A DeepMind exec finally said what the trillion-dollar AI spend is for: machines that improve themselves.

Everyone has been asking what the trillion-dollar AI buildout is actually for. A senior Google DeepMind executive has now given an unusually blunt answer. The spending, he says, is a bet on machines that improve themselves.

Jasjeet Sekhon, DeepMind’s chief strategy officer, made the case at a summit at UC Berkeley. Recursive self-improvement, or RSI, is “becoming a key component of the AI investment thesis,” The Information first reported. RSI is the idea of AI that can rewrite and upgrade itself, generating ever more capable successors without humans in the loop.

The striking part is the candour underneath it. Sekhon admitted that AI revenues “don’t sustain the capital expenditures we’re making so far.” The money is being spent, in other words, on a promise. He argued that betting against it would be unwise, since there are already “the makings of RSI.” He offered a neat analogy: steam engines built the next steam engine.

The new north star

What makes the framing land is what it replaces. For years the industry justified its spending by pointing at artificial general intelligence. Sekhon is effectively swapping one distant goal for another. RSI, on this telling, is the new AGI: the payoff that turns today’s data centres from a cost into the most valuable machines ever built.

The scale is real. Alphabet spent $44.9bn on capital projects in a single quarter, roughly double a year earlier, and lifted its 2026 guidance to as much as $205bn. It has promised a “significant” increase again in 2027. Amazon, Microsoft and Meta are all saying the same. Sekhon likened the effort to something bigger than Apollo or the Manhattan Project.

Some of it is clearly working. Google Cloud revenue jumped 82% in the quarter, with a backlog above $500bn. But the bill is enormous, and Alphabet posted its first-ever negative quarterly free cash flow, about $5.9bn in the red. Spending and revenue are moving at very different speeds.

A bet that may not pay

Sekhon named the risk himself. There could be an “AI air pocket,” he warned, where the expenditure happens but the revenue never arrives. That is the quiet fear under every hyperscaler earnings call, said out loud by the person whose job is to justify the outlay.

And RSI is not a shipping product. It is a research hope with real doubts attached: safety, control and technical feasibility. And whether it is even doable on the timeline executives imply, roughly 2027 to 2028. Rivals are already needling DeepMind over whether it has the self-improvement know-how to get there before OpenAI or Anthropic.

There is a modest version of the claim that is already true. Models can now generate code and, in narrow ways, help improve their own. The leap Sekhon is selling is from that to full, autonomous self-enhancement, and it is a large one.

What he has really done is make the trade explicit. The industry is spending Apollo-sized sums today against a capability that does not yet exist, and may not for years. His honesty is refreshing. It is also, if you are an investor, slightly terrifying.