Google DeepMind’s Jasjeet Sekhon says AI spending is betting on recursive self-improvement

Jasjeet Sekhon says current AI revenue does not yet support the scale of infrastructure investment, but argues RSI could justify it over time.

24d ago

Fact Check

The BlockBeats flash directly confirms the claim's exact figures: 2027 total capex of $1.0403 trillion for the top 11 hyperscalers and $761.3 billion (73%) in memory/storage. The original X source (P Equity Research) independently corroborates the $761 billion at 73% figure and the memory-spend growth trajectory. The only nuance is terminology: the claim says 'storage' while UBS's underlying breakdown is memory products (HBM, DDR, NAND); the aggregate numbers align precisely. All sources are secondary reporting of a UBS research note rather than a UBS primary publication, warranting medium confidence.

Summary

Google is expected to spend about $200 billion this year on AI data centers and devices, even as DeepMind Chief Strategy Officer Jasjeet Sekhon said current AI revenue is not yet enough to cover that outlay. He warned the industry could face a gap in which spending comes first and revenue lags behind. Sekhon said the longer-term rationale for the investment is recursive self-improvement, or RSI, a concept in which AI helps develop the next generation of AI and then keeps improving training methods and the models themselves. He said RSI has become one of the central ideas behind this wave of capital spending, although true RSI has not yet arrived. At present, AI can only optimize parts of algorithms, kernel code and training workflows, and Sekhon expects the capability could emerge within two years.

Terms & Concepts

recursive self-improvement: A proposed AI capability in which systems help improve their own performance and design better versions of themselves.
kernel: A core low-level software component that manages key computing tasks and can affect model training efficiency.