Hyperscaler earnings have a hidden reflexivity problem baked into AI valuations

Jack Farley has put a number on a risk that sits largely unexamined in the AI investment thesis. By his account, 30% of hyperscaler income in early 2026 came not from selling compute, cloud services, or subscriptions, but from marking up the private valuations of AI labs held on their balance sheets. That is a significant share of reported income derived from an accounting treatment rather than cash generation, and it has a property that operational revenue does not: it goes into reverse when sentiment does.

The mechanism Farley describes is straightforward. Hyperscalers have taken equity stakes in AI labs. As those labs’ private valuations rise, the hyperscalers record unrealized gains. Those gains flow into income. When the numbers are good, the arrangement flatters earnings. When the AI trade softens and private valuations are marked down, the same channel runs the other direction. Income falls not because the underlying business deteriorated, but because the valuation on a balance-sheet line item moved.

What makes this worth taking seriously is the word Farley uses to describe it: reflexive. The concern is not just that the income figure is fragile. It is that the fragility is tied to the same sentiment driving the AI trade itself. A sustained selloff in AI equities puts pressure on private lab valuations. Falling private valuations reduce hyperscaler income. Reduced hyperscaler income undermines confidence in the earnings story that supported the AI trade. Each step feeds the next.

If the AI trade starts to unwind a lot these hyperscalers 30% of their income earlier this year is from the markup of the AI labs private valuations on their balance sheet so it's very reflexive Jack Farley

This is distinct from the ordinary cyclical risk that technology investors have long modeled. In a conventional downturn, a cloud provider’s revenue falls because customers cut spending. The causal chain is external. What Farley is describing is a loop that originates inside the financial structure of the AI trade, where the asset and the income statement are entangled. The hyperscaler is simultaneously a vendor to the AI ecosystem and a marked-to-market stakeholder in it.

The 30% figure, if it holds, also reframes the optics of hyperscaler profitability in early 2026. Earnings that looked like evidence of AI monetization at scale were partly, on this account, a reflection of rising private valuations. That does not make the underlying businesses weak, but it does mean that the income line was doing two different jobs at once, and only one of them was operational. Analysts separating signal from noise in those results would need to isolate the markup contribution before drawing conclusions about run-rate earnings power.

Farley’s claim is a single data point from a named observer, not a verified filing or audited figure. It warrants that caveat. But the structural logic he describes is not exotic. Equity stakes in private companies, carried at fair value, produce exactly the income dynamics he outlines. Whether the 30% figure proves precise or approximate, the direction of the argument is grounded in how these balance sheets work. The question worth pressing is not whether this mechanism exists, but how large a role it has played, and what happens to reported earnings if the private valuations that inflated them in early 2026 come under pressure later.