Banks Are Lending $22 Billion Against AI Chips. Here’s the Risk.

A bank knows what to do with a house after a borrower defaults. It can sell the property and recover part of the mortgage. A warehouse full of artificial intelligence chips creates a stranger problem. The hardware is valuable, difficult to replace and capable of generating enormous revenue. It can also lose economic value quickly when a faster generation arrives.

Ten banks are now preparing one of the largest tests of that collateral. The group is providing a reported $22 billion loan to Crux AI, a cloud-computing venture created by Blackstone and Alphabet’s Google. The money will purchase Google’s Tensor Processing Units, or TPUs. The chips and Crux AI’s customer contracts will secure the debt.

The rise of banks lending against AI chips shows how the boom is changing shape. Early investment came from venture capital and the cash reserves of large technology companies. The next phase is drawing in commercial banks, insurance money and bond investors. AI demand now has to produce enough cash to service debt on a schedule.

Inside the $22 billion AI chip loan

The bank group includes Goldman Sachs, Sumitomo Mitsui Banking Corporation, Barclays, BNP Paribas and Bank of Nova Scotia, according to Reuters. More lenders may join through syndication, a process that divides a large loan among several institutions.

Crux AI combines Blackstone’s capital with Google’s technology. Blackstone committed an initial $5 billion in equity. Google will provide its custom chips, software and services. The business plans to sell dedicated computing capacity to AI laboratories, technology companies, corporations and governments.

The venture expects to bring its first 500 megawatts of data-centre capacity online in 2027. That requires far more than processors. A functioning AI cloud needs land, power connections, cooling systems, networking equipment, storage and long-term access to electricity. When Google and Blackstone announced the venture, they presented it as a compute-as-a-service platform: customers would rent access instead of buying and operating the infrastructure themselves.

The reported financing has two layers of security. The TPUs have resale or reuse value, and the customer contracts create expected cash flow. Those contracts matter most. Banks receive repayment in dollars, so they need customers to keep paying for compute.

How banks are lending against AI chips

Asset-backed lending is common. Airlines borrow against aircraft. Manufacturers pledge machinery. A logistics company can finance trucks using the vehicles and their operating revenue. AI chips now belong to the same broad category because they are expensive productive assets.

A high-end accelerator earns money whenever a customer rents it to train a model or run inference, the process that produces an AI system’s answers. Demand has been strong enough for specialized cloud companies to build entire businesses around renting clusters of processors by the hour.

Chip-backed financing was already developing before the Crux deal. In 2024, AI cloud provider Lambda secured a $500 million loan using Nvidia GPUs and the revenue they generated as support. Reuters described it as a special-purpose GPU financing vehicle. By 2026, CoreWeave had arranged an $8.5 billion term-loan facility as its total debt and equity commitments over the preceding year reached about $28 billion.

Crux takes the model to a new scale and applies it to Google’s proprietary hardware. Google describes TPUs as custom accelerators built for machine-learning workloads. They support both training and inference and operate inside a wider system of software, networking and data-centre infrastructure.

The chip therefore resembles an engine inside a power plant. Its value comes from the system around it and the work customers pay it to perform.

Why banks are willing to lend so much

The lenders have several cushions. Blackstone’s $5 billion equity commitment absorbs losses before the debt. Google brings technical expertise, a supply of TPUs and a cloud ecosystem already used by major AI customers. Crux plans to secure customer commitments that make future revenue more visible.

The banks can also spread their exposure. Syndication allows additional lenders to buy pieces of the loan. The reported plan leaves room to replace the bank financing later with longer-term bonds sold to institutional investors. Each step moves some risk away from the original group.

AI compute currently commands high prices because advanced capacity remains scarce. Cloud provider Nebius raised some chip-rental prices twice within three months in 2026, while CoreWeave reported signing short-term capacity contracts at annualized prices of about $40 million per megawatt. Strong pricing makes the projected debt payments easier to support.

The underwriting still depends on details that have not been publicly disclosed. Contract length, cancellation rights, customer credit quality, minimum-use requirements and pricing terms determine how dependable the revenue will be. A five-year commitment from a profitable technology company carries very different risk from a one-year agreement with an AI startup spending investor capital.

Why borrow instead of funding everything with equity?

Debt lets Blackstone and Google build a much larger platform without supplying every dollar themselves. Equity capital takes the first loss and receives the remaining upside. Lenders accept a fixed return in exchange for contractual protections, collateral and priority over the owners if the venture runs into trouble.

The arithmetic rewards the equity investors when Crux earns more on its infrastructure than it pays in interest. Imagine that a financed chip cluster produces a 12% return after operating expenses while its debt costs 7%. The difference increases the return on the smaller equity contribution. If the cluster earns 5%, the interest bill stays in place and the pressure lands on equity much faster.

