Analysis
We read Jefferies's new note on Asian markets as the most concrete numbers yet behind a worry that has circulated all year: the AI infrastructure boom may be running ahead of the revenue that is supposed to justify it. Jefferies estimates that roughly half of the $2.3 trillion in contracted AI infrastructure backlog at Microsoft, Alphabet, Amazon and Oracle is owed by just two customers, OpenAI and Anthropic, The Register reported. That concentration is the crux of the firm's caution: "The AI story will be due for a big reset at some point in the not-so-distant future," Jefferies wrote, adding, "The question now is whether all of the above has been priced fully into markets."
Our read is that the backlog number matters less than where hyperscaler earnings growth is actually coming from. Jefferies found that hyperscalers' annualized earnings rose $222 billion over the four quarters through the second quarter of 2026 -- and that the entire increase traced to "other non-operating income," up $224 billion year-over-year, which the firm says likely reflects gains on investments in unlisted AI model builders. In plain terms: the companies selling compute to OpenAI and Anthropic are also marking up their equity stakes in companies like OpenAI and Anthropic, and that markup is what's driving reported profit growth, not the cloud business itself.
“Our read is that the backlog number matters less than where hyperscaler earnings growth is actually coming from.”
Separately, Bain & Company's 2026 Global Technology Report puts a number on the demand side of the bet: the industry needs roughly $6 trillion in annual AI revenue by 2031 to fund the infrastructure already planned, against existing applications that may generate $1.8 trillion, leaving a $4.2 trillion gap.
Room for disagreement: Bain's own model finds that $1.5 trillion of the gap could be closed by applications that don't fully exist yet, and Jefferies itself calls the long-term AI bet likely sound even while flagging medium-term uncertainty. Falling token prices -- Chinese open-source models like Moonshot AI's Kimi K3 now reportedly near frontier performance at a fraction of the cost -- could also expand the addressable market faster than it erodes incumbent pricing power, the bull case Jefferies's own note doesn't fully rule out.
But the caveat cuts both ways: a market that expands because tokens got cheap is also a market where pricing power keeps collapsing, and Jefferies's own backlog math assumes OpenAI and Anthropic keep paying contracted rates even as competitors undercut them. If either lab renegotiates or slows spending before new revenue arrives, the hyperscalers left holding $2.3 trillion in commitments -- not the labs themselves -- absorb the risk first.