Analysis
A senior Google DeepMind executive told The Information that the industry's unprecedented data-center and compute buildout only makes financial sense as a bet on recursive self-improvement -- AI systems capable of meaningfully accelerating their own future development -- rather than as a bet on any single model generation or product cycle. It's a notably more aggressive public rationale than labs have generally offered for capex that now runs into the hundreds of billions of dollars annually across the industry.
Most public capex commentary from hyperscalers has stuck to a simpler framing: more compute produces better models, better models produce more revenue, and the spend pays for itself on a multi-year horizon tied to product adoption. The RSI framing is a different bet entirely -- it says the spend is justified even if near-term product revenue doesn't obviously cover it, because the real payoff is a step-change in how fast AI systems can improve themselves, at which point today's cost structure becomes irrelevant.
The timing is notable. This same week, every major hyperscaler -- Microsoft, Google, Amazon and Meta -- has signed at least one nuclear power deal, together committing to nearly 10 gigawatts of capacity, enough to power roughly 7 million homes. MediaTek separately committed $5 billion toward its own AI-datacenter push. Real capital is already moving at a scale that's hard to justify on ordinary product-cycle economics alone, which lends some credibility to the idea that RSI, not incremental product improvement, is the thesis actually driving budget approval inside these companies.
“MediaTek separately committed $5 billion toward its own AI-datacenter push.”
For VCs and LPs, the RSI framing changes the diligence question on every AI infrastructure investment: it's no longer just 'does this capex generate a return on this model generation,' it's 'do you believe recursive self-improvement is coming on a timeline that justifies spend that doesn't need to pay for itself in the interim.' That's a genuinely different, higher-variance bet than the one most public capex commentary has implied investors were making.
The bear case is straightforward: if RSI doesn't materialize on anything like the timeline implied, the industry has built out compute and power infrastructure sized for a step-change that didn't arrive, and the resulting overcapacity becomes a multi-year drag on the exact companies now committing gigawatts of nuclear power and tens of billions in chip spend. The Register's own recent commentary on the AI bubble "already popping" is the most direct expression of that skepticism circulating this week.
What to Watch
What to watch: whether other lab executives echo the RSI framing publicly or walk it back, whether capex guidance from Microsoft, Google, Amazon or Meta references anything resembling an RSI timeline in upcoming earnings calls, and whether any lab publishes concrete evidence of models meaningfully accelerating their own training or research pipeline rather than just asserting the thesis.
