Q2 2026 Earnings: Big Tech's AI Capex Race Hits $730 Billion (Microsoft, Meta, Amazon, Google)
FutureX Research · AI Lab · 2026.08.03 · 13 pp · preview 4 pp
Listen · Audio Summary
5-8 min · AI narration in English · abstract + all key findings
Abstract
(Data as of 2026-10-02) In late July 2026, Alphabet (7/22), Microsoft and Meta (7/29), and Amazon (7/30) reported earnings. Their latest 2026 capex guidance sums to roughly $730 billion at midpoints, nearly double 2025's ~$410 billion. Market reactions diverged sharply: Alphabet posted its first negative free-cash-flow quarter since its 2004 IPO and Meta fell 10% the day after reporting, while Microsoft rose 8% after hours on $678 billion of commercial RPO and Amazon crossed a $3 trillion market cap on August 3 after AWS growth accelerated to 37%. This report verifies each company's actual capex and guidance basis, traces where the money goes (servers, memory, power, in-house silicon), maps the transmission to Nvidia, storage, power equipment, and model developers, and presents both sides of the bubble debate. Not investment advice.
Key Findings
- 01Microsoft (7/29, FY26 Q4): quarterly capex of $41B including finance leases; full fiscal-year cash capex of $115.95B, up 80% YoY. Calendar-2026 expectation restated from ~$190B to ~$175B due to a finance-to-operating lease shift (spending plans unchanged, per company); stock rose ~8% after hours.
- 02Meta (7/29, Q2): capex of $31.08B including finance-lease principal, up 83% YoY; raised the 2026 guidance floor for the second time this year, to $135-145B. EPS of $6.18 missed by ~14%, free cash flow fell to $784M, and the stock dropped 10% on 7/30, its worst day of the year.
- 03Alphabet (7/22, Q2): capex of $44.9B, roughly double a year earlier; full-year guidance raised to $195-205B from $180-190B. Quarterly free cash flow was negative $5.9B, the first negative quarter since the 2004 IPO; the company raised $49.6B in equity in June plus $20.3B in senior notes in Q2.
- 04Amazon (7/30, Q2): capex of $54.2B versus $32.1B a year earlier; 2026 cash capex guidance lifted from ~$200B to ~$220B, citing memory prices. AWS grew 37%, the fastest in 18 quarters, with a $496B backlog; the stock jumped ~15% on 7/31 and Amazon crossed a $3T market cap on 8/3.
- 05The four companies' latest 2026 guidance totals roughly $730B, nearly double 2025's ~$410B. Analysts estimate memory will absorb ~30% of hyperscaler AI data-center spend in 2026 (vs ~8% in 2023-24); DDR5 prices are up ~110% this year and HBM capacity is sold out through 2026.
- 06Downstream valuations ballooned in parallel: Anthropic closed a $65B round at a $965B post-money valuation on 5/28 and filed confidentially for an IPO; its revaluation added a $53.4B pre-tax gain to Amazon's Q2 and $98B of (mostly unrealized) equity gains to Alphabet's, inflating headline profits.
The Earnings-Week Timeline: What Each of the Four Reported
Alphabet opened on July 22: Q2 revenue of $119.8B (+24%), Google Cloud accelerating to +82%, but capex of $44.9B (roughly 2x YoY), full-year guidance raised to $195-205B, and free cash flow of negative $5.9B — its first negative quarter since the 2004 IPO. On July 29, Microsoft posted FY26 Q4 revenue of $90B (+18%) with Azure at 43% and quarterly capex of $41B; fiscal-year cash capex reached $115.95B (+80%), while its calendar-2026 expectation was restated from ~$190B to ~$175B on a lease-accounting shift. Meta's quarterly capex hit $31.08B (+83%), guidance rose to $135-145B, EPS missed by ~14%, and free cash flow shrank to $784M. Amazon closed the week on July 30: revenue topped $200B, AWS grew 37% (fastest in 18 quarters), quarterly capex was $54.2B, and 2026 guidance rose to ~$220B on memory prices. The four now guide to roughly $730B combined.
