The AI race is now a cash-flow fight.
The Magnificent 7 are not racing on models anymore. They are racing on concrete, power and borrowed money — $720–740 billion of Big Four capex guided for 2026 alone — while investors have started paying only for the one thing the buildout is consuming: cash cover. In Q3 2026 the Big Four spent more on capex than they made in operating cash. That inversion, not a missed product launch, is why beats now get sold.
Spending tripled in two years. 2027 rises again — nobody will say by how much.
The scale is the story. Big Four capex was ~$228B in 2024, ~$410B in 2025, and is guided to $720–740B in 2026. Add Oracle and the guided total is ~$810–840B. All four hyperscalers told investors 2027 spending rises again, without giving a figure — that refusal to bound 2027 is itself part of the skepticism.
Mixed company definitions. 2024: Doomdayer compilation of 10-Ks citing Bloomberg/Goldman/Morgan Stanley. 2025: FT tally via Luminix. 2026: Luminix Q3 2026 guide (opened).
Goldman Sachs supply-side model: cumulative AI capex 2026–2031, with $765B in 2026 alone.
Goldman Sachs Global Institute via GPU Smith
McKinsey: worldwide data-center investment required by 2030 — $5.2T of it AI-ready.
McKinsey, “The cost of compute,” Apr 28, 2025
JPMorgan: cumulative global AI-related capex by 2030, plus ~$1.5T of AI-related bond issuance over 5 years.
JPMorgan via Crisil Integral IQ, Oct 8, 2026
This is no longer funded like software. UBS — on a broader scope than the Big Four — puts hyperscaler capex at $492B (2025) → $1.009T (2026) → $1.447T (2027) → $1.619T (2028). Morgan Stanley’s CEO said on the Q2 2026 call that 2026 data-center capex is “coming in at about $850B” vs $575B initially projected. Every forecast revision in 2026 went the same direction: up. The question investors are asking is not whether demand exists — it is who bears the cost if the revenue arrives two years after the concrete.
The physical race: gigawatts, not just dollars
Largest sites
Amazon Project Rainier (IN, for Anthropic): ~1,037 MW IT power running Oct 2026 — the largest live AI site — 1,925 MW planned early 2028; >1M Trainium2 chips.
Microsoft Fairwater (WI): 2,263 MW planned by May 2028, 369 MW live. Bloomberg: Microsoft plans ~38 GW by 2032, >3× current footprint.
Meta Hyperion (LA): scaling toward 5 GW, >$50B in the parish; Epoch AI estimates 1,632 MW by Jan 2028, 0 MW running in Oct 2026.
Epoch AI via Mrkt30 / Sotkn, Oct 2026
The power bottleneck
Big Four have contracted 14,103 MW of nuclear capacity — but only 835 MW of new capacity is scheduled before end-2027, and no SMR power before 2030.
Deals: Microsoft–Constellation 835 MW Three Mile Island restart (20-yr PPA); Amazon–Talen up to 1,920 MW (~$18B, 17-yr); Google–Kairos up to 500 MW SMRs by 2035; Meta up to 6.6 GW by 2035 (Vistra/Oklo/TerraPower, Jan 2026).
Axis Intelligence ledger, Nov 3, 2026
Seven companies, four different races.
Pick a name. The pattern to watch: the two companies still generating free cash — Microsoft and, by abstention, Apple — were treated completely differently by the market in October from the three burning it.
Microsoft — the only one still covering its capex
Azure was disclosed in dollars for the first time in Q1 FY27: $33.2B, +48%, after $101.9B in FY2026. AI business crossed a $37B annual run rate, +123% YoY (Q3 FY26 call, Apr 29, 2026 — no newer AI run-rate was found). Commercial RPO hit $677B, +73%. The catch inside the beat: OpenAI arrangements contributed $10.0B of its $92.3B quarterly revenue — demand concentration in one customer.
Maia 200 chip — Jan 2026Fairwater 2,263 MW planned$9.7B / 5-yr IREN contractChip claims (~3× FP4 of Trainium3) are company claims, not independently verified. ~Three-quarters of Q1 FY27 capex went to short-lived assets, mainly CPUs/GPUs bought in — still heavily Nvidia-dependent.
Guidance cut Oct 28, 2026; the step down from $190B to $170B was partly a finance→operating lease reclassification. Capex missed the $50.67B StreetAccount estimate on third-party delays — and the stock was spared for it.
Alphabet — the fastest cloud, the harshest punishment
Google Cloud doubled: $30.7B in Q3, +103%, operating income $11.1B at a 36.2% margin (vs 23.7% a year earlier), backlog $561B. And the stock still had its worst day since the IPO. Why: EPS was hit by an −$8.2B unrealised SpaceX loss, Search slowed to +12%, capex hit $51.4B, FCF went negative for a second straight quarter, there was no buyback in Q3, and 2027 capex was guided “significantly” higher with no number.
TPU v7 “Ironwood”~25% of global AI compute — Epoch AIAnthropic: up to 1M TPUs, >1 GW in 2026Epoch AI (Apr 7, 2026): of ~5M H100-equivalents owned, ~4M on TPUs vs ~1M on Nvidia GPUs — capacity mix, not dollar mix. Alphabet discloses no TPU/Nvidia capex split.
