ED ZITRON: OpenAI and Anthropic face a 2027 to 2028 wall of take-or-pay compute bills they cannot cover from revenue, venture capital, or a strained bond market

The Gist

OpenAI and Anthropic have promised to pay hundreds of billions of dollars for computing power, and most of those bills can't be canceled. The payments get much bigger in 2027 and 2028, like a teaser-rate mortgage resetting. Both companies lose money, so they would need endless new investment or borrowing to pay. This packet steelmans the arguments made by Ed Zitron and Dean Baker in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

OpenAI and Anthropic have signed largely non-cancelable compute commitments whose payments ramp up sharply in 2027 and 2028, much as adjustable-rate mortgages reset in 2007, and neither their current revenue, nor venture capital, nor a strained high-yield bond market offers a credible source for the hundreds of billions of dollars a year those commitments will require.

Premises

1. Anthropic's confidential prospectus, as reported by Reuters, discloses at least $518 billion of compute commitments over about a decade, of which roughly 80 percent, about $413 billion, is non-cancelable or payable regardless of usage.
2. OpenAI's planned compute spending through 2030 has been reported at about $750 billion, and its broader vendor commitments are larger.
3. Many compute contracts require little payment up front and ramp up as capacity is delivered, so the heaviest payments fall in 2027 and 2028; UBS estimates the two labs will account for over $124 billion of Google Cloud revenue alone in 2027.
4. Both companies lose money; Anthropic reportedly had $4.6 billion of revenue and a $42 billion net loss in 2025.
5. Their recent rounds were funded mainly by strategic investors, hyperscalers, and large asset managers, not by venture capital alone, and repeating rounds of $65 billion to $122 billion every year is a demanding assumption.
6. Borrowing the remainder would mean issuing tens of billions of dollars a year of likely sub-investment-grade debt into a market where high-yield spreads have widened and long-term Treasury yields are near their highest in two decades.
7. Adjusted profitability measures exclude the very compute costs at issue, so the binding test is cash flow, and the required annual cash outflows would rival the total annual costs of Microsoft.

Assumptions

• The commitments are spread over seven to ten years and some capacity deals, such as Anthropic's xAI agreements, are cancelable on notice, so the totals are not due at once.
• Microsoft's fiscal 2026 cost of revenue plus operating expenses was about $177 billion, higher than the figure cited in the conversation, which does not change the point of the comparison.
• The combined commitment total varies by source and time horizon; the $1.3 trillion figure is one estimate among several.
• Revenue at both labs has grown fast, and OpenAI reported 77 percent run-rate growth in its third quarter of 2026, so the gap could narrow if that pace continues.
• The exact venture capital share of OpenAI's $122 billion round was not verified, although Amazon, Nvidia, and SoftBank committed about $110 billion of it.

Analysis

Argument type: Inductive.