This is leverage in its most useful and dangerous form. It accelerates construction during strong demand and reduces the margin for forecasting errors. The $5 billion equity commitment gives lenders a cushion, though the public reporting does not disclose enough about the final capital structure, loan covenants or advance rates to calculate its exact protection.

The chips may depreciate faster than the loan

Buildings often remain useful for decades. AI hardware moves through generations much faster. A working chip can lose commercial value when a newer model performs the same task using less time or electricity.

Google’s own product releases show the pace. Its TPU 8i system, announced in 2026, offered 80% better performance per dollar for inference than the prior generation, according to the company. Earlier releases also delivered large gains in speed, memory and efficiency. Those improvements benefit customers and weaken the relative economics of older equipment.

TPUs create another wrinkle because they are custom Google accelerators tied closely to Google’s software and infrastructure. Nvidia GPUs have a large and visible market across many cloud providers. The potential buyer pool for used Google hardware is more specialized. That likely makes recovery values harder to estimate if a lender ever has to seize and resell the chips.

Physical ownership also solves only part of the problem. Chips need compatible servers, high-speed networks, cooling and electricity. Removing them from one data centre and installing them elsewhere costs time and money. Their liquidation value can fall far below the revenue value they had inside an operating cluster.

The customer contracts carry the real weight

The loan resembles project finance more than a pawnshop transaction. Lenders will watch the cash produced by the assets, just as they would for a power plant, toll road or aircraft fleet. The hardware creates capacity; customer payments service the debt.

That arrangement works when Crux keeps the chips busy at profitable rates. Empty capacity still consumes capital. Data centres also carry large fixed expenses for power, cooling, staff and maintenance. A drop in rental prices can reduce margins quickly because those costs remain.

Customer concentration can magnify the damage. A small number of AI labs and hyperscalers account for much of the industry’s spending. Several borrowers, cloud providers and data-centre developers may therefore rely on the same group of customers. One laboratory cutting its expansion plans could affect chip suppliers, landlords, utilities and lenders at once.

This overlap is receiving attention from credit analysts. S&P has estimated that six large hyperscalers could spend more than $7 trillion on AI and data centres through 2030. The agency also warned that financing arrangements are becoming more complex and less transparent, especially when large technology companies help smaller operators borrow by signing leases, purchasing chips or providing other commitments.

What could break the financing model?

The clearest risk is weaker demand. Companies may discover that AI tools generate less revenue or fewer productivity gains than expected. Cheaper open-source models could reduce the amount of expensive computing needed. Competition among cloud providers could push rental prices down even if usage continues growing.

Rapid technological improvement creates a second pressure. Crux could still have years of debt outstanding when newer processors change the price customers will pay for its installed hardware. The chips would remain functional while producing less cash than the original model assumed.

Construction and electricity add execution risk. Crux needs to complete data centres, secure grid connections and bring hundreds of megawatts online. Delays postpone revenue while interest continues accumulating. Power shortages or higher electricity prices can also alter the economics after the loan has been signed.

Refinancing deserves equal attention. The banks may expect institutional investors to replace part of the loan through the investment-grade bond market. That exit depends on future interest rates, credit ratings and investor appetite. Crux may face a higher cost or have to retain more bank debt if bond buyers become cautious.

The risk can travel far beyond the ten banks

The original lenders carry the exposure at closing. Syndication distributes pieces to more banks. A bond refinancing could move it into insurance portfolios, pension funds, credit funds and fixed-income products. Some investors may eventually own AI infrastructure debt without choosing a technology stock.

That chain is familiar from other capital-intensive industries. Banks originate loans, package risk and place it with institutions seeking steady income. The process creates more funding for construction and gives each lender a smaller individual position. It also connects the AI cycle to a wider section of the financial system.

Private funds are already financing data centres, power projects and cloud companies. Their investors face the liquidity and valuation questions described in my article on private credit fund withdrawals. Chip-backed loans add another layer: the collateral can improve in technical performance across the industry while declining in value for the lender holding an older generation.

Does chip-backed lending prove AI is a bubble?

The existence of secured credit answers neither “bubble” nor “boom.” Productive infrastructure is commonly built with debt. If customers keep renting the capacity at profitable rates, the financing may look conservative in hindsight.

The warning signal lies in the assumptions shared across the system. Banks, asset managers, utilities and technology companies are all planning around years of rapidly growing AI demand. Their projects can be individually reasonable and collectively excessive. Useful infrastructure has produced poor investment returns before when too much capacity arrived at once.

The $22 billion Crux loan makes that question measurable. Watch customer contract quality, utilization rates, compute pricing, the speed of chip replacement and the eventual bond refinancing. Those indicators will show whether AI infrastructure is developing into a durable utility or an expensive race to own machines whose earning power fades faster than their debt.

For now, the chips sit at the centre of two promises. Google promises they can produce valuable computation. Crux promises customers will pay for it. The banks are lending against both—and debt collectors still arrive on schedule when technology moves on.