The AI boom's biggest winners aren't selling AI products—they're marking up their bets on the company that does.
The Summary
- Google and Amazon both reported half their recent AI-related profits from unrealized gains on Anthropic stakes, not actual revenue from AI services.
- These are paper gains, billions in shareholder value added from marking up private equity positions they haven't sold.
- The AI profit story investors are celebrating is partly an accounting story, not an operational one.
The Signal
This earnings season revealed something uncomfortable about the AI gold rush. The "blowout AI profits" that sent Google and Amazon stock higher came substantially from unrealized gains on their Anthropic stakes, not from customers actually paying for AI services at scale. Half the AI profit narrative, in other words, is mark-to-model accounting on a private company position.
Both tech giants invested heavily in Anthropic over the past two years, taking equity stakes in exchange for cloud credits and strategic partnership deals. Those stakes are now worth billions more on paper. As Anthropic's perceived value climbs with each funding round and Claude's growing adoption, Google and Amazon mark up their positions. That creates reportable gains. Shareholders see "AI profits." Analysts write bullish notes.
"These are unrealized gains on private equity—billions added to shareholder value from positions they haven't sold."
But here's what matters: unrealized means exactly that. These aren't cash earnings from selling AI infrastructure or agent platforms to enterprise customers. They're not subscription revenue from businesses adopting AI tools at scale. They're accounting entries based on Anthropic's latest valuation, which itself is based on future expectations, not current profitability. It's a feedback loop of optimism, not proof that AI business models work at the scale being priced in.
The real test of AI profits is revenue from operations:
- Are enterprises paying materially more for cloud services because of AI workloads?
- Are AI agent products generating subscription revenue at margins that justify the infrastructure spend?
- Can the companies actually monetize the models, or just the equity positions in the companies building them?
The fact that billions in "AI profits" come from Anthropic stake markups rather than operational AI business suggests we're still early in the monetization curve. Very early. The market is pricing in the future. Fine. But it's worth knowing what's actually driving the numbers you're celebrating.
The Implication
If you're an investor, separate the signal from the accounting. Look past headline AI profit numbers to understand how much is coming from actual AI service revenue versus equity position markups. Unrealized gains can reverse as quickly as they appear if Anthropic's next funding round disappoints or if enterprise AI adoption stalls.
If you're building in this space, take note: the hyperscalers are still figuring out how to turn AI infrastructure into repeatable, high-margin revenue. That means there's still time to carve out specific use cases and prove unit economics before the big players fully figure out monetization. The window is open.