Big tech capital spending falls year on year

What I computed

I pulled each company's "purchases/additions of property and equipment" from SEC EDGAR XBRL company facts (data.sec.gov/api/xbrl/companyfacts/CIK{0001652044, 0001018724, 0001326801, 0000789019}.json, plus Google Inc. CIK 0001288776 for years before 2015), retrieved 2026-10-11. I used the tag PaymentsToAcquirePropertyPlantAndEquipment for Alphabet, Meta and Microsoft, and PaymentsToAcquireProductiveAssets for Amazon. That is Amazon's gross "Purchases of property and equipment" line: for example $35,095M in Q3 2025 and $54,208M in Q2 2026. I converted to calendar quarters by subtracting year-to-date figures where only those were reported; Microsoft's April–June quarter is the annual figure minus nine months.

Combined cash capex for the four companies:

Quarter	Combined	Change from a year earlier
Q3 2025	$97.3B
Q4 2025	$118.6B
Q1 2026	$129.8B	+80%
Q2 2026	$165.1B	+87%

Q2 2026 broke down as Alphabet $44.9B, Amazon $54.2B, Meta $30.1B and Microsoft $35.8B.

History of year-on-year declines (2010–2026). The combined series fell year on year in Q1 2010 (after the financial crisis), Q4 2011 (lumpy building purchase), Q4 2015, Q1 2019 (lumpy comparison) and Q1–Q3 2023 (post-COVID cutback). Every 5-year window since 2009 contains at least one decline, and the longest run without one is 15 quarters. Two of those five episodes came from one-off building purchases when the totals were small, which matters much less now that the total is $165B a quarter.

Current status and outlook

• 2026 plans. All four companies raised or held 2026 guidance in July:
  • Alphabet: $195–205B, with a "significant" increase in 2027 (CFO Dive, 2026-07-30; Yahoo, Aug 2026).
  • Amazon: about $220B, saying capacity will stay short of demand into 2028 (CNBC, 2026-07-30).
  • Meta: $135–145B (Yahoo, Aug 2026).
  • Microsoft: about $175B for calendar 2026 including finance leases after an accounting change, with further growth in fiscal 2027 (CNBC, 2026-07-29). The change does not affect its cash additions line.
• 2027. Consensus is about $1.1–1.2T for the five largest hyperscalers, up 30–50%. Goldman forecasts $1.2T (Bloomberg, 2026-09-25) and Morgan Stanley $1.2–1.4T.
• 2028 deceleration forecasts:
  • Morgan Stanley: data-center capex growth slowing to about 12% in a "digestion phase" (BigGo summary, Sep 2026).
  • UBS: hyperscaler growth of 25% in 2027 and 6% in 2028 (Reuters, 2026-07-16).
  • Morgan Stanley also puts the top five at about $1.4T in 2028, roughly +14% (24/7 Wall St, 2026-07-22).

Near-term horizons. The first eligible quarter is Q3 2026, filed in late October, and it must be compared with $97.3B. With guidance implying about $200B a quarter, the probability of a decline through the Q2 2027 comparison is close to zero. A 2027 decline would need an abrupt reversal while chip orders, construction and roughly $1.7T of cloud backlog are already committed. I put that at about 2.5%.

Paths to YES and NO

Paths to YES:

• Spending flattens in 2028. Spending in 2027 is expected to ramp through the year. If 2028 grows only 6–12% for the full year, as UBS and Morgan Stanley forecast, the late-2028 quarters would be roughly flat against late 2027, so a negative quarter is close to even odds in that scenario.
• Memory costs fall. Memory has driven much of the recent cost increases (Amazon's +$20B) and may be 50–60% of 2027 capex. It is a cyclical market, and when prices fall, dollar capex can fall even if the number of servers installed keeps rising.
• Financing strain. Capex now exceeds operating cash flow, free cash flow is negative at Alphabet and Meta, the Fed raised rates in September 2026, and 10-year Treasury yields are around 5.1% (WSJ and Fortune, Sep 2026). Polymarket put a US recession by the end of 2027 at about 33%.
• Spending moved off the cash line. More leases, joint ventures and rented capacity from smaller cloud providers would shift spending out of the measured line.
• History. Nearly every large capex boom has ended in a decline.

