Update: Big-5 AI Capex Midpoints Hit $835B — Up $33B From the August Vintage

What changed since the August spend update

In mid-August we published the post-Q2 vintage: Big-5 guidance midpoints near $803B, up +$43B from the July research print of ~$760B. That update answered what moved after earnings week. This Q3 refresh answers the next capital-markets question: what moved again in the newest mid-quarter desk vintage, and how large is the incremental YoY / vintage delta once Street and credit desks finish catching the raises?

Three forces push the stack higher between early August and late August. Amazon and Alphabet midpoints climb again on early-Q3 infra commentary and sell-side range resets. Meta steps both the floor and the ceiling of its disclosed band. Oracle’s net-of-prepayment cash figure edges from ~$70B toward ~$75B while gross remains nearer $95–100B. Microsoft’s ~$175B calendar print is unchanged — still an accounting hold from lease reclassification, not a build pause. The dashboard above is built as an Aug→Q3 vintage delta: waterfall contributions, slope levels, a triple-2026 stacked path, a Jul/Aug/Q3 revision multi-line, composition share, and research-house fans.

The headline table: Aug print vs mid-Q3 desk

That +$32.5B / +4.0% revision is the clean vintage delta versus our August update. Cumulative versus July’s $760B research print, the Big-5 midpoint sum is now up +$75B / +9.9% across two refreshes. Strip Microsoft’s accounting hold and hold the economic CY nearer ~$190B, and the mid-Q3 stack sits closer to ~$850B — a +$47.5B economic raise against the August accounting baseline.

Apply the CreditSights-style ~75% AI-attributed haircut and the AI-specific slice of the Big-5 stack moves from roughly $602B → $626B. That remains a convention, not a 10-K line — toggle it in the dashboard so both perimeters stay visible without mixing them in a single headline.

Amazon and Alphabet still do the heavy lifting

Amazon’s step from ~$220B to ~$230B is again the largest absolute company raise in the vintage. It keeps Amazon the single largest program in the stack and extends the pattern we flagged in August: the company that already absorbed more than 100% of operating cash flow in Q2 is still authorizing more dollars into the calendar year. Early-Q3 AWS and campus commentary, plus Street midpoint catch-up, explain most of the +$10B — not a brand-new earnings release.

Alphabet’s third raise of the year — midpoint from $200B to $210B on a $205–215B band — is the cleanest “guidance only goes one direction” story in the set. Combined with Amazon, those two names contribute +$20B of the +$32.5B Big-5 delta before Meta and Oracle’s smaller steps. For desks that still carry February midpoints in models, the cumulative Alphabet path from the low-$180Bs to $210B is now a ~$25B+ annual rewrite in under seven months.

Microsoft’s flat print is still not a build cut

The mid-Q3 desk vintage leaves Microsoft’s calendar-2026 headline at ~$175B — unchanged from the August post-Q2 language. Coverage that treats a flat print as a pause still has the economics backwards. The firm continues to shift more data-center capacity toward operating rather than finance leases, which moves dollars off the capex line without canceling servers, power, or campus construction.

For vintage analysis that means two parallel numbers again:

1. Accounting vintage: Big-5 midpoints $803B → $835B (+$33B).
2. Economic vintage: hold Microsoft near the prior ~$190B run-rate → stack nearer $850B (+$48B vs Aug accounting).

Credit and equity desks that underwrite physical build should prefer the economic frame. Desks that underwrite reported free cash flow and PP&E growth should keep the accounting frame — and then add lease commitments back in the footnotes. Our companion capex intensity update still shows the same firm’s intensity climbing even as the calendar headline softens: the build is not pausing.

Meta’s band and Oracle’s net/gross split keep moving

Meta’s move looks moderate in dollars (+$7.5B at the midpoint) but matters in structure: both the floor and the ceiling stepped (toward $140–150B), and the disclosed total continues to fold in finance-lease principal. For desks that compare Meta to cash-only peers, that definitional gap still bites — another reason the dashboard keeps company toggles adjacent to every panel.

Oracle’s ~$75B figure circulating in the mid-Q3 desk synthesis remains net of customer prepayments. Gross guidance in the $95–100B neighborhood is the better comparator to peers who do not net the same way. We carry $75B in the Big-5 midpoint sum for cash-comparability with the Axis-style synthesis many desks quote — and we flag the gross stack in the table and caveats so readers do not undercount OCI / Stargate-linked build by $20B+.

That definitional gap is exactly why the July research post refused a single “true” AI spend number. The Q3 update does not invent one. It shows which perimeter moved and by how much since August.