Potential Fallacies

• Overreach from 'strained' to 'impossible' (Conclusion versus P5 and P6) — The premises show that each funding channel is difficult, costly, or demanding. The conclusion says no credible source exists. Difficulty does not establish insufficiency without a quantitative comparison of annual needs and annual capacity.
• Incomplete enumeration (false trichotomy) (Conclusion; P5; P6) — The conclusion treats revenue, venture capital, and high-yield bonds as the only funding sources. P5 itself shows that strategic investors and hyperscalers supply most of the money. Omitted are IPOs, private credit, asset-backed and project financing, vendor financing, sovereign capital, equity-for-compute swaps, and contract renegotiation. A universal claim of 'no credible source' needs only one credible counterexample to fail.
• Weak analogy (2007 adjustable-rate mortgage resets) (Conclusion and title) — The analogy fits the mechanism of a scheduled payment step-up. It fits poorly on who is exposed. Mortgage borrowers were millions of dispersed households with no bargaining power. Here there are two growing firms and a few large counterparties who are also investors, who can restructure, and whose own balance sheets suffer if the customer fails. The analogy carries an implication of inevitable collapse that the evidence hasn't shown.
• Scope shift from multi-year totals to annual outflows (P1, P2 to conclusion) — P1 and P2 give multi-year totals and planned spending, while the conclusion speaks of 'hundreds of billions a year.' Annual peak obligations are never derived. Planned spending, total commitments, and non-cancelable obligations are different quantities and are used interchangeably.
• Static or stale-baseline extrapolation (P4, P6 to conclusion) — 2025 revenue and losses, and today's credit conditions, are used to judge capacity to pay in 2027 and 2028. The argument concedes 77 percent run-rate growth but never models a revenue path or states the growth rate that would be needed to close the gap. Credit conditions are cyclical and may differ in 2027 and 2028.
• Inferring cash burn from net loss (P4 and P7) — A $42 billion net loss on $4.6 billion of revenue may include non-cash items such as equity compensation, derivative or valuation remeasurements, and accounting for investor instruments. The argument insists that cash flow is the proper test but uses a net-loss figure to support it.
• Misleading scale comparison (P7 and A2) — Comparing annual outflows with Microsoft's total cost base conveys magnitude but says nothing about fundability. Microsoft's costs are covered by a very large revenue base, and the relevant measure is net cash flow after revenue. The comparison also depends on an annualized payment schedule that has not been provided.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Hyperscalers, Nvidia, and other counterparties are simultaneously suppliers, investors, and customers' creditors. A default would damage their own backlog and equity stakes, so they have strong incentives to defer, restructure, convert obligations to equity, or provide vendor financing. The likelier outcome is renegotiation, dilution, or a valuation reset, not a forced-default cascade.
• Premise 4 (High impact) — Revenue is growing at rates such as 77 percent run-rate growth. Compounding from a small base, even a decaying growth rate could narrow the gap substantially by 2027 and 2028. Trailing 2025 revenue and losses are a stale baseline for a forward-looking solvency claim.
• Conclusion (High impact) — The list of funding sources is incomplete. IPOs and public equity, sovereign wealth funds, private credit, GPU-backed and data-center project financing, SPVs, and hyperscaler balance sheets have all been used in the AI buildout. A claim that no credible source exists fails if even one of these can supply the needed capital.
• Premise 1 (Medium impact) — Reported commitment totals are maximum contractual exposure over about a decade and can include minimum-spend tiers and options. Averaged, $413 billion is about $41 billion a year, which is large but not obviously unfundable for fast-growing firms, and unused capacity may be resellable.
• Conclusion (Medium impact) — The 2007 analogy overstates the likelihood of collapse. Mortgage borrowers were dispersed, powerless, and holding falling collateral, whereas here a few sophisticated, mutually dependent parties negotiate bilaterally, and the scarce compute capacity retains value. Telecom and fiber overbuilds in 1999 to 2002 are a closer precedent and ended in renegotiation, bankruptcies, and write-downs rather than a single systemic collapse.
• Premise 6 (Medium impact) — Credit conditions are cyclical. Spreads and long yields in the current period are a weak predictor of conditions over the next 12 to 24 months, and AI issuance is only one factor in current market strain.
• Premise 4 (Medium impact) — Much of the training compute is discretionary. Both labs can slow new commitments, stage spending against revenue milestones, or cut training spend, which would reduce the real peak burden below headline totals.

Suggested Improvements

• Quantify annual obligations — Build a year-by-year schedule of minimum non-cancelable payments for each lab for 2026 to 2030, separating firm take-or-pay amounts from cancelable and contingent ones, and show how 'hundreds of billions a year' is derived. The central quantitative claim is currently asserted from multi-year totals. A schedule would let readers test the size of the wall and its timing.
• Model revenue and funding scenarios — Project 2027 and 2028 revenue and cash flow under several growth paths (for example, sustained 77 percent growth, decay to 30 percent, flat) and state the breakeven growth rate at which the gap closes. Compare the resulting gap with plausible capacity from each funding source. The argument admits fast growth but never tests whether it could close the gap. Scenario analysis would turn a rhetorical claim into a testable one.
• Address counterparty incentives and renegotiation — Explain why creditors who are also investors would enforce rather than restructure, or concede that restructuring is the likely path and argue the case on dilution, valuation, and supplier losses. This is the strongest counterargument. Engaging it directly would sharpen the thesis and make the 2007 analogy defensible or unnecessary.
• Broaden the funding survey — Evaluate IPO, private credit, asset-backed and project financing, vendor financing, and sovereign capital with estimates of capacity and cost under stress. A universal claim of 'no credible source' requires that the alternatives be assessed and not omitted.
• Separate cash burn from accounting loss — Use cash-flow statement data in place of net loss, and identify non-cash components of the $42 billion figure. The argument's own test is cash flow, so the evidence should be cash flow.
• Calibrate the conclusion — Reframe the claim as a material risk of a funding shortfall, forced restructuring, or dilution in 2027 and 2028, with explicit conditions under which the thesis would be falsified. The premises support elevated risk, not certainty. A calibrated and falsifiable claim is more defensible and more useful to decision-makers.
• Use primary sources and fix comparators — Cite the primary documents (the prospectus, the UBS note) and replace the Microsoft cost comparison with a net-of-revenue measure such as peak-year cash need against projected operating cash flow plus committed financing. This reduces reliance on secondhand reporting and replaces a scale-shock device with a feasibility measure.