Paths to NO: Demand still exceeds supply, and cloud growth is accelerating (Google Cloud +82%, Azure +43%, AWS +37%). Forecasts of future spending have been revised up for three years running, power limits may spread spending out rather than create an overbuild, and replacing ageing chips puts a floor under spending. Long industry forecasts (Dell'Oro >$3T by 2030; PwC) show continued growth. Long booms do happen: US IT investment grew from 1992 to 2000.

Hazard model

I assigned per-quarter hazards and accumulated them:

• About 0–0.3% for each observation quarter through Q2 2027.
• 2027 observations as a whole: about 2.5%.
• 2028 observations: about 24%, rising from 3% to 10% per quarter as the comparisons get harder.
• 2029 observations: about 22%.
• 2030 observations: about 18%.
• Q1–Q3 2031: about 12%.

The hazard falls after 2028 because if no decline has appeared by then, the boom has proved durable. The result is about 26% by March 2029, 42% by March 2030, and 58% by the end of 2031.

Strongest argument against my estimate. In a world where AI is transformative and demand outruns supply, spending could grow every quarter through 2031, which would mean NO. The opposite argument is that 32 quarters without a decline would be twice the longest run on record, which suggests a higher number. I weighed both and landed near 58%.

Question. By the last day of each quarter, has the combined quarterly capital expenditure of Alphabet, Amazon, Meta, and Microsoft been reported as lower than in the same calendar quarter a year earlier?

Definitions.

• Companies: Alphabet Inc., Amazon.com, Inc., Meta Platforms, Inc., and Microsoft Corporation, using each company's consolidated accounts, not just its AI or cloud businesses.
• Capital expenditure: cash purchases of property and equipment in the investing section of the consolidated statement of cash flows, expressed as a positive expenditure. Use “Purchases of property and equipment” for Alphabet, Amazon, and Meta, and “Additions to property and equipment” for Microsoft, or identically scoped successor line items. Do not subtract separately reported proceeds from asset sales or incentives. Exclude noncash asset additions, assets obtained through leases, lease principal payments, acquisitions of businesses, and other investing expenditures outside these line items. This definition controls over company-specific headline capex measures.
• Quarter: the same three calendar months for all four companies, ending March 31, June 30, September 30, or December 31. Microsoft's fiscal-quarter numbering does not control. Use a reported three-month amount where available; otherwise subtract successive fiscal-year-to-date amounts, including annual minus nine-month amounts for a fiscal fourth quarter. Do not compare year-to-date or trailing-twelve-month totals.
• Combined: the sum of the four amounts in nominal US dollars, without inflation or constant-currency adjustment. Lower means strictly less, using the most precise figures published in the qualifying filings; equality does not count.

Resolves YES if qualifying filings available by the horizon establish that this sum for at least one calendar quarter is strictly below the sum for the corresponding calendar quarter one year earlier. All four companies' amounts must be available for both quarters. Apply the Observation period convention: the later observation quarter must end after the opening date. A decline in the aggregate suffices; each company need not individually report a decline, and no attribution to AI is required.

Use the latest filed figures available at the time of the comparison, including restated comparative figures. A subsequent filing or amendment that first establishes a qualifying decline can trigger YES on its publication date. Once YES, later revisions do not reverse it, under the Resolution window convention.

Resolves NO at a horizon if no qualifying comparison has been published by then. Guidance, spending plans, analyst estimates, a quarter-on-quarter decline alone, slower positive year-on-year growth, and incomplete four-company totals do not qualify. A missing report is not a zero expenditure. A name change or legal holding-company successor counts; a sale or spin-off is treated according to the consolidated accounts and their published comparatives, without analyst-created pro forma adjustments.

Resolving source. The companies' Forms 10-Q and 10-K, and amendments, including their consolidated statements of cash flows, filed with the US Securities and Exchange Commission on EDGAR: Alphabet (CIK 1652044), Amazon (CIK 1018724), Meta (CIK 1326801), and Microsoft (CIK 789019). Identical copies on the companies' investor-relations sites count. This is a narrower source requirement than Public confirmation (A2); earnings releases alone do not resolve. Apply Resolution date (A3): the qualifying publication date is the date the last filing necessary to establish the comparison becomes public, not the end of the expenditure quarter. Apply the Resolution window convention; older filings may supply historical comparison inputs but cannot independently trigger resolution.