Ed Zitron's AI Bubble Case: OpenAI, Anthropic and Junk Debt (Oct 2026)

Ed Zitron’s October 2026 position is not that “AI is useless.” It is that the capital cycle around generative AI only exists because one company — OpenAI — became the industry’s myth, largest compute customer, and fundraising machine, and that almost every other number in the trade is downstream of that. He has argued this across the free newsletter The OpenAI Bubble (15 July 2026), the Better Offline podcast, and public interviews. What follows is a structured outline of his case, in paraphrase with short quotes. It is not an independent audit. Several load-bearing figures are his estimates or readings of documents; others he attributes to filings and press. Premium newsletter text is not reproduced.

1. The claim: the AI bubble is an OpenAI bubble

Zitron’s thesis is that talking about a broad “AI market” hides a single point of failure. In his words, “without OpenAI, the AI industry doesn’t exist, and the justification for trillions of dollars of capex evaporates.” The boom, he argues, is not the result of measured return on investment. It is “an episode of cult-like psychosis” in which the mythology of one company inspired “the greatest capital misallocation in history.” He calls an OpenAI failure “the Lehman Brothers of the AI bubble.”

The mechanism he describes is simple. ChatGPT made AI feel inevitable. Hyperscalers then spent as if many OpenAIs would appear and all of them would rent enormous GPU clusters. In his view, that second OpenAI never arrived as an independent business. Anthropic exists, he says, largely as a response to OpenAI — founders from the same shop, plus Google and Amazon funding a rival so they would not be left behind Microsoft. Remove OpenAI from 2020–2024, he argues, and the capex wave does not start.

• Primary statement: The OpenAI Bubble, 15 July 2026 (free).
• Press distillation: Business Insider, 16 July 2026, quoting the Lehman line and the claim that OpenAI is why anyone cares about AI at all.
• Podcast companion: Better Offline, “Monologue: The Year Ahead” (16 January 2026), framed as why the bubble cannot survive the death of OpenAI.

What this means if he is right: the trade is not diversified across models, startups, and cloud software. It is a bet that OpenAI keeps raising, keeps signing compute contracts, and does not visibly break.

2. How the mythology was built

Zitron dates the setup to the 2022 tech bust: a dead IPO market, rate hikes, layoffs, and no new product category. ChatGPT, launched November 2022, arrived as the excuse to spend again. By the end of 2023, he notes, OpenAI had about 100 million weekly users and roughly $108 million in monthly revenue — popular, but not a business that could fund its own infrastructure. Microsoft’s early money, he stresses, was largely Azure credits, not cash that proved independent demand.

He then argues the industry made a category error: because one large compute customer existed, investors assumed a template for many more. “The assumption… was that because one OpenAI existed, many OpenAIs would bloom.” Data-center debt and startup funding, in this telling, were bets on becoming the next OpenAI or the next OpenAI’s landlord. Venture capital alone, he says, could not have built the clusters; only Microsoft, Google, and Amazon had the balance sheets and construction experience. He calls OpenAI and Anthropic “hyperscaler psy-ops”: incubated, loss-subsidized, and then pointed at as proof that outside demand was real.

• The OpenAI Bubble.
• Related free framing of the post-2022 “rot” and lack of new growth ideas is threaded through that essay and earlier Rot Economy writing on the same site.

Implication: the apparent customer is often the landlord’s own portfolio company. That is not the same thing as a market that can stand without the sponsor.

3. The unit economics do not close

Zitron’s June 2026 exclusive, which he says is based on audited documents independently checked by the Financial Times, is the factual spine of the later essays. He reports:

• 2024: about $3.7 billion revenue, $12.4 billion in costs and expenses, operating loss about $8.8 billion, net loss attributable to the company about $5.1 billion after a large “noncontrolling members” adjustment he says he cannot explain.
• 2025: $13.07 billion revenue, $34 billion in costs and expenses, operating loss $20.92 billion. After a conversion-related fair-value charge and further noncontrolling adjustments, net loss attributable to the company about $38.5 billion.
• Related-party flows: SoftBank paid OpenAI $867 million and Microsoft $303 million in 2025; OpenAI’s expenses to Microsoft totaled about $17.2 billion, including $10.59 billion he reads as training/R&D.

His reading: revenue grew, but spending grew faster, and accounting adjustments made the headline loss harder to interpret rather than better. He has repeatedly said he sees no path from those economics to a self-funding lab. In October interviews he has extended the same logic to Anthropic, arguing that 2025 filings show it spent about $2.75 for every dollar of revenue — worse, on his comparison, than OpenAI’s roughly $2.60 — with a large share of sales running through Google and Amazon, which are also competitors and financiers.

• Free write-up: Exclusive: OpenAI Losses Increased Nearly 8X in 2025, 15 June 2026.
• Podcast: Better Offline, “Exclusive: OpenAI Spent $34 Billion In 2025, With Losses Up 8X” (16 June 2026); “Monologue: Anthropic Lost $8bn In 2025” (30 September 2026).
• Public interview summaries: Monetary Matters, 1 October 2026; Risk Takers, 13 October 2026.

He treats “they will grow into the losses” as the belief that has to be proven, not the base case. The commitments he stacks on top — he cites more than $748 billion of remaining performance obligations at Microsoft, Amazon, and Oracle tied to OpenAI, plus other neocloud and chip deals, and an intention to burn on the order of $852 billion by 2030 — only work, in his view, if fundraising never stops.

4. Annualized revenue is a marketing number, and startups are the customers

By October 2026 this is his sharpest media critique. In Annualized Enron Rate (13 October 2026) he argues that “annualized run rate” takes a short, often consumption-based period, multiplies it, and gets treated as recurring revenue. He cites the early-October episode in which a reported ~$70 billion OpenAI ARR was revised to about $50 billion, with the Financial Times describing differences in whether partner sales are grossed up. Markets moved on a metric he says “doesn’t really tell anyone anything.” He also says OpenAI’s actual 2026 revenue projection is on the order of $30 billion, not the annualized figure headlines use.

The mechanism he describes for the revenue that does exist:

• AI startups raise venture money, resell or wrap model access, and let users burn far more tokens than the subscription price.
• Those token bills become OpenAI and Anthropic “revenue.”
• The labs’ growth therefore depends on startups that, on his reading of Ramp data, are a tiny slice of customers producing most enterprise spend.

In Concentration Risk (8 September 2026) he cites Ramp’s lead economist: about 80% of OpenAI and Anthropic enterprise revenue comes from 1% of customers, skewed toward tech and AI products, and that concentration has not improved in three years. He compares those startups to pre-2008 “NINJA” borrowers — demand manufactured by a system desperate to originate volume, not by customers who can pay from their own cash flow. His example is Harvey: large raises against a much smaller annualized revenue base, with a large share of the cash, he argues, flowing onward to the labs.

• Annualized Enron Rate.
• Concentration Risk.
• His X posts on 13 October 2026 summarize the same ARR argument and link the newsletter.

Implication: if venture funding for wrapper startups slows, lab “growth” slows even if ChatGPT usage among consumers does not collapse. He treats that as the dirty secret under the ARR headlines.

5. “Insatiable compute demand” is mostly two signatures

Zitron’s demand argument is that backlog is not the same thing as a market. Hyperscalers report capacity constraints and giant remaining-performance-obligation numbers. He says those contracts are overwhelmingly OpenAI and Anthropic (and, in some cases, Meta) signing capacity they cannot fund from operations. In August 2026 he was publicly arguing that 70% or more of Microsoft, Google, and Amazon AI revenue came from those two labs, and that OpenAI alone was about 70% of Microsoft’s AI revenue and more than 7% of Microsoft’s fiscal 2026 revenue — against, in an IBD interview, Microsoft AI revenue of $34.33 billion of which he attributed $24.1 billion to OpenAI, set against hundreds of billions of capex since 2022.

He has also argued, including on Better Offline (“Monologue: The Pale Horsemen Arrive,” 7 August 2026; “Monologue: AI’s Compute Demand Story Is A Lie”), that outside the labs and the clouds serving them, real third-party compute demand is only a few billion dollars — in some premium framings he has put it under $1 billion. The public version of the claim is consistent even when the exact residual differs: there is no broad base of customers large enough to justify the build.

• Site news note, 5 August 2026: Microsoft disclosures and OpenAI’s share of FY26 AI revenue.
• IBD interview, 23 September 2026, where he says analysts are “declaring victory when they should be declaring failure,” and that the two labs “can’t afford to actually pay for the compute.”
• Where're All The AI Chips?, 22 September 2026.

If this is right, “sold out” data centers are a queue of subsidized buyers, not evidence that inference is a scarce, high-margin utility.

6. The chips are bought faster than they can be plugged in

This is the argument that has dominated his September–October media tour. He says NVIDIA’s revenue growth is largely speculative purchasing by hyperscalers and neoclouds that take years to install GPUs. Citing Bloomberg reporting that Microsoft had about 12 gigawatts of capacity but only about 2 gigawatts centered on AI-specific chips, he argues Microsoft has spent on the order of $265 billion in capex since the start of 2022 to bring a much smaller dollar amount of GPUs actually online, and has used aggregate capacity language to blur the difference between CPUs, storage, and AI accelerators.

His warehouse estimate, repeated on Better Offline (25 September 2026), The Tech Report, and the George Noble panel (5 October 2026), is roughly $200 billion to $350 billion of GPUs sitting in warehouses or in data centers without power. He distinguishes this from dot-com dark fiber: fiber, once lit, became general-purpose infrastructure; these chips depreciate, are model-specific, and earn nothing until power and buildings exist. In Dead Money (29 September 2026) he pairs that with reports of flat-to-down older-GPU rental rates, and argues that when new capacity does come online it is immediately absorbed by OpenAI or Anthropic — which creates the appearance of demand.

He also argues planned capacity — he cites Sightline figures around 190 gigawatts — would require on the order of $1.6 trillion to $2.9 trillion in annual demand to fill, against what he estimates is only about $22 billion of demand outside the two labs. Construction timelines of 18–24 months, slipped Stargate sites, and power shortages are, for him, the physical constraint the stock market is ignoring.

• Where're All The AI Chips?.
• Dead Money.
• Earlier related free piece: NVIDIA Isn't Enron — So What Is It? (December 2025), on Blackwell shipments he could not find installed.

7. Oracle is the clearest single-tenant accident

Zitron’s Oracle argument is the OpenAI-bubble claim in one balance sheet. The public version, laid out in the premium essay How OpenAI Kills Oracle (24 April 2026) and in the Better Offline monologue of the same name, is:

• Stargate was announced as a $500 billion American infrastructure company. He says Stargate LLC was never formed, no government money arrived, and the only real project was a campus in Abilene already underway, with other named sites still mostly land and steel.
• Oracle is building on the order of 7 gigawatts largely for one tenant, funded with project finance that keeps large debts off the corporate balance sheet, against a reported multi-year compute deal on the order of $300 billion.
• Schedules have slipped repeatedly. He argues OpenAI would need something like a 10x revenue increase and well over $150–200 billion more capital to pay Oracle, Amazon, Microsoft, CoreWeave, and others. If it cannot, Oracle’s growth engine and its data-center lenders are the ones holding the contract.

By October he was pointing at Oracle data-center debt trading below par (he has cited an $18 billion complex under pressure) and at higher yields if that debt were issued into today’s market. The podcast line is blunt: Oracle does not get paid if OpenAI cannot pay.

8. Debt is the timer, not a sudden product failure

Zitron’s October view is that the bubble does not need a model to “stop improving.” It needs borrowing to get expensive for credits that cannot earn an investment-grade rating. In Credit Crunch (6 October 2026) he walks through why a selloff in 10-year Treasuries — he cites yields around 5.3% — raises the floor under every data-center loan, then adds company-specific spreads. He names the exposed list as every data-center developer, SoftBank, the labs, and the hyperscalers themselves.

His credit conclusion, stated in that free essay and in the 9 October Tech Report interview on “Credit Crunch”:

• OpenAI and Anthropic would need at least $50 billion a year in debt after any IPO, possibly $50–100 billion across high-yield, private credit, and leveraged loans.
• At negative EBITDA — he uses the 2025 OpenAI operating loss as the illustration — they would be junk, roughly B to CCC, unless a hyperscaler or a government guaranteed them. He has mocked the guarantee structure as “a student loan where your mum or your dad co-signs, except your mum is Sundar Pichai.”
• Hyperscalers, he says, are themselves planning something like $400 billion of bond issuance in 2027. That issuance competes with Treasuries and with the labs’ own future paper.
• Private credit to neoclouds is already, in his phrase to Monetary Matters, a “brewing crisis,” because interest accrues while buildings are late. He has pointed at CoreWeave’s large debt stack and at Goldman counts of lower-rated AI borrowing (reported around $88 billion this year in secondary coverage of his interviews).

He also argues rising rates hit the loop twice: they make the build more expensive, and they make the venture dollars that become lab revenue harder to raise.

• Credit Crunch.
• Premium follow-on, not quoted here: The Hater's Guide To Junk, 9 October 2026.
• Interviews: CEPR Mostly Economics #55 (8 October 2026); Benzinga summary of the Monetary Matters comments (2 October 2026).

9. Why he says this is worse than the dot-com bubble

The January 2026 Better Offline series “AI Is Worse Than The Dot Com Bubble” (parts one through four, 27–30 January) is the historical frame he still uses.

• Dot-com split, in his telling, into flimsy websites and a telecom overbuild. Some of the ideas (delivery, marketplaces) later worked. Dark fiber could be lit for general internet use. The mania was spread across many stocks.
• This cycle, he argues, has the formerly “sensible” companies — Microsoft, Google, Amazon, Meta — doing the irrational spending. GPUs are not general-purpose dark fiber. Depreciation hits earnings even if the chips sit idle. The stock market’s exposure is concentrated in NVIDIA, which he says must keep selling accelerators into customers who are borrowing to buy them.
• His stated endpoints, from the 30 January episode: a 40–90% collapse in NVIDIA revenue over a short period; multi-billion-dollar impairments at the hyperscalers; an existential contraction in venture capital; Oracle in severe distress or worse; and on the order of $200 billion or more in unpaid data-center debt. He has said he hopes he is wrong and that he is “terrified” because the historical analogue was already destructive — jobs, pensions, a lost decade in parts of tech employment — and this one has more debt and fewer real customers.

He rejects the comforting version of the analogy (“the fiber was useful, so this will be too”). Usefulness of a chatbot, in his framework, does not amortize a GPU that was bought for a tenant who cannot pay.

10. What he thinks happens, and who pays

He does not publish a date. He publishes a transmission path.

1. A visible OpenAI break — missed growth, a failed or ugly IPO, an inability to close the next round, or a public cut to free ChatGPT and price hikes — is the psychological event. Until then, he says, NVIDIA keeps shipping and commitments keep being signed.
2. Counterparties that booked OpenAI as revenue or as the tenant — Oracle, CoreWeave and other neoclouds, and the hyperscalers’ AI lines — lose the cash flow those contracts assumed.
3. SoftBank, which he describes as heavily exposed to funding OpenAI, is a direct casualty. So are pensions and insurers holding private-credit and junk paper issued to build the sites.
4. Venture funding for AI startups contracts, which removes the artificial token demand that had been propping up lab revenue.
5. Equity markets reprice NVIDIA and the AI trade. He has said he expects “a violent, punishing effect on the entire stock market” as a precursor to a broader drawdown, not a contained startup bust.

He is explicit that Wall Street placed the bets and that households will hold the losses through 401(k)s, insurance portfolios, and any eventual pressure for a backstop. He has treated hints of a federal guarantee for data-center loans as a tell that the private economics do not work (Better Offline, “AI’s Desperation Era,” November 2025).

On the product itself, the financial case does not require him to claim models do nothing. His separate argument — in The AI Hater's Manifesto (25 August 2026) and in the September piece AI Is Already In Dangerous Hands — is that LLMs are interesting, their cost is inexcusable relative to reliability, and “safety” panic (he has singled out researcher media tours) functions as distraction from cash burn. AGI announcements in September 2026, including Jensen Huang saying AGI had arrived around a still-incomplete Abilene campus, are for him a timing device: change the subject while the S-1 season approaches.

How to read the numbers

Treat three layers differently.

• Figures he says come from audited OpenAI statements, with FT verification: 2024 and 2025 revenue, costs, and losses in the June exclusive. Those are the hardest numbers in his file, and they are still his reading of documents, including adjustments he flags as unclear.
• Figures he attributes to company disclosures and major press: Microsoft AI-revenue mix, OpenAI’s funding round (he cites $122 billion committed at an $852 billion post-money valuation), partner revenue shares, ARR revisions reported by the FT, Axios, and Reuters. Secondary write-ups sometimes compress his arithmetic; the newsletter is the better source.
• Figures that are his estimates: unplugged GPU stock ($200–350 billion), share of global compute, residual demand outside the two labs, and the 40–90% NVIDIA revenue drawdown. He presents these as reconstructions from capex, power, and installation reporting, not as line items in a 10-K.

He has been early, loud, and sometimes ahead of the documents. Anthropic’s 2026 reported growth has been used by critics to say he underestimates the labs; his reply, including in the October Monetary Matters conversation, is that run-rate growth funded by cloud partners and venture-subsidized startups does not answer the cash question. That disagreement is the live argument as of mid-October 2026, not a settled postmortem.

Where to read him, in order

Free essays, which carry the thesis:

• The OpenAI Bubble — 15 July 2026. The single best statement of the claim.
• Exclusive: OpenAI Losses… — 15 June 2026. The financial exhibit.
• Concentration Risk — 8 September 2026.
• Where're All The AI Chips? — 22 September 2026.
• Dead Money — 29 September 2026.
• Credit Crunch — 6 October 2026.
• Annualized Enron Rate — 13 October 2026.

Podcast (public show notes, not transcripts reproduced here): Better Offline — especially the January 2026 dot-com series, the June 2026 OpenAI financials episode, “How OpenAI Kills Oracle” (April 2026), and the September–October monologues on Anthropic’s losses, the data-center overbuild, and Anthropic’s compute burden.

Interviews that restate the same case without newsletter paywalls: Business Insider, 16 July 2026; Investor’s Business Daily, 23 September 2026; CEPR Mostly Economics episode 55, 8 October 2026; the George Noble / Gary Marcus panel, 5 October 2026.