Guidance raised Jul 22, 2026, held Oct 28. Q2 FCF was −$5.9B — first negative quarter since the 2004 IPO. Pivotal Research projects 2026 FCF falls ~90%, $73B → ~$8B.
Amazon — the biggest spender, the deepest cash hole
AWS is accelerating: $46.0B in Q3, +39.3% — fastest in ~5 years — at a 38.7% operating margin, ~$184B annualised, backlog $578B. Management guided AWS growth >40% in both 2027 and 2028 and floated a $1T AWS run rate within 5–10 years. The market focused instead on the funding: trailing-twelve-month FCF of −$42.4B and ~$48B of new debt in the quarter.
Trainium2 / Trainium3 (3nm)Project Rainier — largest live AI site44 GW under development, 16 statesSemiAnalysis (Nov 2026): Trainium was one of the few platforms with near-term forecasts left unchanged. Amazon plans to sell Trainium outside AWS from late 2027 (Luminix).
Net vs gross are both filing-true, different lines (10-K, Feb 6, 2026). Guidance raised Jul 30, 2026 from $200B, held Oct 29; 2027 “meaningfully more,” no figure. Stock fell ~7% after hours Oct 29.
Meta — ads fund it, until they don’t
Advantage+ AI ad solutions passed a $75B annual run rate in Q2 2026 (from $60B in Q3 2025); Q3 revenue $64.60B, +26%, price per ad +14%, ~10M advertisers using generative-AI creative tools. Important hedge: Advantage+ is ad revenue flowing through AI automation, not incremental AI revenue. Off the capex line, the Q3 10-Q shows $454.52B of non-cancelable commitments ($121.41B due 2027) plus up to $41.49B of residual-value guarantees on two off-balance-sheet ventures.
MTIA v2 — inference / rankingHyperion → 5 GWPrometheus ~1 GW, OhioGuidance narrowed Oct 21, 2026. Q3 2025 FCF was +$10.63B — a ~$12.7B year-on-year swing. Barclays projects Meta FCF −90% in 2026, then negative in 2027 and 2028 (via Decoding Discontinuity).
Nvidia — the supplier, and now the risk
Data-center revenue path: $51.2B → $62.3B → $75.2B → $89.0B (Q2 FY27, 92.5% of $96.2B total revenue, +106%). FY2026 revenue $215.9B, data center $193.7B. Q3 FY27 guide: $108.0B ±2% at 74.0% gross margin, assuming zero China data-center compute revenue. But Nvidia is also a financier: $99B of investments + $25B of commitments + $36B of cloud commitments at Q2 FY27, and maximum gross guarantee exposure of $108.5B — dominated by a $105.0B lease guarantee for OpenAI’s Pike County, Ohio campus.
~86% of AI accelerators73–75% gross marginsReports Nov 17 — after report cutoffNvidia’s own capex could not be verified as a clean sourced figure in the research — it is the supplier in this story, not a hyperscaler spender.
Finnhub closes via report index. Founder-CEO Jensen Huang resigned effective Nov 11; CFO Colette Kress is interim CEO. See Part 5.
Apple — winning by not playing
Apple issued no capex guidance — and spent less: FY2026 capex $9.0B, −29% YoY, while R&D rose to $45.8B, +32.5%. FY2026 revenue $478.8B, +15.1%; FCF ~$136.1B funded $86.39B of buybacks plus $15.72B of dividends. Its AI strategy is rental, not construction: the next Apple Foundation Models are based on Google Gemini, with Google Cloud as preferred provider. Siri AI rolled out in beta Oct 15–26, 2026. No AI revenue line is disclosed.
Best Mag 7 YTD: +26.7% to Oct 30A20 Pro — 2nd-gen Neural EngineThe market’s verdict is in the YTD table: the low-capex name led the group while the group underperformed the S&P 500 for the first time since “Mag 7” was coined in 2023.
10-K, results Nov 2, 2026. Q4 capex missed the $3.33B LSEG estimate — on the low side, and nobody punished it.
Tesla — capex doubled in the quarter, then the plan was cut
Q3 2026: revenue $28.05B, flat YoY; operating income $718M, −56% (2.6% margin); R&D $2.76B, +69% on “Super Intelligence”/Optimus/Cybercab. FSD: 1.74M active subscriptions, +67%; robotaxi ~2M cumulative unsupervised miles (~8× QoQ), Cybercab rides in Austin from Sep 3, 2026 — but Musk said robotaxi is “not a massive impact right now.” Unsupervised FSD for customer cars guided to ~Q2 2027. Training compute: Cortex 1 >105 MW + Cortex 2 >120 MW, plan 400 MW by end-2026. No AI revenue line.
AI5 chip — integration ~Apr 2027“Terafab” Austin — $100B+ long term (Musk)Guidance cut Oct 21, 2026 from a mid-year plan of >$25B. Shares fell ~4–5% the next morning on the −56% operating income — then held up relative to peers after the capex cut.