Why the Market Rewarded Microsoft and Amazon but Punished Alphabet and Meta
The same capex escalation earned opposite verdicts within one week. Alphabet fell over 4% after hours and roughly 5-7% cumulatively over subsequent sessions (windows vary), and per CNBC's July 28 report its results sparked a sector-wide sell-off ahead of the other three reports. Meta dropped ~9.6% after hours, then 10% on July 30 — its worst day of the year — and kept sliding into early August. Microsoft, by contrast, jumped ~8% after hours: commercial RPO reached $678B (+84%), Copilot hit 30 million paid seats (~$10B annualized), and capex guidance stayed steady. Amazon surged ~15% on July 31 and crossed a $3 trillion market cap on August 3, the fifth company ever, on the back of a $496B AWS backlog that grew $132B in a single quarter. The logic: markets no longer punish spending per se — they distinguish capex backed by contracted demand (RPO, backlog) from capex paired with deteriorating free cash flow.
Where the Money Goes: Compute, Memory, Power, and In-House Silicon
Four destinations absorb the ~$730B. Compute remains the biggest: Alphabet said ~60% of Q2 spend went to servers and 40% to data centers; Microsoft said about two-thirds went to short-lived assets such as CPUs and GPUs. Memory is the new swing factor: analysts estimate it will take ~30% of hyperscaler AI data-center spend in 2026 versus ~8% in 2023-24; DDR5 is up ~110% this year and HBM capacity at SK Hynix, Samsung, and Micron is sold out through 2026 — Microsoft (~$25B component-price impact), Amazon ($20B guidance raise attributed to memory), and Meta all cited it explicitly. Power is the next constraint: Jassy said capacity will stay short through 2027; Korean power-equipment stocks surged on August 1 (Hyosung Heavy +27.6%, LS Electric +23.0%) and US peers hit new highs. In-house silicon is going mainstream: Amazon's chip business runs above a $25B annualized pace with $225B in multi-year commitments, while Google's TPU won an Anthropic deal for up to one million chips and reportedly sells capacity to Meta.
The Bubble Debate: The Strongest Evidence on Each Side
Bears hold four cards: deteriorating cash flow (Alphabet's quarterly FCF turned negative; Meta's fell to $784M; some analysts estimate Amazon's full-year 2026 FCF could go negative — forecasts are uncertain); debt-funded capex (per FactSet, incremental debt rose from ~9% of hyperscaler capex in FY24 to ~32% by mid-2026, with $100B+ of bonds in six months; Moody's counts ~$662B of signed-but-not-commenced off-balance-sheet data-center leases); the depreciation wall (short-lived assets dominate capex, and useful-life assumptions — including Microsoft's extension for buildings, which some investors question — swing earnings materially); and circularity (giants fund model developers, who rent their compute, whose revaluations then inflate the giants' profits). Bulls counter with contracted demand ($678B RPO, $496B backlog), simultaneous cloud acceleration (43%/37%/82%), real monetization (~$10B Copilot run-rate), and operating cash flows above $150B per quarter combined — fundamentally unlike the leveraged builders of the 2000 telecom bubble. We draw no verdict; the full evidence table sits in the locked sections.
Late-August update: capex keeps climbing, and bond markets feel it (data through 2026-09-02)
The late-August news bears out this report on both counts. The spending is still accelerating, the cash-flow strain is deepening, and the ROI debate has now spilled into the bond market.
Confirmed (Nvidia earnings release, CNBC, Aug 26): Nvidia booked $96.2bn in revenue for the quarter ended July 26, including $89bn from data centers, up 117% year on year, and guided next quarter to roughly $108bn excluding China. Hyperscaler capex guidance is turning into supplier revenue on schedule.
Confirmed (earnings call transcript, Fortune, Aug 26-27): CFO Colette Kress put top-five hyperscaler capex at about $800bn this year and $1.3 trillion in 2027, adding that customer purchase commitments for Nvidia chips more than doubled in a single quarter, from $119bn to $279bn. That sits well above where consensus stood during earnings season.
Confirmed (Bloomberg Aug 17, Axios Aug 25, Fortune Aug 29): the AI borrowing wave is pushing up Treasury yields. Ed Yardeni calls it reverse crowding out, with the adjustment showing up in the risk-free rate rather than in credit spreads; Treasury Secretary Bessent remarked that these companies barely care what they pay to borrow. One caveat: Goldman's estimate of roughly $489bn in global AI-related debt supply this year appears only in a single secondhand account and remains unverified.
