Gavin Baker has a habit of asking uncomfortable questions at the right time. The managing partner and CIO of Atreides Management recently pointed out that roughly 500,000 people currently use agentic AI systems. Then he asked what happens if that number reaches 100 million.
The gap between 500,000 and 100 million is not just a number
To put Baker’s framing in context: 500,000 users represents a rounding error against global internet adoption. Alex Sacerdote flagged in June 2026 that current agentic AI usage sits around 0.1% of the global population. Baker’s hypothetical of 100 million users would push that figure to roughly 1.3% of the world’s 7-8 billion people.
Data-center GPUs are already running at over 70% utilization. The supply side has two primary bottlenecks: power measured in watts, and semiconductors measured in wafers, with TSMC sitting at the center of the latter constraint. Baker has indicated that meaningful relief on the power side is not expected until 2027-2028.
Why this cycle is different from the last bubble people think it resembles
GPU utilization above 70% is not a vanity metric. It reflects real workloads generating real revenue. Early agentic AI models including GPT-5.2, Grok 4.20, and Codex 5.3 are being deployed commercially, meaning the compute consumption driving those utilization rates is tied to actual economic activity.
What the compute shortage means for crypto and digital asset markets
Baker also referenced orbital compute as one innovative solution being explored to address power and physical space limitations. The concept involves deploying compute infrastructure in space to take advantage of solar power and thermal management conditions that are simply not replicable on Earth.
For investors watching AI infrastructure plays, the key variables to track are TSMC wafer capacity expansion timelines, power interconnection queues at major data center markets, and the pace at which agentic AI adoption moves from early adopters toward mainstream enterprise deployment.