At a glance — 2025 actual vs 2026 guidance
| Company | 2025 capex | 2026 guidance | Q3 2026 FCF | Own chip |
|---|---|---|---|---|
| Amazon | $128.3B net | ~$220B | −$42.4B (TTM) | Trainium 2/3 |
| Alphabet | $91.4B | $195–205B | −$9.1B | TPU v7 Ironwood |
| Microsoft | $118.0B cal. | ~$170B | +$18.6B | Maia 200 |
| Meta | $72.22B | $135–145B | −$2.12B | MTIA v2 |
| Tesla | $8.53B | ~$20B | +$1.53B | AI5 |
| Apple | ~$12.7B | No guidance | FY26 ~$136.1B | A20 Pro (on-device) |
| Nvidia | Not verified* | Rev. guide $108B | Reports Nov 17 | Sells to all of the above |
* Nvidia’s own capex was listed as “could not verify” in the research. Custom ASIC shipments are growing ~44.6% YoY in 2026 vs ~16.1% for merchant GPUs, reaching 27.8% of AI server shipments — yet every custom chip still depends on TSMC 3nm + SK Hynix HBM, and Nvidia still holds ~86% of AI accelerators. JPMorgan (Nov 16, 2026): Nvidia GPUs will be only 39% of AI compute OpenAI + Anthropic buy going forward, vs 83% of AI chip revenue in 2023–25.
Why the panels don’t sum to the headline: the −$26.4B inversion in Parts 1 and 3 uses the cash-purchases basis — Big Four Q3 capex of ~$201.9B against ~$175.5B of operating cash flow. The company panels above use each filer’s lease-inclusive capex (Microsoft $45.2B, Alphabet $51.4B, Amazon $77.3B gross, Meta $34.91B), which sums to ~$208.8B for the same quarter. Both totals are correct on their own basis; the ~$7B gap is finance-lease additions, not a discrepancy.
Five arguments. Each one now has a number attached.
1 · The revenue gap — the maths that won’t close
Sequoia’s David Cahn framed it in June 2024: Nvidia’s data-center run-rate ×2 for the rest of the data-center cost ×2 for the gross margin end customers must earn ⇒ ~$600B/yr of end-user AI revenue required, vs a most-generous tally of only a few billion at the time. Re-running Cahn’s method on Nvidia’s latest annualised quarter gives ~$1.4T/yr required in 2026 (AIMojo re-run; ~$775B on FY2026 data-center revenue).
JPMorgan: new annual revenue in perpetuity needed for the buildout through 2030 to earn a 10% return — ≈$34.72/month from every iPhone user.
Bain & Co (Sep 2025): annual revenue needed to fund compute — with an ~$800B shortfall even after AI-related savings.
Recognised AI revenue run-rates as a share of annualised Big Four capex in Q3 2026.
Reuters Breakingviews calculation cited in Luminix Q3 guide
And adoption is the soft underbelly: MIT NANDA’s “GenAI Divide” (2025) found 95% of enterprise GenAI pilots showed no measurable P&L impact — the most-cited adoption-skeptic statistic in the debate.
2 · Free cash flow has broken — and bonds replaced buybacks
Amazon is trailing-twelve-month; others are the quarter. Sources: Luminix Q3 guide; FourWeekMBA. Bank of America projects aggregate FCF deficits across eight hyperscalers/neoclouds of $128B (2026), $335B (2027), $77B (2028) — via 456X, Nov 5, 2026.
In nine months to Sep 30, 2026, Big Four long-term debt went from $206.4B to $410.2B (Amazon $65.6B → $173.1B; Alphabet $46.5B → $125.1B; Meta $58.7B → $80.9B; Microsoft fell before its October bond). The Big Four raised a net $272B in nine months; buybacks fell 83%. Alphabet bought back no shares in Q3 2026. (TechBag, Nov 1, 2026, from 10-Q/10-K XBRL.)
The regime change in one sequence: Big Five bond issuance was ~$17–20B in 2024 → $109B in 2025 (BofA counted $121B — counters differ) → $194B in H1 2026 alone (JPMorgan AM). Goldman expects ~$250B in 2026 and $400B in 2027, funding ~⅓ of capex.
| Deal | Date | Size | Why it matters |
|---|---|---|---|
| Meta bonds, 6 parts | Oct 30, 2025 | $30B | ~$125B order book — then a record; 40-yr at +110bp |
| Meta / Blue Owl JV (off-balance-sheet) | Oct 2025 | ~$27B | Hyperion SPV private credit — debt not on Meta’s balance sheet |
| Alphabet, multi-currency | Feb 9–10, 2026 | ~$32B equiv. | Incl. a £1B 100-year bond — first tech century bond since Motorola in 1997 |
| Alphabet equity raise announced | Jun 1, 2026 | $80B | $30B underwritten offerings + $40B at-the-market + $10B from Berkshire Hathaway |
| Meta equity + convertible | Oct 26–29, 2026 | $37.5B | $25.0B common + $12.5B 6.50% mandatory convertible; plus a planned $60B ATM programme |
| Microsoft bonds, 10 tranches | Oct 29, 2026 | $25B | First US IG bond since 2017; over $100B of orders (>4× covered); 40-yr coupon 6.35% |
| Meta bonds, 5 tranches to 2056 | Nov 2–3, 2026 | $15B | Coupons 5.85–7.5% (Frenzy Capital tracker, Nov 16) |
3 · Depreciation — are GPUs 6-year assets on a 2-year cycle?