Meta's free-cash-flow squeeze thus gets a sequel. Once internal cash runs low, the cost of external funding becomes the binding constraint.
FutureX Position · Open Source Breaks the Deadlock: After the Upward Revisions, Watch the Second Derivative (Xiamen keynote, 2026-09-03)
The keynote reinforces this report's read: spending is still climbing, and the strain has reached the bond market. What it adds is the transmission mechanism and one rule for watching it. On spending, the four global hyperscalers ran $410bn of capex in 2025, guide to $725bn for 2026, and consensus puts 2027 at $920bn (public estimates). The 1996–2001 parallel: telecom investment grew 18% a year while OECD telecom revenue grew only 7.2%, then shrank for seven straight quarters after 2001 (OECD 2003; Richmond Fed 2003). FutureX's read: model-layer revenue is an unreliable denominator, so we do not argue from it. We track only the second derivative of capex. The turn comes when growth itself starts to slow. The mechanism starts with depreciation. Amazon shortened some server lives to 5 years; Meta stretched to 5.5 (company filings). Michael Burry estimates the longer schedules cut 2026–2028 charges by about $176bn (public estimate, not independently verified). Bonds run 7–10 years; the asset's economic life may be 2–3. The keynote then lays out credit moving down four steps: parent balance sheets, customer contracts, GPU collateral, single-project SPVs. The checkable marks (public data): Oracle's remaining performance obligations rose from $138bn to $638bn within a year, with FY2026 free cash flow at −$23.7bn; CoreWeave carried $35.6bn of interest-bearing debt at June 30, 2026; the BIS 2026 annual report says the repricing shock to credit markets could compare with 2008. The keynote does not touch Nvidia's ex-China guidance; that entry in this report stands. Full deck: /reports/open-source-breakthrough (first 5 pages public).
Early-September Update · Verified (data current as of 2026-09-09): Broadcom's AI chip revenue tripled but guidance came in below estimates, and Amazon signed a multi-generation custom-silicon deal with Qualcomm; Big Tech spending keeps rising while market expectations run ahead of it
Confirmed (The Motley Fool, CNBC, September 2, 2026): Broadcom reported fiscal third-quarter 2026 results on September 2. Revenue was $29.6 billion, up 86% from a year earlier. AI chip revenue more than tripled to $16.7 billion. Semiconductor revenue rose 127% to $20.8 billion, capital expenditures were just over $500 million, and free cash flow was $13.7 billion. Figures are as reported by The Motley Fool from the earnings release.
Confirmed (Reuters, The Motley Fool, CNBC, September 2, 2026): Broadcom guided fourth-quarter revenue to $34.8 billion, below the $35.03 billion analysts expected according to LSEG, a gap of about $230 million. The guidance still implies 93% year-over-year growth. Reuters' headline that day said the quarterly revenue forecast came in below estimates, and CNBC reported the shares fell in extended trading.
Confirmed (Qualcomm press release, CNBC, 247wallst.com, September 8, 2026): Qualcomm announced on September 8 a multi-generation collaboration with Amazon Web Services covering customized silicon and optical connectivity for AI data-center infrastructure, with a focus on inference workloads. In a filing with the Securities and Exchange Commission, Qualcomm said it issued Amazon warrants to acquire 25 million shares at $161.26 each, a total of $4 billion. CNBC reported Qualcomm shares up 3% in morning trading; 247wallst.com reported a 5% gain the same day and noted that the agreement discloses no committed volume or delivery timeline.
Implications for this report's conclusions: The week reinforces the late-August finding that spending keeps moving up and the strain passes through to suppliers. Broadcom's tripled AI chip revenue shows the four hyperscalers are still adding orders. A guidance shortfall of only $230 million was enough to send the shares lower after hours, which shows the market is pricing acceleration rather than absolute volume; that evidence belongs in the bubble-debate section. The Amazon-Qualcomm agreement shows custom silicon taking a share of GPU budgets, consistent with the "second derivative" section, though the effect cannot be sized because no volumes were disclosed. One reservation to add: Bloomberg reported on September 8 that the Google-Blackstone data-center venture faces delays at several sites; this update verified only the headline, not the article body. There is a lag between capex guidance and delivered capacity, and the report's confidence on the 2026 build-out pace should be lowered accordingly.
