OpenAI’s Head of Compute Infrastructure, Sachin Katti, dropped a blunt assessment during a recent podcast: demand for computing power in the AI sector is vastly outpacing supply, and every new unit of capacity gets consumed the moment it comes online. His warning carries a particularly uncomfortable corollary. Slowing down the expansion of capacity doesn’t just delay progress. It actively makes the shortage worse.

The numbers behind the crunch

OpenAI’s available compute power sat at 0.2 GW in 2023. By 2024, that figure tripled to 0.6 GW. The target for 2025 is approximately 1.9 GW, which would represent another roughly 3x jump in a single year.

OpenAI is spending around $50 billion annually on compute resources, primarily through its partnership with Microsoft. The exponential growth isn’t coming from one place, either. It’s the compounding effect of more users, more AI agents, and more always-on applications all demanding inference simultaneously.

OpenAI isn’t alone in sounding the alarm

Katti’s warning echoes sentiments previously expressed by OpenAI CEO Sam Altman and CFO Sarah Friar, both of whom have flagged chronic compute shortages as a bottleneck that hinders product rollouts and reliability.

Data center power needs for AI workloads are projected to reach tens of gigawatts globally in the coming years. OpenAI has responded by diversifying its hardware partnerships beyond its historically deep relationship with Nvidia. The company is now working with AMD, Broadcom, and Cerebras to source alternative computing hardware. The company is also investing in hardware co-design, essentially working with chipmakers to build custom silicon optimized for its specific workloads rather than relying on general-purpose GPUs.

Where crypto enters the picture

The chronic compute shortage is generating real interest in decentralized GPU networks as a potential complementary solution to traditional centralized data centers. Platforms like io.net, built on Solana, are positioning themselves squarely in this niche.

For certain inference workloads, fine-tuning tasks, and less latency-sensitive applications, decentralized compute could carve out a meaningful role. When a single company is spending $50 billion annually on compute and still can’t get enough, even capturing a small fraction of that demand represents a massive addressable market for decentralized alternatives.

What this means for investors

For crypto investors specifically, the narrative around decentralized physical infrastructure networks, commonly called DePIN, gets a tangible demand-side catalyst. The projects that solve for reliability and enterprise-grade security will separate themselves from those that remain interesting experiments.

As AI companies diversify away from Nvidia’s dominance, the resulting fragmentation in hardware could benefit decentralized networks, since they can abstract away the complexity of managing different chip architectures behind a single interface.