ED ZITRON: Going public will not solve OpenAI's and Anthropic's funding needs, because without cash flow or assets they would likely be rated deep junk

The Gist

Ed Zitron argues that listing on the stock market will not fix OpenAI's and Anthropic's money problems. They lose money, own almost no physical assets, and owe hundreds of billions in compute contracts, so credit raters would likely call their debt junk. Floating-rate loans could become crushing if rates rise, as happened to other companies. This packet steelmans the arguments Ed Zitron makes in this episode, in their best light, for analysis; it is not an endorsement of his conclusions.

Conclusion

A stock market listing will not by itself solve the funding needs of OpenAI and Anthropic, because both have negative cash flow and operating earnings, own few assets that could serve as collateral, depend on a few products and customers, and carry hundreds of billions of dollars of non-cancelable commitments that creditors treat like debt; on standard credit criteria they would likely be rated well into junk, which limits how much they can borrow each year and at what cost, and floating-rate leveraged loans would add a rate risk that has already overwhelmed other leveraged software firms.

Premises

1. Anthropic reported a 2025 operating loss of about $8 billion, OpenAI has reported large operating losses and cash burn in 2026, and neither expects sustained profitability soon.
2. Neither firm owns its data centers or chips; both lease compute, so they have little hard collateral apart from intellectual property.
3. Anthropic has disclosed about $413 billion of non-cancelable compute obligations over roughly a decade, and creditors treat such take-or-pay obligations as debt equivalents.
4. Credit ratings rely heavily on earnings before interest, taxes, depreciation, and amortization and on cash flow, so firms with negative operating earnings rarely receive ratings above the single-B or CCC range; by contrast, SpaceX obtained investment-grade ratings in June 2026 on the strength of its profitable Starlink business, and CoreWeave, which has positive EBITDA, is rated B+ and Ba3.
5. Rating agencies also penalize concentration, and nearly a quarter of Anthropic's 2025 revenue came from two customers.
6. The labs are expected to need $50 billion to $100 billion of outside funding a year, and once public they cannot rely on private venture rounds at their recent scale.
7. Selling new shares is limited by the size of the public float and by price pressure, and Anthropic's $15 billion revolving credit line is a contingency facility rather than a source of annual funding.
8. Leveraged loans float over a base rate, and when rates rose from 2022 onward borrowers such as Pluralsight and Medallia saw interest costs surge and were handed to their private credit lenders.
9. A listing that draws too little demand at its target valuation may be withdrawn, and a withdrawn listing is hard to revive.

Assumptions

• OpenAI's reported first-quarter 2026 cash burn was about $3.7 billion, with an operating loss near $9.3 billion; the $12 billion figure cited in the episode matches a non-cash accounting charge rather than compute spending.
• Default means a missed payment, not negative cash flow; the labs hold large cash balances (OpenAI reported more than $73 billion at the end of March 2026), and the CCC category sits above CC, C, and default on standard rating scales.
• Bankers for both labs are reported to be lobbying for investment-grade ratings after their listings, so the rating outcome is uncertain.
• The labs already raise large sums through partner-backed structures, such as Google's backstop of leases underpinning a $35 billion chip financing for Anthropic and Broadcom's backstop of up to $42 billion of loans for Anthropic's chip leases, which eases the constraint on their own credit while shifting risk to the backstopping firms.
• The $50 billion to $100 billion annual need and the estimate that public share sales could raise about $20 billion a year are the speaker's own estimates.
• Figures from OpenAI's 2025 audited results that rely on the speaker's own reporting were not independently verified.

Analysis

Argument type: Inductive.

