The Capex Answer Is In: No Slowdown — Big Four at ~$725B and Rising — as the Debate Shifts From "Will They Spend" to "What Return on $1.75T of Debt"
The open question after Dario Amodei's Sept 12–13 slowdown call (backed by Musk and Altman) was whether it would translate into measurable capex changes. Two weeks later, the evidence says no — spending guidance keeps rising, and the argument has moved to a new battlefield: how the spending gets financed, and what return it can possibly earn.
Capex keeps going up, not down. The big four hyperscalers are on track for roughly $725B in combined 2026 capex, up ~77% from ~$410B in 2025 (Goldman Sachs June 2026 tally: Amazon ~$200B, Alphabet $180–190B, Meta $125–145B, Microsoft ~$190B FY26) (gpusmith capex tracker; Gate.com summary). Meta raised guidance to $125–145B with CFO Susan Li explicitly citing "higher component pricing this year" — i.e., memory-cost inflation is pushing capex up, not demand weakness down (Stocktwits/earnings coverage). Demand-side barometers agree: Anthropic's run-rate hit $65B (July 2026), surpassing OpenAI's $40B, and H100 rental prices have recovered from $2.20 to ~$2.80/GPU-hour — one tracker puts the latest print at $3.28, +22% in a month (BigGo Finance; ScanX).
But the financing structure is where the new bear case lives. Morgan Stanley estimates AI-compute expansion needs $1.75 trillion of external debt financing in 2026–28 (55% of all external financing), with private credit the largest source at $700B and traditional bank lending just 2% — banks won't take 3–5-year GPU collateral. Big-Five bond issuance: ~$24B/yr (2023–24) → $114.5B (2025) → $186.1B through mid-September 2026 alone, with Goldman projecting $250B this year and $400B in 2027. Oracle is the cautionary tale: operating cash flow turned negative and it was downgraded to BBB-, one notch above junk (BigGo Finance). The Big Five carry $3.18T in off-balance-sheet obligations vs. $1.66T on-balance-sheet debt; Nvidia's own procurement commitments are $344B (TSMC, memory).
Burry escalates with the accounting angle. In a Sept 25 blog post, Michael Burry warned Big Tech's AI boom "could produce massive write-offs by 2028 or 2029," flagging net capital investment at 2.07% of GDP — the highest in four decades outside the March 2000 Nasdaq peak — and zeroed in on off-balance-sheet commitments: Alphabet ~$900B, Meta ~$700B (approaching $1T), Microsoft's uncommenced leases tripled to >$300B, Amazon's up 81% in nine months to ~$267B with long-term debt doubling in six months to $128.9B. On Microsoft's extension of data-center useful lives from 15 to 25 years: "Too cute by half." "The most profitable companies in the world, except Apple, are betting everything on this as their free cash flow turns negative and the borrowing pace increases" (AOL/FinanceBuzz on Burry's post). S&P Global Ratings projects five of six hyperscalers (incl. SpaceX) post negative free cash flow in 2027 — Microsoft the exception — as combined capex exceeds $1.3T, with recovery not until 2029.
The return math is the killer number. Goldman's framework: for the six hyperscalers to earn a 15% ROIC on Phase-2 (2026–27) capex of $1.73T, they need **$1.42T of cumulative revenue in 2028–30** — while Phase-2 D&A alone runs $276B/yr, ~1.7x the EBIT those assets must generate. If real GPU life is 3 years instead of the assumed 5, the mature-phase revenue requirement jumps from ~$3.7T to ~$6T/yr — 14.8% of 2032 US GDP (BigGo Finance). Meanwhile credit markets are already diverging: investment-grade spreads under 1%, but CCC-and-below high-yield spreads have breached 10%, approaching their March 2025 peak — the AI tail is being repriced long before the majors feel it.
Why it matters: the Amodei slowdown call is, so far, noise at the capex line — the opposite happened (guidance up, component costs up, debt issuance at 5x the 2023–24 pace). The thesis to hold instead is Burry's and S&P's: the constraint isn't demand (Anthropic $65B run-rate, H100 rentals rising) — it's that the funding model (SPVs, private credit, residual-value guarantees — see Nvidia's $500B Platform Is Still at MOU Stage — While the MediaTek Convert Shows the "Nvidia as Financier" Pattern Expanding) has converted an equity boom into a credit cycle, and credit cycles end with spreads, not blog posts. Watch the CCC-spread divergence and the 2027 negative-FCF projections against actual prints. This sits alongside Nvidia Answers the Bubble Debate With a Record $150B Buyback as Chip Names Finally Pull Back and the demand-side bull case in Intel's Comeback Goes Vertical: +250% YTD, +14.3% Last Week — SK Hynix Ohio Talks, Muse-Driven Agentic Demand, and a $150B Mark-to-Market on Its Old NAND Business.