---

Recent Findings Supplement (October 2026)

Ed Zitron's core thesis, reiterated and updated in October 2026 content, holds that the AI industry is not a broad technological revolution but a concentrated "OpenAI bubble" (with Anthropic as a close parallel). This bubble relies on hype around unsustainable metrics, circular spending by VC-backed startups, massive unprofitable compute commitments, and debt financing for data centers that lacks genuine underlying demand or ROI. Recent developments, particularly around revenue revisions, leaked financials, and debt pressures, reinforce his view of systemic fragility.[1][2]

Zitron draws from his Where's Your Ed At newsletter (e.g., the October 13 "Annualized Enron Rate" post and October 6 "Credit Crunch"), Better Offline podcast monologues and episodes (e.g., October 2026 discussions on Anthropic's burdens and data center overbuild), and interviews/panels (e.g., Mostly Economics on October 8 and a Risk Takers discussion on October 13). He emphasizes audited or leaked numbers over promotional "run rates," highlighting how Big Tech subsidies and press amplification sustain the illusion.[3][4]

Focus remains on post-April 2026 developments, especially October 2026 updates showing metric revisions, financial leaks, and rising financing costs.

Fabricated "Annualized Run Rate" Metrics Inflate the Bubble

Zitron argues that AI revenue hype rests on non-GAAP, inconsistent "annualized run rates" (ARR) that annualize short, often non-recurring consumption periods using arbitrary methods (e.g., trailing 28 days × 13 or a single day's revenue × 365). These obscure actual economics and enable narrative control, as seen in the early October 2026 market reaction when OpenAI's reported ~$70B ARR was revised downward by $20B to ~$50B due to gross-vs-net accounting differences with partners like Microsoft.[1][5]

• OpenAI and Anthropic use differing methods (OpenAI nets partner cuts; Anthropic grosses some), leading to incomparable figures that journalists and investors treat as fact; Anthropic's July 2026 $65B ARR claim lacked clear definition or period.[5]
• Leaked Anthropic prospectus details (reported September–October 2026) show 2025 audited revenue of ~$4.6B against operating losses exceeding $8B (or higher net losses including non-cash items), with nearly half of sales routed through Google and Amazon.[4][4]
• Broader industry examples include AI startups reporting inflated ARRs while burning VC cash on tokens; actual quarterly revenues remain opaque.[5]

This means competitors or entrants cannot rely on headline growth figures for valuation or strategy; demand rigorous, audited quarterly data and model scenarios where promotional metrics collapse, exposing true burn rates.

OpenAI (and Anthropic) as the Single Point of Failure

Zitron's longstanding claim—that the AI ecosystem is an "OpenAI bubble" propped up by one company's mythology and commitments—gained fresh emphasis in 2026 coverage, with OpenAI positioned as the potential "Lehman Brothers of AI." Its scale justifies hyperscaler capex, chip demand, and startup activity; a slowdown or failure cascades.[6][2]

• OpenAI and Anthropic together represent the bulk of AI compute demand; ~80% of their enterprise revenue comes from the top 1% of customers (mostly VC-funded AI startups).[2]
• Compute commitments remain enormous (hundreds of billions to $1.3T range across the pair in various reports), far outstripping plausible revenue trajectories without continuous external support.[7]
• October 2026 panels and newsletters stress that hyperscaler AI revenues (e.g., portions of Microsoft Azure) are heavily tied to these labs, proving limited organic demand.[4]

For new entrants, this implies building independent demand or diversified revenue is essential; betting on ecosystem subsidies risks total exposure if the core labs face scrutiny or capital constraints.

Debt, Private Credit, and Unplugged Infrastructure Create a Brewing Crisis

Recent October 2026 commentary highlights how AI data center financing—via private credit, junk bonds, and complex deals—is becoming costlier amid rising Treasury yields (10-year notes above 5.3%) and selective lenders. Zitron calls this a "brewing crisis," with CoreWeave alone reporting $35.6B in debt by mid-2026 and estimates of $200–350B in unpowered or warehoused GPUs.[8][7]

• "Credit Crunch" newsletter (October 6) details how negative EBITDA at OpenAI/Anthropic makes investment-grade ratings unlikely, forcing junk-rated issuance of $50B+ annually—structurally difficult at scale.[9]
• Amazon and other hyperscalers use creative structures (e.g., chip sale-leasebacks); private credit fills gaps but signals distress as traditional markets balk.[8]
• GPU rental prices and utilization data are questioned, with real-world deals showing lower effective rates than promotional figures.[5]

Competitors must stress-test capital structures against higher-for-longer rates and potential writedowns; over-reliance on debt-fueled capacity risks insolvency if utilization or pricing fails to materialize.

Circular Spending and Lack of Measurable ROI Sustain Artificial Growth

Zitron details how VC dollars flow to AI startups that burn multiples of revenue on OpenAI/Anthropic tokens, creating a feedback loop that inflates lab revenues while masking weak end-user economics. Token-based billing shifts have prompted enterprise pushback on costs vs. returns.[5][10]

• AI startups often spend 3–10× revenue on tokens, enabling "free" or subsidized user consumption that wouldn't exist otherwise.[5]
• No clear productivity or ROI evidence emerges despite trillions in projected spend; enterprises report costs rising faster than benefits.[8]
• October 2026 content ties this to broader "rot economy" dynamics, where Big Tech uses AI narratives to justify capex amid slowing core growth.[2]

Entrants should prioritize verifiable use cases with clear payback periods rather than hype-driven adoption; subsidized token economics are not scalable without perpetual capital inflows.

Investor and Systemic Risks Extend Beyond Tech

Zitron warns that pension funds, retail investors (via 401(k)s and indices), and counterparties will bear losses from a unwind, comparable to past crises but amplified by opacity. Recent podcasts emphasize Wall Street bets versus Main Street consequences.[3]

• Concentration in NVIDIA (heavy client dependence) and data center debt creates contagion risks across finance.[2]
• Memory/RAM price spikes from AI deals have already affected consumers, illustrating spillover.[11]

Those entering or competing in adjacent spaces (infrastructure, applications, or finance) should model tail risks of correlated failures and diversify away from pure AI capex exposure.

These arguments draw primarily from Zitron's October 2026 output, with links including https://www.wheresyoured.at/annualized-enron-rate/ (Oct 13), https://www.wheresyoured.at/credit-crunch/ (Oct 6), and related podcast episodes. Earlier 2026 pieces (e.g., "The OpenAI Bubble" in July) provide foundational framing but are supplemented by fresh data on revisions and debt. Actual paywalled text is paraphrased only.

Ed Zitron’s October 2026 debt writing is two newsletters, not one. The load-bearing piece is the free Oct. 6 essay Credit Crunch. The junk-market sequel, Premium: The Hater’s Guide To Junk (Oct. 9), is paywalled after a short public preview, and that preview is where almost every checkable junk-market figure lives.

• Free, full text retrieved: Credit Crunch, Oct. 6, 2026. Zitron labeled it a free newsletter on X. [1]
• Paywalled after the intro: Premium: The Hater’s Guide To Junk, Oct. 9, 2026. The page cuts at “Read the full story / Sign up.” [2] [3]

The mechanism he is arguing is not “AI companies borrowed a lot.” It is that the base rate and the credit spread are rising at the same time, and the borrowers who actually need the money — neoclouds, project SPVs, and eventually OpenAI and Anthropic — do not have the cash flow or hard assets that investment-grade markets require. Hyperscaler co-signs (Meta on a CoreWeave facility, take-or-pay leases, customer-payment SPVs) are the only way those structures clear at tolerable rates. Without that co-sign, he says the buildout prices at 9–14%.

What Credit Crunch actually claims

The free essay has three stacked claims.

First, the 10-year Treasury is the floor under every AI loan, and that floor had moved above 5.3%. He uses 5.33% as the live CNBC quote and says a 6 bp move already matters; a ~120 bp move versus the ~4.13% 10-year at the time of an earlier Oracle deal is, on his math, more than $7 billion of extra interest on an $18 billion Oracle issue, before any spread widening. [4] [4]

Second, even “investment grade” AI paper is being priced like junk, and true junk is already painful. He says Oracle debt was trading near the high-yield index’s ~8.2% effective yield, with Oracle spreads having moved from 105–165 bp over Treasuries at issue to 171–282 bp. CoreWeave, by contrast, was at 672–882 bp, with a four-year yield of 11.53% and a six-year yield of 13.49%. The workaround is a hyperscaler backstop: CoreWeave’s March 2026 $8.5 billion delayed-draw term loan, rated investment grade because the counterparty was Meta, priced at SOFR + 2.25%. He then cites UBS for the claim that CoreWeave still needs $102 billion of debt through 2030, which cannot all be Meta-wrapped. [4]

Third, OpenAI and Anthropic cannot become repeat IG issuers on their own numbers. He puts their post-IPO need at $50–100 billion a year, argues a standalone rating would be B to CCC+, and says CLO eligibility and high-yield capacity collapse as the rating falls. The collateral problem is structural: they do not own the data centers or the chips, training is an operating expense rather than capex, and take-or-pay contracts are debt equivalents. His worked example is OpenAI’s 2025 audited figures as he previously reported them: $34 billion of expenses, $13.07 billion of losses, EBITDA of −$20.92 billion, margin −160%. [4] [4]

Lender selectivity is sourced to The Information: Société Générale, SMBC, and MUFG “becoming more selective”; CleanSpark had to give concessions on a Meta-tied deal; Blue Owl agreed to up to $10 billion with Primary Digital Infrastructure in a 15-minute meeting. “Virgin” projects with no hyperscaler co-sign, he says, face 9–14%. He also flags SB Energy’s need to raise $174 billion even with an Nvidia backstop. [4]

What the junk piece adds — and what is behind the wall

The public preview of The Hater’s Guide To Junk is mostly a frame, plus one important correction. Zitron retracted a Bloomberg high-yield series because he had excluded matured bonds and overstated growth since 2011. The replacement figures, which he attributes to SIFMA, are $2.19 trillion of total US corporate issuance through September 2026, $303.4 billion of that in high-yield, versus $222.1 billion of high-yield issuance in 2011, or $329.6 billion inflation-adjusted. [2]

The preview then restates the Credit Crunch thesis in junk-market language: Anthropic and OpenAI are CCC / Moody’s Caa candidates; they would need $50–100 billion a year across high-yield, private credit, leveraged loans, and revolvers; CoreWeave alone needs $102 billion through 2030 (UBS, via the free essay); hyperscalers alone are expected to issue about $400 billion in 2027. He uses Paramount Skydance’s Warner deal as the crowding-out exhibit: a $52 billion bond-and-loan sale, $12.4 billion of it junk, with pricing as high as 9.1% on the 10-year. Oracle, he says, sits on the last rung of investment grade; a further downgrade would force IG and pension mandates to dump the paper, which matters because he puts Larry Ellison’s stock-collateralized personal loans at $30 billion or more, including a further $9.2 billion pledge. The S&P language he quotes is that OpenAI is roughly half of $638 billion of Oracle remaining performance obligations. [2]

Anything after “Welcome to the Hater’s Guide To Junk” — CLO capacity tables, named lender books, default math — was not readable without a subscription. Secondary write-ups found in this search only repeat the public preview. Treat the paywalled half as unread.

Claim, figure, his source, verification

Claim	Figure	Source he cites	Independent status
10-year Treasury has broken above 5.3% and is the floor for all AI debt	5.3%+; example quote 5.33%	CNBC US10Y page, as of the essay	Verified in level, not the exact tick. FRED DGS10 was 5.27% on Oct. 6 and 5.31% on Oct. 5; intraday highs that week printed about 5.33–5.35%. [5] [6] [7]
Oracle’s earlier $18B deal would cost >$7B more interest if repriced today	>$7 billion extra interest; 10-year was 4.13% at issue; spreads then 105–165 bp, now 171–282 bp	His prior Dead Money essay; Reuters/FT on distressed Oracle project debt	Partially mismatched. The Reuters/FT story he links is $18 billion of Project Jupiter construction loans, quoted by Santander and Jefferies at 89–91 cents, not a corporate bond sale and not 84 cents. A separate Zitron tweet he cites says 84 cents. The 4.13% issue-date yield and the $7B interest delta were not re-derived here. [8] [9]
High-yield index average yield ~8.2%, and Oracle “trades like junk”	8.2%	FRED series BAMLH0A0HYM2EY	Verified as a market level. ICE BofA US High Yield effective yield was 8.2% on Oct. 7 (Janus Henderson) and 8.08–8.13% on FRED over Oct. 7–13. That is an index yield, not a print on a named Oracle bond. [10] [11]
CoreWeave unsecured/conventional debt is deeply wide	Spreads 672–882 bp; 4-year yield 11.53%; 6-year yield 13.49%	Unstated terminal/market prices in the essay	Directionally confirmed, exact prints not independently matched. By Oct. 14, CNBC and Bloomberg had CoreWeave 2031/2032 notes near 15% and below 80 cents. His early-October snapshot is consistent with a selloff that then got worse; the 11.53/13.49 figures were not found in a primary TRACE or Bloomberg reprint. [12] [13]
Hyperscaler co-sign is the only cheap structure	$8.5B DDTL, IG, SOFR + 2.25%, Meta as the contracted customer	CoreWeave investor release, March 31, 2026	Verified. Company release: $8.5B delayed-draw, Moody’s A3 / DBRS A (low), floating tranche SOFR + 2.25%, fixed tranche ~5.9%, borrower is CoreWeave Compute Acquisition Co. VIII, LLC — an SPV, not the parent. Bloomberg had already reported the loan as backed by Meta contracts of up to $14.2B plus a further >$5B agreement. Reuters: anchored by Blackstone Credit & Insurance. [14] [15] [16]
CoreWeave must keep borrowing on its own credit	$102B of debt through 2030	“Analysts at UBS”	Not independently verified. No UBS note was retrieved. The figure appears only as repeated from Zitron.
Big project lenders are turning selective	SocGen, SMBC, MUFG “more selective”; CleanSpark concessions; Blue Owl $10B decision in 15 minutes	The Information	Partly verified via secondary reprints of that reporting, not the paywalled original. A KuCoin/The Information summary matches the bank names and says CleanSpark’s Meta-tied deal priced at 98.5 cents on ~$2.3B, coupon 7.875%. CleanSpark later disclosed closing $2.276B of 7.875% senior secured notes due 2031. Bloomberg had initial price talk near 8.5%, about 2 points over the BB average. The 15-minute Blue Owl anecdote was not independently confirmed. [17] [18] [19]
SMBC in 7 data-center deals, MUFG in 17; both in SoftBank’s $15B 2025 bridge	7 / 17 / $15B	His own end-2025 tally; archive.is for the SoftBank loan	Not re-verified. This is his prior count, not a regulatory exposure table.
Virgin projects price at 9–14%	9–14%	His inference from Treasury + spread regime, referring back to the September premium debt series	His estimate, not a new-issue print. Consistent with where CoreWeave and CCC paper later traded, but not a primary clearing yield for a named “virgin” SPV.
SB Energy still has to raise enormous project debt even with Nvidia	$174B	WSJ, “The IPO That Asks You to Buy AI Hype Today…”	Verified in substance against the S-1, with a wording caveat. SB Energy’s Sept. 1, 2026 S-1 reports ~$178B of backlog-associated capex, of which ~$174B is the data-center segment. That is capex still to fund, not a signed $174B debt commitment. Nvidia’s disclosed guarantee on the Ohio campus was up to $105B, not a full backstop of the $174B. [20] [21]
Hyperscalers will issue ~$400B of bonds in 2027	$400B	Yahoo Finance write-up of Goldman Sachs	Verified as Goldman’s earlier forecast; later revised up. Reuters, July 29, 2026: Goldman expected ~$250B in 2026 and $400B in 2027 from the five hyperscalers. Subsequent Goldman notes raised the 2027 figure to $420B. [22] [23]
Anthropic non-cancelable compute book	$413B over 7–10 years; separately $161.2B of Broadcom TPU leases	YouTube clip for the $413B; Reuters for the $161.2B and the $518B total	$518B and ~80% non-cancelable are Reuters, from a confidential prospectus Reuters reviewed — not a public filing. Reuters: at least $111.1B Google + $110B Amazon + $31.4B Microsoft, “regardless of usage,” plus ~$161.2B largely non-cancelable Broadcom equipment leases. Those four lines sum to ~$413.7B, which is also ~80% of $518B. Zitron’s “$413 billion” is a reconstruction of that subset, not a number Reuters prints as “$413 billion.” [24] [25]
OpenAI 2025 EBITDA	−$20.92B; margin −160%; $34B expenses; $13.07B losses	His own prior “audited 2025 financials” essay	Not independently verified. No public OpenAI audit was retrieved. Later in the same essay he also calls $20.92B the 2025 loss, which does not match the earlier “losses were $13.07B, EBITDA was −$20.92B” sentence. Treat the EBITDA math as his calculation, and note the internal slip.
Anthropic 2025 operating loss	$8B (he rounds)	Reuters prospectus story	Verified as reported by Reuters from the confidential prospectus: operating loss widened to $8.06B in 2025; revenue ~$4.6B; net loss $42B, mostly accounting charges. Not a public SEC filing. [25]
Anthropic raised >$95B in 2026 and may seek a $60–100B IPO	>$95B; IPO talk $60–100B; ~5% float; maybe ~$20B/year of ATM capacity	Yahoo; his float estimate	$95B is a round-sum of reported 2026 raises, not a filing total confirmed here. The 5% float and $20B ATM cap are explicitly his guess.
Standalone lab rating would be junk	B to CCC+ / CCC / Moody’s Caa	His application of S&P/Fitch/Moody’s public rating definitions; S&P concentration criteria	Judgment, not a rating. No agency has published an OpenAI or Anthropic issuer rating in the sources retrieved. The rating-definition pages he links are real; the mapping onto these issuers is his.
OpenAI compute spend through 2030	at least $750B	WSJ	Verified as a WSJ-reported company projection, originally from July 2026, and still being cited by WSJ in October. Separate FT reporting in September put compute-and-infrastructure spending at about $856B by end-2030 and cash burn at $278B. Those are different concepts; he cites the WSJ $750B figure. [26] [27] [28]
Combined lab obligations the hyperscalers are counting on	$1.3T	His aggregation	His sum. $518B (Anthropic, Reuters) + $750B (OpenAI, WSJ) = $1.27T, so the round number is in range of those two reports, but it is not a single primary total.
Analysts expect the two labs to deliver $444B of hyperscaler revenue in three years	$444B	UBS, Barclays, Wells Fargo, as summarized in his earlier Concentration Risk piece	Not independently verified.
Non-lab compute demand is tiny	at best ~$22B	His premium essay	Not independently verified.
Uninstalled Nvidia GPUs	>$200B in warehouses; Morgan Stanley, via Bryce Elder at the FT, says more than half of 2026–28 GPU sales have nowhere to plug in	FT	FT link is real; the Morgan Stanley number was not re-opened past the paywall. Status: cited, not re-read.
Broadcom raised $60B to fund Anthropic chip leases	$60B	Yahoo Finance	Not re-verified from a Broadcom filing in this pass. Reuters does confirm the $161.2B lease obligation on Anthropic’s side.
Ellison stock-backed loans	another $9.2B pledged; his count ~$30B total; margin-call zone “somewhere around $60” a share	Yahoo Finance	Not re-verified. The $9.2B and $30B are his reading of Yahoo; the $60 share level is his inference.
Oracle rating	last rung of investment grade; S&P says OpenAI is ~half of $638B RPO	S&P Global	The rating is verified. Reuters and the FT both report S&P cut Oracle to BBB− in July 2026, one notch above junk. The $638B RPO / “half is OpenAI” sentence is quoted by Zitron from S&P; the S&P article itself was not re-opened here. [8] [9]
Junk-market crowding: Paramount/Warner	$52B bond/loan sale; $12.4B junk; as much as 9.1% on the 10-year	Bloomberg, Oct. 2, 2026	Size largely verified; the 9.1% and $12.4B were not. Wolf Street, citing the same financing, puts it at $52B of debt sales, $44B of bonds, $12B junk. [29]
US corporate issuance, SIFMA correction in the junk preview	$2.19T total through September; $303.4B high-yield; 2011 HY $222.1B, or $329.6B inflation-adjusted	SIFMA, after he retracted a Bloomberg series	Total issuance verified; HY split not confirmed from the public SIFMA page. SIFMA’s Oct. 14, 2026 update: issuance through September $2,195.1 billion, +24.5% year over year. The $303.4B high-yield breakout is in his correction and was not on the summary page retrieved. [30]
CLOs buy 60–75% of corporate loans	60–75%	UCLA Anderson Review	Not re-opened. Standard market description; not checked against a current CLO trustee report.
Railroad comparison	~$250B inflation-adjusted total 19th-century railroad investment vs $400B of hyperscaler issuance in 2027 alone	His “rough calculations”; Yahoo/1873 railroad piece	His inflation adjustment, not a primary historical series.