Hyperscalers depreciate servers/GPUs over 5–6 years while Nvidia’s product cycle is 2–3 years. Michael Burry (X, Nov 10, 2025, as “Cassandra Unchained”): hyperscalers will “understate depreciation by $176 billion 2026–2028”; by 2028 Oracle would overstate earnings 26.9%, Meta 20.8%. Hedges the report attaches: Burry has a disclosed short position and never published his model. Reverse-engineering (Luminix): true 4-year lives ⇒ ~$132B extra depreciation 2026–28; 3-year lives ⇒ ~$263B — Burry’s $176B sits between, implying a ~4-year value life.
Cutting one year from assumed equipment lives would erase ~$36B of the Big Four’s combined 2027 net income.
Reuters Breakingviews, cited in Luminix Q3 guide
Microsoft Q1 FY27 depreciation: $7.1B → $11.1B, vs revenue +19%. AWS depreciation +67% YoY in Q2 2026 ($4.84B → $8.08B). The bill is already arriving.
TON618; Luminix Eisman tracking
The life extensions are documented: Meta extended most server lives to 5.5 years in 2025, adding +$2.59B of net income (+$1.00/share); Alphabet ~3→6 years, Microsoft to 6, Amazon 4→6 then a subset back to 5 in early 2025. In July 2026 Microsoft extended data-center building lives 15→25 years and shifted leases from finance (counted in capex) to operating (not counted) — operating-lease additions jumped $1.4B → $6.5B YoY. Jim Chanos’s version: amortising CoreWeave’s ~$19.1B capital employed over even 7 years gives ~$2.73B of economic depreciation vs ~$2.62B annualised EBITDA — ~0% ROIC. Nvidia’s rebuttal (late 2025): customers see 4–6-year utilisation in practice.
4 · Circular financing — the money goes round
2026 analyses put interlocking AI arrangements at $800B+; Burry’s figure is $879B of “commitments circling through Nvidia.” The Bank for International Settlements (Bulletin No. 137, “Circular relationships among AI firms,” Oct 1, 2026) mapped the loops directly, and found 46.4% of AI-to-AI deal value in 2021–25 involved firms that also buy from / sell to each other — a structure it warns raises opacity and correlated exposure. Nvidia’s receivables tell the same story at company level: DSO rose 45 → 60 days on “extended payment terms on large, multi-quarter agreements.”
| Loop | Terms | Date |
|---|---|---|
| Nvidia → OpenAI | LOI to invest up to $100B per GW deployed (≥10 GW). Huang, Feb 1, 2026: “never a commitment.” Became a $30B equity check in OpenAI’s round — announced at $110B (Amazon $50B, SoftBank $30B, Nvidia $30B), closed Mar 31, 2026 upsized to $122B committed at $852B post-money | Sep 2025 – Mar 2026 |
| OpenAI → Microsoft | $250B Azure purchase commitment; Microsoft holds ~27% of OpenAI | Oct 28, 2025 |
| Microsoft + Nvidia → Anthropic | Up to $15B equity (Nvidia $10B, Microsoft $5B); Anthropic pledges $30B of Azure compute | Nov 18, 2025 |
| Nvidia → CoreWeave | ~7% equity; ≥$6.3B capacity order obliging Nvidia to buy unsold capacity; +$2B Class A at $87.20 (Jan 2026). CoreWeave holds ~$22.4B of OpenAI contracts | 2025 – Jan 2026 |
| AMD → OpenAI | 6 GW of AMD GPUs + warrants for up to 160M AMD shares at $0.01, vesting per GW deployed | Oct 2025 |
| OpenAI → Oracle | ~$300B compute over 5 years; ~45% of Oracle’s $664B backlog | 2025 |
OpenAI: ~$1.4T of infrastructure commitments over 8 years (Altman, Nov 6, 2025) against >$20B ARR at the time. By Sep 2026: ~$50B ARR (company net method), Q3 operating loss $10.9B (~−109% margin), and an investor plan showing −$278B of negative FCF over 2026–2030 (FT via Luminix).
Anthropic: run rate ~$81B at end-Oct 2026 against a draft-prospectus ~$450B of mostly non-cancelable cloud commitments over 7–10 years. Open question: cited totals conflict by scope — ~$450B (prospectus cloud) vs ~$518B (broader stack) vs ~$750B (incl. negotiations) vs ~$900B (one S-1/A reading, ≈11× run rate). The research does not reconcile them; neither do we.
