Is AI a bubble? Eight signs in the filings

The question behind the question

The question “is AI a bubble?” is really two questions. The first is whether the technology is real, and the filings leave little doubt that the demand is: the cloud businesses of Amazon, Microsoft and Alphabet are growing faster than they were a year ago, and the companies selling chips, memory and power report record revenue. The second is whether the spending has run ahead of what that demand can pay for, and whether the money funding it is fragile. That is the question bubbles are made of, and it is the one the filings can help answer.

Bubbles are rarely about whether a technology matters. Railways, electricity and the internet all mattered, and each still had a period when capital poured in faster than revenue could justify, financed in ways that only worked while prices kept rising. The signs show up in companies’ own reports before they show up in share prices: cash flows that no longer cover spending, suppliers lending to customers, commitments recorded outside the balance sheet, and accounting choices that make the economics look better than the cash does.

So rather than a prediction, this page tracks eight of those signs, each computed from the same SEC data as the rest of the AI stack map. For each one it shows the latest reading, the reading a year earlier where the filings allow a comparison, and what would make it more worrying. As of the latest filings, 4 of the eight look more stretched than a year earlier; the rest are either steady or too new to compare.

The case that it is getting ahead of itself

The case that the build-out is getting ahead of itself starts with cash. In Q2 2026, Microsoft, Amazon, Alphabet, Meta and Oracle together spent $182B on capital projects against $186B of operating cash flow: 97% of it, up from 72% a year earlier and 35% at the low point in Q3 2023. Amazon and Oracle are already spending more than they generate, with free cash flow of −$11.6B and −$28.7B over the latest four quarters.

The spending is also growing faster than the revenue it is meant to earn. Capex rose 87% on the same quarter a year earlier, while revenue at AWS, Microsoft Cloud and Google Cloud combined rose 38%. The cost of all that equipment has barely started to arrive: the five buyers’ capex is 3.9 times their depreciation, so depreciation will keep climbing for years even if spending stops growing.

Then there is what sits outside the cash-flow statement. Five companies disclose $1.11T of data-center leases signed but not yet started. Suppliers have extended $342B of equity and credit to their own customers, from Nvidia’s guarantees of OpenAI’s leases to Amazon’s stakes in both OpenAI and Anthropic. And 13 of the 25 financing links on our money-loop map run to or from those two private labs, whose ability to pay depends on raising more money.

The case that it isn’t

The case against starts with revenue. The core cloud businesses grew +38% over the latest year, faster than the +24% of the year before: demand is accelerating, not fading. AWS earned $54.7B of operating income over its latest four quarters and Google Cloud $31.8B, +198% on the year before, so the largest sellers of AI capacity are making money on it, not just revenue.

The buyers are also among the most profitable companies in the world, not start-ups. Three of the five still fund their spending entirely from operating cash flow, and many of the leases and power contracts behind the build-out run for fifteen or twenty years, which makes the capacity harder to cancel but also gives the buyers time to fill it.

And some of the accounting has moved in the conservative direction. After years of lengthening how long they depreciate servers, which flatters profits, Amazon has shortened its server life again, although Meta and Oracle extended again in 2025. A company that expects its hardware to become obsolete quickly shortens the life; one that extends it is betting the hardware keeps earning.

Capex as a share of operating cash flow

Capex took 97% of the five buyers’ operating cash flow in Q2 2026, against 72% a year earlier. Below 100%, the spending is paid for out of the business; above it, someone else has to lend or invest.

Why it matters: Spending paid from cash flow depends only on the business. Spending that needs borrowing also depends on lenders, which ties the build-out to credit markets staying open.

What would make it worse: A reading above 100% for the five buyers combined, meaning they borrow or run down cash to keep building.

Capex growth minus cloud revenue growth

In Q2 2026, capex was 87% higher than a year earlier and core cloud revenue 38% higher. A year before that, the same comparison was +75% against +24%. The capex payback calculator shows how long this spending takes to earn back at today’s cloud growth.

Why it matters: Capacity has to be sold. When spending grows much faster than the cloud revenue it is meant to produce, the gap has to close later, either through faster revenue growth or slower spending.

What would make it worse: Cloud revenue growth slowing while capex growth holds up: the gap widening from both sides.

Depreciation still to come

The five buyers recorded $165B of depreciation and amortization over their latest four quarters, about twice what they recorded two years earlier, while spending $566B on capex. The depreciation calculator shows how this year’s capex becomes next year’s cost under different useful lives.

Why it matters: Today’s capex becomes tomorrow’s depreciation. The further capex runs ahead of depreciation, the larger the cost wave still to reach income statements.