Mid-to-Late-September Update · Verified (data current as of 2026-09-25): Spending keeps accelerating, financing strain lands on single projects, and the market prices capex only when a product shows up
Company disclosure (Oracle FY2027 Q1 results, released after the close on September 10, 2026, via Yahoo Finance): Quarterly revenue was $19.35 billion, up 30% year on year; OCI revenue was $7.39 billion, up 121%; RPO rose $209 billion year on year to $664 billion, about 46%. Oracle booked more than $30 billion of AI cloud contracts in the quarter and raised its full-year revenue target to at least $90 billion; shares rose 6.8% after hours. Quarterly capex was $28.5 billion against $8.5 billion a year earlier (Bloomberg, September 24).
Reported (Bloomberg via Yahoo Finance, September 24, 2026): Oracle sent a force majeure notice to the Blue Owl unit developing the Project Jupiter data center in New Mexico, allowing it to defer payments if the site misses its planned 2028 start. Bloomberg cited community pushback and objections from environmental groups; a Reuters source the same day cited possible power delays. About $18 billion of debt attached to the project was already trading at stressed levels. Oracle said the project remains on schedule; the stock closed down about 4% on September 24.
Verified (Bloomberg via Yahoo Finance, priced September 18, 2026, reported September 19): CleanSpark sold $2.28 billion of five-year high-yield notes for a data center in Sandersville, Georgia, at an 8.25% yield and 98.5 cents on the dollar, with about $10 billion of orders, more than four times the deal. The site carries a $6.6 billion, 20-year lease with Meta subsidiary Anvilan LLC, Meta guarantees rent and operating expenses, and operations are expected to begin in Q4 2027. It is the first junk bond for a Meta-linked data center.
Public market data (Yahoo Finance, September 22, 2026): Meta shares rose 11% on September 21 to close at $741. Wells Fargo analyst Ken Gawrelski lifted his price target from $640 to $796, pointing to Meta's AI assistant Muse, which had a record 264,000 US downloads on September 19 and 448,000 daily active users the day before.
Effect on this report's conclusions: "Spending keeps accelerating" is reinforced: Oracle's quarterly capex more than tripled year on year, and JPMorgan CEO Jamie Dimon said on September 21 in India that hyperscaler spending rose from roughly $300 billion last year to about $700 billion this year and could reach $1 trillion next year (Yahoo Finance, September 22). "Pressure spills into the bond market" needs narrower wording: strain has landed on single projects (the $18 billion of Project Jupiter debt and the force majeure notice) while a lease-guaranteed high-yield deal still drew four times coverage, and Reuters wrote on September 22 that investment-grade credit has split along AI lines. The bond market is pricing project by project, not closing. "Watch the second derivative" gains a Meta sample: the company was under pressure over capex in earnings season and was repriced once a product showed usage numbers. The market pays for product; spending alone earns no premium.
Late-September to Early-October Update · Verified (data current as of 2026-10-02): Micron posts $54.23 billion in quarterly revenue and guides to about $61.5 billion, about 80% of Anthropic's reported $518 billion-plus compute commitments are said to be non-cancellable, and Meta's R&D tax credits rise from about $700 million to about $3.9 billion in two years
Per company disclosure (Micron results, CNBC, September 30, 2026): After the close on September 30, Micron reported fiscal fourth-quarter revenue of $54.23 billion, above the LSEG consensus of $51.07 billion, versus $11.32 billion a year earlier. DRAM revenue rose 343% to $39.8 billion, 73% of sales, and net income was $37.7 billion. For the next fiscal quarter Micron guided to revenue of about $61.5 billion and adjusted EPS of $38.15, both above the LSEG analyst expectations of $57 billion and $35.40.
Reported (Reuters, September 29, 2026): A non-public draft of Anthropic's prospectus, seen by Reuters, shows at least about $518 billion in compute and infrastructure commitments over roughly a decade, including at least $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft, plus about $161.2 billion in Broadcom-related equipment leases; around 80% is non-cancellable or payable regardless of actual usage. The same report puts 2025 revenue at nearly $4.6 billion.
Reported (The New York Times, September 30, 2026, as relayed by Yahoo Finance): According to Meta's securities filings, its federal research tax credits rose from about $700 million in 2023 to about $2 billion in 2024 and about $3.9 billion in 2025. The Times reported that Meta treats its AI data centers as pilot models under tax rules and claims the credit on the Nvidia chips inside them, a practice it dates to late 2024; an accounting academic quoted by the Times said Meta's own accountants have flagged to investors that the treatment risks being overturned by the IRS. Meta spokesperson Andy Stone said the company uses incentives Congress created to encourage domestic investment.
