The tens of billions Google is pouring into AI infrastructure aren’t just about serving more chatbot queries. A Google DeepMind executive told The Information that the industry’s unprecedented capital spending is really a bet on something more specific: RSI, or recursive self-improvement. In other words, AI systems that help build better AI systems, compounding on themselves.
That framing matters. It reframes the spending debate from “are we overbuilding data centers” to “are we funding a flywheel.” And it tells you how the people closest to the frontier actually justify the numbers.
What RSI actually means
Recursive self-improvement is the idea that AI can accelerate its own progress. Models help design better chips, write better training code, generate better synthetic data, and run more experiments than any human team could. Each gain feeds the next one.
If you believe that loop is real and near, then today’s capex looks less like a cost and more like the down payment on a self-reinforcing engine. If you don’t, it looks like the biggest infrastructure gamble in tech history.
That’s the tension The Information’s reporting exposes. The DeepMind view is that spending is rational precisely because the payoff isn’t linear. You’re not buying a bigger calculator. You’re buying the thing that builds the next builder.
Why this lands now
Three things make this the moment RSI moves from research-paper talk to boardroom logic:
- The capex numbers stopped making sense on old math. Google, Microsoft, Amazon, and Meta are each running annual infrastructure budgets that dwarf entire national research programs. Standard demand forecasts don’t cover it. RSI does.
- Coding agents are the first proof point. AI that writes and reviews code is already compressing engineering timelines inside these labs. That’s the earliest, most visible piece of the loop.
- The competitive clock is loud. If a rival’s AI starts improving itself even slightly faster than yours, the gap compounds. Nobody wants to be the lab that underspent and got left behind.
What stands out here is the honesty of the bet. Executives are no longer selling capex purely on cloud revenue or ad targeting. They’re selling it on the possibility that the technology bootstraps itself.
The skeptic’s case
This deserves a counterweight. RSI has been predicted before, and “AI improving AI” has real limits. Data quality caps out. Compute hits physical and power constraints. Model gains have shown signs of diminishing returns per dollar. A self-improvement story is also conveniently unfalsifiable in the short term, which makes it a useful narrative for justifying spend that hasn’t paid off yet.
So read the RSI framing as a thesis, not a fact. It’s the belief driving the checkbook, held by people with obvious incentive to believe it. That doesn’t make it wrong. It makes it worth watching for evidence rather than taking on faith.
What to watch, and what to do
For practitioners and businesses trying to plan around this, a few practical moves:
- Track the proof points, not the promises. Watch how fast labs ship internal tooling gains, model release cadence, and cost-per-capability curves. Those are the real RSI tells. Marketing language isn’t.
- Assume compute stays expensive and contested. If the frontier labs believe in the flywheel, they’ll keep buying every chip and megawatt available. Budget your own AI plans around scarce, pricey capacity, not falling prices.
- Build on capability, not on one vendor’s roadmap. If self-improvement accelerates, model quality could jump in steps rather than smooth lines. Architect your products to swap models easily so you can ride upgrades instead of being locked to last year’s version.
- Watch the power story. RSI-scale compute runs into energy limits fast. Grid capacity, nuclear deals, and data center siting become the real constraints on how far this bet can run.
The next year or two will test whether RSI is a genuine engine or an expensive article of faith. If the loop is real, the labs spending the most now will pull away hard. If it stalls, this becomes the cautionary chapter in every future capex debate. Either way, the spending has a thesis now, and it has a name. More detail is available in the original reporting from The Information.