IT investment just passed its dot-com peak — and three of the four biggest spenders stopped paying for it out of cash

Two things are true at once, and almost every argument about an AI bubble picks one and ignores the other.

The first: the money is real and it is enormous. The four biggest spenders put $510.7 billion into capital equipment in the twelve months to June 2026 — about 1.57% of the entire US economy, and roughly 32% of everything America spent on information-processing equipment and software. That line item is now 4.97% of GDP, the highest since the series began in 1947, above the dot-com peak.

The second: it has stopped being paid for out of earnings. A year ago, Meta, Alphabet and Amazon funded about a quarter of their capital spending with borrowed money. In the first half of 2026 they funded almost two-thirds of it that way. Amazon went from 1.3% to 68%. Alphabet stopped buying back its own stock and raised $105.8 billion of outside capital in six months instead.

What this report went looking for was the mechanism that breaks. It did not find the one everybody names. It found a different one, and it is not in the numbers on the income statement — it is in the obligations sitting just off the edge of the balance sheet.

What this run did differently, and why it matters

This question was researched once before, on 2026-08-21, and that pass is the reason this one exists. Web fetching was blocked in that environment, so no agent opened a single primary filing — every figure was search-surfaced text about a document. Of 45 load-bearing claims, only 3 survived unchanged: 21 needed correcting, 12 could not be verified at all, and 9 were refuted outright. The failure modes were consistent: stale figures quoted in the present tense, numbers misattributed between banks, and a dozen outlets restating one analyst note being mistaken for a dozen sources.

This run opened 83 primary documents. Of 33 load-bearing claims, every one was checked against a primary source, and the tally is: 3 confirmed clean, 30 confirmed with a correction, 0 refuted, 0 unverifiable.

That difference — 9 refuted and 12 unverifiable, versus none of either — is the single best argument for the method this section is built on. The corrections that remain are refinements of magnitude and basis, not demolitions. Where the earlier pass concluded that a power-and-transformer stall was the likeliest outcome, the physical data does not support it, and that conclusion has been dropped.

The finding: the risk moved, it did not disappear

Start with the thing that is genuinely unprecedented. Private fixed investment in information-processing equipment and software reached 4.97% of GDP in 2026Q2 — the highest of the 318 quarters BEA has published since 1947, above the 4.46% of the dot-com peak in 2000Q4 and far above 1995Q1's 3.24%. The top four readings in seventy-nine years are the last four quarters.

Worth being precise about the deflator, because it cuts the opposite way from the usual objection: IT prices fall while the GDP deflator rises, so a real-terms comparison flatters today rather than 2000. In real terms the gap is roughly 3.6×, against 1.11× in nominal terms. Recomputing in real terms reinforces the headline rather than reversing it.

Against that, the four biggest spenders put $510.70 billion into capital equipment in the twelve months to June 2026 — 1.57% of GDP at an annual rate, and 31.7% of the entire national information-processing equipment-and-software line. Two caveats travel with that 31.7%, one of which the original claim had backwards: data-centre shells and land are counted as structures, not IT equipment, so they cannot appear in the denominator at all — and excluding finance leases makes the capex figure too small, not too large. Microsoft alone added $24.6B of finance-lease right-of-use assets in FY2026 on top of the $115.9B, with a further $329.1B of datacenter leases not yet commenced.

Where the money now comes from

This is the change that happened inside twelve months, and it is the spine of the thesis.

That is the share of cash capital spending covered by long-term debt proceeds — $148.1B against $228.1B in the first half of 2026. Alphabet is the one that fell, which matters: the shift is not uniform, and stating it as a clean industry-wide inversion overstates it.

Two other filings sharpen the picture from opposite ends. Oracle funded roughly 98% of a 162% increase in capex with borrowed money — $46.09B of notes and term loans plus $3.35B of short-term capex financing plus $4.95B of mandatory convertible preferred, against zero debt issuance in FY2024. Its borrowings went $92.6B to $129.5B in one year and free cash flow was −$23.69B.

And Microsoft is the counterexample, at the largest scale of all: FY2026 capex of $115,948M with "Proceeds from issuance of debt: 0" — its second consecutive year at zero — while still repurchasing $22.3B of stock and paying $26.4B of dividends out of $182.9B of operating cash flow. Any thesis that says "the industry can no longer fund this from earnings" has to explain the biggest spender doing exactly that.

Alphabet sits in between, and its behaviour is the most interesting: it took repurchases to zero in H1 2026 (from $28.31B) while leaving its $70.0B authorisation untouched, and raised $30.50B of common stock — including a $10.0B private placement to Berkshire Hathaway — plus $19.06B of 6.25% mandatory convertible preferred, on top of $56.23B of debt. $105.8B of external capital in six months against $80.6B of capex.

The part that is not on the balance sheet

Here is the actual fragility, and it is not a valuation.