How to read the gaps

Three patterns matter more than any single unverified basis-point print.

He is careful about market levels and sloppy about instrument identity. The 10-year above 5.3%, the high-yield index near 8.2%, the CoreWeave IG SPV at SOFR+225, the S&P BBB− on Oracle, the Goldman $400B (later $420B), and the Reuters prospectus math on Anthropic all check out. The “$18 billion at 84 cents” line does not match the FT story he links, which is 89–91 cents on project loans. That is the clearest miss.

The largest forward numbers are either bank forecasts or his sums. $400B, $102B, $444B, $1.3T, and the B/CCC rating are not filings. The $174B is real S-1 capex, misread if treated as debt already raised. The $413B is a fair reconstruction of Reuters’ non-cancelable subset, not a figure in the prospectus as Reuters printed it.

The junk essay cannot be audited past the preview. If the claim is about CLO capacity, named bookrunners, or default probabilities inside The Hater’s Guide To Junk, it is behind the paywall and was not in any independent reprint found here.

Adjacent October commentary, not these two essays: a Benzinga write-up of a Monetary Matters interview (Oct. 2) has Zitron calling private credit a “brewing crisis,” citing CoreWeave debt of $35.6 billion as of June 30 and a ~13% conventional-debt cost that Benzinga flags as his calculation. Goldman’s $88 billion of lower-rated AI borrowing year-to-date, which appears in that interview coverage, does match a separate Goldman leveraged-finance tally reported elsewhere ($88B in 2026 vs about $20B the year before). That interview is not Credit Crunch and not the junk guide. [31] [32]

---

Recent Findings Supplement (October 2026)

Ed Zitron's primary October 2026 pieces are the free newsletter "Credit Crunch" (published October 6, 2026, at https://www.wheresyoured.at/credit-crunch/) and the premium follow-up "The Hater's Guide To Junk" (published October 9, 2026, at https://www.wheresyoured.at/premium-the-haters-guide-to-junk/).[1][2]

"Credit Crunch" focuses on rising 10-year Treasury yields (above 5.3%), widening spreads, lender selectivity, and the challenges of funding AI data centers and AI labs (especially Anthropic and OpenAI) with junk or high-yield debt amid negative EBITDA, lack of collateral, and circular demand. It references prior premium debt analyses and teases junk-debt specifics. "The Hater's Guide To Junk" expands on high-yield/speculative-grade debt markets, leveraged loans for data centers and software, and why AI labs will rely on them post-IPO due to poor fundamentals.[1][2]

Both are on his Substack ("Where's Your Ed At"). The free "Credit Crunch" post is largely accessible (with some premium archive links paywalled); the junk guide is fully paywalled for non-subscribers. Related commentary appeared in interviews/podcasts (e.g., Tech Report on October 9–10 discussing the "Credit Crunch" article) and X posts promoting them.[3][4]

Zitron traces many figures to outlets like The Information, Reuters, FT, Goldman Sachs, SIFMA, FRED, WSJ, Bloomberg, company filings, and UBS. Independent sources (e.g., Benzinga/Reuters reports from early October 2026) corroborate several macro figures on issuance volumes, yields, and selectivity, though lab-specific projections (e.g., exact annual debt needs or ratings) are Zitron's estimates or inferences.[5][6]

Table of key claims/figures from the pieces (focus on post-4/15/2026 developments; only new or updated data included):

• Claim/Figure: Companies (e.g., Anthropic/OpenAI) need ≥$50B+ in debt annually, likely at junk (B to CCC+) ratings due to negative EBITDA and lack of assets/collateral.
  
  Zitron's Source: His analysis; references Anthropic's $95B+ VC in 2026 and non-cancelable contracts (~$413B for Anthropic); FT/Reuters on ratings dynamics.
  
  Independent Verification: Partially supported; Goldman Sachs cited ~$88B lower-rated AI borrowing YTD 2026 (up sharply from ~$20B junk/leveraged in first 11 months of 2025 per Neuberger Berman/Reuters). Morgan Stanley: $23B AI-related debt in September 2026 (down ~50% MoM). No direct independent confirmation of exact $50B–$100B lab needs or precise ratings.[5][6]

• Claim/Figure: 10-year Treasuries at 5.3%+ (selling off), raising baseline borrowing costs for all AI debt; hyperscalers expected to issue $400B bonds in 2027.
  
  Zitron's Source: CNBC/FRED-style benchmarks; Goldman Sachs via Yahoo Finance.
  
  Independent Verification: Consistent with contemporaneous reporting on Treasury yields and Goldman forecasts for Big Tech bond issuance to fund AI capex.[1]

• Claim/Figure: CoreWeave: spreads 672–882 bps; yields ~11.53% (4-year) to 13.49% (6-year); $35.6B debt as of June 30, 2026; $8.5B Blackstone-anchored facility; needs ~$102B debt through 2030 (UBS).
  
  Zitron's Source: Market data/FRED high-yield index (~8.2% avg); company reports/filings; UBS.
  
  Independent Verification: Matches Benzinga/Reuters coverage (Oct 2026) citing CoreWeave's debt load and private credit reliance; later X/market commentary noted bonds trading at higher yields (e.g., 14.5% on some 2032 paper).[5][7]

• Claim/Figure: Oracle $18B data center debt trading distressed (e.g., at 84 cents on the dollar in some reports); would cost billions more in interest today due to rate/sentiment shifts.
  
  Zitron's Source: Reuters/FT reporting on Oracle bonds.
  
  Independent Verification: Directly corroborated in Oct 2026 coverage (e.g., FT/Reuters on Oracle data center debt under pressure).[8]

• Claim/Figure: Lenders (SMBC, MUFG, SocGen) becoming "more selective" on data center projects; "virgin" (non-hyperscaler-backed) projects facing 9–14% rates.
  
  Zitron's Source: The Information (cracks in debt boom; CleanSpark concessions example).
  
  Independent Verification: The Information reporting cited across coverage; aligns with broader Reuters/Bloomberg notes on selectivity and higher costs for lower-rated issuers (BB+ at 9–10%, lower at up to 14–15%).[8][6]

• Claim/Figure: High-yield issuance: ~$303.4B through Sept 2026 (SIFMA), up from prior years; AI/data centers a major driver of recent supply.
  
  Zitron's Source: SIFMA data.
  
  Independent Verification: Consistent with Reuters/BNP Paribas data in Oct 2026 reports (~$40B AI infrastructure high-yield YTD 2026 vs. $12B all of 2025).[2][6]

• Claim/Figure: Anthropic/OpenAI negative EBITDA (example: OpenAI ~negative 160% margin on $13B+ losses); precludes investment-grade ratings without hyperscaler/government backstops; reliance on take-or-pay contracts as debt equivalents.
  
  Zitron's Source: His prior analyses of audited financials; Reuters/FT on contracts and "adjusted" metrics.
  
  Independent Verification: Aligns with reporting on lab losses, contract scale (e.g., Anthropic's hundreds of billions in commitments), and skepticism around run-rate metrics; no contradiction in sources.[9]

• Claim/Figure: Examples of structures: Amazon considering $8B Nvidia chip sale-leaseback SPV; CoreWeave GPU-backed facilities rated investment-grade via hyperscaler counterparties (e.g., Meta); SB Energy ~$174B needed for OpenAI-linked project.
  
  Zitron's Source: Company/news reports (e.g., WSJ on SB Energy).
  
  Independent Verification: Amazon/CoreWeave structures noted in early Oct 2026 Benzinga coverage; broader SPV/off-balance-sheet themes in prior Nikkei-style reports.[5]

Additional context and verification notes: Zitron emphasizes mechanisms like Treasury-driven base rate increases compounding company-specific sentiment deterioration (e.g., power delays, overbuild risks), lender herd behavior, and circular demand (labs as primary offtakers). Independent Oct 2026 sources (Reuters, The Information, Goldman/Neuberger via media) confirm the surge-then-slowdown in AI-related high-yield/leveraged issuance, rising costs for lower-rated borrowers, and selectivity by banks like SMBC/MUFG. Lab-specific debt projections and ratings outcomes remain forward-looking estimates without direct primary filings yet. No major contradictions found in available reporting; some figures (e.g., exact spreads/yields) are market snapshots subject to rapid change.[6]

For full details, the free post provides substantial excerpts; premium content requires subscription. Related X activity from Zitron (@edzitron) links and discusses these pieces.

Ed Zitron’s failure case is not a single “the models stop working” story. It is a cash-flow and contract story: two loss-making labs have signed take-or-pay compute bills that only close if they keep raising extraordinary amounts of equity (and, after any IPO, junk-rated debt), while the cloud, chip, neocloud, and private-credit complex has already booked those bills as future growth. The mechanism is circular. Hyperscalers fund the labs, the labs spend the money back on the hyperscalers’ clouds, and the clouds report that spend as AI revenue. If the next raise fails, the circle stops paying.

1. The failure paths he actually describes

Zitron’s August 18, 2026 essay, What Happens If OpenAI Dies?, treats collapse as a funding event, not a product event. He argues OpenAI has only two real outcomes: it becomes the largest company in history, or it runs out of money. The path to the second is a sequence, not a date. [1]

Funding gap, then a failed raise. OpenAI closed (or was still closing) a $122 billion round at an $852 billion post-money valuation in March 2026, of which Zitron says it had received at least $50 billion, including SoftBank tranches, with a further SoftBank tranche due October 1, 2026. His point is that one round does not retire the problem. More usage means more inference, so the company needs another mega-round every year, from a thinner pool of buyers. He cites SoftBank borrowing against its OpenAI stake and Jensen Huang suggesting Nvidia’s investment that year was likely its last. In Credit Crunch (October 6, 2026) he extends the same logic to both labs: outside a hyperscaler or government guarantee, deeply negative EBITDA would leave them rated around B to CCC+, which he argues cannot support the $50–100 billion a year of debt they would need after going public. [2] [3]

Compute commitments that do not shrink when demand does. These are largely take-or-pay or non-cancellable. Zitron’s July 2026 accounting puts OpenAI behind about $748 billion of remaining performance obligations at Microsoft, Amazon, and Oracle, plus at least another $70 billion at CoreWeave, Cerebras, Nebius, Lambda, IREN and similar names, with a plan to spend $50 billion or more on compute in 2026 alone and to burn at least $852 billion through the end of 2030. The Wall Street Journal separately reported in July 2026 that OpenAI had raised its own compute-spending plan through 2030 to about $750 billion, up from roughly $600 billion earlier in the year (Reuters had reported the $600 billion figure in February). [2] [4] [5]

Revenue that does not cover burn, and inference that scales with success. On the audited 2025 statements Zitron published and the Financial Times confirmed, OpenAI booked $13.07 billion of revenue against $34 billion of costs and a $20.92 billion operating loss. Net loss attributable to the company was about $38.5 billion, swollen by a one-time fair-value charge on the nonprofit-to-for-profit conversion (roughly $30 billion on the FT’s account; other write-ups of the same documents put the charge at $41.55 billion). Strip non-cash items and a person familiar told the FT the underlying loss was closer to $8 billion. That is still a business that spent far more than it earned, and Zitron’s mechanism is that the gap does not close by growing: cost of revenue (serving the models) rose from $2.65 billion in 2024 to $7.5 billion in 2025, and R&D hit $19.18 billion, of which $10.59 billion was paid to Microsoft. The Information, as cited by Zitron, put Q1 2026 at $5.7 billion of revenue against $12.1 billion of compute and training cost. More customers, on this model, means a larger cash hole unless prices rise enough to offset tokens — and he argues both labs are cutting prices to hold share. [6] [7] [1]

A failed or delayed IPO as the liquidity event that never arrives. OpenAI confidentially filed and, per the New York Times as cited by Zitron, was leaning toward a 2027 listing after advisers questioned a $1 trillion valuation. Anthropic’s draft prospectus, reviewed by Reuters in late September 2026, is the document Zitron treats as the industry’s receipts: about $4.6 billion of 2025 revenue, an operating loss above $8 billion, a $42 billion net loss (about $34 billion of it a non-cash financing remeasurement), and at least $518 billion of cloud, compute, and infrastructure commitments over roughly a decade, about 80 percent non-cancellable or payable regardless of use. His normative claim — that neither company should be allowed to list — is an opinion. His operational claim is narrower: public markets and junk-debt markets will not fund the commitments at the prices the contracts assume, and an Anthropic listing that puts the losses on the page first makes OpenAI’s own float harder. [8] [9] [1]

He also argues the revenue base is brittle underneath. In October 2026 he wrote that about 80 percent of both labs’ revenue comes from 1 percent of customers, mostly venture-backed AI startups that burn several times their own subscription revenue in tokens and therefore need perpetual VC infusions to keep paying API bills. If that layer dies first, the labs lose the customers that were supposed to grow into the take-or-pay contracts. [10]

Three end-states if the raise fails. In the August essay he names three, all destructive, none clean:

1. Microsoft absorbs OpenAI. Microsoft cancels the contracts OpenAI signed with everyone else, restates away the roughly $250 billion of incremental Azure spend it had talked up, and accepts GPU impairments rather than keep paying Google, Oracle, and Amazon. Free ChatGPT is stripped back or killed; Plus and Pro move toward metered token billing.
2. OpenAI is allowed to die, with Altman as the scapegoat. It short-changes something on the order of a trillion dollars of compute contracts. CoreWeave, Cerebras, and other offtake-dependent names lose the cash flow that services their debt. Zitron’s point is narrative as much as legal: markets will treat OpenAI as the Lehman of the AI trade even if Anthropic’s contracts are just as large relative to its cash.
3. A distressed merger with Anthropic. He treats this as a bag-pump that lets founders exit blame, not a fix. Combined non-cancellable compute still has to be paid, and the two companies are already each other’s price competitors. [1] [2]

He is explicit that he does not have a calendar date. The trigger he watches is a down round, a delayed raise, or a renegotiation of take-or-pay terms — and he expects the break to come only after data-center debt and venture capital have already been largely drawn. [11]

2. What reputable reporting actually shows on revenue, losses, and commitments

These figures are not interchangeable. Run rate is usually a recent month or four weeks times 12. Booked revenue is what the income statement records. Gross marketplace revenue (Anthropic’s method) counts the full cloud-partner sale; net (OpenAI’s method) counts only the lab’s share. Several headline gaps in October 2026 were accounting, not a sudden collapse in usage.

