Silicon as Collateral: Ten Banks Hand Crux AI a $22 Billion Loan to Buy Google's TPUs

The most important number in AI this week is not a parameter count or a benchmark score. It is a loan covenant. Bloomberg News reported that a group of ten banks is providing a $22 billion loan to Crux AI, the cloud-computing venture backed by Blackstone and Alphabet — and that the debt will be secured by the market value of the very chips it buys, Google’s Tensor Processing Units, alongside Crux AI’s customer contracts. Reuters, Yahoo Finance and multiple market outlets picked up the story within hours.

It is, by most accounts, the largest chip-backed financing ever assembled, and it marks the moment when AI compute stopped being an exotic line item and became a financeable asset class — with everything that implies for both the boom and its downside.

What Crux AI actually is

Crux AI is the “neocloud” joint venture that Google and Blackstone formed in May 2026. Blackstone committed an initial $5 billion in equity to build a US-based compute-as-a-service business that would rent out Google Cloud’s TPUs rather than the Nvidia GPUs that power rivals like CoreWeave. The venture’s pitch is straightforward: hyperscale data center capacity, operations and networking wrapped around Google’s in-house silicon, aimed at model developers who want an alternative to the Nvidia supply chain.

The scale ambitions were clear from the start. The joint venture targeted its first 500 megawatts of compute capacity online by 2027, with reporting around the launch describing a total bet in the neighborhood of $25 billion. Earlier this month the venture got its name — Crux AI — and made a telling hire: Alan Duong, previously Meta’s head of data center engineering, brought in to lead a recruiting push ahead of a planned ramp of roughly 2 gigawatts of capacity per year.

Analysts covering the financing note the strategic logic: Crux AI adds neocloud compute capacity that could serve customers of the type OpenAI and Anthropic have become, positioning the venture as a direct challenger to CoreWeave in the market for rented AI infrastructure.

Why the structure of this loan matters more than its size

The headline figure — $22 billion from ten banks — is striking, but the collateral is the real story. The debt will be backed by the market value of the TPUs Crux AI purchases and by the venture’s customer contracts. In other words, the banks are underwriting two things at once: the residual value of specialized AI silicon, and the credibility of long-term compute purchase agreements.

This is the mortgage-ization of AI hardware. Blackstone President Jon Gray has argued publicly that AI compute will come to be seen as a “financeable asset class” in the same way mortgage lenders look at homes. The $22 billion Crux deal is the most concrete expression of that thesis so far — a structure where the chip itself is the house.

The precedent is spreading fast. AI-related loans have gone from essentially zero three years ago to more than $200 billion outstanding. Oracle locked in what was then the largest AI debt deal in history — a $38 billion package to fund data centers powering OpenAI. JPMorgan and Goldman Sachs are now lining up to finance billions in European AI debt. And earlier this year, GPU-backed financing reached investment-grade credit ratings for the first time, a signal that credit markets now treat AI hardware as core infrastructure rather than speculative equipment.

The risk hiding inside the collateral

Here is where the story turns uneasy. The collateral structure that makes the loan possible is also its weakest point, because AI chips depreciate on a schedule that has nothing to do with the repayment calendar.

Rental rates for widely deployed AI chips have fallen an estimated 70 to 90 percent since 2023 as each hardware generation obsoletes the last. A TPU fleet valued at $22 billion at closing does not stay worth $22 billion for long if the successor generation arrives mid-loan. Litigation specialists at Quinn Emanuel warned in a client alert this spring that collateral securing AI data center debt is “vulnerable to rapid deterioration,” and that the disputes which follow deterioration will be among the most contested in structured finance. Chicago Booth’s review of AI debt posed the question bluntly: loans secured by chips work exactly like mortgages — until the house burns down faster than the amortization schedule.

The second leg of the collateral, customer contracts, carries its own concentration risk. Neocloud economics depend on a handful of enormous model developers signing multi-year take-or-pay agreements. If the anchor customers’ demand forecasts prove optimistic — or if their own funding rounds compress — the contracts underpinning the loan reprice in a hurry. Bloomberg’s own mapping of the AI financing web, which traced a circular flow of capital on the order of hundreds of billions of dollars, illustrates how intertwined these counterparties have become: the chip vendor’s parent is the cloud rival, the equity holder is the landlord, and the customers are companies whose valuations assume the compute will be there.

None of this makes the Crux loan irrational. The venture has a deep-pocketed sponsor in Blackstone, a technology partner in Google with the strongest non-Nvidia silicon track record in the industry, and a demand environment in which frontier labs are signing compute contracts years in advance. Compared with the vendor-financed loops elsewhere in the market, a straight bank loan against chips and contracts is almost refreshingly legible.

What it means going forward

Three consequences are worth watching.

First, the neocloud wars now have a second balance sheet. CoreWeave built its position on Nvidia GPUs and Nvidia’s own investment; Crux AI arrives with private-equity equity and investment-bank debt behind Google silicon. Model developers gain genuine negotiating leverage between the two camps, which is precisely why Google structured the venture at arm’s length in the first place.

Second, expect the collateralized chip obligation to become a template. Once one $22 billion deal prices and syndicates successfully, every neocloud, sovereign cloud and hyperscaler subsidiary will bring a variant to market. The “CCO” could do for AI infrastructure what collateralized debt obligations did for housing — channeling enormous amounts of capital efficiently, and amplifying systemic risk in precisely the same way if hardware values turn.

Third, the deal hardens Google’s TPU ecosystem as a genuine alternative to CUDA. Banks do not lend $22 billion against assets they expect to be stranded. Their credit committees have now effectively certified Google’s silicon as durable collateral — a commercial validation that no benchmark score could buy.

The Crux AI loan is being reported as an infrastructure story, but it is really a financial-innovation story. The AI buildout has reached the stage where its constraints are no longer chip fabrication or power interconnects alone, but the creativity of credit markets. Ten banks just decided that a warehouse full of TPUs is as good as gold. Whether that judgment ages well will tell us a great deal about how the AI decade ends — or compounds.