Potential Fallacies

• Conflating price with availability (Conclusion; P4) — A junk rating raises borrowing costs and limits some buyers, but high-yield and private-credit markets routinely lend large sums to junk-rated issuers, CoreWeave being one example. The step from 'junk-rated' to 'limited in how much they can borrow' is asserted rather than shown, and it needs an analysis of market capacity against the stated need.
• Weak analogy / thin reference class (P4; P8) — Pluralsight and Medallia were sponsor-backed leveraged-buyout software firms with thin cash cushions. CoreWeave owns hard assets and SpaceX has a profitable segment. These comparisons illustrate mechanisms but do not predict how the labs, with large cash balances and strategic backers, would fare. The failures are also cited without a base rate of floating-rate borrowers that did not fail, and the labs are not shown to be planning leveraged loans.
• Overconfident forecast / hasty generalization (Title; P4; Conclusion) — A general tendency (negative-EBITDA firms 'rarely' rate above single-B or CCC) is applied to two unrated firms. Agencies also weigh liquidity, growth, and partner support, and the assumptions concede the outcome is uncertain. 'Deep junk' in the title goes beyond the body, which supports at most single-B to CCC, and A2 notes that CCC is above default.
• Reliance on unverified single-source estimates (P6; P7) — The funding gap that drives the conclusion depends on the $50–100B need and ~$20B equity capacity, which are the speaker's own estimates with no shown method. They also treat the need as fixed, although it can flex with funding availability and with compute contract terms.
• Incomplete dichotomy of funding sources (Conclusion; P6–P7) — The analysis centers on IPO equity plus corporate credit. Convertibles, follow-on offerings, strategic and vendor investment, and SPV or partner-guaranteed financing, partly acknowledged in A4, are the channels most relevant to the current AI funding environment and are not integrated into the conclusion.
• Weak-version framing (straw man) (Conclusion) — Rebutting the claim that a listing alone solves funding targets a position few hold. The 'by itself' qualifier makes the thesis safe but leaves open the more useful question of how much a listing helps within a mixed capital structure.
• Generalizing from one lab to both (P3; P5; Conclusion) — The concentration and the $413B commitment figures are disclosed for Anthropic. The conclusion extends them to OpenAI without equivalent evidence.
• Compounding uncertain steps (Overall structure) — The chain runs from negative EBITDA to a junk rating to constrained borrowing to an unsolved funding need. Each step is probabilistic, so the joint likelihood is lower than the confident framing suggests. The losses, the commitments, and the rating heuristic also feed the same leverage metrics and should not be counted as fully independent support.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Rating level is not the binding constraint. Loss-making growth companies such as Amazon, Tesla, Uber, and Netflix funded tens of billions through equity, convertibles, and high-yield debt. Public equity investors price growth and strategic position rather than EBITDA, and a liquid listed stock also enables follow-ons, stock-based deals, and at-the-market issuance.
• Conclusion; Premise 4 (High impact) — A junk rating is a pricing and capacity constraint, not a bar to borrowing. High-yield and private-credit markets lend heavily to B and CCC issuers, and CoreWeave has raised large sums at B+/Ba3. The argument shows higher cost but does not show that annual needs cannot be met.
• Premise 6; Assumption 4 (High impact) — Partner-backed and off-balance-sheet structures (Google and Broadcom backstops, vendor financing, SPVs) already supply large sums outside the labs' own credit, and suppliers who need the labs alive have incentives to keep supporting them. The standalone-credit gap may therefore be irrelevant to real funding capacity, though risk concentrates in the guarantors.
• Premise 4 (Medium impact) — Agencies are forward-looking and weigh liquidity, growth, contracted revenue, and partner support. Large post-IPO cash piles could support ratings above the CCC range. Bankers are lobbying for investment grade, and agency criteria may adapt for hypergrowth firms.
• Premise 3 (Medium impact) — Agencies typically adjust for leases and some unconditional purchase obligations only partially, depending on enforceability, delivery conditions, and revenue matching. Treating the full $413B as debt overstates leverage, and suppliers can renegotiate or resell capacity.
• Premise 6 (Medium impact) — The $50–100B need is not fixed. Spending is discretionary and can be staged, renegotiated, or cut in response to funding conditions, and revenue growth could bring operating earnings positive sooner than a static snapshot implies.
• Premise 8 (Medium impact) — The labs are equity-funded with large cash balances and can use fixed-rate bonds, converts, or hedging. The Pluralsight and Medallia analogy applies only to a financing choice the labs are not shown to make.

Suggested Improvements

• Quantify the funding gap — Derive the annual need and the equity and debt capacity from stated methods, show ranges and sensitivity, and net out partner-backed financing and existing cash. The conclusion depends on a gap that currently rests on single-source point estimates. Showing that the gap persists across plausible ranges would make it robust.
• Separate price from availability — Compare the labs' annual need with realistic single-issuer capacity in high-yield, private-credit, and convertible markets at B/CCC pricing, and state the cost at which the need is met. This turns an ambiguous 'limits borrowing' claim into a testable one and answers the strongest objection.
• Ground the rating forecast — Cite published agency methodologies on take-or-pay and unconditional purchase obligations, and loss-making rated precedents with large cash and sponsor support. State the rating range with conditions and drop 'deep junk' unless supported. It replaces a heuristic with documented practice and aligns the headline with the body, which supports single-B to CCC at most.
• Engage the mixed-financing view — Address converts, follow-ons, strategic equity, and backstopped asset-level financing directly and argue why they cannot close the gap, or what they would cost in dilution and counterparty concentration. It answers the steel-man objection that an IPO is one component of a stack, and it makes use of the argument's own concession on backstops.
• Use better comparators and base rates — Add loss-making growth issuers that raised heavily, such as Amazon, Tesla, and Uber, alongside the failures, and a base rate for floating-rate borrower distress. Use evidence on both labs, not Anthropic alone. It reduces cherry-picking concerns and clarifies when the Pluralsight and Medallia mechanism applies.
• Add the time dimension — Model ratings and funding capacity 12–36 months after listing under alternative revenue and margin paths and interest-rate regimes. Ratings are forward-looking, and the labs are fast-growing, so a static snapshot may overstate the constraint.