OpenAI, audited 2025 (Zitron documents, confirmed by the Financial Times):

• Revenue: $13.07 billion, up from $3.7 billion in 2024. By year-end it was generating nearly $2 billion a month.
• Costs and expenses: $34 billion. R&D $19.18 billion. Sales and marketing about $5.7–6 billion. Cost of revenue $7.5 billion.
• Operating loss: $20.92 billion (160 percent of revenue, versus 237 percent in 2024).
• Net loss attributable to the company: about $38.5 billion, dominated by the conversion-related fair-value charge. Underlying loss nearer $8 billion if that charge and other non-cash items are excluded, per a person familiar cited by the FT.
• Payments to Microsoft in 2025: $17.2 billion. Microsoft paid OpenAI $303 million. SoftBank paid OpenAI $867 million. Year-end assets a little over $50 billion, about half cash. [6] [7] [12]

OpenAI, 2026 run rate (not audited full-year revenue): The Financial Times reported on October 8, 2026 that OpenAI told investors annualized revenue was approaching $50 billion at the end of September, about $20 billion below a ~$70 billion figure that had circulated the week before. The gap was largely investors trying to gross OpenAI up to Anthropic’s partner-sales method. Bloomberg reported the company still expected to reach $70 billion or more in annualized revenue by year-end 2026. Recognized 2026 revenue is smaller than either run rate; one account of investor materials put booked 2026 revenue around $36 billion, with cumulative negative free cash flow of roughly $280 billion from 2026 to 2030 if the company’s own plan holds. Q2 2026 quarterly revenue was reported at $6.7 billion, behind Anthropic. [13] [14] [15]

Anthropic, draft IPO prospectus seen by Reuters (late September 2026):

• 2025 revenue: nearly $4.6 billion ($4.59 billion in some write-ups), up from about $386–400 million in 2024, roughly twelvefold.
• Operating loss: $8.06 billion, up from $2.98 billion. Compute and infrastructure spend: $7.33 billion, more than half of $12.65 billion in operating expense.
• Net loss: about $42 billion, of which roughly $34 billion was a non-cash charge on financing that could convert into shares.
• Cash and short-term investments at December 31, 2025: $20.28 billion.
• Two unnamed customers each about 12 percent of 2025 revenue. Forty-seven percent of sales, about $2.16 billion, ran through Amazon and Google cloud marketplaces; Reuters calculated the platforms kept roughly 16 cents on those dollars.
• Commitments: at least $518 billion over the coming years with six partners. Named pieces reported from the filing: Google $111.1 billion, Amazon $110 billion, Microsoft $31.4 billion, Broadcom-related equipment leases about $161.2 billion. About 80 percent non-cancellable or payable regardless of usage. The filing, per Reuters: “If our actual spend falls short, we must pay Google the difference.” Zitron’s later shorthand of $413 billion non-cancellable is his reading of that 80 percent, not a separate Reuters line. [8] [16] [17]

Anthropic, 2026 operating updates (company-to-investor, not the 2025 audit): Bloomberg reported a run rate past $65 billion by the end of July 2026, and investors were told Q2 revenue topped $11.5 billion. Reuters sources have also cited a path toward roughly $100 billion annualized by year-end, and a 2028 forecast of $190–200 billion that bankers were using to justify a listing above $2 trillion. Zitron’s response is that a higher run rate does not retire the cash-flow problem if every extra dollar of usage requires more committed compute, and that “annualized revenue” can be a selected window rather than a contract. He had said in June that Anthropic would not sustain a then-reported ~$47 billion run rate; the later print was faster than that call. [18] [19]

Commitment stack, as reported rather than as Zitron sums it:

Counterparty	Reported OpenAI exposure	Reported Anthropic exposure
Oracle	~$300 billion over about five years, starting 2027, tied to multi-gigawatt capacity (WSJ)	Not the core of the $518 billion stack
Microsoft	$250 billion Azure commitment from the October 2025 recapitalization; Microsoft booked $24.1 billion of FY2026 revenue from OpenAI	$31.4 billion over 7–10 years
Amazon	Expanded cloud deal reported at $138 billion over eight years, including Trainium capacity; an earlier $38 billion piece was also reported	$110 billion
Google	Smaller direct role than for Anthropic	$111.1 billion, shortfall payable
Broadcom	Custom “Jalapeño” chips; financing terms disputed in trade press	$161.2 billion of largely non-cancellable equipment leases
Neoclouds	CoreWeave expansions reported in the tens of billions (one tally: up to $22.4 billion across 2025 contracts)	CoreWeave and others sit inside the broader stack; not the Reuters headline lines

OpenAI’s own earlier public boast of $1.4 trillion to build 30 gigawatts was a capacity ambition, later walked toward the $600 billion then $750 billion compute-spend plans. Those are not the same number as signed RPOs. [20] [4] [21] [22]

Zitron’s $852 billion “burn through 2030” and his claim that the two labs are 70–89 percent of hyperscaler AI revenue are his syntheses. The 70 percent figure for Microsoft has a public anchor: Microsoft’s own filing showed $24.1 billion from OpenAI in the year ended June 2026, and Bloomberg estimated that was likely around 70 percent of Microsoft’s AI sales if the AI business finished the year near $34 billion. The wider “70 percent of Microsoft, Google, and Amazon AI revenue” claim is Zitron plus sell-side notes he cites (UBS on Google Cloud, Barclays on AWS), not a hyperscaler disclosure. [23] [24]

3. Why the economics do not self-correct

The non-obvious part of his argument is that growth makes the funding gap worse before it makes it better. Inference is a cash cost that arrives with every token. Training and reserved capacity are a contractual cost that arrives whether the tokens show up or not. A lab that wins share by cutting price, or by letting coding-agent users burn thousands of dollars of compute on a $200 subscription, increases both. Zitron’s ratio for 2025 — Anthropic spending $2.75 to make $1, OpenAI $2.60 — is his compression of the audited cost stacks, not a GAAP line. The direction matches the filings: operating losses widened in dollars even as they improved as a percentage of revenue at OpenAI. [25] [6]

The second trap is customer quality. If most enterprise revenue is a thin slice of VC-backed startups reselling subsidized tokens, then the labs’ “demand” is itself a funding round one layer down. When that layer cannot raise, API spend falls, but the Google shortfall clause and the Oracle start-date in 2027 do not. That is why he calls the commitments the subprime mortgages of the buildout: the payment obligation is fixed, the end-borrower is not investment-grade, and the capacity has not all been energized yet, so the pain is still mostly in the future. [26] [10]

A third trap is the IPO itself. Listing converts a private funding treadmill into a public one, but it also publishes the losses. Zitron’s credit argument is that debt markets price EBITDA and hard assets, not “annualized revenue.” Labs with negative EBITDA and few hard assets, he says, cannot roll $50 billion-plus a year at anything but punitive spreads — and spreads were already rising in October 2026 as Treasury yields moved up and AI issuance competed for the same buyers. Whether the 10-year was specifically at 5.3 percent on the day he wrote is his market snapshot; the mechanism (higher base rates plus wider credit spreads on AI paper) is what matters for the thesis. [3]

4. Knock-on map: who gets hit, and how

Microsoft. Least likely to go bankrupt, most exposed to a narrative and guidance break. It disclosed $24.1 billion of fiscal 2026 revenue from OpenAI, about 7 percent of $331.8 billion total revenue, and OpenAI owed it $6.0 billion at June 30. Bloomberg’s estimate put that at roughly 70 percent of AI sales. Commercial remaining performance obligations were $678 billion; the CFO said backlog excluding OpenAI grew only 25 percent, so most of the contracted-growth story is the lab. In Zitron’s absorption scenario, Microsoft keeps the models, kills rival-cloud contracts, and impairs GPUs. In the death scenario, it loses the related-party revenue, marks down a stake it has already had to explain in non-GAAP bridges, and has to tell investors the AI ramp was mostly one customer’s compute bill. It can absorb that. The stock multiple built on “AI cloud growth” cannot, on his telling. [23] [27] [1]

Oracle. The cleanest single-name credit transmission. The WSJ reported a roughly $300 billion, five-year computing contract starting in 2027. S&P, cutting Oracle to BBB- on July 9, 2026 — one notch above junk — estimated OpenAI was about half of a $638 billion RPO book and called OpenAI a key credit risk: if OpenAI cannot pay, Oracle is left with data-center leases it may be unable to exit or re-let on worse terms. By the August 31, 2026 quarter, RPO was $664 billion, quarterly capex was $28.5 billion, and fiscal 2027 capex guidance was $90–95 billion. S&P’s cash-flow forecast was a free-operating-cash-flow deficit near $42 billion in fiscal 2027. The contract does not start paying in size until 2027, so Oracle is financing the build before the customer’s checks arrive. Zitron’s line is that a payment stoppage leaves Oracle unable to service the debt it took on to build for one customer. CDS spreads in the fall of 2026 were already trading that risk; one October account put Oracle CDS near 261 basis points. [20] [28] [29] [30]

Amazon. Dual role: cloud vendor and Anthropic backer. Reuters’ prospectus read puts $110 billion of Anthropic commitments with Amazon, and trade reporting put an expanded OpenAI–AWS deal at $138 billion over eight years. Amazon can fund unusual structures in a way Oracle cannot. The hit Zitron describes is not insolvency. It is AWS growth that was Anthropic’s bill, plus equity in a lab that may renegotiate, plus Trainium capacity built for a customer that can walk toward Nvidia, Google, or its own campuses. Barclays work cited by Zitron put OpenAI plus Anthropic at 13 percent of AWS this year and 18 percent next — material, not Microsoft-like. [16] [4] [24]

Google. Same circular structure, tighter shortfall language. $111.1 billion of Anthropic spend from April 2026 to July 2033, payable even if usage misses. Google is also a distribution channel (it grosses up Anthropic’s revenue and takes a cut) and a model competitor. UBS figures Zitron cites had the two labs at 27 percent of Google Cloud revenue in 2026 and more than 48 percent in 2027. If that estimate is even directionally right, a lab renegotiation is a Cloud segment miss, not a rounding error, and Google still has to decide whether to keep subsidizing a rival’s training runs to protect the backlog. [17] [24]

Nvidia. The demand is one step upstream of the labs. Nvidia does not need OpenAI to be its direct invoice; it needs Microsoft, Oracle, CoreWeave, and Amazon to keep ordering GPUs because the labs will pay for them. In recent filings, two direct customers were 39 percent of one quarter’s revenue, and three customers were about 44 percent of fiscal first-half sales in trade summaries of the 10-Q — concentration in integrators and hyperscalers, not a named “54 percent from three clients” line. (That 54 percent figure appears in secondary summaries of Zitron, not in the Nvidia filing excerpts retrieved here.) The circular piece that is documented: Nvidia has invested in OpenAI, and in August 2026 agreed to guarantee up to $105 billion of lease and power obligations tied to an OpenAI-affiliate campus. Zitron’s extreme case is GPU sales falling as much as 90 percent once debt-funded capex stops. The milder, still serious case is a volume air pocket: capacity already sold into 2026–27, then a stall in the next architecture (Vera Rubin and beyond) because the only buyers who were absorbing “all available” clusters can no longer prepay. He also argues a large stock of GPUs is installed slowly or sitting unpowered — his range, not a company disclosure, is $200–350 billion unplugged — so a demand break hits new orders before it hits depreciation schedules. [31] [32] [33] [3]

CoreWeave-type neoclouds. This is where the offtake is the entire credit. CoreWeave’s model is borrow to buy Nvidia GPUs, pledge the chips and a customer contract, rent the capacity. By June 30, 2026, debt plus lease obligations were about $52 billion; the contracted backlog was about $104 billion, with a large share tied to Microsoft, Meta, and OpenAI. The Financial Times reported the company had raised more than $35 billion in 2026 across chip-backed loans, high-yield bonds, converts, and equity, and was courting life-insurer capital it called effectively unlimited — but only if more of the paper can be rated investment grade. On October 14, 2026, yields on its 2031 and 2032 unsecured notes were about 14.8–14.9 percent, with the 2032s trading under 80 cents and yields touching 15 percent; the stock fell about 16 percent. A chip-backed loan works only if the offtaker pays and the data center is online in time to invoice. Zitron’s line, echoed by the bond move after the OpenAI run-rate revision: delay in energizing capacity compounds interest; a lab that renegotiates leaves the lender holding depreciating GPUs and a half-built hall. Nebius, IREN, Lambda, and Crusoe-style developers sit in the same structure at smaller scale. [34] [35] [36] [18]

Data-center developers and project SPVs. The campus is financed against a lease or a take-or-pay that has not started. Oracle had hundreds of billions of leases signed but not commenced. Zitron has pointed to an $18 billion Oracle data-center debt slice trading around 84 cents on the dollar, and to SPVs below par, as the market already separating “contract signed” from “cash collectible.” If the lab fails, the developer does not get a new hyperscaler tenant at the same rent: Microsoft, in his absorption scenario, will not pay Oracle’s and Google’s contracts, and unsold gigawatts compete with each other. Half-finished buildings still owe construction debt and power deposits. Delays are cash burn before the first GPU invoice — his Abilene example is that only a slice of planned megawatts was actually earning, while the capital structure assumed the full campus. [3] [21] [37]

Private credit. This is the layer that turned GPU purchase orders into loans. Structures include chip-backed facilities (CoreWeave’s $8.5 billion Meta-offtake loan was fully investment-grade, a template the industry wants to repeat), Broadcom equipment leases that Anthropic cannot easily cancel, residual-value support from Nvidia on a fraction of GPU sales, and insurer or private-credit money behind SPVs. Goldman Sachs has counted tens of billions of lower-rated AI-related borrowing in 2026; one October account put nearly $500 billion of AI-related debt priced in the year across Oracle, Broadcom, SpaceX and others. Zitron’s “brewing crisis” is not that every loan defaults on day one. It is that the collateral — a specific GPU generation in a specific building, leased to a junk-rated lab — is worth less the moment the offtaker’s raise slips, and that refinancing that paper at 13–15 percent while Treasuries are elevated eats the spread the original model assumed. Private credit does not mark to market like a bond ETF, so the stress can sit inside funds until a redemption, a ratings migration, or a sponsor that cannot extend. [18] [30] [34]

5. Contagion chain

TRIGGER
  OpenAI or Anthropic misses a raise, delays an IPO, or renegotiates take-or-pay
  (cash burn > new equity; junk-debt market will not fund $50B+/year)
        |
        v
LAB LAYER
  Token subsidies cut or free tiers killed
  VC-backed AI startups (the 1% of customers = ~80% of lab revenue, per Zitron)
  lose API credit and fail their own raises
        |
        +---------------------------+---------------------------+
        v                           v                           v
  MICROSOFT                   ORACLE                      GOOGLE / AMAZON
  $24.1B FY26 related-party   ~$300B contract starts      Anthropic $111.1B / $110B
  revenue stalls              2027; ~half of RPO          shortfall clauses hit
  RPO growth ex-OpenAI        S&P already BBB-,           Cloud "AI growth" misses
  was only ~25%               OpenAI = key credit risk    (UBS: up to ~48% of GCP
  Absorption path: cancel     Stuck with leases and       2027 in Zitron's cite)
  rival-cloud contracts,      negative FCF (~$42B
  impair GPUs, restate        FY27 forecast)
  $250B Azure promise
        |                           |                           |
        +---------------------------+---------------------------+
                                    v
                         NEOCLOUDS + CAMPUS SPVs
                         CoreWeave ~$52B debt/leases,
                         ~$104B backlog, bonds ~15% yield
                         Chip-backed loans lose offtake
                         Developers: unfinished GW, leases
                         signed but not commenced
                                    |
                                    v
                         PRIVATE CREDIT / INSURERS
                         GPU collateral marks down
                         SPVs below par; refinancing
                         gap vs original underwriting
                         Insurer "unlimited" bid only
                         if ratings hold — they don't
                                    |
                                    v
                         NVIDIA + CUSTOM SILICON
                         Hyperscaler and neocloud orders
                         pause after current deliveries
                         Customer concentration
                         (top 2–3 direct customers ~39–44%)
                         Circular supports reverse:
                         equity stakes, $105B lease
                         guarantee, residual-value wraps
                                    |
                                    v
                         PUBLIC MARKETS
                         AI-capex premium in MSFT, AMZN,
                         GOOGL, ORCL, NVDA, CRWV reprices
                         Zitron's endpoint: a violent
                         drawdown once the market accepts
                         the two-lab demand story was
                         the demand story

What would break the chain, on the evidence rather than on his rhetoric: a lab that actually produces free cash flow before the 2027 Oracle ramp and the Google/Amazon shortfall windows, a hyperscaler that quietly backstops the contracts (which is absorption by another name), or enough non-lab enterprise demand to re-let the halls at similar rents. Public filings so far show the first has not happened. S&P has already written down the probability of the third for Oracle. The second is the scenario Zitron thinks Microsoft might choose — and even that scenario, in his telling, zeroes out everyone who is not Microsoft.

6. What a competitor or a creditor should actually watch

The load-bearing variable is not the next model demo. It is whether the next OpenAI or Anthropic financing prints at a flat or up valuation, in cash, with no seller-financed cloud credits masquerading as equity — and whether any take-or-pay is amended. Secondary tells already in the market: Oracle’s rating and CDS, CoreWeave’s unsecured yields versus its chip-backed investment-grade loans, Microsoft’s related-party receivable ($6 billion at June 30, 2026), and the gap between run rate and cash. A lab can report a $50–65 billion annualized figure and still be unable to fund a $50 billion compute year if gross margin on tokens is thin and the contracts are gross obligations.

Two caveats belong in any use of this map. First, Zitron is a partisan analyst. Some of his totals ($852 billion burn, $1.3 trillion combined obligations, 90 percent Nvidia revenue drop, 54 percent Nvidia client share) are his arithmetic on top of leaked documents and sell-side notes, not line items in an SEC filing. Where the FT, Reuters, the WSJ, Bloomberg, or S&P have confirmed a number, that number is the one to underwrite. Second, Anthropic’s 2026 revenue acceleration was faster than his mid-year skepticism, and both labs are still growing. Growth does not, on the contracts as Reuters and the WSJ have described them, cancel the bill. It is how the bill gets larger.

