Why financing is not yet the limit on AI compute

Epoch AI asked whether money, rather than chips or power, will be the thing that caps AI compute, and used Anthropic's $50 billion infrastructure program as the test case. Campbell Hutcheson's answer is that financing is not the near-term bottleneck. Anthropic raised close to $50 billion in debt while its revenue was under $9 billion annualized last November, a gap that would normally not support borrowing at that scale. Epoch's analysis works through how the deal holds together anyway.

The answer is who absorbs the risk. On the compute side, investors led by Apollo and Blackstone put up $34.5 billion to buy Google TPU systems, with Broadcom backstopping $30 billion and Anthropic repaying through five-year leases. On the datacenter side, $15.2 billion in project finance covers construction across five sites, with Google agreeing to cover rent if the operator, Fluidstack, defaults. Together that funds more than 1 GW of TPU systems and 1.43 GW of datacenter capacity. The lenders lean on established suppliers to de-risk the deal rather than on a young lab's balance sheet. Epoch notes the same players plan to scale this structure past 20 GW of deployments through 2028, and with Anthropic's revenue reaching $47 billion by May, larger labs should find debt markets easier to tap, not harder.

Why it matters

If you are trying to read where the compute buildout goes next, watch the financing structures, not just the chip orders: vendor backstops from firms like Broadcom and Google are what let labs borrow years ahead of revenue, and that is the lever that keeps scaling going, or stalls it if a backstop breaks.