally large enough to make the spending look economically justifiable, even if barely. That is a meaningful shift from the early phase of the boom, when the market mostly talked about potential and supply constraints.

The next question is whether AI companies can move from paying for depreciation to paying for the full stack: power, labor, financing, replacement cycles, and profit. If they can keep revenue above those costs while chip prices and demand stay firm, the buildout keeps its logic. If not, the current capex wave gets much harder to defend.

For now, the data says the AI boom is still expensive, but it is no longer living entirely on faith. The number to watch next quarter is simple: does AI revenue stay above depreciation, or does the hardware bill catch back up?