---

Recent Findings Supplement (October 2026)

Ed Zitron's core thesis (detailed in his October 6 and October 13, 2026, analyses) is that OpenAI and Anthropic function as highly leveraged bets on perpetual massive capital raises and sustained explosive growth to cover fixed, largely non-cancelable compute commitments totaling around $1.3 trillion. Their deeply negative EBITDA (OpenAI at negative ~160% in 2025; Anthropic worse on a spend-per-revenue basis) and lack of hard assets make post-IPO debt raises of the required $50–100 billion per year impossible at investment-grade rates, likely resulting in junk ratings (B or CCC+).[1][2]

Zitron outlines three primary failure paths for OpenAI specifically: absorption by Microsoft (with Altman potentially sidelined), outright collapse (Altman as scapegoat amid funding exhaustion), or a value-destructive merger with Anthropic. Stronger reported demand does not resolve the issue because higher usage drives even greater compute needs and capital requirements, creating a cash-flow trap rather than a revenue solution.[3][1]

Recent audited or prospectus-backed figures (leaked/reported September–October 2026) show both companies remain deeply unprofitable despite rapid top-line growth, with commitments far exceeding current scale.[4][5]

• Anthropic (leaked draft S-1/prospectus, reported late September 2026): 2025 revenue $4.59–4.6 billion (12x growth from ~$386 million in 2024); operating loss $8.06 billion; net loss ~$42 billion (largely non-cash accounting items); compute/infrastructure spend $7.33 billion. Spend ratio: $2.75 per $1 of revenue (worse than OpenAI’s reported $2.60). Q2 2026 revenue ~$11.5–11.6 billion; Q3 ~$16.8 billion (45% sequential); annualized run rate ~$65 billion by end of July 2026. $518 billion in cloud/compute/infrastructure commitments over ~10 years (~80% non-cancelable or take-or-pay style, with shortfall penalties to partners like Google/Amazon). Breakdown includes Broadcom equipment leases ~$161.2 billion, Google ~$111.1 billion, Amazon ~$110 billion, Microsoft ~$31.4 billion. Two customers each ~12% of 2025 revenue; ~47% of revenue routed through cloud partners (with ~16% fees). IPO targeted post-U.S. midterms (November 2026 window discussed), with valuation aspirations ~$1.8–2 trillion.[4][6][7]
• OpenAI: 2025 revenue ~$13.07 billion with ~$34 billion in expenses/costs and $20.92 billion operating loss. 1H 2026 revenue ~$12.4 billion. Q2 2026 revenue $6.7 billion. Annualized run rate corrected to approaching $50 billion by end of September 2026 (down ~$20 billion from prior investor-derived ~$70 billion estimates due to gross-vs-net partner sales accounting differences with Anthropic); full-year 2026 revenue projected ~$35 billion. Compute commitments ~$750 billion through 2030 (or ~$665 billion cited at end-2025); internal projections include $856 billion infrastructure spend and $278 billion negative free cash flow from 2026–2030 (revenues projected to rise from ~$36 billion in 2026 to $350 billion in 2030, but expenses outpace). $122 billion funding round (March 2026) expected to run out around 2028.[8][9][10]

Knock-on effects center on extreme revenue concentration: OpenAI and Anthropic reportedly account for 70%+ (in some cases 80%+) of AI-related revenues at hyperscalers and the vast majority of frontier compute demand. Hyperscalers require ~$308 billion in annual AI revenue to break even on 2026–2027 capex but have only ~$183 billion currently, with 64% tied to these two labs. Unmet commitments would trigger immediate revenue misses, capex pauses or writedowns, and earnings pressure at Microsoft, Google, Amazon, and Oracle.[1][11]

Nvidia and Broadcom face potential revenue drops of up to 90% if lab demand collapses, given their heavy exposure to this concentrated buying. Neoclouds (CoreWeave et al.) and data-center developers face funding squeezes, higher borrowing costs (e.g., Zitron’s ~13% estimate for conventional CoreWeave debt), project delays, and distressed debt trading (examples include Oracle data-center bonds). Private credit faces a “brewing crisis” from layered exposure via loans, leases, and SPVs tied to these brittle customers, compounded by rising Treasury yields (~5.3% on 10-year notes) and souring credit markets.[3][12]

Text-form contagion chain (Zitron’s October 2026 framing):

Labs exhaust VC/debt capacity or hit junk-rate walls and cannot meet non-cancelable $1.3T commitments → Hyperscalers (MSFT/GOOG/AMZN/ORCL) report massive AI-revenue shortfalls and slow/reverse capex → Nvidia/Broadcom see order cancellations and 80–90% demand collapse → Neoclouds/CoreWeave-type players and data-center developers face covenant breaches, refinancing failures, and project halts amid high private-credit costs/delays → Private credit and related SPVs (Blackstone, etc.) incur losses on AI-tied debt; broader equity/credit contagion hits Big Tech earnings, pensions, and 401(k)s → Systemic confidence shock resembling an “Enron-Lehman hybrid.”

Recent market reactions (early October 2026) included dips in Nvidia, Oracle, and related names following OpenAI’s revenue correction reports, illustrating the sensitivity.[13]

These developments are drawn exclusively from post-April 15, 2026 reporting, primarily October 2026 leaks, filings coverage, and Zitron’s analyses. Figures reflect reported or projected values at the time and remain subject to audit or revision upon full public filings.

Ed Zitron’s checkable calls from late 2025 through mid-October 2026 mostly failed on timing and on how large AI revenue would get. The pattern is the same one that started in 2024: he treats a slowdown, a stall, or a credit crack as imminent and permanent, then the next quarter prints higher revenue, a new model, or another funding close. What has started to show up — and only recently — is stress in the debt that funds the buildout, not a collapse of Nvidia, OpenAI, or Anthropic.

How to read the record

Zitron writes a thesis, not a set of dated bets. The testable pieces are the ones with a deadline, a number, or a named company outcome: bubble pop by a quarter, a lab that cannot hit a revenue run rate, a neocloud that cannot survive, a coding startup that cannot exit, models that cannot improve, chips that cannot be plugged in. Conditional lines (“if OpenAI fails, it is Lehman”) are not yet scored.

Dan Luu’s September 1, 2026 review is the cleanest public ledger of the earlier calls. It marks essentially every resolved capability and timing prediction through late 2025 as wrong. [1] The primary sources below are used where they exist; Luu is used only where he attributes a specific quote.

Scorecard: late 2025 through October 15, 2026

1. The bubble pops no later than Q2 2026.

Date: October 2025, as attributed by Dan Luu.

Outcome: No cascade, no Nvidia growth break, no Silicon Valley flush by June 30, 2026. Nvidia set a new intraday high of $243.37 on October 6, 2026, and closed October 14 at $221.30, up about 23% year to date, with a market cap around $5.3–5.5 trillion. [2] [3]

Verdict: Wrong, if that “no later than” wording is accurate. Luu treats the phrasing as a hard upper bound. [1]

2. The bubble bursts within about 18 months, as a succession of startup failures rather than one Bear Stearns moment.

Date: October 16, 2025, Ars Live. He also said Nvidia’s growth is what inflates the bubble, and that a growth slowdown would burst it. He warned of a depression once markets accept that tech does not grow forever. [4]

Outcome: The 18-month window runs to roughly April 2027, so the dated call is not closed. The mechanism he named has not fired. Nvidia’s fiscal Q2 2027 revenue (quarter ended late July 2026) was $96.2 billion, up 106% year over year, with data-center revenue of $89.0 billion; guidance for the next quarter was $108 billion. Sequential growth is cooling (about 18% in that quarter, guidance implying about 12% next), but year-over-year growth accelerated, it did not break. [5] [6]

Verdict: Unresolved on the 18-month clock. Wrong so far on “Nvidia growth slows, then the market breaks,” and wrong so far on a depression.

3. “These models aren’t getting better. What we’re seeing today is pretty much what they’re always gonna be like.”

Date: November 2025. [1]

Outcome: OpenAI shipped GPT-5.6 in July 2026, GPT-6 Astra on September 3, 2026, and GPT-6 Sol and Luna on September 22, with the company claiming large accuracy gains and lower prices on the cheaper tiers. [7] [8] [9]

Verdict: Wrong. This continues a 2024–2025 streak of “peak AI / hit a wall” calls that Luu also scores as wrong. [1]

4. Anthropic cannot sustain the growth needed to support a roughly $47 billion annualized run rate.

Date: June 2026, as reported by Benzinga from a Zitron interview. [10]

Outcome: Anthropic told investors its run rate had passed $65 billion by late July 2026. Quarterly revenue was about $4.7 billion in Q1, $11.5–11.6 billion in Q2, and $16.8 billion in Q3, up 45% from the prior quarter. At a pre-IPO meeting on October 14 it also claimed adjusted profitability in Q3. Full-year 2025 revenue was about $4.6 billion, with an operating loss above $8 billion. [11] [12] [13]

Verdict: Wrong, and off in magnitude. A single 2026 quarter already exceeded a third of the 2027 revenue figure he had called “laughable” in February 2025. [1]

5. OpenAI’s published forecasts are absurd, and growth will slow sharply.

Date: October 2024 for the specific claim that $11.6 billion in 2025 (and $100 billion in 2029) was close to a “financial crime,” with growth set to slow into 2025. Included because it is the baseline his 2025–2026 commentary still rests on. [1]

Outcome: Audited 2025 revenue was $13.07 billion, above the figure he mocked, with a net loss around $38.5 billion (Zitron’s reporting; later cited by Quartz and others, and he says the Financial Times confirmed the statements). By the end of 2025 the annualized run rate was about $20 billion. By the end of September 2026 OpenAI told investors the run rate was approaching $50 billion, and it told them it expects $70 billion or more by year-end. It raised $122 billion in March 2026 at an $852 billion post-money valuation and is in talks for another $30 billion or more at a $1.4 trillion pre-money valuation. The IPO has slipped to 2027. [14] [15] [16]

Verdict: Wrong on the 2025 revenue ceiling and on an imminent growth stall. Partly right that headline run rates are slippery: the $70 billion figure circulating in late September was about $20 billion above the ~$50 billion number OpenAI itself gave investors a week later, largely because investors had grossed up partner sales the way Anthropic does and OpenAI does not. That gap knocked Nvidia, Oracle, and the Nasdaq on October 8–9. It was an accounting discrepancy, not a revenue collapse. [17] [15]

6. Cursor cannot be sold at a price that does not incinerate cash; a $10 billion exit is implausible.

Date: August 2025. [1]

Outcome: SpaceX agreed in June 2026 to buy Anysphere (Cursor) for $60 billion in stock and closed the deal on August 14, 2026. Reported ARR was about $4 billion by mid-2026, up from $2 billion in February. [18] [19] [20]

Verdict: Wrong. The exit was six times the ceiling he treated as implausible.

7. CoreWeave cannot survive six months except by fundraising.

Date: March 2025. [1]

Outcome: It is still operating in October 2026. The six-month death call was early. The credit call aged better, much later. On October 14, 2026, CoreWeave’s unsecured bonds due 2032 traded below 80 cents on the dollar, a level Bloomberg treats as distressed, with yields near 15% versus a 9.625% coupon issued in June. The stock fell 15.5% to $65.78. CEO Mike Intrator told staff the market was acting “irrationally” and pointed to a $104 billion revenue backlog. Interest expense was about $640 million in Q2 2026. Debt figures in circulation for mid-2026 range from about $35.6 billion of reported debt (Zitron, citing the June 30 balance sheet) to roughly $50–52 billion including lease obligations. No default has been announced. [21] [22] [10]

Verdict: Wrong on survival within six months. Partly right, and early, on the idea that CoreWeave’s model is a duration mismatch — long debt, slow data-center delivery, concentrated customers — that the bond market would eventually reprice.

8. Millions of Blackwell GPUs are sitting in warehouses; a large share of announced data centers will not be finished on the promised schedule; power is the binding constraint.

Date: November 24, 2025 (“Hater’s Guide to NVIDIA,” which he later summarized as arguing that 6 million Blackwell GPUs could not all be installed and powered). Restated in September–October 2026 as $200–350 billion of GPUs unplugged, and as Microsoft holding far more chip purchase power than energized AI load. [23] [24]

Outcome: Flagship Stargate Abilene is partly live, not dead. Epoch AI put operating capacity at about 0.3 GW in April 2026 against a 1.2 GW plan, with full campus completion targeted for Q4 2026. Oracle said in September 2026 that about 75% of the capacity it is providing there had been delivered to OpenAI. Other named Stargate sites were still at zero energized capacity, with completion targets mostly in 2028. A 900 MW expansion next to Abilene was dropped by OpenAI and reallocated. The $200–350 billion “unplugged” stockpile is Zitron’s estimate. It has not been confirmed by an independent inventory. [25] [26] [27]

Verdict: Partly right on delays and on power lagging chip purchases. Unresolved on the dollar magnitude of idle GPUs, and wrong if read as “these campuses will not be built.” Abilene is operating and still ramping.

9. OpenAI’s failure would be the Lehman Brothers of the AI bubble, with a violent market drawdown.

Date: July 15–17, 2026. He did not give a date. He noted SoftBank’s third $10 billion tranche was due October 1, 2026, and argued a stoppage of payments to Oracle and CoreWeave would break their debt math. [14] [28]

Outcome: SoftBank paid the tranche on October 1, funded with proceeds from an $11.1 billion high-yield bond sale (coupons up to 9.75%), completing a $30 billion follow-on and taking cumulative investment to $64.6 billion, about 13% ownership. It also repaid and canceled its bridge facility. OpenAI has not collapsed. It is still raising. [29] [30]

Verdict: Unresolved as a conditional. The near-term funding tripwire he flagged was cleared, at a high cost of debt for SoftBank rather than by a default.

10. AI private credit is already a brewing crisis; OpenAI and Anthropic cannot raise $50–100 billion a year at anything but junk rates; hyperscaler capex cannot be earned back.

Date: September–October 2026 newsletters and interviews (“Credit Crunch,” Monetary Matters, IBD). [10] [31] [32]

Outcome: Credit stress is visible and spreading at the margin, not systemic yet. CoreWeave’s bonds are the clearest print (above). Oracle fell 8.8% on October 14 to $130.55, and one tally puts the stock 59.6% below its September 10, 2025 record and down 32.4% in 2026, with long-term debt cited near $160 billion. Hyperscalers have not cut the build. Microsoft’s fiscal 2026 capital spending was $145.3 billion on the earnings-call definition ($115.9 billion of cash additions to property and equipment). In July it guided calendar 2026 capex to about $175 billion, down from its own prior $190 billion figure only because it lengthened assumed data-center life from 15 to 25 years and reclassified leases. It said the spending plan itself was unchanged, and Azure grew 43% in the June quarter. [33] [34] [35]

Verdict: Partly right on the direction of credit — junk spreads, CoreWeave distress, Oracle’s equity drawdown, SoftBank paying 8.6–9.75% to fund an equity check. Wrong so far on an imminent capex stop or earnings disaster at Microsoft, Google, or Amazon. Unresolved on whether the labs can keep refinancing $50–100 billion a year. They have kept raising equity; they have not yet had to live on standalone junk debt at that scale.

Where he was early, and where the magnitude was off

The timing error is consistent, and it predates the window in the question.

• From February 2024 through November 2025 he repeatedly called peak capability: data exhausted, models not getting more powerful, o1 as proof OpenAI was out of ideas, DeepSeek as the commoditization of LLMs, “generative AI isn’t going to do much more than it does today.” Each of those was scored wrong against later model releases and retained pricing power at OpenAI and Anthropic. [1]
• Revenue ceilings were not slightly low. They were off by multiples. The October 2024 sneer at $11.6 billion of 2025 OpenAI revenue missed a year that came in at $13.07 billion of recognized revenue and ended near a $20 billion run rate. The February 2025 sneer at a $34.5 billion Anthropic forecast for 2027 missed a company that was already at a $65 billion run rate in July 2026 and booked $16.8 billion in a single quarter. [1] [11]
• Survival calls were early by years, not months. CoreWeave’s March 2025 six-month death sentence failed. The credit stress he described showed up in October 2026, after an IPO and after tens of billions more of debt, and it is a repricing, not a bankruptcy.
• Bubble-pop dates have slid. Peak AI in early 2024, three quarters to prove itself, no later than Q2 2026 in the Luu ledger, 18 months from October 2025 in the Ars interview, and as of October 2026 still “when the money runs out,” with no new hard date. [4] [1] [32]

The magnitude error and the timing error are the same mistake. He models AI demand as a small, circular, soon-to-stall pool of venture-subsidized token spend. Recognized revenue at the two labs has compounded through that forecast. Losses have compounded too — that part of the ledger is real — but the stall he needed in order for the capex math to break on his schedule did not arrive.

What he got directionally right, without the crash

Three claims are holding up as descriptions, not as crash dates.

Circular revenue and sloppy ARR are not a conspiracy theory. Microsoft has disclosed on the order of $24.1 billion of fiscal 2026 revenue tied to OpenAI, which Zitron and Bloomberg-based writeups put at roughly 70% of Microsoft’s AI revenue and more than 7% of company revenue. The October 8 run-rate revision — ~$50 billion versus a ~$70 billion figure built to match Anthropic’s gross method — is exactly the kind of non-GAAP fog he has hammered. [36] [37] [15] The miss for him is treating that fog as proof the underlying demand is fake. Both labs’ quarterly revenue is still rising fast on the numbers reporters have actually seen.

Power and construction lag chip announcements. Abilene’s ramp was slower than early 2025 talk, other Stargate sites are 2027–2028 stories, and Oracle has had to issue force majeure language on at least one related project while saying the 2028 date still holds. [26] [25] That supports his “you cannot energize a press release” point. It does not yet support “hundreds of billions of GPUs will be written off.”

The bond market is the channel he said to watch, and this month it blinked. CoreWeave at a 15% yield, Oracle’s equity cut by about a third in 2026, SoftBank’s record junk deal at coupons near 10% to fund an OpenAI check, and Goldman’s count of $88 billion in lower-rated AI-related borrowing this year are all in the public record as of October 14–15. [10] [22] [30] [35] No major AI name has defaulted. Nvidia just made a high. The crisis he describes is a credit repricing at the edge of the trade, not the earnings collapse he has dated and re-dated since 2024.

What this implies

Someone trading Zitron’s calendar — sell Nvidia because growth must break, fade Anthropic because $47 billion of run rate cannot hold, fade Cursor because a $10 billion exit is fantasy — has been run over by the numbers through October 15, 2026. Someone who used him as a map of where the fragility sits — neocloud duration mismatch, Oracle’s tenant concentration, gross-versus-net ARR, chips bought ahead of power — has a live set of risks that the October 14 tape just marked to market.

The open items, in order of how soon they can be scored: whether CoreWeave refinances without a default or a heavily dilutive raise; whether Abilene actually hits the rest of its 1.2 GW by the end of 2026; whether OpenAI’s year-end run rate reaches the $70 billion it just told investors to expect, on its own net accounting; whether Anthropic’s adjusted Q3 profit survives a public S-1; and whether the 18-month bubble call from October 16, 2025 fails on schedule around April 2027 the way the Q2 2026 call already has. Until one of those breaks, the scorecard is a record of early and oversized bearish calls, plus a narrower and more recent hit on credit stress that has not yet become the crash.

---

Recent Findings Supplement (October 2026)

Ed Zitron (newsletter Where’s Your Ed At, podcast Better Offline) has been a prominent AI skeptic, arguing that generative AI progress has stalled, hyperscaler capex vastly exceeds real demand, much of Nvidia’s sales represent warehoused/uninstalled GPUs, AI revenues are inflated or circular (heavily reliant on loss-making OpenAI/Anthropic), and a financing/debt-driven bust is imminent.[1][2][3]

His specific, checkable predictions from late 2025 into 2026 have largely been tested against outcomes as of mid-October 2026. The broader picture shows sustained (and accelerating) hyperscaler capex, record Nvidia data-center revenues and guidance, ongoing massive fundraising by AI labs, creative financing arrangements, and no major defaults, project cancellations, or market collapse—despite rising debt concerns and opaque capacity reporting.[4][5][6]

Zitron’s earlier timing calls (2024–early 2025) on model stagnation, revenue peaks, and bubble burst were consistently too early or off in magnitude; demand and investment have proven more persistent than predicted. Recent 2026 commentary focuses more on structural overbuild, unverified numbers, and impending cash-flow/debt issues rather than hard new deadlines.[7][7]

Scorecard of Specific Predictions (Late 2025–Oct 2026)

Bubble burst timing (“no later than Q2 2026” or similar near-term collapse via cascade of startup failures/fundraising collapse): Oct 2025 (and reiterated/nuanced into 2026).[8][7]

• Outcome: No burst as of Oct 15, 2026. Hyperscalers (Microsoft, Alphabet, Amazon, Meta) guiding ~$733B capex for 2026 (midpoints); Nvidia data-center revenue hit $89B in one recent quarter with strong sequential/YoY growth and high guidance ($108B total revenue next quarter); AI labs continuing to raise capital; no reported wave of cancellations or defaults.[4][6]
• Verdict: Wrong (strong testable upper bound missed; demand and spending have not collapsed).

