A five-year, $125.2 billion TPU lease sits at the centre of Anthropic's compute plan, and the company helping to finance it is the one supplying the hardware. Anthropic flagged that dual role as a potential conflict in its IPO filing. That filing reportedly targets a valuation of about $965 billion.
That disclosure is the newest step in a pattern that began with the neoclouds. First, compute buyers were financed by the companies selling them compute. Then the same loop reached the frontier labs' equity. Now it sits inside a frontier lab's debt stack. If the listing goes ahead, public investors will be buying more than a model business. They will also take on a layered credit structure in which the chip supplier is a customer-facing vendor, a credit enhancer and a possible future shareholder all at once.
Step one: the vendor backstops the buyer
The first version of this loop was easy to see because it was written into contracts. CoreWeave's capacity agreement with Nvidia obligates Nvidia to buy datacenter capacity that the company's own customers do not fully use. In effect, a chip vendor guaranteed demand for capacity built on its own chips. That made the buyer easier to lend to, and the buyer then placed more orders.
Neocloud buildouts were financed by their own counterparties. The same lease counted as revenue, held up the equity mark and served as collateral for the debt. When one contract does all three jobs, it is hard to tell whether the business is profitable or just well-financed. The structure works while capacity stays scarce. It is untested when capacity stops being scarce.
Step two: the vendor buys equity in the buyer
The second version moved the loop from capacity contracts into equity. Nvidia invested in funding rounds for Nscale and Nebius and said it would invest up to $100 billion in OpenAI. The pattern runs both ways. Equity goes from the chip vendor to the buyer, and the buyer's spending comes back as hardware demand. The vendor books revenue that its own balance sheet helped pay for.
At this stage the loop was still mostly private. Private funding rounds set valuations, private lenders set terms, and the people pricing the circularity were the people taking part in it. Neither of the two largest labs had yet asked public-market investors to price a capital structure that includes the supplier.
Step three: the vendor underwrites the lab's debt
The Anthropic structure, as reported this week, has more layers than either earlier version. The base is a five-year, $125.2 billion TPU compute lease commitment. Banks including Bank of America, Citigroup and Morgan Stanley are arranging $42 billion in senior secured loans, with residual-value support from Broadcom. Another $18 billion would come from subordinated debt. Blackstone reportedly contributes $9 billion of that, and other investors supply the rest. The full package reaches up to $60 billion.
The reporting changed over the week, and the change matters. The October 2 account, sourced to Reuters' reading of the IPO filing, described Broadcom as providing up to $42 billion in financing. The October 5 account, sourced to Bloomberg, describes that $42 billion as bank-arranged senior debt backed by Broadcom's residual-value commitment. In that version, Broadcom is not the direct lender. The later version is more specific and is the one this column relies on. Broadcom's support should be read as backing for the senior debt against equipment depreciation, not as a guarantee of all $60 billion.
[L2] For scale: per the AI Market Watch index, which covers about 5,000 companies and is not a census, Anthropic's total funding stands at $132.3B. The infrastructure package, if fully assembled, would add debt equal to nearly half that figure. The package is still being put together. The reporting does not show that the full amount has closed, and the $42 billion figure is a ceiling, not money already paid out.
Broadcom now holds the depreciation risk
The core of the mechanism is the residual-value support. A lender to a compute buyer worries most about what the equipment is worth if the borrower stops paying. Accelerators lose value as newer generations ship. A lease that looks well collateralized in year one can be under-collateralized by year four. When the supplier promises a floor on resale value, it takes that tail risk off the banks. That is why the senior debt can be arranged at this size.
The cost of that promise moves to the supplier. If Anthropic's compute demand falls short of the lease and the TPUs come back to market, the supplier has to make up the gap between the guaranteed value and the market price. Part of the supplier's AI revenue therefore stops being a plain hardware sale. It becomes revenue whose final value depends on the customer's credit. The supplier also gets something for taking that risk. The residual-value guarantor role reportedly positions it to capture chip remarketing fees and preferred status in future AI infrastructure debt deals.
Anthropic's own disclosure names the friction. The company said Broadcom's position as both hardware supplier and financing partner creates potential conflicts, because Broadcom's pricing and equipment decisions affect Anthropic's compute access. The counterparty setting the price of the chips is also the one whose guarantee makes the debt financeable. An ordinary vendor sets chip prices without the buyer's debt financing depending on its guarantee.