Research houses: Street and credit climb; GS GI holds the layer path

Company midpoints are the observed near-term object. Research scenarios remain the forward object — and they still do not all re-cut on the same week.

• Street consensus for 2026 hyperscaler gross has climbed from roughly ~$790B after Q2 toward ~$820B, still trailing the company midpoint sum as catch-up continues.
• CreditSights-style aggregates that sat near ~$800B in the August print now sit closer to ~$830B.
• Goldman Sachs Global Institute’s Tracking Trillions all-in AI infra baseline ($765B in 2026, $1.01T in 2027) remains a layer framework, not a company-guidance rollupit did not need to jump just because Amazon and Alphabet each raised another $10B. The dashboard’s GI panel keeps that path visible so readers do not force a fake contradiction.
• Goldman Sachs Investment Research’s hyperscaler 2026 path edges from ~$780B toward ~$800B in secondary reporting, while the 2027 base (~$1.14T) and bull (~$1.4T) remain the trillion-dollar headline pair.

Toggle the scenario year control between 2026 and 2027. On 2026 you see Aug-vs-Q3 bars. On 2027 you see the longer fan where IR’s hyperscaler base can still sit above GI’s all-in total because the perimeters differ — the same paradox documented in the July spend map and the chips-and-data-centers breakdown.

The cumulative path now matters more than any single week

YoY context still dwarfs either vintage delta. Against the research post’s 2025 Big-5 stack (~$344B), the mid-Q3 ~$835B 2026 guide is roughly a +143% step-up. The +$33B August-to-Q3 revision is large in absolute dollars and small next to that year-over-year cliff — which is why markets can treat a mid-single-digit vintage raise as “more of the same” even while financing residuals scream that the same is no longer self-funding.

What is new is the cumulative revision path. From July’s $760B to August’s ~$803B to mid-Q3’s ~$835B, company midpoints have rewritten the year by +$75B without a single full-year print yet. Street, CreditSights, and GS IR paths all slope upward across the same three vintages in the dashboard’s revision panel — a visual that matters more for model hygiene than any one week’s headline.

Pair this update with the AI financing research for the credit-channel view and with the intensity update for the revenue-share view. Dollars, intensity, and funding remain three faces of one cycle. When absolute guides keep rising while absorption stays near full OCF — as Q2’s ~99% Big-4 print showed — the marginal dollar is still a capital-markets dollar.

Caveats and methodology

1. Desk vintage ≠ audited guidance. Mid-Q3 figures blend company midpoints, call language still in force, and sell-side / Axis-style synthesis; they can miss by tens of billions by year-end.
2. Definitions differ. Meta and Microsoft fold finance leases differently than Amazon/Alphabet cash PP&E. Oracle’s net-of-prepay figure is not peer-comparable to gross guides without adjustment.
3. Microsoft’s flat $175B is treated as an accounting hold. Economic build is closer to unchanged-to-higher; do not narrate it as a demand rollover without lease footnotes.
4. AI-attributed (~75%) is a research convention applied uniformly for interactionactual AI shares differ by company and year.
5. GS Global Institute figures are a sensitivity framework, not Goldman Sachs Investment Research point forecasts. GS IR figures come from public secondary reporting of research notes.
6. Street and CreditSights levels are neighborhood aggregates from public commentary, not a single broker model.
7. Totals may not sum across houses because perimeters differ (leases, power, non-hyperscaler buyers, geography).
8. This post is explanatory data journalism, not investment advice.

Primary synthesis: Axis Intelligence Research AI Capex Tracker (late-Aug / mid-Q3 2026 refresh of primary filings and desk midpoints); prior theme vintages in AI capex spend update (Aug) and AI capex spend research; Goldman Sachs Global Institute Tracking Trillions; Street / CreditSights aggregate commentary via public reporting.

What to watch into Q3 earnings and year-end

Three coincident signals will tell you whether the vintage delta stabilizes or accelerates again: (1) Q3 earnings language vs these midpoints — especially whether Amazon’s ~$230B and Alphabet’s $205–215B hold after GPU delivery timing; (2) Microsoft lease mix — if operating-lease migration continues, reported CY capex can keep undershooting economic build; (3) absorption and issuance — if the Big-4 cohort stays near 100% of OCF while absolute guides rise, the cycle’s marginal dollar remains a credit-market dollar. Until the next official earnings vintage, the live Big-5 midpoint sum is not August’s ~$803B. It is ~$835B — and on an economic lease-adjusted read, closer to ~$850B.