Models have hit a wall / “aren’t getting better” / “pretty much what they’re always gonna be like”: Multiple statements into late 2025.[7]

• Outcome: Sustained high demand for advanced GPUs (Blackwell ramp, future Vera Rubin expectations) and hyperscaler spend continuing at elevated levels into 2026–2027 guidance imply ongoing perceived value and progress.[9]
• Verdict: Wrong.

OpenAI/Anthropic revenue forecasts “laughable” or absurdly high; growth already slowing dramatically: Late 2024 into 2025, with ongoing skepticism into 2026.[7]

• Outcome: Labs showing high (if loss-making) run rates; Anthropic leaked S-1 details ~$4.6B audited revenue with >$8B losses and heavy reliance on resellers; OpenAI run-rate revisions downward by ~$20B noted in Oct 2026 commentary, but overall ecosystem spend and Nvidia sales remain robust.[10]
• Verdict: Partly right (losses and circularity confirmed; magnitude/timing of slowdown overstated relative to continued investment).

~$200–350B (or ~50%) of AI chips/GPUs sitting in warehouses or unpowered data centers; hyperscalers (e.g., Microsoft $265B capex since ~2022 yielding only ~$50B operational GPUs at ~2GW AI-specific capacity vs. vague “12GW” claims): Sep 22, 2026 article and subsequent Oct 2026 interviews/podcasts.[2][3]

• Outcome: Zitron’s estimate derived from capex vs. operational GW figures, construction-in-progress (CIP) balances (e.g., Oracle, Amazon, Google), and vague capacity language. No independent audit or refutation in recent sources; aligns with reports of power/construction delays and financing creativity (e.g., Amazon $8B chip sale-leaseback).[11][12]
• Verdict: Unresolved/plausible (his analysis; supports overbuild thesis but unverified by third-party data in available sources).

70%+ of hyperscaler AI/cloud revenues from OpenAI/Anthropic (who cannot sustainably pay); broader debt/financing “brewing crisis” with higher costs, delays, and eventual project cancellations when “money runs out”: Aug–Oct 2026 pieces and appearances.[11][13]

• Outcome: Concentration and loss-making nature of labs widely referenced; debt issuance rising (Goldman Sachs ~$88B lower-rated AI borrowing); creative structures (Amazon chip deals, CoreWeave financing); no major cancellations or defaults reported yet, though credit concerns noted.[11]
• Verdict: Right on concentration and losses/circularity; wrong/partly right on imminent crisis timing and severity (system still functioning with ongoing raises and spend).

Additional notes on past timing calls (pre-late 2025, for context): Multiple 2024–early 2025 predictions on model progress stalling, specific revenue targets being unrealistic, CoreWeave failing quickly, Gemini user targets, etc., were rated wrong in independent audits. Zitron has been repeatedly too early on stagnation or bust and off on the persistence/magnitude of demand and investment.[7][7]

Recent Developments (Post-4/15/2026 Focus)

• Nvidia earnings/capex: Strong ongoing results with data-center dominance (e.g., recent quarter $89B data-center revenue, high guidance, supply commitments rising to $279B). Hyperscaler IT capex share for Nvidia rising (11% early 2023 → 38% Q2 2026, expected 44% end-2027). No signs of demand collapse.[14][6]
• Hyperscaler spending: 2026 guidance ~$733B (four majors); continued acceleration in Q3 2026 infrastructure spend.[4]
• Financing/Overbuild signals: Amazon exploring $8B Nvidia chip sale-leaseback; rising private credit/junk debt for AI; Zitron and others highlighting uninstalled assets and power/construction bottlenecks. No bust materialized.[11]
• AI lab finances: Anthropic S-1 leaks and OpenAI run-rate revisions underscore losses but also continued capital access.[10]

Implications for competitors/entrants: Zitron’s core thesis on circular demand, opaque metrics, and eventual cash-flow mismatch remains a live risk to monitor (especially debt markets and actual utilization), even as his repeated timing predictions have missed. Actual outcomes favor continued buildout and Nvidia strength in the near term, with potential for sharper correction if power, financing costs, or lab runway issues intensify. Independent verification of utilization/capacity claims would strengthen or refute the overbuild narrative. Sources are primarily recent news, earnings references, and Zitron’s own output (cross-checked where possible).

The strongest critiques of Ed Zitron do not deny the gap between AI capital spending and proven third-party profit. They deny that the gap implies fraud, imminent insolvency of the cash-rich hyperscalers, or a 2008-style socialization of losses. His thesis, as of late 2026, is more specific than “AI is hype”: OpenAI and Anthropic cannot fund the compute they have committed to; most hyperscaler “AI revenue” is those two labs recycling cloud spend; inference is still too expensive relative to usefulness; GPUs are not durable infrastructure the way fiber and rail were; and private credit plus off-balance-sheet leases will push losses onto pensions and insurers. The better rebuttals hit some of those claims with operating data and miss others by arguing against the 2024 version of his case.

Enterprise adoption is real. That does not close his revenue-concentration claim.

Kelsey Piper’s April 2026 critique in The Argument is the cleanest version of the “he is arguing with a market that no longer exists” case. Zitron’s 2024 articles leaned on earnings calls in which companies said AI was not yet moving revenue, and on the prediction that models “cannot do much more than [they are] currently doing” and “isn’t getting much more efficient.” By 2026 that empirical base had flipped: progress from 2024 to 2026 was faster than from 2022 to 2024, and paying enterprise use was no longer anecdotal. Piper’s point is that once businesses are paying, the bubble case has to be about whether that revenue can ever cover the buildout, or whether use is a fad—not about whether anyone is using the product. She argues Zitron mostly stopped making that harder case and shifted toward fraud allegations he does not prove. [1]

The adoption numbers behind that shift are concrete, and they are concentrated in a few workflows rather than a general productivity miracle.

• Andreessen Horowitz reported that about 29% of the Fortune 500 and roughly 19% of the Global 2000 were live, paying customers of a leading AI startup, defined as a converted pilot under a top-down contract—not a failed pilot count. Coding, support, and search dominate the revenue momentum. [2]
• Ramp’s expense data showed Anthropic, for the first time, with more verified business customers than OpenAI among its clients: 34.4% versus 32.3% in May 2026, up from 9% for Anthropic a year earlier. [3]
• Cloud platforms that actually sell the capacity are still accelerating. Microsoft said Azure and other cloud services grew 43% in the quarter ended June 30, 2026, with full-year Azure revenue above $100 billion, commercial remaining performance obligations of $678 billion (up 84%), and Microsoft 365 Copilot past 30 million paid seats. Google Cloud grew 82% in the same quarter; AWS grew 37% to $42.2 billion. Management at all three attributed the acceleration to AI infrastructure and AI solutions demand, not only to the two labs. [4][5]

Where this engages Zitron: it kills the 2024 claim that enterprises are not buying. Where it misses him: he has already moved the target. His later claim is that hyperscaler AI revenue is overwhelmingly OpenAI and Anthropic—he has cited Barclays, Wells Fargo, and UBS estimates that a large majority of some clouds’ AI revenue, and a large share of future Google Cloud revenue, is those two customers—and that outside the labs, third-party generative-AI software revenue is still single-digit billions against hundreds of billions of capex. [6][7] Cloud growth and Copilot seats do not, by themselves, refute customer concentration. McKinsey’s 2026 survey is closer to his ROI point than to the bulls’: 80% of respondents say AI improved their own productivity, 44% say AI is scaling across the enterprise (up from 38%), but the share reporting any EBIT impact is stuck at 37%, and “high performers” attributing at least 5% of EBIT to AI are still about 6%. About one in five say token costs are already constraining use. [8] Individual usefulness and enterprise profit are not the same number, and Zitron’s thesis lives in that gap.

For someone competing here, the implication is narrow: sell into coding, support, and search, where customers already pay, and do not underwrite a business on “AI will show up in EBIT next year.” McKinsey’s own value-curve work estimates roughly $2.5 trillion a year of economic opportunity in automating or augmenting existing work, but says about $760 billion in annual profit pools would be needed to justify roughly $7 trillion of infrastructure capital by 2030. That is a demand case, not a proof that today’s capex is already earned. [9]

Inference costs collapsed. Zitron’s 2024 efficiency claim is wrong. His 2026 cash-flow claim is only partly answered.

The cost data is the sharpest factual rebuttal in the set. Epoch AI’s September 2026 paper, “The Plunging Price of Thought,” finds that the cost of hitting a fixed benchmark score has fallen about 47% per quarter since 2023, or roughly 13 times per year—faster than DNA sequencing or batteries on Epoch’s comparison. On GPQA Diamond, OpenAI’s o3 scored 75% in January 2025 at about $0.30 per question; GPT-5.6 Luna matched that score for $0.0004, a 725-fold drop in under 18 months. The decline is fastest while a capability is new (about 66% per quarter) and slows later (about 32%). [10][11][12] Stanford’s AI Index series, which Piper and others still cite, showed a GPT-3.5-class system falling from $20 per million tokens in November 2022 to $0.07 by October 2024, more than 280 times. Piper cites a separate estimate that GPT-4-level intelligence is about one-thousandth as expensive as at release. [13][1]

That directly falsifies Zitron’s summer 2024 line that generative AI “isn’t getting much more efficient.” It also changes which uses clear a hurdle rate. It does not, by itself, make the labs’ income statements work, because customers respond to cheaper tokens by using far more of them. Epoch’s own caveat is the right one: a forecast of total spending still needs usage, task difficulty, and the number of model calls. Agents multiply calls. Piper notes that Zitron treats the existence of a $200-a-month subscriber who burns $1,000–$2,000 of compute in a week as proof the whole product is “deeply unprofitable,” which confuses a loss-leader cohort with the blended book—an all-you-can-eat restaurant can lose money on the largest appetites and still clear a margin. [1]

SemiAnalysis’s October 2026 estimates are the industry-defender version of the same mechanism. It puts Anthropic’s blended inference gross margin in the mid-60% range by mid-2026, up from an estimated negative 94% in 2024, with the API book (roughly 75–85% of ARR in that analysis) above 80%. Subscriptions are about 10% of revenue and more than 40% of inference compute, which drags revenue per megawatt. These are model-based estimates, not audited GAAP. A more conservative normalization from Grizzly Bulls puts economic gross margin, after partner revenue share and before treating training as a period cost, closer to 50–60%, because the oft-cited “above 80%” figure was reported before those items. [14][15]

Where this engages Zitron: unit cost is not stuck, and a usage-based API business can look like software once utilization and hardware efficiency improve. Where it misses him: he is now arguing about cash, not cost per token. If every cheaper token is consumed by a longer agent trace, and if training plus committed capacity still exceed contribution margin, falling prices can widen the funding hole even as they prove the technology works. Goldman’s September 2026 framing is closer to his math than to a victory lap: the bank expected the top five U.S. hyperscalers to spend about $800 billion on AI infrastructure in 2026, against AI-related cloud revenue only about $70 billion above pre-AI trends, and roughly $300 billion needed to break even—an annual shortfall it put near $230 billion. [16] Cheap inference is necessary. It is not sufficient.

Hyperscalers can borrow. They can no longer honestly claim the buildout is self-funded.

The “pristine balance sheet” defense is real and time-stamped. Jared Gross, head of institutional portfolio strategy at JPMorgan Asset Management, told Bloomberg in October 2026 that this is a boom, not a bubble: a bubble implies fantasy, whereas this is a real industry, even if some firms are overvalued and some will fail. His load-bearing point on credit is initial conditions. A large share of the new debt is being issued by hyperscalers that entered the cycle with revenue engines and balance sheets “pristine to a degree that we’ve almost never seen.” Spread widening, in that view, is price discovery against an uncertain payback, not evidence the issuers cannot service the paper. [17]

A June 2026 JPMorgan estimate put global AI-related capex at $5.5 trillion through 2030, hyperscaler capex at $650 billion in 2026 and above $1.1 trillion in 2027, with operating cash flow still projected to surpass $900 billion by 2027. The bank read elevated leverage as a choice: borrow while terms are good, delever later. [18]

That self-funding story is already breaking at the margin, which is the part of Zitron’s case the cash-flow bulls underweight. Epoch AI’s analysis of SEC filings, published in June 2026, found aggregate cash capex at Microsoft, Amazon, Alphabet, Meta, and Oracle on track to overtake combined operating cash flow around the third quarter of 2026. Operating cash flow was growing about 23% a year and cash capex about 70%. Oracle had already crossed; Amazon was crossing; Alphabet, Meta, and Microsoft were later on the same trend. All five remained profitable. The crossover means continued capex growth has to be financed by cash drawdown, debt, or equity—not by free cash flow. [19] Amazon’s free cash flow had already fallen to negative $7.6 billion for the twelve months ended June 30, 2026, from positive $18.2 billion a year earlier, as property and equipment spending rose $66.1 billion. [20] S&P Global’s “Hyper 5” note put five-year historical capex at $1.1 trillion and analyst estimates at another $5.3 trillion through 2030, and named the risk directly: the transition from cash-funded to debt-funded investment before returns validate the spend. [20]

Where this engages Zitron: Microsoft, Google, Amazon, and Meta are not startups that “run out of money” the way a venture-backed lab does. Their advertising, productivity-software, retail, and non-AI cloud franchises can service a lot of debt even if AI returns are late. Where it misses him: “can borrow” is not “the AI business pays for itself.” His claim is that the incremental cash is circular—labs funded by the same platforms pay those platforms for compute—and that the core franchises are being spent down to support customers who cannot stand alone. Elite Currensea’s useful distinction is that markets may not be betting on the labs at all. They may be betting that the public winners are the toll roads. That is a different trade from the one Zitron is attacking, and treating them as one organism is his main analytical overreach. [21]

Anthropic is a better demand story than OpenAI. It is not yet a rebuttal of the commitment math.

The industry-defender case that “Anthropic is the serious one” has more evidence in 2026 than Zitron allows, and less than the ARR headlines imply.

Reported run rates are not comparable. Bloomberg reported in October 2026 that Anthropic books the gross amount of sales through cloud partners such as Amazon, while OpenAI books only its share of certain partner sales. Anthropic’s annualized revenue was reported at $65 billion by the end of July. A widely circulated $70 billion OpenAI figure was, per the Financial Times, an investor attempt to gross OpenAI up to an apples-to-apples comparison. OpenAI told investors its own annualized revenue was approaching $50 billion at the end of September, with full-year 2026 recognized revenue discussed around $35 billion. On a net basis OpenAI was still growing quickly—run-rate revenue up 77% in the third quarter and enterprise revenue up 107%, according to figures attributed to its investor disclosures—but the horse race was partly an accounting race. [22][23][24]

The product mechanism behind Anthropic’s lead is specific: Claude Code, launched in May 2025, opened enterprise coding and agent spend. Dealroom’s August 2026 compilation put Anthropic’s annualized revenue at $74.3 billion on July 22 versus OpenAI at $41.3 billion, with Claude Code the inflection. Those figures inherit the gross-versus-net problem and should not be treated as audited revenue. [25] Ramp’s business-customer crossover is harder to dismiss as a booking convention, because it is observed spend, not a lab’s run-rate definition. [3]

Zitron’s counter, which the bulls often skip, is the 2025 income statement and the forward commitments, not the 2026 growth rate. He has argued from a leaked prospectus that Anthropic spent about $2.75 to make $1 in 2025—worse than his read of OpenAI—and that compute commitments run into the hundreds of billions, much of it with Amazon and Google, who are also competitors and financiers. A decoding of the draft prospectus put 2025 revenue at $4.59 billion, of which $2.16 billion of usage was billed by Amazon and Google. [26][27] Growth since then can be both real and insufficient to retire those commitments if each dollar of new revenue requires another dollar of capacity the labs do not yet earn.

Where this engages Zitron: lumping Anthropic with OpenAI as equally hollow misses a coding-led enterprise business with improving inference margins and a broader paying-customer base. Where it misses him: an unaudited annualized run rate, especially one that grosses up partner sales, is not cash available to pay Oracle, Amazon, Google, or Broadcom. His claim is about the funding gap and circular demand, not about whether Claude is useful to engineers.

Railroads and dark fiber explain why society can win while the people who financed the peak lose. They do not rescue GPU equity.

Jeff Bezos’s “industrial bubble” line, as used by Gene Marks, is the historically literate bull case: railways, fiber, and factories get overbuilt; investors in the overbuild lose money; the assets remain and lower the cost of the next economy. Marks’s August 2026 Guardian column applies that to data centers specifically. Unlike a failed biotech partnership, a data center does not vanish if the clinical trial—in this case, the AI profit pool—disappoints. CBRE’s H2 2025 North America report, which he cites, had developers adding 36% capacity while vacancy fell to 1.4%. Microsoft’s AI Diffusion Report put generative-AI use at 17.8% of the world’s working-age population, which he reads as early, not saturated. [28]

The analogy is useful and, on Zitron’s actual objection, incomplete. Railroad rights-of-way and buried fiber last decades. The expensive part of an AI data center is the accelerator, and power—not glass—is now the scarce input. A strand of fiber laid in 1999 carries today’s photons once the optics at the ends are swapped; a GPU is the endpoint. An older chip that still works can have a negative opportunity cost if it occupies a megawatt a newer chip would use more profitably. Meta’s own filings have put server and network lives around five to five-and-a-half years, against 25–30 years for buildings. That is the distinction Zitron insists on when he says this is not dark fiber, and independent technical writing reaches the same split: the railroad analogy fits physical scale, subsidies, and geography; the fiber analogy fits financial risk and technological obsolescence; neither fits a depreciating, power-hungry endpoint. [29][30]

The partial rebuttal is utilization, not durability. Structure Research argues that circular financing is not, by itself, a bubble call; the tests are GPU utilization, contract pricing, renewals, and revenue per megawatt. It cites continued rental demand for A100s and H100s on inference and batch work after they leave frontier training. CoreWeave has disclosed an A100 contract running into 2029—about nine years after Nvidia announced the chip in May 2020—and management has said pricing on prior-generation SKUs is at or above where it was years ago. Oracle has reported renewals or resales of older GPU capacity at a premium, with fleet utilization near 98% in at least one cited quarter. [31][32][33] That challenges a strict two-to-three-year economic life. It does not make a GPU into fiber. Shells, substations, and fiber laterals can be reused. The chips that absorbed most of the capex cannot be upgraded in place.

