Gavin Baker says GPU contract repricing can fund hyperscaler AI capex

This is a July 28 X post by investor Gavin Baker arguing that the market is overreacting to wider credit spreads for hyperscalers such as Microsoft, Google, Amazon, and Meta. His thesis is that current cash flow understates what these companies will earn because much of their GPU capacity is still priced under older contracts. As those contracts expire and reset at higher rates, he expects operating cash flow to accelerate enough to fund the AI buildout, reduce the need for debt, and pull credit spreads back in.

Credit spreads measure how much extra compensation investors demand to lend to a company. Wider spreads mean the market sees more risk. Baker thinks that signal is too pessimistic because investors are comparing enormous AI spending with cash flow that has not yet benefited from higher GPU prices.

Baker's key evidence is the gap between current GPU rental prices and older contracts. If buyers are prepared to pay at least twice as much when those contracts expire, hyperscalers are currently earning less from their installed capacity than they could at today's prices.

Every contract that resets at a higher price lifts revenue from GPU capacity that is already installed. Baker therefore expects growth and operating cash flow to accelerate without requiring the same increase in physical capacity. His 50% Q2 figure mixes estimates with reported results and represents his model, not a completed quarter for every company.

Baker starts with 25–35 gigawatts of new 2028 capacity and a cost of $60 billion per gigawatt. He compares the resulting capex estimate with $1.3–$1.4 trillion of expected operating cash flow. The market's concern is that the remaining gap must be funded with debt.

His counterargument is that the cash-flow estimate is too low. If GPU contracts reprice and cloud growth accelerates, he expects the projected $100–$700 billion debt gap to shrink or disappear. He also argues that even the upper end would add less than one turn of leverage, while support from cash-rich Nvidia and Broadcom could improve financing terms.

The figures should be read as Baker's rough ranges rather than reconciled arithmetic: 35 gigawatts at $60 billion equals $2.1 trillion, not $2.2 trillion, and the printed capex, cash-flow, and debt-gap ranges do not align perfectly at every endpoint.

The BofA chart shows semiconductor free cash flow rising while hyperscaler free cash flow falls under the weight of AI spending. Baker says it is backward-looking because it misses accelerating demand from AI labs and inference providers, as well as the future price increases embedded in expiring GPU contracts. The green projection on the chart is Baker's annotation, not BofA's.

Customers that locked in GPU capacity during 2024 or 2025 are paying less than today's rental rates, so Baker calls them the current winners. Hyperscalers are the other side of those contracts. When the agreements reset, he expects more of the economic value to move back to the infrastructure owners, lifting cash flow and reducing perceived credit risk.

A credit-default swap is insurance against a borrower failing to repay. Buying more protection pushes its price higher and can make the market appear more worried. Baker argues that this signal can be distorted, so he gives more weight to hyperscalers' balance sheets, interest coverage, and future cash generation.

Baker's conclusion is not that the AI buildout is risk-free. He thinks the financial risk is being overstated, while the physical execution risk is real: data centres need grid connections, generation, equipment, and enough power to run the GPUs already being purchased.