5 · The market has started voting — beats get sold for capex
| Event | Result | Reaction | Driver cited |
|---|---|---|---|
| Alphabet Q2 · Jul 22, 2026 | Revenue $119.8B (+24%) beat; Cloud +82% | −7.1% next session | Capex raised to $195–205B; FCF −$5.9B |
| Meta Q3 · Oct 21, 2026 | Revenue $64.60B (+26%) | −7.0% | Capex $135–145B; FCF −$2.12B |
| Alphabet Q3 · Oct 28, 2026 | Cloud +103%, backlog $561B | −12.64% (~$500B) | FCF −$9.1B, no buyback, 2027 capex “significantly” higher |
| Amazon Q3 · Oct 29, 2026 | Operating income $28.1B beat; AWS +39.3% | ~−7% after hours | TTM FCF −$42.4B; ~$48B new debt in quarter |
| Microsoft Q1 FY27 · Oct 28, 2026 | Revenue $92.3B (+19%) beat; Azure +48% | −0.2% close | The exception: FCF stayed +$18.6B; capex missed low |
| Nvidia Q4 FY26 / Q3 FY26 | Both beat + raised guide | −5.46% / −3.15% | Even the supplier gets sold on beats |
Group de-rating: Mag 7 was 34.5% of the S&P 500 (Oct 1, 2026) at a median P/E of 34.8 vs 25.1 for other profitable members. Nvidia’s P/E fell ~38% during 2026 to ~28.5 trailing / 24.9 forward — because earnings outran the price. Year-to-date to Oct 30, the Mag 7 weighted return was +6.2%, underperforming the S&P 500 for the first time since 2023: Apple +26.7% (best), Nvidia +9.7%, Alphabet −10.5%, Tesla −11% (secondary compilation, MarsBit/HTX, Nov 4, 2026). Reuters (Oct 30): the Nasdaq-100 was set for a >$2T October market-value loss.
What fund managers say vs do: BofA Global Fund Manager Survey, October 2026 — 39% named AI hyperscaler capex the most likely source of a systemic credit event; 30% now expect a hyperscaler capex cut within 6 months, up from 14% in September; 45% expect an AI bubble in 2027. In September, a net 33% said companies are overinvesting; the most crowded trade was long global semiconductors (53% in September, 45% in October; long Mag 7 itself polled 16%) — yet in August managers were a net 56% overweight equities. Survey size is reported inconsistently (161 panellists / $453B vs 185 / $512B) — flagged in the research. No hard hedge-fund positioning dataset on AI shorts was found; only survey self-reporting and disclosed individual positions (Burry, Eisman’s partial hedges, Chanos).
The bulls’ numbers are real too. That’s what makes this hard.
Demand is contracted, not hoped for
Backlog stack ≈ $1.82T: Microsoft RPO $677B (+73%; the portion beyond 12 months +98%, weighted duration 2.3 yrs), Google Cloud backlog $561B (>55% to convert within 24 months), AWS backlog $578B (from $496B in June). Cloud growth is accelerating: Q3 Azure +48%, Google Cloud +103%, AWS +39.3% — and Google Cloud’s margin expanded 23.7% → 36.2% while doubling. Microsoft guided Azure up to 49–50% for the December quarter while saying “demand exceeds supply.”
Usage is exploding
Google processed >3.2 quadrillion tokens/month in May 2026, ~7× YoY (480T in May 2025), ~19B tokens/min via API. OpenRouter volume hit 25T tokens/week in May 2026, 5× in six months. McKinsey State of AI 2026 (1,719 respondents, 97 countries): 89% report regular AI use in at least one function. Rental rates for older GPUs rose ~30–35% in 2026 — inconsistent with a glut. Every Q3 constraint cited was supply, none was idle capacity.
Token figures are company-reported, unaudited, differently defined. Bulls cite McKinsey’s 89%; bears cite the flat 37% who say AI contributed to EBIT — and MIT’s 95%.
Both sides are using true numbers. The bulls own the demand side — backlog, cloud acceleration, token growth. The bears own the funding side — negative FCF, $410B of long-term debt, depreciation that hasn’t landed yet, and commitments ($1.7T of long-term leases + $1.9T of unconditional purchase commitments at tier-one hyperscalers, $685B maturing before end-2027, per Goldman via 456X) that make the spending hard to stop even if demand wobbles. The WSJ counts ~$3T of off-balance-sheet purchase/lease commitments across nine AI-spending firms, ~3× their traditional debt + leases. This is a timing bet, not a fraud-or-fantasy bet: does contracted revenue convert before the depreciation and debt service peak?
November broke the story in three places at once.
1 · SemiAnalysis cut the future — mostly on price, not units
Model update dated Nov 13, 2026, reported Nov 16: global AI accelerator capex cut 28% for 2028 ($1.054T → $758B) and 40% for 2030 ($1.531T → $923B, −$608B); 2027 cut 19% to $627B; 2029 cut 41% to $796B. Nvidia 2030 shipments cut 66% on a compute-revenue basis; Nvidia is 66% of the total silicon-area reduction. Drivers: slowing ARR growth at frontier labs in Q3, tighter financing for projects without hyperscaler commitments, and ~5.8 GW of commissioning delays including US midterm-related regulation.
