Limits on AI financing are needed before the buildout grows further - Carlyle’s Jason Thomas

Carlyle Group CG is urging private credit lenders to establish explicit concentration limits on artificial intelligence financing before the buildout accelerates further, warning that the industry risks repeating the same overexposure mistakes it made with software-as-a-service companies earlier this year.

Jason Thomas, Carlyle's head of global research and investment strategy, said in an interview with CNBC that a potential 40% concentration in AI data centers represents far too much risk for the asset class.

Thomas pointed to the February release of Anthropic's ANTHROPIC Cloud Code as a cautionary example, noting that SaaS exposure had reached 20% to 40% of portfolios depending on how it was defined.

"Now it seems that this experience could be relived as it relates to AI data centers," he said, adding that the private credit industry appears to have learned little from that volatility.

The scale of potential exposure is substantial. Thomas outlined a $5T AI infrastructure buildout expected over the coming years, with approximately $1T of that financing projected to come from private credit.

Based on current assets under management and growth expectations, that would translate to roughly 40% concentration—a level he considers excessive given recent market turbulence.

Thomas also raised concerns about the underlying economics of these investments. While compute backlogs currently stand at $2.5T in remaining performance obligations, he calculated only about $375B in net economic value when discounted to present value and compared against the actual costs of GPU purchases, data center construction, and power infrastructure.

"That margin, 15 to 17% present value, could compress very quickly," he warned, particularly if pre-profitability AI labs lose access to capital markets needed to honor lease commitments.

He also drew comparisons to the 2005–2008 period, noting that lenders today are relying heavily on residual value guarantees from hyperscalers and semiconductor manufacturers—similar to assumptions made about subprime collateral before the financial crisis.

"That's very reminiscent of what people said in 2005," Thomas observed, cautioning that documentation varies across deals and workouts remain untested.

Thomas recommended a 15% concentration cap for private credit exposure to AI-related assets, emphasizing the need for ex-ante limits rather than post-facto risk management.

"To have no exposure would be silly, particularly given the range of possibilities," he acknowledged but stressed that incremental risk beyond that threshold should command hefty premiums.