DeepMind says AI spending bets on self-improving machines

Jasjeet Sekhon, Google DeepMind’s chief strategy officer, offered an unusually direct explanation for the AI industry’s extraordinary spending: it is betting on machines that can improve themselves.

Speaking at a summit at UC Berkeley, Sekhon said recursive self-improvement (RSI) is becoming a central part of the industry’s investment thesis. The Information first reported his comments.

RSI describes AI systems that can rewrite or upgrade themselves, producing more capable successors with less or no human involvement. In practical terms, the bet is that today’s data centers will eventually generate the systems that make tomorrow’s data centers more valuable.

“AI revenues don’t sustain the capital expenditures we’re making so far.”

From AGI to recursive self-improvement

For years, the industry defended its escalating infrastructure budgets by pointing to artificial general intelligence, or AGI, as the eventual payoff. Sekhon’s framing shifts that target. RSI becomes the mechanism that could turn massive computing investments into a self-reinforcing capability.

Recommended reading

China’s open models squeeze US AI on price

His analogy was simple: steam engines built the next steam engine. The implication is that AI systems could eventually contribute to the design, training, and improvement of their successors, accelerating progress beyond what human researchers can manage alone.

There is a limited version of this process already. Current models can generate code and, in narrow ways, help improve AI systems. But the leap from those capabilities to autonomous, general-purpose self-enhancement is substantial. RSI is not a product DeepMind is shipping; it remains a research goal with unresolved questions about safety, control, and technical feasibility.

Sekhon reportedly suggested that the industry’s implied timeline is roughly 2027 to 2028. The source also noted that rivals are questioning whether DeepMind can develop the necessary self-improvement capabilities before OpenAI or Anthropic.

The cost of the bet

The spending behind the thesis is already enormous. Alphabet spent $44.9 billion on capital projects in a single quarter, roughly twice as much as a year earlier, and raised its 2026 capital-spending guidance to as much as $205 billion. The company has also promised another “significant” increase in 2027.

Amazon, Microsoft, and Meta are pursuing similarly aggressive AI infrastructure strategies. Sekhon compared the effort with projects larger than Apollo or the Manhattan Project, underscoring how far the industry’s financial commitments have moved beyond ordinary product development.

Some returns are visible. Google Cloud revenue rose 82% in the quarter, while its backlog exceeded $500 billion. But Alphabet also recorded its first-ever negative quarterly free cash flow, at approximately $5.9 billion below zero.

That gap between spending and revenue is the central financial risk. The infrastructure is being built now, while the capability expected to justify it may not arrive for years — or at all.

An investment thesis without a finished product

Sekhon himself identified the danger of an “AI air pocket”: a period in which companies continue spending heavily but revenue fails to catch up. His admission makes the industry’s underlying trade-off unusually clear. Current AI income does not yet support the scale of capital expenditure, so investors are being asked to fund a future capability rather than a present business model.

The unresolved issue is not whether models can assist with pieces of their own development. They already can. The question is whether that narrow assistance can become reliable, autonomous self-improvement on the timeline executives imply.

That makes RSI a more concrete explanation for the spending than a vague promise of AGI, but not a safer one. The industry is committing Apollo-scale sums to a capability that does not yet exist, with its technical, financial, and safety risks still open.