Scenario Tests

• Both labs list and receive single-B or CCC-range ratings after listing, with borrowing costs well above investment-grade peers. (Supports) — This confirms the rating claim and the cost penalty. It still does not by itself show that annual needs go unmet, because junk-rated issuers can raise sizeable sums at higher prices.
• A lab lists successfully at a high valuation and raises large follow-on equity and convertibles, supplemented by hyperscaler backstops and strategic investment. (Challenges) — Funding needs are met through a layered mix with the rating largely irrelevant, undermining the 'limits how much they can borrow' framing. The weak form (an IPO alone is not enough) still holds.
• Revenue growth outpaces compute costs and operating earnings turn positive within a year or two of listing. (Challenges) — The core premise linking negative earnings to junk ratings weakens, and agencies could reach better outcomes, possibly with improved ratings.
• A listing draws weak demand at the target valuation, is pulled, and later attempts struggle, in a period of tight credit markets. (Supports) — This illustrates the withdrawal and float constraints, though failure would stem partly from market conditions rather than rating mechanics.
• Agencies give only partial debt-equivalent credit to take-or-pay compute obligations and weigh the labs' large cash balances and partner support. (Challenges) — Leverage metrics come out lower than implied, ratings are less severe, and the case for 'well into junk' is weakened.
• A labs financing relies on floating-rate leveraged loans during a renewed rate rise, with limited hedging. (Supports) — The Pluralsight and Medallia mechanism applies, since interest costs would swell against negative operating earnings. This is conditional on the labs choosing that structure.

Coherence & Relevance

The argument is a convergent, inductive case. Several largely independent considerations (earnings, collateral, debt-like commitments, equity limits, rate risk) point toward a constrained and costly funding path. The hedge 'by itself' makes the core claim defensible. Coherence weakens where the argument moves from sub-investment-grade ratings to a quantity ceiling on borrowing, and where it relies on a self-estimated funding gap while the conceded assumptions (cash balances, partner backstops, uncertain ratings) are not integrated into the conclusion. The strongest defensible version is that an IPO is one limited source in a financing mix that will likely be costly and dependent on partner support, not that the labs cannot fund themselves.

• Anthropic reported a 2025 operating loss of about $8 billion, OpenAI has reported large operating losses and cash burn in 2026, and neither expects sustained profitability soon. (Strong) — Establishes weak credit metrics but not funding failure. Large cash balances and the burn correction mean it bears on runway and cost of capital more than on near-term inability to fund.
• Neither firm owns its data centers or chips; both lease compute, so they have little hard collateral apart from intellectual property. (Moderate) — Matters for secured and asset-backed lending but less for equity or unsecured funding, and partner backstops substitute for the labs' own collateral.
• Anthropic has disclosed about $413 billion of non-cancelable compute obligations over roughly a decade, and creditors treat such take-or-pay obligations as debt equivalents. (Strong) — Debt-equivalent treatment is asserted without a methodology, is only partly applicable, and the evidence concerns Anthropic rather than both labs.
• Credit ratings rely heavily on earnings before interest, taxes, depreciation, and amortization and on cash flow, so firms with negative operating earnings rarely receive ratings above the single-B or CCC range; by contrast, SpaceX obtained investment-grade ratings in June 2026 on the strength of its profitable Starlink business, and CoreWeave, which has positive EBITDA, is rated B+ and Ba3. (Strong) — The comparators are imperfect and bracket rather than pinpoint the outcome. The step from a rating to a funding ceiling is not demonstrated, and agency flexibility for strategic and liquid…
• Rating agencies also penalize concentration, and nearly a quarter of Anthropic's 2025 revenue came from two customers. (Weak) — A minor, moderate-level rating factor, with no quantification of the penalty and no equivalent evidence for OpenAI.
• The labs are expected to need $50 billion to $100 billion of outside funding a year, and once public they cannot rely on private venture rounds at their recent scale. (Strong) — Essential to the claim that funding remains unsolved, but the figures are single-source estimates, treated as fixed, and the argument omits other public-company funding channels.
• Selling new shares is limited by the size of the public float and by price pressure, and Anthropic's $15 billion revolving credit line is a contingency facility rather than a source of annual funding. (Strong) — The equity ceiling is an unshown estimate and ignores follow-ons and convertibles. The revolver point is a correct but limited clarification.
• Leveraged loans float over a base rate, and when rates rose from 2022 onward borrowers such as Pluralsight and Medallia saw interest costs surge and were handed to their private credit lenders. (Weak) — Conditional on the labs borrowing through unhedged floating-rate loans, which is not shown, and the precedents involve different capital structures.
• A listing that draws too little demand at its target valuation may be withdrawn, and a withdrawn listing is hard to revive. (Weak) — A generic possibility with no evidence of weak demand. It bears on whether a listing happens, not on whether a successful one solves funding.

View this argument on LogicFirst.ai