A Google DeepMind executive is projecting that AI-related capital expenditures could hit the trillion-dollar mark by 2026. The key variable in that equation isn’t just bigger models or faster chips. It’s something called recursive self-improvement, or RSI, which the executive considers the new AGI.
RSI is the point where an AI system gets good enough to start making itself better, without waiting for humans to do the tuning.
The numbers behind the bet
The combined capex guidance from the four major hyperscalers, Alphabet, Microsoft, Amazon, and Meta, points to roughly $725 billion in collective investment for 2026. That’s a 77% jump from approximately $410 billion in 2025.
Alphabet alone revised its 2026 capex forecast to between $195 billion and $205 billion.
Goldman Sachs has layered on its own estimates, pegging annual AI-related capex at around $765 billion for 2026. The bank’s cumulative forecast stretches to about $7.6 trillion from 2026 through 2031.
Why RSI changes the calculus
DeepMind CEO Demis Hassabis has referenced RSI as a crucial focus for the lab’s development roadmap in 2026. The concept sits alongside multi-agent systems and traditional scaling as pathways not just toward AGI, but toward what researchers call Artificial Superintelligence, or ASI.
AGI is roughly defined as AI that can perform any intellectual task a human can. ASI goes further: intelligence that surpasses human capability across every domain.
What this means for crypto and digital infrastructure
AI data centers are already competing with Bitcoin miners for power capacity in key markets like Texas and the Nordics. A 77% year-over-year increase in hyperscaler spending means more facilities, more cooling, and more strain on power grids.
For crypto investors specifically, the Goldman Sachs estimate of $7.6 trillion in cumulative AI capex through 2031 creates a sustained demand environment for the raw inputs of computation: energy, semiconductors, and networking infrastructure.