Stress pulls obligations forward while cutting off the funding
The clause that most affects the risk case is the default provision. Payment or performance defaults could accelerate some lease obligations and restrict access to the financing. If the lab struggles, the bill comes due sooner and the credit line meant to pay it shrinks at the same moment. Both effects are triggered by the same default. It is a structure built to amplify stress.
The second clause is conversion. Some of the financing instruments may convert into Anthropic equity under certain conditions. The reporting does not give the conversion price, the triggers or which tranches are convertible. For public shareholders, the gap matters. Conversion that happens under stress dilutes holders at the worst point in the cycle and could make creditors, possibly including the supplier, significant equity owners.
This is the counter-signal to the thesis, and it cuts both ways. If the eventual public offering document shows the facility mostly undrawn, conversion rights tightly limited and acceleration triggers set well below realistic stress levels, then this is ordinary vendor finance at unusual scale and the circularity is cosmetic. If it shows heavy drawdowns, broad conversion rights and cross-default links to the lease, the $965 billion target is pricing a credit structure as much as a model company. Nothing in this week's reporting settles which case applies. The reporting rests on a confidential filing that has not been published.
The other lab is postponing the same test
OpenAI is taking the opposite route on timing. It is reportedly discussing a pre-IPO round of at least $30 billion at a valuation of roughly $1.4 trillion. Bloomberg described the round as a bridge to an expected public debut next year. It would follow a $122 billion raise in March at an $852 billion valuation, which had been expected to be OpenAI's final private round. CEO Sam Altman has reportedly ruled out a 2026 IPO to prioritize AI safety.
The revenue case behind the bridge is a snapshot. Bloomberg attributed a 70% increase in run-rate revenue since July, reaching $40 billion in August, to a strategic refocus that includes coding. Run-rate revenue annualizes a recent period. It is not revenue earned over a full year, and a private bridge round lets that number support a valuation for another year without public-market scrutiny. The round itself is unconfirmed, and OpenAI did not respond to TechCrunch's request for comment.
Taken together, the two largest labs are funding compute through counterparty capital and private credit, not just through public equity. One is about to put that structure in front of public investors. The other is paying to delay doing so. The industry's bull case says capital for frontier training is itself the moat. It now has to account for the fact that much of that capital comes from the companies selling the compute.
Vendors will finance to win, and unbacked labs will pay for the gap
The likely next step is competitive. A frontier lab choosing between accelerator suppliers will weigh financing terms alongside chip performance and availability. A supplier willing to back residual value can win a lab whose balance sheet could not otherwise carry a nine-figure-per-month compute commitment. Other hardware vendors will face pressure to match the offer, which spreads the depreciation risk across more supplier balance sheets.
Labs without a supplier taking that risk carry it themselves. Reflection AI has signed deals collectively worth more than $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029. [L2] Per the AI Market Watch index, which tracks about 5,000 companies, Reflection's total funding stands at $4.63B. Its compute commitments therefore exceed its recorded funding. Vendor money reaches Reflection only indirectly: Nebius is one of the neoclouds whose rounds Nvidia has invested in. The reporting shows no residual-value floor under Reflection's capacity. A lab in that position pays for equipment-value risk through its lease rates or its own balance sheet, whatever its claims about inference efficiency.
What diligence now has to cover
For anyone underwriting the Anthropic offering, revenue growth is still necessary but no longer enough. The filing needs to be read for:
- Drawdowns: the drawn versus undrawn split on each tranche.
- Residual-value strike: the level at which the supplier's guarantee applies.
- Conversion terms: which tranches convert, at what price and on what triggers.
- Acceleration triggers: what counts as a performance default.
- Supplier influence: how much the supplier's equipment and pricing decisions can affect compute access, as the conflict disclosure already admits.
The supplier's shareholders face the mirror-image question. Part of its AI revenue now depends on its customer's ability to keep paying.
Notes. The open question this week's reporting cannot answer is where the residual-value floor sits relative to expected TPU resale prices in years four and five of the lease. That one number decides whether the supplier has sold chips with a modest warranty attached or has taken on the lab's hardware risk.