What would make it worse: Operating margins at the cloud segments falling as depreciation catches up, without a matching rise in revenue.

Buyers with negative free cash flow

Free cash flow over each buyer’s latest four quarters (operating cash flow minus capex): Microsoft $60.0B, Amazon −$11.6B, Alphabet $19.7B, Meta $28.6B, Oracle −$28.7B. Negative free cash flow means the gap is being filled by borrowing, cash on hand or leases.

Why it matters: A buyer with negative free cash flow is funding the build-out with debt, leases or equity. That works while lenders are willing and stops quickly when they aren’t.

What would make it worse: A third buyer turning negative, or new bond issues and lease financing growing faster than capex.

Leases signed but not yet started

Microsoft, Meta, Oracle, Alphabet and CoreWeave disclose leases that have been signed but have not yet started, totalling $1.11T. They are not yet capex and not yet on balance sheets, so the usual spending figures understate how much has been promised.

Why it matters: These leases are not on balance sheets or in capex yet, but they are commitments. They show how much of the build-out is already promised beyond what the cash-flow statements show.

What would make it worse: The total rising faster than revenue, or leases being renegotiated or cancelled.

Seller-financed demand

Our money-loop map lists 10 investments, loans and guarantees totalling $342B that suppliers have extended to companies that buy from them. 13 links in all run to or from OpenAI and Anthropic; see who funds OpenAI and who funds Anthropic.

Why it matters: When suppliers invest in, lend to or guarantee their customers, part of the demand they report is demand they helped pay for. It is a normal tool, but it inflates how healthy demand looks.

What would make it worse: Guarantees being called, investment write-downs, or new links that are larger than the revenue they support.

Backlog far ahead of revenue

Remaining performance obligations, the contracted revenue a company has not yet earned, are the strongest evidence of future demand, and in AI they have become very large relative to revenue at a few companies. Oracle’s 10-Q says how slowly its backlog turns into revenue:

Why it matters: A large backlog is good news only if the customers behind it can pay. When most of it depends on a few customers that are themselves funded by investors, its quality matters more than its size.

What would make it worse: Backlog shrinking without converting to revenue, or a large customer renegotiating.

• Oracle: Share of backlog due within a year

  “Remaining performance obligations were $664 billion as of August 31, 2026, of which we expect to recognize approximately 13% as revenues over the next twelve months, 37% over the subsequent month 13 to month 36, 34% over the subsequent month 37 to month 60 and the remainder thereafter.”

  Oracle 10-Q, filed September 11, 2026

Server useful lives

The latest useful-life changes in the buyers’ filings: Microsoft 4 → 6 years (July 2022); Amazon 6 → 5 years (January 1, 2025); Alphabet 4 → 6 years (January 2023); Meta 5 → 5.5 years (fiscal year 2025); Oracle 5 → 6 years (fiscal 2025).

Why it matters: Longer depreciation lives raise reported profits without changing cash. Extending them while AI hardware turns over faster flatters earnings; shortening them is the conservative direction.

What would make it worse: A new extension by any of the large buyers, or write-downs of older accelerators.

If it bursts, what would that look like?

If the build-out has run ahead of demand, the filings suggest the pressure would show up in a fairly predictable order. First, a large customer would slow its purchases or renegotiate: the backlogs at Oracle and CoreWeave are concentrated enough that a single change would show. Second, the sellers who financed that demand would take the losses first, through guarantees called, capacity buy-backs triggered or investments written down. Third, the buyers would cut capex, which their cash flows would allow in a quarter or two, and the effect would run down the stack to chips, memory, servers and power equipment, in roughly that order of speed.

None of the indicators on this page predict when that would happen, and some may never trigger. What they do is tell you where to look. A bubble tends to deflate when funding becomes harder to get, so the indicators tied to funding (negative free cash flow, seller financing and dependence on the two private labs) are the ones to watch most closely. The quarterly AI Stack Report updates every one of them after the five largest buyers file.

History rhymes, up to a point

The parallel investors reach for most often is the late-1990s telecom build-out, when equipment makers lent heavily to the carriers that bought their gear and much of the fiber laid then went unused for years. Some of the patterns on this page are the same: suppliers financing customers, capacity built ahead of revenue and long commitments signed on optimistic forecasts.

There are also real differences. The telecom buyers were often young companies funded by debt, while today’s largest buyers are highly profitable and fund most of their spending from cash. The demand is also visible in revenue now rather than projected: cloud revenue is growing and the cloud segments are profitable. Where the comparison is closest is at the edges, among the AI labs and GPU clouds funded by investors and suppliers rather than by their own profits.