Leases signed and not yet commenced, at 30 June 2026: Microsoft $329.1B, Meta $279.0B, Amazon $137.2B, Alphabet $85.2B. Roughly $830 billion of contracted obligation that appears on no balance sheet because the leases have not started. For scale, that is many multiples of every data-centre securitization outstanding combined — the structure everyone watches is not where the exposure is.

Then the structures themselves. Meta's Louisiana venture — never named in the filing, just "the Venture" — is a 20% membership interest in roughly $27B of estimated development cost, with an aggregate initial lease commitment of ~$12.31B commencing 2029, and residual value guarantees with an aggregate threshold of ~$28 billion. Meta carries it at $2.92B and discloses a maximum exposure to loss of $46.03B. A 15.8× gap between the carrying value and the company's own statement of what it could lose.

The reason it stays off the balance sheet is a single narrow judgment, quoted from the filing: decisions about "remarketing the data center campus" were determined to have the most significant impact on the Venture's economic performance, and "we do not have the power to direct" them. Meta simultaneously provides the venture's construction management and property management services.

Ernst & Young made that a critical audit matter — the third of three in the FY2025 audit — citing the "significant judgment required in determining the activities that most significantly affect the VIE's economic performance." That is an auditor stating publicly that the answer is contestable. A second structure on the same template, in El Paso, was disclosed as a subsequent event with a further ~$13B of residual value guarantees.

Alphabet's version is different in form and the same in kind: credit-derivative backstops to data-centre entities went from $16.94B to $43.79B in six months — 2.6× — with a recognised fair-value liability of just $815M. On default, Alphabet "retain[s] the right to assume the underlying leases for internal use or to sublease to third parties."

What is not wrong, and this is most of the bear case

Three widely-repeated arguments did not survive contact with the filings.

Depreciation games are not live. Amazon shortened a subset of servers from six years to five effective 1 January 2025, taking $1.4B of extra depreciation and $1.0B of net income to do it, citing AI-driven technology pace. Microsoft is unchanged at "two to six years." Meta is the only lengthener — most servers to 5.5 years, worth $2.92B and $1.00 a share — and that was eighteen months ago and is already in the base.

The power wall is a rationing system, not a wall. The cohort test is in the "headline vs the filing" section above; the summary is that half the promised capacity arrived, the median slip was four months, and 0.7% was cancelled. The equipment market is genuinely tight — GE Vernova is booking gas turbine orders roughly 4:1 against shipments, and shipments actually fell year over year — and Siemens Energy's Grid Technologies book-to-bill of 1.48 on a €51bn backlog says transformers are queued and priced years out. But queued is not cancelled.

Credit stress is real and confined. The CCC & lower high-yield spread was 10.35% on 20 August 2026, the widest print of 2026 — but broad high yield was 2.75%, near its 2026 tight, and investment grade was 0.82%. And the CCC level is still below the 11.37% reached in April 2025, so this is a 2026 high rather than a multi-year extreme. The dispersion is the signal: the marginal borrower is being repriced while the index is not.

CoreWeave is the marginal borrower made concrete: $72.0B of total liabilities on $5.0B of equity, debt 90% recourse, and net interest expense of $640M — 24.9% of revenue — against an operating loss of just $(49)M. Revenue grew 112%; the net loss more than doubled to $(626)M, and interest is the largest single contributor. The neocloud problem is the balance sheet, not the gross margin. Its disclosed $5.1B of OEM and software license financing at 9–11% is vendor financing named as such.

Risk — each isolated, do not blur

1. Off-balance-sheet obligation (dominant). ~$830B of uncommenced leases, ~$41B of announced residual value guarantees at Meta alone, $43.8B of Alphabet backstops carried at $815M.
2. The consolidation judgment. One auditor has already flagged it. If a structure of this type consolidates, the change lands on the balance sheet of the largest companies in the index at once.
3. Funding-mix reversal. 25% to 65% in a year, with Amazon's trailing free cash flow now negative.
4. Backlog conversion. Oracle's $638B at ~12% inside twelve months is the weakest near-term conversion of any filer opened.
5. Marginal-borrower credit. Confined to CCC today; the transmission risk is that it stops being confined.
6. The productivity gap. Utilization-adjusted TFP negative while investment hits a 79-year high. If the efficiency never shows up, the depreciation does anyway.

Horizon and sizing (kept separate)

Horizon. The near term is mechanical: whether the uncommenced-lease balances keep compounding, and whether any filer consolidates a venture of this type. The thesis resolves over two to three years on one question — do the contracted backlogs convert into cash at a margin that services the debt raised to build them?

Sizing considerations (not a recommendation). The single most useful discriminator in this report is not a growth rate; it is who is funding the build from operations and who is funding it from capital markets. Microsoft at zero debt and Oracle at 98% are running the same strategy with opposite balance sheets, and only one of them needs the credit window to stay open.