Google DeepMind Strategy Chief: Massive AI Spending Is a Bet on Self-Improving Models

The chief strategy officer of Google DeepMind said out loud what the tech industry had preferred to keep quiet: current AI revenue doesn't justify the scale of spending that Google and its rivals are pouring into infrastructure. Jasjeet Sekhon admitted at a conference in Berkeley that the whole game is about something else entirely, the moment when AI starts building increasingly capable versions of itself.

A Bet on Self-Improvement

The acronym RSI, which Sekhon used, stands for recursive self-improvement: artificial intelligence designing, training, and improving its own successive, more capable versions, needing less and less human involvement along the way. It's long been treated as a distant research goal, but according to Sekhon, it's exactly what's behind Google's decision to spend roughly $200 billion this year on data centers, chips, and computing power.

Sekhon compared the current moment to the industrial revolution, noting that steam engines were once used to build better steam engines. In his framing, today's language models are set to play a similar role for their successors. They can already write code and partially improve their own components, but full autonomous capacity for self-development remains, for now, an aspiration.

This is the biggest scientific bet our civilization has ever made. - Jasjeet Sekhon, Chief Strategy Officer, Google DeepMind

AI revenue doesn't yet justify the scale of the capital spending we're taking on. - Jasjeet Sekhon, Chief Strategy Officer, Google DeepMind

Numbers That Don't Add Up

The admission is significant because it comes from the person responsible for strategy at one of the biggest players in the AI market, at a time when investors are asking ever more loudly when the massive outlays will start paying off. Google Cloud posted 82 percent revenue growth last quarter, and its backlog topped $500 billion, but Alphabet as a whole logged its first-ever negative quarterly free cash flow, at around negative $5.9 billion.

Sekhon also warned of a scenario he called a revenue vacuum or air pocket, a situation where investment keeps flowing while the expected returns from new AI applications fail to materialize. It's a risk market analysts have been flagging for months, watching tech company valuations climb faster than their actual AI revenue.

Who Is Behind These Words

Jasjeet Sekhon joined Google DeepMind relatively recently, moving over from the hedge fund Bridgewater Associates, where he led AI-driven investment strategy. His current role at DeepMind covers exactly this: the company's long-term strategy for the development of general artificial intelligence, which gives his remarks extra weight. This wasn't an offhand comment, but the voice of someone setting the giant's investment priorities.

Sekhon's interviewer at the Berkeley panel was Dawn Song, a researcher focused on AI system security, which added another dimension to the discussion. Sekhon also raised specific risks tied to a potential breakthrough in model self-improvement, including the asymmetry between attackers and defenders in cybersecurity and the risk of AI-assisted pathogen design.

What Happens to the Forecasts

The 2027-2028 window mentioned in connection with a possible breakthrough in recursive self-improvement has come up before in statements from other researchers tied to leading AI labs, including DeepMind itself and OpenAI. Sekhon cautioned, though, that RSI remains a research hypothesis for now, not a finished product, and that serious doubts remain about its feasibility, control, and safety.

For readers following the debate over AI infrastructure costs, this is another signal that the numbers cited around global data center spending don't stem from current AI service profitability, but from a bet on technology that has yet to be built. That raises questions about how long markets will keep financing that bet before demanding proof of its payoff, and how badly a delay or failure of RSI would hit the valuations of companies like Alphabet, which is already booking negative cash flow despite rising cloud revenue.