Blackstone says AI infrastructure debt fears are overstated

Lending against the AI buildout

Blackstone manages $1.3 trillion in assets and says it owns more computing capacity than any company outside the largest cloud providers. Its portfolio spans emerging cloud services and data centers, giving the firm capacity it can use to support other businesses.

Khaira says that mix distinguishes Blackstone from traditional venture firms: it can finance computing capacity with flexible capital, from debt to equity. One example is Crux AI, a new cloud provider created with Google to commercialize Google’s tensor processing units, or TPUs. Blackstone committed $5 billion to the project.

The firm has also backed Etched, a four-year-old chip startup that drew investor attention for its rapid development. More broadly, newer cloud providers rent AI infrastructure powered by GPUs.

Blackstone’s lending model weighs a borrower’s customer contracts alongside its equipment. GPUs can serve as collateral, while the creditworthiness of customers such as Microsoft or Meta helps shape the risk assessment.

The first such deal came in 2023, with CoreWeave. Soon after ChatGPT launched, demand for the cloud provider’s services surged. Khaira recalled that CoreWeave had “$30 million in revenue and minimal profit” when it sought a $2.3 billion loan from Blackstone.

The deal became a template for billions of dollars in later loans to CoreWeave and competing cloud providers. Earlier this year, Blackstone and Coatue also led the arrangement of a $10 billion credit line for Australian GPU provider Firmus.

The collateral question

The market has grown less comfortable with the debt. CoreWeave’s debt has reached about $50 billion, and its shares have fallen more than 20% in recent days. Firmus abandoned plans for an Australian Stock Exchange IPO at a target valuation of about $30 billion after investors proved unwilling to accept that price, Khaira said.

In a recent CNBC interview, Altimeter Capital founder Brad Gerstner argued that CoreWeave’s GPUs no longer provide sufficient collateral, despite once helping the company secure cheap debt. Khaira pointed to a wider tightening in credit conditions: interest rates rose in September, inflation came in above forecasts, and lenders are demanding higher spreads for AI infrastructure financing. Borrowing costs have risen even for cash-rich, profitable technology companies such as Meta and Microsoft.

Khaira argues that this is not a 2008-style setup. Banks, he said, are not themselves in fragile financial condition, and much of the data-center construction is for the largest cloud providers. In his view, that means the debt ultimately rests on the credit of some of the world’s most reliable companies.

I think that is a narrower defense than it first appears. The quality of an end customer matters, but it does not answer how much the infrastructure is worth if demand, pricing or utilization falls short of expectations. The announcement is quiet about how Blackstone would value GPU collateral under those conditions — precisely the point raised by Gerstner.

Building capacity, finding its use

Demand for computing remains strong, and AI companies still lack the capacity to run their models. That demand is driving data-center construction. But the pace of adoption inside businesses is less consistent.

Blackstone is also trying to help portfolio companies put AI to work. This year, it formed Ode with Anthropic, a $1.5 billion joint venture with Anthropic and Hellman & Friedman. The venture will provide engineers to help companies adopt AI and integrate it into their workflows, starting with 280 companies in Blackstone’s portfolio.

Asked whether those companies are already seeing meaningful returns, Khaira said they are, but unevenly. He described adoption as “jagged,” a pattern that can make it look as if infrastructure is being built faster than the market can absorb it.

That unevenness is the tension in Blackstone’s case: demand for computing may justify more capacity, while the value of the debt financing it depends on still turns on how quickly customers can make that capacity pay.