Impact on this report's thesis: Micron's numbers reinforce our view that the money is flowing to memory. DRAM revenue up 343% is consistent with the estimate we cited earlier that memory's share of AI data-center spending is rising, and CNBC notes that the shortage has pushed prices up, which suggests part of capex dollar growth is price. Anthropic's commitments to Google, Amazon and Microsoft total about $252.5 billion, giving cloud backlogs a named source, but the roughly 80% non-cancellable terms concentrate demand risk in a single model company with nearly $4.6 billion of 2025 revenue. Meta's R&D tax credit is a cash-flow offset this report had not covered; if it is curtailed, free-cash-flow pressure grows. Thresholds: if the four companies raise 2026 guidance again in their next earnings round and Micron delivers about $61.5 billion next quarter, our view that spending keeps accelerating holds; guidance cuts or falling memory prices would call for a reassessment.
Key Questions
How much will Big Tech spend on AI capex in 2026?
Roughly $730B at guidance midpoints, nearly double 2025's ~$410B: Amazon ~$220B, Alphabet $195-205B, Microsoft ~$175B (calendar year), Meta $135-145B (data as of 2026-08-03).
Why did Meta drop after Q2 2026 earnings while Microsoft and Amazon surged?
Markets rewarded spend backed by revenue: Meta's EPS of $6.18 missed by ~14% with $784M free cash flow, falling 10% on 7/30; Microsoft rose ~8% on $678B commercial RPO; Amazon jumped ~15% on 7/31 after AWS grew 37%, crossing a $3T market cap on 8/3.
Where is the AI capex money going, and who benefits?
Mainly servers, memory, power, and in-house silicon. Analysts estimate memory will absorb ~30% of hyperscaler AI data-center spend in 2026 (vs ~8% in 2023-24); DDR5 is up ~110% this year and HBM capacity is sold out through 2026, benefiting Nvidia, storage, and power-equipment makers. Not investment advice.
Sourcing and standards
Compiled from public sources; data current as of 2026.08.03. The text separates verified facts, reported claims, our own estimates and disputed points, and states the derivation behind every estimate. When we get something wrong, the correction is written into the report body with the original call left visible, and logged publicly.
Research standards & corrections →📄 Full Report
Full report: 13 pages · provided to professional investors & partners only
This is the public preview. The full report includes the sections below. For compliance reasons it isn't posted publicly or offered as a free download. To request a copy, contact the FutureX team.
- 🔒Breaking Down $730B: The Guidance Ledger, Three Rounds of Raises, and the Lease-Accounting Trap
- 🔒Depreciation-Wall Sensitivity: What One Year of Server-Life Assumption Does to Earnings
- 🔒When Memory Hits 30% of Capex: The Storage Price Chain and the 2027 Supply Gap
- 🔒Anthropic at $965 Billion and 'Circular Gains': Cleaning Up Big Tech's Income Statements
- 🔒Power Is the Next Bottleneck: Equipment Backlogs, Korea's Rally, and a Watchlist
- 🔒A VC Lens: Where the Arms Race Spills Over to the Model and Application Layers
Where we stand on this
The bubble sits in the middle layer that must keep refinancing: the FutureX seven-dimension bubble scorecard
The AI bubble sits in the middle layer that must keep refinancing: compute leasing on borrowed money, circular financing, GPU-backed debt. FutureX Capital has scored seven dimensions daily since March 2026; the August reading lit six, and liquidity triggered on August 17. The scorecard answers one question: which layer cannot be bought at the consensus price.
Full argument and falsification tests →Open source breaks the deadlock: the cost curve has a new author
The dividing line of this AI cycle is the cost curve, and open source pushes it down. A closed model's moat is capital expenditure that must be refinanced; an open model's moat is a living community that needs none. The curve's author has changed from US closed labs to Chinese open models; value moves to applications and open-source chokepoints.
Full argument and falsification tests →Questions people ask next
Related Research
Industry research from FutureX Capital's AI Lab, compiled from public information; not investment advice; contains no fund performance, AUM, or offer to raise capital.