The market treats rising hyperscaler capex guidance as proof of AI demand. The capital-cycle read runs the other way. Equity peaks have tended to arrive before the peak in net investment. The commitments disclosed in the filings of Microsoft, Amazon, Oracle, Meta and Alphabet suggest the investment peak is still a couple of years out. If that ordering holds, the next two years of record spending are not evidence that the stock cycle has room to run. They are evidence that the spending is locked in whether or not it earns its return.
That is the core of the Cassandra Unchained "Capital Cycle IQ & Forensic Files" series on the five largest hyperscalers, as framed in its abridged edition. The abridged text available here is a preview: it carries the framework and the headline aggregate, not the company-by-company forensic findings. What follows develops the argument the preview makes and marks where the full evidence is not visible.
Investment Peaks Trail Market Peaks
The series anchors on one series: net capital investment by S&P 500 companies as a share of US GDP. Its central observation is that market tops came before tops in that ratio. This is the capital cycle in its plainest form. High returns attract capital. Capital arrives through multi-year construction programs, so supply keeps arriving after the returns that justified it have started to fade. Prices, which discount the future, turn first. The investment line, which reflects decisions made two or three years earlier, turns last.
The implication for the AI build is uncomfortable for anyone using capex as a timing tool. A hyperscaler raising its spending guide tells you what management committed to when its outlook was strongest. It says little about the return on the marginal dollar that comes online in 2027. The author's claim is blunt: "We are near one of those moments again."
The ordering is a pattern that has often held. It is not a law. The lag between the equity peak and the investment peak has varied, and some investment booms (the 1990s fiber build, for instance) ran for years before prices broke. The pattern is useful as a warning about what capex guidance cannot tell you. It does not set a date.
Three Trillion Dollars That Never Shows Up in the Capex Line
The more important disclosure sits below the headline capex figure. The prior installment of the series totaled roughly $3 trillion in aggregate commitments across the five companies. That total combines purchase obligations, future lease payments, guarantees backing third-party debt, and other off-balance-sheet potential liabilities.
This is where guidance language and filings diverge. Quarterly capex guidance is a flow number for the next twelve months, and management can trim it. The commitment notes in the 10-K and 10-Q describe obligations that stretch years forward and are much harder to unwind:
- Purchase obligations cover chips, power equipment and construction contracts already signed. These convert into capex as they are delivered.
- Leases not yet commenced are data-center and land leases that are signed but do not appear on the balance sheet until the asset is handed over. They become right-of-use assets and lease liabilities later, often after the build is finished.
- Guarantees of third-party debt arise when a hyperscaler stands behind financing raised by a developer, joint venture or special-purpose vehicle. The debt sits elsewhere. Under adverse conditions the risk comes home.
Together these explain the author's conclusion that the filings point to "a couple more years of record investment" in AI data centers. The pipeline has been contracted. The capex line of the next eight quarters is largely an accounting translation of decisions already made.
That changes how to read a strong print. When a hyperscaler beats on cloud revenue and reaffirms its spend, consensus hears demand confirmation. The forensic reading is that the company has little practical option except to reaffirm. The interesting question is not whether capex rises. It is whether revenue rises fast enough to cover the depreciation and lease expense those commitments will generate once the assets go into service.
Where the Stress Actually Lands
Aggregate figures hide the dispersion that matters. Five companies are funding the same build from very different positions, and the series' stated purpose is to read earnings reports, filings and conference calls for "clues as to how stressed, or not, each hyperscaler really is." The per-company verdicts are behind the paywall, so the sketch below reflects publicly known funding structures rather than the author's conclusions.
Microsoft and Alphabet fund the build largely from operating cash flow, with net cash or modest net debt. For them, commitments are a margin question, not a solvency question. The risk is a slow grind of depreciation into operating margins, not a funding squeeze.
Amazon sits between the extremes. Retail and AWS cash flow is large, but the company's historical habit of running free cash flow close to zero during heavy build phases means the AI program pushes against a thinner cushion.
Meta has been the most visible user of off-balance-sheet structures. Its large Louisiana data-center project was reported to be financed through a joint vehicle with an outside investor, with Meta as tenant and with support provisions tied to the asset's value. This is the third category above in practice: the debt is not Meta's on paper, but the economic exposure is not gone.
Oracle is the clearest stress case. It has a smaller cash base than the other four, a heavier existing debt load, and a backlog concentrated in a handful of large AI customers. Its spending has caught up with its cash generation.
When capex absorbs operating cash flow, the residual funding comes from debt markets. That moves the AI build from a question for equity holders into one that credit investors must price. Oracle's position does not prove the cycle has turned. It does show where the cycle would show strain first: at the company with the least internal funding and the most concentrated counterparties.
The Case Against the Read
The strongest objection is that this cycle is contracted on the demand side too. Hyperscalers report large remaining performance obligations, and several have described capacity constraints rather than idle capacity. If customers have signed multi-year take-or-pay agreements, the supply arriving in 2027 is pre-sold. On that view, the classic capital-cycle failure, where capacity floods in after demand has peaked, does not apply.
The objection has real force for Microsoft and Alphabet, whose balance sheets could absorb a disappointing return on the marginal data center without impairing the franchise. A capital cycle that ends in lower returns on capital is not the same as one that ends in distress.
Two observations weaken the objection. First, backlog concentration cuts both ways. A contract with a single AI lab that is itself funded by fresh equity and vendor financing is only as good as that lab's next raise. Second, purchase obligations and guarantees are fixed in a way customer demand is not. A backlog can be renegotiated, delayed or cancelled. A guarantee on a developer's debt cannot be renegotiated on the same terms. The asymmetry lies between flexible revenue and rigid commitments, and that asymmetry is what the forensic approach is designed to find.
There is a further limitation. Commitment totals bundle obligations of very different quality. Some purchase obligations carry cancellation clauses. Some leases can be sublet. A $3 trillion headline overstates the truly fixed portion by an amount the preview does not quantify.
What Would Settle It
The thesis is testable in the next few reporting cycles, and the evidence will appear in places other than the capex guide.
The thesis strengthens if commitment balances in the filings keep growing faster than reported cloud and AI revenue. It also strengthens if depreciation growth begins to outrun revenue growth as 2025 and 2026 builds enter service, and if more financing migrates into vehicles where the hyperscaler is a guarantor rather than a borrower. On the credit side, wider spreads on Oracle debt relative to the stronger issuers would show the market pricing the dispersion described above.
It weakens if free cash flow after lease payments holds steady at Microsoft, Alphabet and Amazon while commitments rise. That would mean revenue is absorbing the build in real time.
The decision variable is not whether the hyperscalers spend more next year; the filings say they will. It is whether the return on that spending shows up before the depreciation does. Capital-cycle history suggests the stocks will answer that question before the capex line does.
