TrueSeeker Finance · Verified claim report Case fcac8e87ac · 2026-09-27

§ Claim under review · Fact

"OpenAI and Anthropic now owe about $2.5 trillion that doesn't show up on their balance sheets. That's roughly 7% of the entire U.S. national debt. It's also roughly as much as $GOOG, $META, $AMZN, $MSFT and $ORCL owe combined. It isn't bank debt. It's the money they've promised to pay for chips, data centers and cloud computing over the next several years. Because it isn't a loan, it doesn't show up as debt on paper, but the bills still come due... You don't need to read every AI lab's balance sheet to see where this is going. Autopilot runs the strategies that invest in AI automatically."

Circulating claim, as submitted.

Verdict

Partially accurate but misleading

Confidence

Medium
§

Summary

This post takes something real and rounds it up. OpenAI and Anthropic have signed enormous multi-year contracts to buy chips, data centre capacity and cloud compute, and those contracts are not bank loans. The widely reported figures are about $1.4 trillion for OpenAI, a number from late 2025 that the company has since reframed as roughly $600 billion to $750 billion of compute spend through 2030, and up to about $517 billion for Anthropic as reported in September 2026. Those add to roughly $1.9 trillion, and no source found states the $2.5 trillion total the post claims. Against US federal debt of about $40.07 trillion on 23 September 2026, $2.5 trillion is 6.2%, not 7%. The comparison to what Alphabet, Meta, Amazon, Microsoft and Oracle owe combined is not supported by anything located: those five issued $121 billion of bonds in all of 2025 and about $159 billion through early June 2026, which is hundreds of billions, not trillions. Also worth knowing: neither OpenAI nor Anthropic publishes audited accounts, so there is no public balance sheet for these amounts to be "off," and at least one of the deals is documented as contingent on the buyer's ongoing commercial performance rather than fixed. General information only, not financial advice.

§

The readings

key figures from the evidence
2.5 $T

claimed combined OpenAI + Anthropic commitments

1.9 $T

sum of sourceable OpenAI ($1.4T) and Anthropic ($517B) figures

6.2 %

actual share of US national debt ($40.07T, Sept 2026) that $2.5T represents

§

Why this verdict

The underlying phenomenon is real and documented: OpenAI's reported $1.4 trillion of infrastructure commitments (as of late 2025) and Anthropic's reported up-to-$517 billion of compute agreements (as of September 2026) are large multi-year purchase obligations owed largely to Nvidia, Oracle, Microsoft, Amazon and Google, and they are not bank debt. But the specific $2.5 trillion total appears in no retrieved source and sits roughly 30% above the sum of the sourceable components; the comparison to the combined borrowings of the five named companies is unsupported and points strongly the wrong way given that those five issued $121 billion of bonds in 2025 and about $159 billion through early June 2026; and the "7% of the national debt" framing is 6.2% against Treasury's $40.07 trillion as of 2026-09-23. I considered Accurate and rejected it because the headline number cannot be sourced and the Big Tech comparison fails; I considered False and rejected it because the central components are real, substantial and correctly characterised as non-loan obligations; I considered Unverified and rejected it because enough of the claim is documented to grade the framing rather than throw the whole thing out. Confidence is capped at Medium because search budget ran out before the five companies' filed debt totals or any derivation of the $2.5 trillion figure could be retrieved. ---
§

Evidence

The submitted text is an advertisement built on an X post published about two days before the verdict date. The post's wording ("$2.5 trillion of off balance sheet debt. 7% of the National Debt. Roughly as much as $GOOG $META $AMZN $MSFT and $ORCL combined") is reproduced almost verbatim in the ad.

On the components:

  • OpenAI. The widely circulated figure is $1.4 trillion of infrastructure commitments, discussed publicly by Sam Altman from around November 2025, alongside a separate $1.15 trillion estimate covering hardware and cloud agreements through 2035. In February 2026 CNBC reported that OpenAI told investors it was targeting roughly $600 billion of total compute spend by 2030, explicitly to clarify plans after the $1.4 trillion figure circulated. In July 2026 the WSJ reported OpenAI had lifted planned compute spending to about $750 billion through 2030, a projected-spend figure distinct from both the $1.4 trillion and $1.15 trillion numbers.
  • Anthropic. Anthropic's own newsroom confirms an expanded Google and Broadcom compute partnership and a November 2025 commitment to invest $50 billion in US computing infrastructure. Secondary and tertiary sources put Anthropic's aggregate compute agreements at "up to $517 billion" across roughly 14.8 gigawatts, including a $30 billion Azure purchase commitment. The $517 billion total is a press aggregation, not a company-published number.
  • Contingency. Reporting on Broadcom's securities filing states that Anthropic's ability to draw on the additional compute depends on its ongoing commercial performance. This is characteristic of capacity agreements: headline values are ceilings tied to build-out and draw-down, not fixed amortising obligations.
  • US federal debt. Federal debt passed $40 trillion on 2026-08-18 and stood at roughly $40.07 trillion on 2026-09-23 per Treasury's Debt to the Penny series (retrieved via mirrors of the Treasury record; the Treasury dataset page confirms data current through 2026-09-24).
  • Big Tech borrowing. The five named companies issued $121 billion of bonds in all of 2025 and about $159 billion through early June 2026. Goldman Sachs estimated $489 billion of total AI-related debt issuance in 2026 to date, of which about 40% came directly from hyperscalers. No retrieved source states a combined outstanding-debt figure anywhere near $2.5 trillion for those five issuers.

No retrieved source, primary or secondary, states a combined OpenAI-plus-Anthropic commitment total of $2.5 trillion. The best-documented headline components sum to roughly $1.9 trillion ($1.4tn plus $517bn), and those two numbers are measured on different bases and over different horizons.


§

Findings

✓ What's accurate 6

  • OpenAI and Anthropic have each entered very large multi-year commitments for chips, data centre capacity and cloud compute. Anthropic's own site confirms an expanded Google and Broadcom compute partnership and a $50 billion US infrastructure commitment made in November 2025.
  • The $1.4 trillion figure for OpenAI infrastructure commitments is real and was discussed publicly by the company's chief executive from around November 2025.
  • Press aggregations putting Anthropic's compute agreements at up to $517 billion, including a $30 billion Azure commitment, were published in September 2026.
  • It is correct that purchase and capacity commitments of this kind are not bank loans and are not recorded as debt on a balance sheet. For public filers they appear in contractual-obligation disclosures.
  • It is correct that the counterparties include Nvidia, Oracle, Microsoft, Amazon and Google, the same firms whose revenues and share prices are tied to AI demand. Anthropic's Google and Broadcom deals and OpenAI's Oracle, Microsoft, Amazon and Nvidia agreements are all documented.
  • US federal debt was roughly $40.07 trillion as of 2026-09-23, so a $2.5 trillion figure would be in the single-digit-percentage range of it.

≈ What's misleading 7

  • The post states $2.5 trillion as a settled amount. The best-documented public components sum to about $1.9 trillion ($1.4tn for OpenAI plus $517bn for Anthropic, as of September 2026), and no retrieved source states $2.5 trillion for these two companies. The headline number is about 30% above the sum of the figures that can actually be sourced.
  • No period is given. OpenAI's $1.4 trillion has been described as spanning roughly eight to ten years, and a related vendor-by-vendor estimate covers 2025 to 2035. "Over the next several years" compresses a decade-scale figure into a near-term liability.
  • **Omitted qualifier (contingency):** reporting on Broadcom's securities filing states Anthropic's ability to draw on additional compute depends on its ongoing commercial performance. Headline deal values of this type are ceilings tied to capacity being built and drawn, not fixed amounts certain to be paid. The post presents them as bills that simply "come due."
  • The $1.4 trillion figure dates from late 2025. CNBC reported in February 2026 that OpenAI told investors its compute target was around $600 billion by 2030, and the WSJ reported in July 2026 a planned $750 billion through 2030. Using the largest and oldest framing without the company's own subsequent figures presents a superseded number as the current one.
  • "owe... that doesn't show up on their balance sheets" treats contracted future purchases as debt. Neither company publishes audited financial statements, so there is no public balance sheet on which anything could or could not appear. The remaining disagreement is definitional: whether multi-year capacity contracts should be read as debt-equivalent obligations is a live analytical debate, not a reported fact.
  • **Percent arithmetic:** $2.5 trillion against federal debt of $40.07 trillion as of 2026-09-23 is 6.2%, not 7%. Seven percent would be $2.80 trillion. The gap is small but the number was stated as if precise.
  • **Unsupported comparison:** the claim that the total roughly equals what Alphabet, Meta, Amazon, Microsoft and Oracle owe combined is not supported by anything retrieved. The scale evidence points the other way: those five issued $121 billion of bonds in all of 2025 and about $159 billion through early June 2026, and Goldman estimated $489 billion of total AI-related debt issuance across the whole market in 2026 to date. Combined borrowings of that group are documented in the hundreds of billions, not trillions. This element of the claim appears to overstate by a large multiple, though the exact combined figure was not retrieved.