The cut is mostly price, not units — 2030 wafer area falls only 15% and HBM bits 12% vs the 40% dollar cut, as HBM prices normalise. The 2026 forecast barely changed ($392B base), revised spending still grows +60% in 2027, Broadcom’s cumulative ASIC revenue was cut only 11%, and Trainium forecasts were left unchanged. SemiAnalysis explicitly said the model did not factor in Huang’s resignation. The primary report is paywalled; figures are via BigGo Finance / Longbridge / CoinCamps summaries.
Market reaction, Mon Nov 16: Philadelphia Semiconductor Index −5.28%, SanDisk −8.6%, Micron −7.62%, AMD −5.58%, Nvidia −4.12% to $156.14, Nasdaq Composite −2.49% to 24,509.76 (now >10% below its Oct 6 high — a correction), VIX +19.84% to 23.13. Two confounders the same day: a new Commerce Dept rule blocking Chinese firms from renting US AI chips in third-country data centers, and pre-earnings caution.
2 · Nvidia lost its founder — a governance shock on top of a capex debate
Nov 11: the board accepted Jensen Huang’s resignation as president, CEO and director, effective immediately (announced Nov 12). CFO Colette Kress (CFO since 2013) became interim CEO while keeping the CFO role; Harvey C. Jones became the first-ever board chair; no timeline or shortlist for a permanent CEO was given. The company said operations, commitments and roadmap continue without interruption. The resignation followed reports and an independent-director investigation described by the NYT and Reuters — these are attributed allegations, not findings; Reuters said it had not independently verified the foreign-intelligence allegation, and Beijing denies it.
The stock: −4.27% Nov 9, intraday low $148.80 Nov 12, −12.75% in the week to Nov 13 (−$573.34B — the largest weekly wipeout on record, per AInvest), $156.14 on Nov 16, −34.7% below the Oct 6 record close of $239.24. Notably, the semiconductor index rose 8.1% in the week Nvidia fell 12.75% — that week’s hit was Nvidia-specific. At least 10 analyst downgrades from Buy to Hold followed (FactSet via CNBC), all leadership-driven, not capex-driven: Baird $500 → $200, Bernstein $400 → $200, Citi $285 → $225, Stifel $315 → $205, Oppenheimer withdrew its target entirely. HSBC’s prior target is reported as both $365 and $375 — a conflict flagged in the research. Cantor Fitzgerald held Overweight at $250. On Nov 16 CNBC reported the board was not considering bringing Huang back, despite an FT report of return discussions and a >180,000-signature petition.
3 · Data centers became a ballot issue
Nov 3 midterms: Trenton, OH voters banned data centers >25 MW with ~80% of the vote (threatening an ~1,000 MW Amazon campus); ~67% in Piqua, OH backed a charter amendment aimed at Meta’s 1.4 GW campus and recalled the mayor and two commissioners. New governors in TX, OH and GA pledged moratoria/pauses. Cleanview tally: 158.6 GW of planned capacity — ~41% of the national pipeline — sits in states whose new governors promised pauses (TX 88.4 GW, OH 21.9, GA 20.2, NM 11, NY 9.9). Anthropic’s prospectus (filed Oct 26) already listed moratoria as a risk factor.
Cancellation tallies disagree by scope — treat them as ranges, not a single count: Programs.com counts 69 large projects cancelled since Jan 2024 (15,910 MW on the 26 reporting size; $165B on the 32 reporting cost); Sue Data Centers counts 164 cancelled/withdrawn US sites, 14.1 GW, $69.9B; Data Center Watch counts 75 projects (~$130B) blocked/delayed in Q1 2026 and 45 (~$68B) in Q2. Counterpoint (Kenanga, Nov 10): the ~400 GW planned pipeline is heavily inflated and chip supply is ~21% short of even the reduced pipeline, so a ~50 GW cancellation wave would not immediately cut chip orders; Samsung/SK Hynix reported no customer cancellations after the election. On Nov 4, hyperscalers actually rallied on a slower-capex/scarcity read (Microsoft ~+6.7%) while power and equipment names sold off.
Execution is slipping inside the prints too: Microsoft’s $5.5B capex miss was third-party capacity not delivered on time (pushed to H1 2027); Microsoft reportedly did not exercise an option for ~700 MW of extra Nscale capacity in Texas and withdrew from Nscale’s Essex, England project after its grid connection slipped into the 2030s. Gartner predicts 80% of neoclouds will merge or fail by 2030.
What decides it — and what we still don’t know.
Watch these, in order
1. Nvidia, Nov 17. First print under Kress, against its own $108.0B ±2% guide; LSEG consensus ~$109.1–109.2B revenue, EPS ~$2.47–2.49; options implied move ~8.3–8.6%. It falls after this report’s cutoff — no result is claimed here.
2. Cash cover, not growth. October proved the market’s new rule: Microsoft missed capex low and was spared; Alphabet doubled Cloud and lost ~$500B. The next prints will be read FCF-first.
3. Whether 2027 gets a number. All four hyperscalers guided 2027 up without a figure. A bounded 2027 guide — or the capex cut 30% of fund managers now expect — reprices everything.
4. Depreciation prints. Microsoft depreciation already +56% YoY. Each quarter tests the 5–6-year life assumption against a 2–3-year product cycle.
Open questions — stated plainly
• Nvidia’s own capex: not verified as a clean sourced figure.