Scenario Tests

• Revenue growth continues near current rates (for example, 60 to 80 percent a year) through 2027 and 2028 as enterprise and coding demand scales into the commitments. (Challenges) — The funding gap narrows substantially and the commitments become a demanding but serviceable fixed cost. The 'cannot cover' thesis would be falsified, though the cash-flow framing would still be useful for monitoring.
• Hyperscaler and chip counterparties renegotiate, defer, or convert take-or-pay obligations into equity to protect their own investments. (Challenges) — The 'non-cancelable' premise loses practical force. The failure mode becomes dilution, a valuation reset, or consolidation, not a hard default, and the 2007 analogy weakens.
• Revenue growth decelerates sharply to well below the pace needed, while equity markets and credit markets stay tight and strategic investors become reluctant to repeat mega-rounds. (Supports) — The cash gap would be real and large in peak years. Distressed raises, down rounds, restructured contracts, or supplier write-downs become likely, validating the core risk warning.
• An IPO window opens or high-yield spreads tighten, or private-credit and asset-backed structures scale up to finance data-center capacity. (Challenges) — P5 and P6 fail as descriptions of the 2027 and 2028 financing environment, and the claim that no credible source exists is refuted.
• Data-center and power delivery delays push capacity ramps, and therefore payments, from 2027 and 2028 into later years. (Challenges) — The timing of the 'wall' shifts, and the cliff view becomes a gentler slope, though the total obligation remains.
• One lab sustains stronger revenue growth while the other falls behind, and the weaker lab's commitments are restructured or acquired by a hyperscaler. (Neutral) — The thesis is partially vindicated for one lab and not the other. This suggests the 'both labs' framing is too symmetrical and that outcomes will be differentiated.

Coherence & Relevance

The argument has a clear cumulative structure: obligations are large and back-loaded, the firms are loss-making, and each funding channel is strained. It correctly identifies cash flow and payment timing as the real test. The premises collectively support a conclusion of material liquidity and refinancing risk. They do not support the categorical conclusion that the bills cannot be covered. There is no quantitative bridge between multi-year totals and 'hundreds of billions a year', and no comparison of annual obligations with projected revenue and financing capacity. The author's assumptions about spread-out and cancelable commitments, varying totals, and rapid revenue growth are not integrated into the conclusion. The premises are broadly coherent with each other but fall short of the strength of the conclusion.

• Anthropic's confidential prospectus, as reported by Reuters, discloses at least $518 billion of compute commitments over about a decade, of which roughly 80 percent, about $413 billion, is non-cancelable or payable regardless of usage. (Strong) — Establishes the size of binding obligations for one lab, but not annual timing or ability to renegotiate, and does not extend automatically to OpenAI.
• OpenAI's planned compute spending through 2030 has been reported at about $750 billion, and its broader vendor commitments are larger. (Moderate) — Planned spending is not binding commitment, no figure or source is given for broader commitments, and there is no non-cancelable share for OpenAI.
• Many compute contracts require little payment up front and ramp up as capacity is delivered, so the heaviest payments fall in 2027 and 2028; UBS estimates the two labs will account for over $124 billion of Google Cloud revenue alone in 2027. (Strong) — Supports timing and scale, but vendor revenue forecasts are not a payment schedule, and delivery slippage could move the peak. A1 spreads totals over 7 to 10 years, which fits a ramp more than a sharp wall.
• Both companies lose money; Anthropic reportedly had $4.6 billion of revenue and a $42 billion net loss in 2025. (Moderate) — Trailing and possibly including non-cash items. It does not model 2027 and 2028 revenue despite the acknowledged growth in A4.
• Their recent rounds were funded mainly by strategic investors, hyperscalers, and large asset managers, not by venture capital alone, and repeating rounds of $65 billion to $122 billion every year is a demanding assumption. (Moderate) — 'Demanding' is not shown to mean 'unavailable'. Because strategics are funding their own customers, this premise partly undercuts the 'venture capital' framing in the conclusion, and it ignores equity from other sources.
• Borrowing the remainder would mean issuing tens of billions of dollars a year of likely sub-investment-grade debt into a market where high-yield spreads have widened and long-term Treasury yields are near their highest in two decades. (Moderate) — Uses current market conditions to forecast 2027 and 2028, assumes a rating, and omits secured, guaranteed, private-credit, and vendor-financed structures. Also, 'tens of billions a year' is smaller than the 'hundreds of billions a year' in the conclusion.
• Adjusted profitability measures exclude the very compute costs at issue, so the binding test is cash flow, and the required annual cash outflows would rival the total annual costs of Microsoft. (Moderate) — The cash-flow test is relevant and sound. The Microsoft comparison conveys scale but not fundability, depends on an underived annual figure, and does not net against revenue.

View this argument on LogicFirst.ai