Historical pattern, stated plainly: railroads and telecom were directionally right, and many of the companies that financed the peak still went bankrupt. The surplus capacity then made the next generation of applications cheap. A critique that stops at “the infrastructure will be useful” has conceded Zitron’s investor claim and answered a different question—social return, not creditor return.

Credit stress is real at the edge and limited at the core. “Not Lehman” is not “no losses.”

Morningstar DBRS’s October 6, 2026 commentary is the clearest ratings-house version of limited systemic risk. Amazon, Meta, Alphabet, and Microsoft will keep borrowing, credit metrics will weaken, and technology has risen to about 8% of U.S. investment-grade corporate bond issuance year to date in 2026, from about 4% in 2023. DBRS still sees stable credit strength. The four retain scale, market position, cash generation, and cash balances. A real hit to ratings, in its view, requires a significant weakening of the AI cycle such that earnings come in well below forecast, which it does not expect in the near term. Even in a stress case, it argues they have the firepower to restore credit quality. Capex could exceed $1 trillion a year by 2027. [34]

Advisor Perspectives’ reading of the CDS tape is more differentiated, and more useful against Zitron’s tendency to treat “Big Tech” as one credit. Oracle’s five-year CDS moved from below 50 basis points toward roughly 200, with debt-to-equity near 4 times. Microsoft, Amazon, Alphabet, and Meta were described with debt-to-equity between 0.18 times and 0.51 times, ratings in the AA to AAA range, and spreads that stayed relatively flat versus the broader AA index. The conclusion: the scary chart is mostly an Oracle chart. Off-balance-sheet leases and take-or-pay GPU contracts complicate that—Nikkei-traced obligations and Meta’s disclosed future leases (hundreds of billions against a much smaller on-balance-sheet debt stock) mean traditional leverage ratios understate the buildout—but they do not convert Microsoft into a neocloud. [35][36]

Marks’s accounting point is narrower than his headline. Special-purpose vehicles that keep data-center debt off the sponsor’s balance sheet are financial engineering with real assets and, today, heavy disclosure. They are not Enron’s fictitious trading book. Some projects will be worth less than cost. That is why the structures exist: to place residual risk with investors who are paid to hold it. Calling every SPV a fraud, which is where Zitron’s 2026 rhetoric sometimes goes, is a category error. [28]

The miss is the periphery Zitron actually emphasizes. The IMF’s October 2026 Global Financial Stability Report reviewed 14 disclosed deals since 2023 in which hyperscalers and other investors put about $300 billion of equity into financially weaker AI firms, alongside $650 billion of debt, in exchange for about $1.4 trillion of future revenue commitments. It called the structure wrong-way risk: the equity and debt lose value in the same state of the world in which the contracted revenue fails. Hyperscalers’ share of global investment-grade bond issuance went from under 1% in 2024 to about 10%. The Fund’s warning is not that Microsoft defaults. It is that a reassessment of AI profits, amplified by leverage, concentration, and circular deals, can tighten financial conditions with global spillovers. [37][38] Bank of America’s August 2026 Global Fund Manager Survey had already named AI hyperscaler capex as the most cited source of a systemic credit event. Apollo’s Torsten Slok has argued that hyperscaler CDS widening is a repricing of fundamentals—debt-funded capex, negative free cash flow, depreciating assets—not dealer hedging of new issuance, because bank CDS did not move with it. [39][40]

So the credit rebuttal holds if the question is “will Amazon’s senior bonds be the next Lehman?” It does not hold if the question is Zitron’s actual one: whether private credit, pension-funded vehicles, neoclouds, and Oracle-style concentrated counterparties can transmit losses even while the four cash franchises remain investment grade. Those are different balance sheets. Conflating them is how both sides overclaim.

What a serious critique has to concede

The best counterarguments, stacked, look like this. Models got cheaper and more capable on a schedule Zitron said was impossible. Enterprises pay, especially for coding. Anthropic’s customer and inference-margin trajectory is better than a pure subsidy story. The four largest U.S. platforms can issue debt against franchises that predate ChatGPT, and their senior credit is not a systemic bomb on current ratings views. Data-center shells and power interconnects can outlast a bad vintage of chips, the way fiber outlasted WorldCom. GPU rental markets are not yet behaving like a glut.

What those arguments do not do is make the capex self-liquidating. Goldman’s shortfall estimate, Epoch’s cash-flow crossover, McKinsey’s flat EBIT share, gross-versus-net ARR, and the IMF’s map of circular commitments are all compatible with a technology that works and a financing cycle that still destroys capital at the margin. Zitron is weakest when he treats every large number as fraud and every lab as identical. He is strongest when he forces the bull case to say, in cash rather than in run rate, who pays for capacity the labs have promised and the hyperscalers have not yet earned back from anyone else.

---

Recent Findings Supplement (October 2026)

Critiques of Ed Zitron’s AI-skeptical thesis (unsustainable hype, weak real demand/revenue, rising or stagnant costs, fragile hyperscaler finances, overstated lab prospects, and bubble-like systemic risks akin to railroads/telecom) draw on post-April 2026 data showing faster-than-expected adoption, collapsing unit costs, resilient financing capacity, and differentiated credit profiles. These responses emphasize verifiable outcomes over run-rate extrapolations or spreadsheet projections, while noting that Zitron’s forecasts have repeatedly missed on user growth, profitability trajectories, and cost curves.[1][2]

Enterprise adoption and revenue growth have exceeded several of Zitron’s earlier benchmarks, with hyperscalers and labs reporting accelerating usage and monetization that defenders tie to measurable productivity gains rather than circular spending.

• Meta’s revenue rose from $135 billion in 2023 to $201 billion in 2025, with operating profit climbing from $47 billion to $83 billion; Alphabet showed parallel gains to $403 billion revenue and $129 billion operating profit. First-half 2026 figures continued the upward trend.[1]
• Gemini reached 750 million users (exceeding cited targets) and later reports indicated 1 billion users, while OpenAI passed prior revenue-run-rate goals and reported enterprise revenue growth of 107% in Q3 2026 alongside overall run-rate expansion of 77%.[1][3]
• Counterarguments note that Zitron’s chain-of-inference critiques (e.g., linking third-party estimates or executive moves directly to company failure) often lack demonstrated mechanisms, and actual outcomes contradict claims of stalled or nonexistent demand.[2]

This data implies that enterprise willingness to pay has proven durable enough to support continued scaling, challenging narratives that adoption is forced or ROI-absent; competitors can point to usage metrics and partner revenue as evidence of product-market fit rather than pure speculation.

Inference costs have fallen dramatically—by roughly 13-fold annually since 2023 according to Epoch AI’s September 2026 analysis—directly contradicting assertions that costs “never decrease” or that efficiency gains are illusory.

• Achieving fixed benchmark performance (e.g., GPQA Diamond at 75% accuracy) dropped from ~$0.30 per query (o3, early 2025) to $0.0004 (later models), a 725-fold reduction; average quarterly declines reached ~47% across benchmarks, outpacing historical rates for DNA sequencing, batteries, or computing by large margins.[4][5]
• OpenAI implemented price cuts of up to 80%; Goldman Sachs-linked analysis showed LLM prices collapsing ~95% in three years, comparable to 15 years of PC price declines.[6]
• Defenders highlight that per-token or per-capability deflation enables new workloads (agents, complex workflows), even as total spend rises with usage volume—framing this as classic tech economics rather than a bubble signal.[7]

Rapid cost deflation supports arguments that the technology is becoming economically viable at scale faster than skeptics projected, lowering barriers for enterprise experimentation and reducing the capital intensity needed per unit of output.

Hyperscalers retain substantial cash-flow capacity and balance-sheet strength to self-fund or bridge capex despite FCF turning negative in 2026, with credit analysts emphasizing scale and market dominance as buffers against Zitron-style collapse scenarios.

• Aggregate capex for the top five hyperscalers is projected to exceed combined operating cash flow for the first time in 2026 (~$600–800 billion range), with FCF forecasts turning negative (e.g., Bank of America estimates shifting from +$180 billion in 2025 to –$64 billion in 2026); debt issuance has risen (e.g., $204 billion YTD for four majors).[8][9]
• However, Morningstar DBRS notes that while credit metrics weaken modestly, “exceptional scale, market dominance, and cash generation capabilities provide substantial downside protection,” allowing quick restoration of credit quality; hyperscalers represent only 10% of global IG bond issuance (up from <1% in 2024) but maintain strong business-risk profiles.[10]
• Microsoft has signaled confidence in demand justifying ~$190 billion 2026 capex; defenders contrast this with pure-play AI firms by noting diversified cash flows from core businesses.[11]

The shift to external financing is real but occurs from positions of strength, suggesting limited systemic spillover risk compared to more leveraged historical bubbles; entrants or competitors must still navigate power and supply constraints but face well-capitalized incumbents.

Anthropic has demonstrated competitive or leading momentum versus OpenAI in recent quarters, with strong enterprise traction and run-rate growth that undercuts claims of concentrated weakness or inevitable underperformance.

• Anthropic reported Q2 2026 quarterly revenue of $11.5 billion (surpassing OpenAI’s $6.7 billion) and a $65 billion annualized run rate by end-July; Q3 revenue reached $16.8 billion (45% QoQ growth), with targets of $110 billion run rate by year-end.[12][13]
• Accounting differences exist (Anthropic books gross cloud-partner sales; OpenAI nets its share), but both show rapid enterprise expansion and investor willingness to fund at scale; OpenAI’s net run rate approached $50 billion by September with full-year 2026 recognized revenue projected around $35 billion.[12]
• Defenders cite these figures and user/adoption metrics as evidence that the leading labs are not “frauds” or structurally doomed but are executing on differentiated demand.[14]

Divergent accounting and growth trajectories highlight that the two leaders are not interchangeable, with Anthropic’s recent outperformance in certain metrics supporting views of a viable multi-player market rather than a zero-sum bubble.

Credit-market analysis points to manageable rather than systemic risks for hyperscalers, with strong profiles and differentiation from weaker AI pure-plays limiting contagion potential despite rising leverage and circular financing elements.

• IMF’s October 2026 Global Financial Stability Report flags AI correction as a risk (alongside yields and cyber threats) due to ~$1.4 trillion in circular deals and hyperscaler bond issuance share, but notes banking-sector resilience even in stress scenarios and differentiation in market reactions.[15][16]
• Counter-views (e.g., Morningstar) stress that downside remains limited for major players because of cash-generation scale and ability to adjust; private-credit concentration warnings exist but coexist with evidence of prudent structuring and overall market absorption capacity.[10]
• Railroads/telecom analogies are engaged by noting real current usage and productivity signals versus purely speculative infrastructure, plus faster cost deflation and diversified revenue bases today.[17]

Overall, recent data tilts toward a view of high-stakes but self-correcting investment rather than imminent systemic failure, with the strongest counters grounded in audited or reported financials, benchmarked cost curves, and usage metrics that have outpaced prior skeptical forecasts. Additional granular enterprise ROI studies or audited lab filings would further strengthen these rebuttals.

Zitron’s core claim is a cash-flow mismatch, not just hype. I’m checking the latest revenue, GPU, and balance-sheet numbers that would have to be wrong for that claim to hold.
The Nvidia figures conflict across sources, so I’m checking primary filings and the Microsoft, Google, and Amazon AI revenue lines that actually pay for the chips.
Zitron’s strongest point is an accounting gap, not a demand mirage. The version that says the money is about to run out, and that hyperscalers cannot survive the mismatch, is the part October 2026 has already started to break.

What the thesis actually is

As of mid-October 2026, Ed Zitron’s case is not “nobody uses chatbots.” It is that the industry is spending on a scale that end-customer economics cannot support, and that the numbers used to deny that are circular and unaudited. In a September 23 Investor’s Business Daily interview he put the mechanism bluntly: Big Tech has already spent over $1 trillion of capex, plans roughly another trillion next year, and most of the visible AI revenue still comes from OpenAI and Anthropic, companies the same hyperscalers also have to fund. He has also argued that annualized run rates are “made-up numbers,” that a large share of lab revenue is startups burning venture dollars on tokens, and that private-credit financing of data centers is a “brewing crisis.” The strong form of the call is a cascade: lab cash runs out, neoclouds default, hyperscalers take catastrophic earnings misses, and NVIDIA revenue could fall by as much as 90 percent. [1] [2] [3]

That diagnosis of a gap is partly right. The prediction that the system must therefore seize up on a near-term clock is where the case is overstated.

Revenue is real, growing, and still not the revenue he is scoring against

The cleanest counter to “tens of billions against trillions” is that the revenue base has moved. OpenAI told investors its own annualized revenue was approaching $50 billion at the end of September 2026, up from about $20 billion at the end of 2025, with 77 percent run-rate growth in the third quarter and 107 percent growth in enterprise. The widely cited ~$70 billion figure was investors grossing OpenAI up onto Anthropic’s accounting, not OpenAI’s number. Recognized full-year 2026 revenue is still expected around $35 billion, so the run rate overstates cash already earned. [4] [5] [6]

Anthropic’s leaked draft prospectus is harsher and more useful. Revenue grew twelvefold to nearly $4.6 billion in 2025. Operating loss was $8.06 billion. The headline net loss of about $42 billion included roughly $34 billion of non-cash accounting tied to financing instruments, not cash burned running the business. Compute and infrastructure cost $7.33 billion. By the end of July 2026 its gross annualized revenue was reported at $65 billion, and preliminary second-quarter revenue was reported around $11.5–11.6 billion. Some accounts say the company reached adjusted operating profit in the second quarter; those reports disagree, and the filing is not a public audited S-1. [7] [8] [9] Nathan Benaich’s State of AI Report 2026 put the two labs’ combined reported run rates at about $105 billion by late summer, up from about $30 billion at the start of the year. [10]

The more important revenue is not the labs. It is the clouds that sell the compute:

• Microsoft’s fiscal 2026 revenue was $331.8 billion, up 18 percent, with operating income of $155.2 billion, up 21 percent. Azure topped $100 billion for the first time, up 41 percent, and grew 43 percent in the June quarter. Microsoft 365 Copilot passed 30 million paid seats. Commercial remaining performance obligations hit $678 billion, up 84 percent. OpenAI contributed $24.1 billion, about 7 percent of Microsoft and roughly a quarter of Azure, so concentration is real. Excluding OpenAI, the backlog still grew 25 percent. [11] [12] [13]
• Google Cloud grew 82 percent year over year to $24.8 billion in the June quarter, with operating income of $8.8 billion and a 35.6 percent operating margin, more than triple the year-earlier profit. Backlog was $514 billion. [14] [15]
• AWS grew 37 percent to $42.2 billion in that quarter, its fastest growth in 18 quarters, and management cited a $169 billion annualized run rate. [14] [16]

Zitron is right that ARR is a bad yardstick and that partner-gross versus partner-net makes leaderboards incomparable. He is wrong if “revenue” is treated as fake because it is not yet equal to capex. Cloud AI revenue is showing up in audited public filings, with margins, and with contracted backlog. Goldman Sachs’ September 25, 2026 note is the fair version of his gap: the top five hyperscalers need roughly $300 billion of annual AI revenue to break even on about $800 billion of 2026 AI infrastructure spend, while AI cloud revenue is only about $70 billion above the pre-AI trend, a roughly $230 billion annual shortfall. That is a returns problem. It is not evidence that the revenue line is invented. [17]

Enterprise profit capture is the soft underbelly, and here Zitron’s skepticism has better footing than the lab run rates. McKinsey’s August 2026 survey of 1,719 respondents found that 80 percent report personal productivity gains, but only 37 percent attribute any EBIT impact to AI, unchanged from the prior year. High performers, those attributing at least 5 percent of EBIT to AI, were still about 6 percent. Usage is scaling. Corporate P&Ls are not, yet. [18]

GPU demand is contracted, not hoped for

NVIDIA’s fiscal second quarter, ended July 26, 2026, is the hardest fact against “the chips are a story.” Revenue was $96.2 billion, up 106 percent year over year and 18 percent sequentially. Data Center was $89.0 billion, up 117 percent. GAAP gross margin was 75.0 percent. Guidance for the October quarter was $108 billion, plus or minus 2 percent, with no China data-center compute assumed. Supply and capacity commitments rose to $279 billion from $119 billion the prior quarter. [19] [20]

Customers are paying before delivery. In the first half of fiscal 2027 NVIDIA recorded $15.6 billion of customer advances, more than double the year-earlier period, and recognized $13.0 billion of that as revenue. That is the opposite of a warehouse of unsold GPUs. [21]

The bearish reading of the same filing is also real, and Zitron is closer to right on this slice. Accounts receivable rose to $63.1 billion. Five direct customers were 70 percent of that balance. NVIDIA has extended investment-grade buyers terms of 90 days to a year, and in August 2026 it agreed to guarantee up to $105 billion of lease and power obligations for an OpenAI-affiliated campus in Ohio. Demand is not imaginary. It is concentrated, partly vendor-financed, and partly circular. A 90 percent revenue collapse would require those contracted buyers to walk away from prepaid capacity and from their own cloud backlogs at the same time. Nothing in the July quarter shows that. [21]

Zitron’s claim of $200–350 billion of GPUs sitting unpowered is an estimate, not a filing. Power and shell delays can strand capacity. They have not yet shown up as unsold NVIDIA product.

Debt does not hit the companies he needs it to hit

The balance-sheet split is the main reason “the money will run out” does not travel from CoreWeave to Microsoft.