• Microsoft’s AI run-rate after Apr 2026 ($37B): no update found in the Q1 FY27 materials reviewed.
• Anthropic’s total commitments: $450B–$900B depending on scope — unreconciled in the sources.
• Hedge-fund positioning on AI shorts: no hard dataset found — only surveys and disclosed individual positions.
• Anthropic’s IPO pricing and Nvidia’s Nov 17 results: after the research cutoff.
• No clean sell-side downgrade of Microsoft/Alphabet/Meta/Amazon on capex was found in Oct–Nov 2026 — the downgrade wave hit Nvidia (leadership) and neoclouds. The spending skeptics were fund managers, not brokers. Ranmore’s Sean Peche (Nov 13) is the named bearish voice; Bernstein pushed back the same week.
The race is real — $1.82T of backlog, cloud growth accelerating at all three hyperscalers, token usage up ~7× in a year. The skepticism is also real — and it is no longer about whether AI works. It is about a buildout that in one quarter outspent the operating cash of the four richest companies in tech, funded the gap with $272B of net new borrowing in nine months, pushed buybacks down 83%, and booked the assets over 5–6 years while the chips inside them turn over in 2–3. Investors aren’t betting against AI. They’re asking who gets paid first: the bondholders, or the shareholders who used to get the buyback.
Educational only. Not financial advice.
Line-by-line review — Nov 17, 2026.
Every load-bearing figure on this page was re-checked against its cited source or an independent one — company filings and call transcripts where they exist, the BIS bulletin itself, the survey write-up, and the closing-price reports. Result: the core of the page holds. Seven items were wrong, incomplete or misattributed and have been corrected in place; they are listed first so nothing was changed silently. A second pass later the same day re-read the corrected page end to end and applied three further fixes, listed after the first seven.
Corrections applied in this review
1. BofA “most crowded trade.” The page said 53% of fund managers called long Mag 7 the most crowded trade in September. The survey write-up (Atrani) says the 53% September reading — 45% in October — was long global semiconductors; long Mag 7 polled 16% in October. Sentence rewritten.
2. OpenAI’s funding round. The circular-financing table called it “OpenAI’s $110B round … closed Mar 31, 2026 at $852B post-money.” The $110B was the announced size (Feb 27); the round closed upsized at $122B committed — $730B pre-money + $122B = the $852B post-money the page printed. Wording corrected.
3. Alphabet’s $80B equity raise. The bond table’s breakdown (“$40B at-the-market + $10B from Berkshire”) summed to $50B of an $80B raise. The missing piece — $30B of underwritten offerings — is restored.
4. BIS attribution. The 46.4% circular-deal figure was credited to the BIS Annual Economic Report of Jun 28, 2026. The figure is from BIS Bulletin No. 137, “Circular relationships among AI firms,” Oct 1, 2026, and the sentence now reflects what the bulletin actually says (opacity and correlated-exposure risk) rather than a “bust/collapse” framing it does not use.
5. Microsoft bond demand. “~$115B of orders (~4.6×)” could not be confirmed; Reuters via TechBag reports “more than $100B of orders” and Luminix “more than four times oversubscribed.” Softened to match.
6. Alphabet guidance path. The panel opened the path at $175–185B. Contemporaneous coverage from February through June 2026 puts the standing guide at $180–190B before the Jul 22 raise to $195–205B; the $175–185B starting point could not be confirmed and was replaced.
7. Meta year-ago FCF. Q3 2025 FCF printed as +$10.62B; the underlying report and Meta’s definition give +$10.63B. Fixed.
8. Second pass — Nvidia’s drawdown, stated two ways. The company panel printed ~−35% below the Oct 6 record close and Part 5 printed ~33–35%. From the page’s own verified figures — the $156.14 close on Nov 16 against the $239.24 record close of Oct 6 — the drawdown is −34.7%. Both places now print −34.7%.
9. Second pass — the two capex bases now reconcile on the page. The headline inversion (−$26.4B) uses the cash-purchases basis (~$201.9B), while the Part 2 panels use lease-inclusive capex and sum to ~$208.8B for the same quarter. A note under the at-a-glance table now states both bases and the ~$7B finance-lease gap, so the panels and the headline no longer appear to contradict each other.
10. Second pass — Microsoft guidance note. The panel’s source line read “the $190B→ cut was partly a finance→operating lease reclassification,” with the destination missing. It now reads “the step down from $190B to $170B.” The same pass fixed two presentation defects: anchored sections now land below the sticky nav instead of under it, and the company tabs gained full keyboard support (arrow keys, Home/End) with proper tab/tabpanel semantics.
Confirmed against source — the load-bearing claims
Spending & cash: Big Four 2026 guidance $720–740B and every Q3 figure in the company panels (Luminix, opened in full, incl. its embedded company reports); the −$26.4B inversion — capex ~$201.9B vs operating cash ~$175.5B, only Microsoft covered (FourWeekMBA’s own filing arithmetic); long-term debt $206.4B → $410.2B with each company’s split, buybacks $81.0B → $13.6B (−83%), net $272B raised (TechBag, from 10-Q/10-K XBRL); Alphabet’s £1B 100-year bond, Feb 9–10, 2026, ~$32B total, first tech century bond since Motorola 1997 (Bloomberg/Reuters coverage).
