AI Supercycle Forensic Monitor

← swipe to see all columns →

Read-through: the aggregate 2026 guide now sits near $725B–$800B across the big four (Moody's carries $785B for 2026 and near $1T for 2027), roughly a 77% step-up on 2025's ~$410B. The forensic tell this quarter is not the headline spend — it is that operating cash flow stopped covering it. Alphabet went FCF-negative, Meta's FCF collapsed to under $1B, and Microsoft's relief rally came from holding the spending envelope flat while extending useful lives. That is the accounting-arbitrage channel this monitor tracks, now visible in the headline numbers rather than the footnotes.

Amazon closes the set — and splits it. Two of the big four are now free-cash-flow negative in the same quarter. Amazon raised its 2026 plan to $220B and named memory prices, not incremental demand, as the reason: input-cost inflation inside the capex line degrades incremental returns even if revenue lands exactly on plan. It is funding the gap with debt. But unlike the dot-com analogue, the revenue side did not crack — AWS accelerated for a fifth straight quarter, expanded margin 650bps during peak capex, and carries $496B of contracted backlog with capacity described as insufficient through 2027. Accelerating, profitable, contracted revenue funded by negative free cash flow is a different failure mode than the one this monitor was originally built for, where revenue decelerated first. The tripwire is therefore not AWS growth. It is the pair: CIP converting to depreciating PP&E without a matching revenue step, and the enterprise middle of Jassy's "barbell" failing to inflect on the timeline those five-year capacity contracts assume — a slope he explicitly declined to underwrite on the call: "I don't know if the trajectory of that middle part of the barbell will be the same wildly steep trajectory that we've seen with the current barbelled AI Labs piece."

Meta is the exception that qualifies all of the above — the first name where both legs moved together. At Amazon, Microsoft and Alphabet the pattern this quarter is uniform: the financial leg deteriorated while the demand leg accelerated. Meta breaks it. Ad impressions — the volume measure underneath the ad engine that funds the capex — decelerated from 18% YoY in Q4 2025 to 19% in Q1 to 14% in Q2, and revenue growth stepped down from 33% to 28% held up by a 12% rise in price per ad rather than by units. Against that, quarterly capex nearly doubled year-on-year to $31.1B, the full-year floor was raised, free cash flow fell to $784M and EPS missed by roughly 14%. That is spending materially more to sell modestly less, on top of a balance sheet already carrying $27B of CIP and a 5.5-year server-life extension deferring $2.9B a year. It does not overturn the read on the other three — Azure at +43% and AWS at +36.7% are not decelerating — but it is the first genuine demand-side crack in the cohort, and it appears at the name with the least contracted revenue to fall back on. Where AWS and Azure sell capacity against multi-year backlogs ($496B and $678B respectively), Meta is spending against an advertising market it must re-win every quarter.

The financing channel is now live. On the same day Microsoft and Meta reported, the FOMC held at 3.50–3.75% on a 9–3 vote, with Hammack, Kashkari and Logan all dissenting in favour of a hike — the first directionally-unified three-way dissent since September 2016. The 10-year closed at 4.657% and the 30-year pushed past 5.19%, its highest since 2007. This matters here for one reason: the cohort has just crossed from internally-funded to externally-funded expansion. Amazon's CFO confirmed debt issuance this year; Alphabet and Meta are both burning through cash generation. Rising long rates raise the hurdle on every incremental capex dollar precisely as incremental returns compress — Microsoft's AI iROIC now runs 17.7% against a 28.7% pre-AI baseline. For the leveraged neo-clouds already borrowing above 10%, it compresses debt-service coverage directly. An accounting story becomes a credit story at the point where the marginal dollar has to be borrowed.

Sources: company Q2 2026 earnings releases & calls (Alphabet, Microsoft, Meta, Amazon), CNBC, S&P Global Market Intelligence, Moody's hyperscaler capex forecast. Note: Amazon's $53.1B is net property & equipment purchases while Microsoft's $41.0B includes finance leases — indicative, not like-for-like.

1. The disclosure that was supposed to be impossible

The standing assumption — including in earlier versions of this monitor — was that no issuer would ever have to disclose circular revenue, because the equity stakes are deliberately structured to stay non-voting and below the significant-influence threshold that triggers ASC 850 related-party disclosure. That assumption failed at Microsoft. The October 2025 recapitalisation converted its holding into an ~25% as-converted equity-method interest, which triggered ASC 850, which compelled the exact figure the structure was meant to keep private:

"As an equity method investee, OpenAI is a related party as defined in … ASC 850 … For fiscal year 2026, we recorded revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments, of $24.1 billion, and accounts receivable from OpenAI as of June 30, 2026 was $6.0 billion." Microsoft Corporation, FY2026 Form 10-K

No triangulation, no analyst inference: a named counterparty and an exact number, alongside $13.0B of funding commitments of which $11.9B is funded. Microsoft is the exception, and that is now measured rather than assumed. The three peers holding comparable positions were each checked on their most recent 2026 10-Q, and none of Amazon, Alphabet or Nvidia discloses ASC 850, a related party, or equity-method treatment for its lab stake. Amazon states the engineering plainly — Anthropic is held as "nonvoting preferred stock" under the measurement alternative. Nvidia's only equity-method item is $1.0B of infrastructure funds, and OpenAI is not named in its 10-Q at all. So the rule is neither that ASC 850 never fires nor that it does: the avoidance structure holds at three of four filers and fails only when a recapitalisation pushes a stake across the line. That makes Microsoft's $24.1B a natural experiment for what the other three are not required to report.

