Google DeepMind Executive: AI Capital Expenditure Bets on Recursive Self-Improvement
Alina Collins
Google DeepMind's chief strategy officer Jasjeet Sekhon says recursive self-improvement (RSI) — AI that can build better versions of itself — is the "core investment thesis" behind the industry's unprecedented capital spending; he concedes current revenue cannot yet support the outlay, calling it "the biggest scientific bet in the history of human civilization."
What exactly is all this money betting on?
Sekhon told the Agentic AI Summit at UC Berkeley that RSI — recursive self-improvement, where AI automatically creates a better version of itself — is the "core investment thesis" behind today's AI capital spending.
This means → tech giants are not simply buying hardware. They are wagering on a capability that does not yet exist: AI that improves itself in a self-reinforcing loop.
Google plans to spend roughly $200 billion this year on AI data centers and equipment, with further increases next year.
Has the spending outrun the revenue?
Sekhon acknowledged that AI currently generates revenue "not sufficient to support the capital expenditure we are undertaking."
He described the risk as an "AI air pocket" — spending has already happened, but revenue has not arrived on schedule. In plain terms = the money is out the door, but the revenue engine has not started turning yet; there is a vacuum in between.
He characterized the industry's total outlay as "the biggest scientific bet in the history of human civilization" — exceeding the Apollo program, the Manhattan Project, and the buildout of the internet combined.
How close is RSI right now?
Sekhon noted that true RSI has not been achieved, but "precursors" are already visible — AI companies use models to help design components of other models.
He drew a steam-engine analogy: using a steam engine to build the next steam engine is unremarkable. This means → today's "AI helping build AI" is tool-level assistance, far short of genuine self-iteration.
True RSI requires a model to independently redesign its own architecture and develop entirely new models, forming a continuous loop. DeepMind researcher Oriol Vinyals and OpenAI co-founder Wojciech Zaremba said they expect RSI between 2027 and 2028.
Why has the industry stopped talking about AGI and started talking about RSI?
RSI is replacing AGI (artificial general intelligence) as the new frame of reference in industry discussions. This reflects a shift from "Can AI be as smart as a human?" to "Can AI make itself smarter?"
Put simply = AGI asks about the destination — is AI powerful enough? RSI asks about the engine — can AI accelerate on its own? The latter determines when the former arrives.
Sekhon also flagged AI's potential risks in cyberattacks and bioattacks. UC Berkeley professor Dawn Song, who recently joined Meta's "superintelligence" unit, said AI will in the near term "benefit attackers more."
Content is for reference only, not financial advice.