Companies: Microsoft — Azure $33.2B +48%, RPO $677B +73%, OpenAI $10.0B of revenue, capex $45.2B vs the $50.67B StreetAccount estimate, and the $37B AI run rate +123% (Microsoft’s own Apr 29, 2026 call transcript). Alphabet — Cloud $30.7B +103% at a 36.2% margin, capex $51.4B, FCF −$9.1B, Q2 FCF −$5.9B, the −$8.2B SpaceX-driven other-income loss, −12.64% on Oct 29 (~$500B; worst day since the 2004 IPO). Amazon — AWS $46.0B +39.3%, backlog $578B, TTM FCF −$42.4B. Meta — capex $34.91B incl. lease principal (106% of operating cash), FCF −$2.12B, $454.5B of non-cancelable commitments, $41.49B residual guarantees, Blue Owl JV ~$27B. Apple — the full FY2026 panel against the 10-K (revenue $478.8B, capex $9.041B, FCF $136.1B, buybacks $86.39B, R&D $45.8B). Tesla — revenue $28.05B flat, operating income $718M −56%, R&D $2.76B +69%, FSD 1.74M +67%, FCF +$1.53B, capex guide cut to ~$20B.
The skepticism & November: BIS 46.4% (bulletin text); BofA FMS October — 39% systemic-credit-event, 30% expecting capex cuts, 45% expecting a 2027 bubble, net 33% “overinvesting” in September, net 56% overweight equities in August (Atrani write-up; its own survey note gives 161 panellists / $453B, one side of the size conflict already flagged on this page); SemiAnalysis cuts — 2028 −28% to $758B, 2030 −40% to $923B, 2027 −19%, 2029 −41%, the price-not-units nuance (2030 wafer area −15%, HBM −12%), Broadcom ASIC −11%, Trainium unchanged, and the explicit note that Huang’s resignation was not factored in (BigGo’s summary of the paywalled primary); Huang’s resignation, Kress as interim CEO, Jones as first chair (Nvidia’s release; TechCrunch); the target cuts — Baird $500→$200, Bernstein $400→$200, Citi $285→$225, Stifel $315→$205, Oppenheimer’s target withdrawn; Anthropic’s ~$81B run rate at Oct 31 and its ~$900B S-1/A commitment stack; OpenAI’s ~$50B run rate; Mag 7 +6.2% YTD to Oct 30 with Apple +26.7% best, Nvidia +9.7%, Alphabet −10.5%, Tesla −11% (MarsBit/HTX compilation); the moratoria record — Trenton ~80% vote on a ~1,000 MW Amazon campus, Piqua ~67% with the mayor and two commissioners recalled, Cleanview’s 158.6 GW / 41%, and all three cancellation tallies (Programs.com, Sue Data Centers, Data Center Watch); Nov 16 closes — SOX −5.28%, Nasdaq −2.49% to 24,509.76, Nvidia −4.12% to $156.14, Micron −7.62%, AMD −5.58%. Internal arithmetic also re-footed: the backlog stack sums to $1.816T ≈ $1.82T, the four debt splits sum to the $410.2B total, and both charts’ bar widths are proportional to the printed figures.
Conflicts and limits that stay on the record
• VIX, Nov 16: printed here as +19.84% to 23.13 (CNBC-based closing reports); Yahoo-based tallies print a 22.91 close (+18.7%). Vendors disagree; the page keeps the CNBC-side figure.
• HSBC’s prior Nvidia target ($365 vs $375) and Anthropic’s commitment total ($450B–$900B by scope) remain unreconciled in the sources, as flagged in Parts 3 and 5.
• Not independently re-verified in this pass — carried on their cited sources as printed: the 2024/2025 capex totals and the forecaster stack (Goldman $7.6T, UBS path, JPMorgan $5.5T and ~$650B/yr), Epoch AI’s site-level MW figures, the Axis nuclear ledger, Meta’s Nov 2–3 bond tranches (Frenzy tracker), BofA’s projected FCF deficits via 456X, Nvidia’s −12.75% week (−$573.34B) and −4.27% on Nov 9, Amazon’s ~−7% after-hours move on Oct 29, and the pre-2026 reaction rows in the Part 3 table (Nvidia Q4 FY26 / Q3 FY26). The long-documented 2024–25 items — Sequoia’s $600B question, Bain’s $2T / $800B shortfall, MIT NANDA’s 95%, McKinsey’s $6.7T, Burry’s $176B depreciation claim, Meta’s Oct 2025 $30B bond — match the public record as printed.
• Nvidia’s own capex and Apple’s AI revenue line remain not verified / not disclosed, exactly as Parts 2 and 6 state.
Educational only. Not financial advice.
Get it in your inbox
The Basic briefing. Before the bell.
Every edition of The Degenerates Journal, delivered free — the Morning Tape before the bell, the Final Hour at 5:30 PM, and the Sunday edition. No noise, no spam — unsubscribe anytime.