AI-generated content. This article was researched and written by an automated AI editorial system and published without prior human review. Every factual claim is checked against cited primary sources before publication, but no journalist read this page before you did — treat it accordingly, and report anything that looks wrong. How this works ›

Some links on this page are affiliate links. We may earn a commission at no extra cost to you.
Updated: Aug 4, 2026
·
anthropicgooglecomputefinancinginfrastructure

Anthropic's compute is being bought with somebody else's money — and kept off its balance sheet before an IPO

TL;DR: The machinery behind Anthropic’s compute is now documented, and it is unusual. Google has assembled a chip-financing programme reported at more than $150bn — the Financial Times puts the interconnected contracts nearer $200bn, a figure that traces to its investigation alone. Structure: a special-purpose vehicle funded by Apollo and Blackstone private credit buys Google/Broadcom TPUs and leases them to Anthropic. First tranche ~$35bn, agreed June 2026, priced around 5.75% — the largest private credit deal on record; Blackstone is already shopping a second. Broadcom provides a deficiency guarantee on roughly $31bn of senior debt and S&P downgraded it as a result. Morgan Stanley advises Broadcom, arranges the debt, and lends to the investors buying it. Bitcoin miners supply the buildings — TeraWulf’s 20-year lease is reported at ~$19bn. And the hardware stays off Anthropic’s balance sheet, which matters because Anthropic filed confidentially for an IPO on 1 June.

The structure

Strip away the headline numbers and the arrangement is straightforward to describe, if not to unwind.

Anthropic needs an enormous quantity of compute. Rather than buying it — which would require capital it does not have and would load debt onto a balance sheet it will shortly show to public markets — a special-purpose vehicle buys the hardware instead. The SPV is funded by private credit from Apollo and Blackstone. It purchases TPUs designed by Google and manufactured by Broadcom, then leases them to Anthropic. Lease income repays the debt.

The first tranche was roughly $35bn, agreed in June 2026 at a price around 5.75%, and was reported as the largest private credit transaction ever completed. Broadcom, Apollo and Blackstone formalised the approach this year under a partnership called AI XPV Platform. Blackstone has since held early discussions with investors about a second package of comparable scale.

Three further participants complete the picture:

Google sits at both ends: chip supplier and guarantor, while also holding an equity stake in Anthropic reported at around 14% — on top of the $40bn investment disclosed in April and the roughly $200bn in Google Cloud commitments Anthropic made over five years, first reported in May.

Why the off-balance-sheet detail is the story

Anthropic filed confidentially for an IPO on 1 June 2026. That single fact reframes everything above.

Hardware held inside an SPV and paid for through lease income does not appear as debt on Anthropic’s own balance sheet. Public markets punish heavily indebted companies, particularly ones without profits. A structure that converts an enormous capital obligation into an operating expense materially changes how the company presents to investors.

This is legal, well-precedented, and not by itself sinister. Aircraft and jet engines have been financed almost exactly this way for decades — the comparison to Boeing and GE’s leasing structures is one the arrangers themselves make. Nobody thinks an airline is hiding something because it leases its fleet.

But the analogy is doing work that deserves examination. A Boeing 787 has a thirty-year service life, a deep secondary market, and residual value that can be estimated with confidence. A TPU generation has a useful frontier life measured in a few years, no meaningful secondary market, and residual value that depends entirely on whether the next generation makes it obsolete. The financing structure assumes durable value in an asset class whose defining characteristic is rapid obsolescence.

That is the mismatch worth watching, and it is not hypothetical: the obligations are long-dated — TeraWulf’s lease runs twenty years — while model generations turn over in months. Anthropic has shipped Opus 4.7, Opus 4.8, Fable 5, Mythos 5 and Opus 5 since April.

The S&P downgrade is the hard evidence

Most commentary about AI infrastructure risk is speculative. This part is not.

S&P downgraded Broadcom over the leverage implications of its guarantee. A ratings agency looked at a vendor backstopping $31bn of a customer’s chip debt and concluded the vendor had become measurably riskier.

That matters because it settles an argument. Vendor guarantees in these structures are sometimes described as financial engineering that merely relabels risk. The downgrade demonstrates the opposite: the risk moved, it landed on Broadcom, and an independent party priced it. Whatever else this structure does, it is not costless.

The concentration of roles compounds it. Morgan Stanley advises the guarantor, arranges the debt that guarantee supports, and finances the investors buying that debt. Each role is unremarkable alone. Together they mean one institution’s judgement is embedded at three points in the same transaction — a structure that works well until it is tested, which is a sentence with history behind it.

Why this matters for anyone buying AI tools

It is why Claude got cheaper. Opus 5 shipped in July at roughly half Fable 5’s price. Price cuts of that size are not marketing decisions; they follow from compute secured at scale and financed cheaply. The structure described here is the direct upstream cause of the pricing in the Claude review.