? What's uncertain 5

  • The origin of the $2.5 trillion aggregate. No source giving that total, or showing how it was built, was located.
  • The current combined balance-sheet debt of the five named companies. Only issuance flows were retrieved, not outstanding stock, so the size of the overstatement in the comparison cannot be quantified precisely.
  • Anthropic's $517 billion total is a press aggregation across multiple deals. No company-published consolidated commitment figure was found.
  • Which OpenAI figure is currently authoritative. The $1.4 trillion commitments number, the $600 billion-to-2030 investor figure and the $750 billion-to-2030 planned-spend figure measure different things over different horizons, and reconciling them from public reporting was not possible.
  • How much of either company's commitments is cancellable, contingent on milestones, or conditioned on commercial performance. Only one such condition, on the Broadcom side of the Anthropic deal, was documented.
Distortion flags exaggeration omitted qualifier date context mismatch definitional dispute
§

Sources

11 of 11 linked to records
[1]

Anthropic newsroom, "Anthropic expands Google and Broadcom compute deal"

primary company investor/press record
https://www.anthropic.com/news/google-broadcom-partnership-compute ↗
[2]

US Treasury Fiscal Data, Debt to the Penny dataset page (series confirmed current through 09/24/2026; dataset landing page retrieved, not the value row)

primary official statistical record
https://fiscaldata.treasury.gov/datasets/debt-to-the-penny/ ↗
[3]

CNBC, "OpenAI resets spending expectations, tells investors compute target is around $600 billion by 2030" (2026-02-20)

secondary press of record
https://www.cnbc.com/2026/02/20/openai-resets-spend-expectations-targets-around-600-billion-by-2030.html ↗
[4]

Reuters/Yahoo Finance, "OpenAI lifts planned compute spending to $750 billion through 2030: WSJ" (2026-07-22)

secondary wire/press of record relaying WSJ
https://finance.yahoo.com/technology/ai/articles/openai-lifts-planned-compute-spending-144917731.html ↗
[5]

Yahoo Finance, "Anthropic announces deal with Google, Broadcom, says revenue has tripled" (2026-04-07), reporting Broadcom's securities filing language on contingency

secondary press of record
https://finance.yahoo.com/sectors/technology/articles/anthropic-google-broadcom-tpu-deal-113234906.html ↗
[6]

CNBC, "Google and Anthropic announce cloud deal worth tens of billions of dollars" (2025-10-23)

secondary press of record
https://www.cnbc.com/2025/10/23/anthropic-google-cloud-deal-tpu.html ↗
[7]

Yahoo Finance / Goldman Sachs research note relayed, "Tech's AI debt boom, in one chart" (2026-07-23): ~$489bn AI-related debt issued in 2026 to date, ~40% directly by hyperscalers

secondary press of record relaying sell-side research
https://finance.yahoo.com/technology/article/techs-ai-debt-boom-in-one-chart-143849995.html ↗
[8]

Cryptobriefing, "Alphabet, Amazon, Meta, Microsoft, Oracle issue record $159B in bonds" (2026-06-11/12): $159bn issued by those five through early June 2026 versus $121bn for all of 2025

secondary specialist press
https://cryptobriefing.com/tech-giants-record-159b-bonds-ai/ ↗
[9]

Remio summary, "Anthropic Compute Deals Reach $517 Billion" (approx. 2026-09)

tertiary unknown-authority aggregator
https://www.remio.ai/post/anthropic-compute-deals-reach-517-billion-reversing-its-cautious-compute-stance ↗
[10]

Originating post: X/@rdd147, 2026-09-25 (approx.), wording near-identical to the submitted text

tertiary social, not evidence
https://x.com/rdd147/status/2103503713340194931 ↗
[11]

Tomasz Tunguz, "OpenAI's $1 Trillion Infrastructure Spend" (2025-11-03): $1.15tn across seven vendors, 2025-2035

tertiary analyst blog
https://tomtunguz.com/openai-hardware-spending-2025-2035/ ↗
How links are chosen. A source is linked only when the address comes from the investigation's own retrieval or from a registry lookup (PubMed, Crossref) that matches the citation's title and year. Author lists shown as registry-verified come from the registry record, not from the report text. Citations that cannot be matched are labeled, never guessed.
This is one case on the record See the full case, browse the archive, and search every checked claim on TrueSeeker Finance Open on finance.trueseeker.com →