Microsoft, Amazon, Alphabet, and Meta still sit in a different credit world from the neoclouds. An August 2026 Advisor Perspectives review of the credit data put their debt-to-equity ratios between 0.18x and 0.51x, with ratings in the AA to AAA range and spreads that had not blown out. Oracle was the outlier: five-year CDS from under 50 basis points to about 200, debt-to-equity near 4x, bonds trading like junk despite an investment-grade rating. [22]

Microsoft’s own books show how the insulation works. Long-term debt fell to $31.1 billion at June 30, 2026, down 22.6 percent. Cash and short-term investments were still about $77 billion. Operating income was $155 billion. The obligation that matters is not the bond line. It is leases: long-term leases of about $79 billion, and data-center leases not yet commenced of $329.1 billion, starting between fiscal 2027 and 2033, some subject to conditions. Microsoft also reclassified future data-center leases from finance leases to operating leases, which cut reported calendar-2026 capex guidance from about $190 billion to about $175 billion without changing the build. That is accounting flexibility a junk issuer does not have. Fiscal Q4 free cash flow was $19.6 billion, down 23 percent, but still positive and above estimates. [23] [24] [25] [26]

The group as a whole is stretching. 2026 capex guidance is about $175 billion at Microsoft, about $220 billion at Amazon, $195–205 billion at Alphabet, and $130–145 billion at Meta, roughly $720–745 billion combined before Oracle. Debt’s share of hyperscaler capex rose from about 9 percent in fiscal 2024 to about 32 percent on a trailing basis by mid-2026. Epoch AI’s reading of filings has aggregate cash capex on track to exceed operating cash flow around the third quarter of 2026. Goldman counted nearly $500 billion of AI-related debt issuance in 2026 by early August. [27] [28] [26]

That is a cost-of-capital problem. It is not a solvency problem for companies that still generate more than $100 billion of operating income. They can fund a multi-year gap, slow the build, or reprice capacity to other customers if a lab stumbles. The entities that cannot are the ones already showing stress. CoreWeave’s total liabilities were about $52 billion by June 2026, with standalone debt around $35 billion and quarterly interest of $640 million, more than double a year earlier. On October 14, yields on its 9.625 percent bonds due 2032 hit 15 percent and the stock fell more than 15 percent. Oracle fell 8.8 percent the same day. [29] [30]

Zitron’s error is treating that credit fracture as if it were Microsoft’s balance sheet. A CoreWeave or Oracle problem can hurt NVIDIA’s customer mix and force contract renegotiation. It does not, by itself, make Sundar Pichai or Satya Nadella unable to pay for GPUs.

The labs can still raise, and they can still rewrite the bill

Capital markets have not closed. OpenAI closed $122 billion of committed capital in March 2026 at an $852 billion post-money valuation. As of September 29 it was in early talks to raise at least another $30 billion at about a $1.4 trillion pre-money valuation, as bridge financing after Sam Altman pushed an IPO out of 2026. Terms can change. The round is not closed. The existence of investor-led demand at a higher mark, six months after the largest private round in the sector’s history, is incompatible with “they cannot raise.” [31] [32] [33]

Anthropic raised $65 billion at a $965 billion valuation in May, confidentially filed, and is discussed at $1.8–2 trillion, with a roadshow reported for the week of November 9. The same draft that bulls use for the valuation is the document that makes Zitron’s commitment point: at least $518 billion of cloud, compute, and equipment obligations, about 80 percent non-cancelable or payable regardless of use, including about $111 billion with Google, $110 billion with Amazon, $161 billion of Broadcom-linked equipment leases, and $31.4 billion with Microsoft. The filing says that if spend falls short, Anthropic must pay Google the difference. [7] [34] [33]

Those commitments are the real risk. They are also renegotiable in a way a bond covenant is not, because the counterparties are the same companies that own the capacity, hold equity, and want the labs as anchor tenants. A restructuring would be ugly for Anthropic’s valuation. It would not automatically zero the GPUs. Microsoft, Amazon, and Google can reallocate unused capacity to other cloud customers, which is exactly what a $514 billion Google Cloud backlog and a $678 billion Microsoft backlog are for. The circularity cuts both ways: the hyperscalers are on the hook, and they are also the residual buyers.

The track record is a sequence of missed ceilings

Dan Luu’s September 1, 2026 review of resolved Zitron predictions is the cleanest public scorecard, and the pattern is specific ceilings stated with high confidence, then missed:

• February and March 2024: generative AI had hit the upper limit of what it could do. Wrong.
• October 2024: OpenAI’s forecasts, including $11.6 billion for 2025, were so absurd they were close to a “financial crime.” OpenAI’s 2025 revenue was $13.07 billion. Wrong on the number he chose to mock. The Financial Times, looking at the same materials Zitron shared, also calculated the cash loss nearer $8 billion on $13.07 billion of revenue after a conversion-related write-down, not the larger burn figure he emphasized. [35] [36]
• February 2025: Anthropic’s $34.5 billion revenue forecast for 2027 was “laughable.” By mid-2026 the company’s reported run rate was already above that. Wrong, or at minimum no longer laughable. [35]
• February 2025: Google’s 500 million Gemini-user target was so unrealistic Sundar Pichai should be fired. Gemini later cleared 650–750 million. Wrong. [37] [38]
• October 2025: asked when the bubble pops, “no later than Q2 2026.” That date has passed. Wrong as a testable upper bound. [35]
• November 2025: the models are not getting better and are roughly what they will always be. Wrong on the capability record since. [35]

Kelsey Piper’s April 2026 critique in The Argument is methodological rather than a prediction scorecard. She argues that the 2024–25 case was a business case, companies were not using the tools and would run out of cash, and that the 2026 case shifted toward Enron- or FTX-style fraud allegations once usage and revenue made the older case harder. She also notes he has analogized OpenAI to FTX while conceding he does not have proof that free tokens are booked as revenue. That is a real method problem: stacking unproven fraud claims so each one props up the next. [39]

Timothy B. Lee documented a concrete source failure. A Zitron spreadsheet meant to show Anthropic’s 2025 revenue was only $3.6 billion, implying the company’s statements did not add up, skipped February 1–10, double-counted March 1–10, treated August 21–October 21 as one month, and omitted October 21–November 1. Fixing the date errors produced roughly $4.3–4.5 billion, in line with later reported figures. Lee said he emailed the errors and got no response. A linked spreadsheet is not a check. [40] [37]

None of this means the capex gap is fake. It means timelines, capability ceilings, and “the filings don’t match” claims have a poor hit rate, and that high rhetorical certainty has not updated when the numbers did. In Vanity Fair he brushed off the boy-who-cried-wolf question by saying 2024 was a long time ago and he had been naive about how long the system would tolerate the gap. That is an admission that the mechanism can persist longer than the call. [41]

Incentives reward the unbroken call

Zitron is not a short seller with a disclosed position. He is a publisher. Where’s Your Ed At, the Better Offline podcast, and a social audience of about 156,000 followers monetize a crisp, certain bear case. Nuance does not book the CNBC, Bloomberg, and YouTube circuit he has been on through 2026. Dan Luu’s review quotes the standard attention mechanism: crisp positions with no room for doubt outperform “the market could go either way.” [42] [35]

WIRED’s October 2025 profile documented the other incentive tension. He runs EZPR. At the time of that profile he said the firm had four clients and that he did not want to pitch generative AI, while the piece also described him pitching an LLM-connected client and representing automation products such as DoNotPay. He told WIRED the AI criticism had not cost him PR business. The conflict is narrower than “he is paid by OpenAI.” It is that the public product is anti-hype certainty, and the private product has included AI-adjacent clients. That does not falsify a capex spreadsheet. It does mean the brand cannot easily publish “I was early, and the gap is narrowing.” [43]

A separate, quieter bias is definitional. If revenue must be profitable, diversified, non-circular, and large enough to cover a year of capex in that same year, then almost any infrastructure build fails the test until it is finished. Telecom dark fiber and early AWS both failed that test. Zitron’s answer, in a June 2026 interview, is that those analogies are “hogwash” and that the industry needs at least $200 billion a year of AI-compute revenue to make sense. On current cloud and lab figures, that bar is no longer absurd. It is also not yet cleared on Goldman’s $300 billion break-even math. The disagreement is now about slope and who holds the residual risk, not about whether revenue exists. [44]

What would falsify him in the next 6–12 months

Through October 2027, the strong thesis dies if the following hold together. None of them requires AGI.

• NVIDIA prints the $108 billion October-quarter guide, or something close, keeps gross margin around 74 percent, and does not guide a sequential air pocket. A second quarter of rising backlog or customer prepayments would make “unsold GPUs in warehouses” very hard to maintain. [19]
• OpenAI closes the reported $30 billion-plus round at or above a $1 trillion valuation, and Anthropic either lists above $1 trillion or raises again without a down round. Failure to close, or a pulled IPO, would strengthen him.
• Microsoft, Amazon, and Alphabet keep cloud growth well above 20 percent and do not cut 2027 capex by a large fraction while talking about excess capacity. A voluntary slowdown because returns are late is not the same as a cash crisis. A guidance cut paired with rising credit spreads at the AA/AAA names would be.
• Incremental AI cloud revenue moves materially off Goldman’s ~$70 billion-above-trend figure toward the $150 billion-plus range, shrinking the $230 billion annual gap. If that gap is unchanged or wider after another year of 30–40 percent cloud growth, his returns math wins even if nobody goes bankrupt. [17]
• Audited or prospectus-quality lab numbers show cash losses narrowing, or at least growing slower than revenue, rather than compute bills compounding faster than sales. Anthropic’s claimed adjusted operating profit quarter, if it survives the S-1, is an early version of this. A cash runway under 12 months without new capital would revive the “money runs out” line for the labs specifically.
• McKinsey-style surveys show EBIT attribution rising well above 37 percent, or the high-performer share moving off 6 percent. If individual productivity stays high and enterprise profit attribution stays flat through 2027, the “useful but not worth the capex” case survives.

What would confirm the strong thesis instead: a NVIDIA guide-down, a hyperscaler capex cut of the sort that implies cancelled orders rather than lease reclassification, a failed OpenAI round, an Anthropic listing pulled or priced as a disappointment, and a CoreWeave-style credit event that forces Microsoft, Amazon, or Google to take visible losses on guarantees. One neocloud bond at 15 percent is not that event. A cluster of them, plus an Oracle contract rewrite that hits hyperscaler earnings, would be.

How strong the case is

Split the thesis, because the blended version is what makes him sound either prophetic or unserious.

The narrow case is moderately strong, about 6.5 out of 10. ARR is a sloppy metric. Lab GAAP revenue is far below run rate. Anthropic’s $518 billion of largely non-cancelable commitments against $4.6 billion of 2025 revenue is an extraordinary mismatch. Goldman’s ~$230 billion annual revenue gap is a mainstream version of his arithmetic. Enterprise EBIT attribution is flat. Credit stress is real at CoreWeave and Oracle. Vendor financing and guarantees make the demand partly circular. Those points do not require believing the revenue is fraudulent.

The strong case is weak, about 3 out of 10. The bubble did not pop by the second quarter of 2026. Models kept improving. Lab revenue and cloud revenue both accelerated. NVIDIA is supply-constrained at 75 percent margins with a $279 billion commitment book. Microsoft cut traditional debt while printing $155 billion of operating income. OpenAI raised $122 billion in March and is in talks for more at a higher valuation. The “money runs out” mechanism does not survive contact with AA and AAA balance sheets that can choose to spend. Collapsing those facts into an imminent 90 percent NVIDIA drawdown requires the same style of ceiling call that has already failed.

Overall, as a forecast of systemic unwind in the next year, the case is about 4 out of 10. As a standing warning that returns are late, commitments are too rigid, and the weakest financed intermediaries can break without the technology being fake, it is the part worth keeping. The error is treating a timing and capital-structure problem as if missed deadlines and fortress balance sheets have not already changed the odds.

---

Recent Findings Supplement (October 2026)

OpenAI and Anthropic have demonstrated accelerating revenue run rates and enterprise traction in Q3 2026 that exceed many prior projections, directly challenging narratives of unsustainable cash burn without corresponding monetization.[1][2]

• OpenAI reported an annualized revenue run rate approaching $50 billion at the end of September 2026 (net of partner shares, per investor documents reviewed by FT/Bloomberg), up from roughly $40 billion in August and more than double the ~$20 billion figure at the end of 2025. Enterprise revenue run rate grew 107% in Q3, with overall run rate growth at 77%; consumer additions in Q3 alone exceeded all of 2025. The company projects $70 billion or more annualized by year-end 2026, driven primarily by enterprise.[3][4]
• Anthropic posted Q3 2026 revenue of $16.8 billion (up 45% sequentially from $11.6 billion in Q2 and $4.7 billion in Q1), achieving adjusted profitability in the quarter. Its annualized run rate had already surpassed $65 billion by late July 2026 (from ~$47 billion in May and ~$9 billion at end-2025), with 2025 full-year revenue at ~$4.6 billion (12x growth from prior year). Analyst estimates from investor meetings point to a potential $110 billion run rate by year-end.[2][5]
• These figures reflect real paid usage growth (enterprise subscriptions, coding tools, API) rather than solely hype-driven projections, though annualized run rates are non-GAAP and exclude full cost structures.

This traction implies that demand-side monetization is scaling faster than some skeptics anticipated, reducing the immediacy of a revenue cliff even amid high inference and training costs.[6]

NVIDIA’s results and hyperscaler capex guidance confirm sustained, broad-based GPU and infrastructure demand into late 2026, with diversification beyond any single lab.[7]

• NVIDIA reported fiscal 2026 revenue of $215 billion (up 65% YoY), with data center at $193.7 billion; Q2 fiscal 2027 data center revenue reached $89 billion (up 117% YoY) on Blackwell ramp, hyperscale strength, and expanding customers (AI labs, enterprises, sovereigns). CFO highlighted “fastest product ramp” expectations for next-generation Vera Rubin.[8][9]
• Outside China, demand remains robust; in China, NVIDIA’s data-center AI accelerator share fell to 34% (shipments down 44% YoY to 678k units) as Huawei rose to 28%, but total Chinese accelerator market grew and non-GPU accelerators surged.[10]
• Hyperscalers (Amazon, Microsoft, Alphabet, Meta, Oracle) are guiding combined 2026 capex in the $600–820 billion range (potentially >$1 trillion in 2027), with AI infrastructure as the driver; Amazon raised its 2026 capex outlook to ~$220 billion.[11][12]

Persistent hardware demand and diversified buyers indicate that GPU utilization and buildout are not stalling, countering claims of imminent overcapacity or demand collapse.[8]

Hyperscalers and AI labs are using a mix of public bonds, private credit, SPVs, leases, and vendor backstops that distribute risk and leverage their scale/cash flows, insulating core operations from immediate distress.[13]

• AI-related corporate debt issuance reached ~$450–570 billion in 2026 (multiple estimates), with hyperscalers accounting for a large share (e.g., Amazon’s $54 billion bond sale); tech IG bond issuance rose to ~8% of total U.S. IG. Aggregate lease/commitment liabilities exceed $1.5 trillion (much off-balance-sheet).[12][14]
• Structures include SPVs/joint ventures (e.g., Meta data center with private credit), sale-leasebacks, and vendor financing (NVIDIA backstopping >$100 billion for an OpenAI Ohio project; Broadcom packages for Anthropic/OpenAI). Credit metrics are weakening but protected by scale, dominance, and cash generation; analysts note downside risk is limited.[13][15]
• OpenAI closed a $122 billion round in March 2026 at $852 billion post-money (Amazon, NVIDIA, SoftBank, MGX among backers) and is now in talks for $30 billion+ at ~$1.4 trillion pre-money.[1]

These mechanisms allow continued spending without immediate balance-sheet collapse for the largest players, shifting some risk to specialized financiers while preserving flexibility.[16]

Analyses from September–October 2026 document a pattern of high-confidence predictions by Zitron on AI timelines, company viability, and market outcomes that did not materialize, alongside questions about methodology and selective framing.[17]

• Documented misses include forecasts of Meta as a “dying company” (actual revenue/profit growth through 2025–H1 2026), AI bubble popping “no later than Q2 2026,” models hitting a permanent capability wall by late 2025, and collapses at specific firms (e.g., Cohere, Cursor). OpenAI and Anthropic revenue exceeded or approached targets he questioned.[18][18]
• Critiques highlight spreadsheet errors in revenue estimates (e.g., Anthropic), treatment of temporary constraints (data, scaling) as permanent, and a rhetorical style of near-certainty on falsifiable claims. Outlets and analysts (e.g., Dan Luu compilations) note audience incentives for contrarian positioning via newsletter/podcast.[17][19]

Zitron continues to highlight valid concerns around un-audited ARR figures, warehouse GPU inventory, and financing costs (e.g., October 2026 commentary on private credit as a “brewing crisis”), but the track record of mistimed calls weakens the predictive power of his broader thesis.[20]

Both labs are successfully raising or preparing to raise capital at escalating valuations, with Anthropic advancing toward a potential >$2 trillion IPO (targeted November 2026 or later) and OpenAI pursuing a large private bridge round ahead of a 2027 public listing.[21]

• Anthropic’s pre-IPO investor meetings highlighted Q3 results and adjusted profitability; OpenAI’s latest round involves UAE funds (e.g., MGX) and strategic partners.[2][1]
• These outcomes show continued access to private (and soon public) capital despite losses and commitments ($518 billion+ for Anthropic infrastructure; similar scale for OpenAI).

Successful execution here would further insulate the ecosystem by providing equity-like funding and liquidity, testing whether revenue growth can eventually support the capital structure.[22]

Indicators that would falsify or significantly weaken Zitron’s core view (unsustainable economics, overhyped demand, inevitable reckoning) in the next 6–12 months include: sustained or accelerating revenue growth into 2027 with improving unit economics; hyperscaler capex delivering measurable ROI or utilization without major impairments; successful high-valuation IPOs or further raises without dilution spirals; and NVIDIA maintaining or growing data-center revenue amid Blackwell/Rubin ramps.[23]

Conversely, sharp revenue deceleration, widespread project delays/cancellations, or credit events in private financing vehicles would strengthen it.[24]

Overall, Zitron’s case on financing risks, opacity in metrics, and long-term ROI uncertainty retains substantive elements and aligns with some analyst caution, but recent revenue traction, demand persistence, and capital access indicate it is overstated in its immediacy and breadth as of October 2026; the evidence tilts toward continued (if volatile) expansion rather than near-term collapse.[6]

New data since mid-2026 has consistently shown faster monetization and investor willingness to fund the buildout than the most pessimistic timelines projected.