2. The headline commitments mostly have no issuer source

The very large round numbers in circulation — Oracle's $300B OpenAI contract, AMD's ~$90B — could not be traced to any filing by either party. To test that properly rather than assert it, an automated pass read 212 documents covering every Oracle and AMD 10-K, 10-Q and 8-K accession since mid-2025, exhibits included, plus both companies' full XBRL fact sets. The $300 billion figure returns zero hits; so does Stargate. AMD's money figures return zero hits, including inside the warrant instrument itself and the earnings slides — AMD names OpenAI in sixteen documents and never once attaches a dollar amount to it. Oracle's sole appearance of the word is a product-catalogue reference in an Ellison quote listing available models, not a customer disclosure.

Both figures were carried by this register until verification and have been withdrawn. Removing them cuts the measurable ceiling from $25.1B to $5.62B per quarter, and what remains is only Amazon's two commitments — the sole arrangements stated verbatim in a filing. The rule this establishes is worth carrying forward:

That $5.62B is a ceiling, not a measurement. It is the maximum arithmetically consistent with what has been disclosed, assuming commitments are drawn evenly — which is not how capacity contracts actually run. The true figure is unknown and lower.

3. The penny warrant is an industry instrument

A warrant struck at one cent is not an option — it is stock delivered on a condition. The same $0.01 strike recurs across AMD→OpenAI, AMD→Meta (a second 160-million-share grant issued February 2026 on identical terms), Google→TeraWulf, Google→Cipher and CoreWeave→Core Scientific. Alphabet's own 10-Q shows the aggregate behind these: $43.8B of credit derivatives and $7.6B of financial guarantees provided as backstops, with $24.1B more agreed — and the credit-derivative line nearly tripled in six months, from $16.9B. Nvidia classifies its analogous $3.5B exposure the same way. Two suppliers, independently, describing the underwriting of their customers' fixed costs as writing credit protection.

Read the linkage carefully — it is inference, not disclosure. Alphabet names no counterparty, so the $43.8B cannot be shown to contain the TeraWulf and Cipher deals or to say how much of it they represent. The matching contract language supports only that these arrangements are of the same kind as what the aggregate is made of. Disclosure here is asymmetric by size: the landlords name Google and quantify their warrants precisely; Google names no one and discloses no warrants received. Neither side is incomplete on its own terms. Only reading both ends together reconstructs the arrangement.

4. What the margin test can and cannot see

If round-trip revenue were priced concessionally, segment margin should dilute. It did not: AWS operating margin expanded ~645bps year over year while growth accelerated 17%→37%, and Google Cloud margin roughly doubled as growth accelerated to 82%. Microsoft is the exception, and its Intelligent Cloud compression (41.47%→39.66%) has a capex and depreciation explanation visible in its own cost line rather than a pricing one.

That evidence is narrower than it first appears, and the qualification is not cosmetic. The test detects concessions delivered as price. It is blind to concessions delivered as equity, in two distinct ways. On timing: AMD states that its OpenAI and Meta warrants "did not have an impact on the Company's financial statements" — nothing has vested, the instruments are carried as a liability pending equity classification, and no income-statement presentation is disclosed anywhere. AMD's margin therefore carries no information about them in either direction. On form: an equity investment in a counterparty is an investing item that never touches revenue or cost of sales in any period, so for the Nvidia, Amazon and Microsoft structures the margin test has no power by construction — not merely no signal yet. Cash conversion and investing outflows are the instruments that see that channel; margin is not.

One figure from the peer review belongs here rather than in a footnote. Amazon recognised a $50.5B upward adjustment on its Anthropic holding in Q2 2026 alone, "reflect[ing] observable changes in prices." That is larger than Microsoft's entire disclosed annual OpenAI revenue, it runs through earnings, and it is driven by a private funding round rather than by any operating result.

Method, and what would falsify this

Every figure above is either quoted verbatim from an SEC filing or marked as an inference. Deal terms sourced only to press reporting are held separately and are not quotable until an issuer confirms them. Findings that were disproven are retained with their retraction rather than deleted — the register currently carries eight refuting findings against twelve confirming ones, and the single most important row in it is the one that refuted the premise the exercise began with.

The conclusions were then put to an adversarial review panel instructed to attack rather than confirm them. It returned a verdict of kill on the negative findings, on the grounds that they rested on narrative text alone — which is what prompted the 212-document exhibit and XBRL sweep described above. The claims survived it; one overstatement about Oracle was corrected as a result, and the margin-test scope in section 4 was narrowed after the panel showed the original wording claimed more than the filings support.

Open items, stated as such. Microsoft's reported 15→25 year useful-life extension is not yet in a filing — the FY2026 10-K still reads "five to 15 years" — so it awaits the Q1 FY27 10-Q. AMD's first vesting tranche will force a warrant presentation policy into the open and is the point at which the margin test regains power there. And the single most useful forward signal from the whole exercise: watch for any other holder crossing into equity-method treatment, because that is the moment the numbers stop being private.

← swipe to see all columns →

Figure 1 — Comparative infrastructure spending. A single hyperscaler's 2026 guide is already 2× the entire 1998–2002 Dot-Com telecom build-out; the big four combined are roughly 8×.