It is also the fragility. Pricing built on lease economics inherits the exposure of those economics. If a second tranche prices materially worse, if a ratings action forces a restructuring, or if a TPU generation depreciates faster than modelled, the pressure surfaces downstream — in rate limits, in tier changes, in pricing. Nothing suggests that is imminent. It is simply the mechanism by which financial-structure risk becomes a product experience.

The template is spreading. This is being replicated across the sector, and analysts have started asking about concentration in hardware-backed structures generally. Anthropic’s 2GW AMD deal and the compute arrangements visible in SpaceX’s S-1 belong to the same family. When one financing pattern underpins the compute of most frontier labs simultaneously, its failure modes stop being company-specific.

Vertical entanglement keeps deepening. Google supplies the chips, guarantees part of the structure, hosts the workloads, and owns ~14% of the customer. Anthropic competes with Google’s own Gemini. That is not a conflict anyone is hiding, but it is a dependency worth pricing when assessing the durability of a vendor relationship.

Honest caveats

The ~$200bn aggregate is single-source. It comes from the Financial Times’ investigation, based on filings it reviewed and people involved. It has not been independently confirmed by another outlet with its own document access, and the article was not directly accessible for this piece. The $150bn+ programme figure and the $35bn tranche are corroborated across multiple outlets and are the numbers to rely on.

Several figures are reported rather than filed. The ~5.75% pricing, the ~$31bn guarantee and the ~14% equity stake come from financial press reporting, not from public filings this article inspected.

Off-balance-sheet is not a synonym for hidden. Anthropic’s eventual S-1 will disclose lease obligations; securities law requires it. The observation is about how the headline balance sheet reads, not about concealment.

None of this indicates distress. Anthropic is growing fast and has raised at a $965bn valuation. The structure exists because demand exceeds what conventional financing supports, which is a story about scale rather than weakness.

The obsolescence argument is analysis, not fact. Chips retired from frontier training retain real value for inference, and useful economic life may prove considerably longer than frontier-competitive life. That is a genuine counterargument and it may well be right.

What to do about it

If you use Claude: nothing. Service and pricing are unaffected, and the structure is currently the reason pricing is favourable rather than a threat to it.

If you are making a multi-year vendor commitment: treat compute financing as part of vendor due diligence, the way you would treat any supplier’s balance sheet. Every frontier lab now has one of these arrangements; the question is not whether but how concentrated.

If you are watching the Anthropic IPO: the lease obligations in the eventual public S-1 are the disclosure to read first. The gap between the headline balance sheet and the total contracted compute commitment is where the real capital structure is.


Related: Anthropic’s confidential S-1 filing · The AMD 2-gigawatt deal · Claude Opus 5 at half of Fable’s price

Frequently asked questions

What is actually being financed?

Google-designed, Broadcom-manufactured TPUs for Anthropic's training and inference. Rather than Anthropic buying the chips, a special-purpose vehicle funded by private credit purchases them and leases the hardware to Anthropic, repaid out of lease income. The first tranche was roughly $35bn, agreed in June 2026 and priced at about 5.75% — reported as the largest private credit transaction on record.

Why does the off-balance-sheet treatment matter?

Because Anthropic filed confidentially for an IPO on 1 June 2026. Hardware held in an SPV and paid for through lease income does not appear as debt on Anthropic's own balance sheet. Public markets penalise heavily indebted companies, so the structure materially changes how the company presents at listing. This is legal and well-established — aircraft and engines have been financed this way for decades — but it means the headline balance sheet understates the compute obligations behind the business.

What is Broadcom's exposure?

Broadcom provides a deficiency guarantee covering roughly $31bn of the senior debt and has committed to buy back unsold capacity. That guarantee is what allowed the debt to price at investment grade and reach institutional investors. It also has a cost: S&P downgraded Broadcom over the leverage implications, which is the clearest evidence available that vendor backstops in AI infrastructure transfer real risk rather than merely reallocating paperwork.

What do bitcoin miners have to do with it?

They own the buildings and the power. Several listed miners have pivoted to AI hosting because the margins are better and demand is steadier than mining rewards. TeraWulf signed a 20-year lease with Anthropic reported at roughly $19bn of contracted revenue, Hut 8 has a $7bn Google-backed arrangement, and Google backed a $15bn Anthropic data centre in Texas. Fluidstack operates hosting infrastructure in the structure.

Does any of this change what I pay for Claude?

Not today, and that is rather the point — it is why prices have been falling. Opus 5 shipped at roughly half of Fable 5's price in July, which is only possible with compute secured at scale and financed cheaply. The risk runs the other way: pricing built on lease economics is exposed if those economics change, and the obligations are long-dated while model generations are short.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.