The thesis is elegant: AI compute demand is exploding; centralized cloud providers cannot keep up; therefore, distributed GPU networks will capture meaningful share. Each step is plausible. The conclusion does not follow automatically.
The supply-demand gap is real
H100 delivery times were measured in months throughout 2023 and into 2024. Frontier model training requires clusters of thousands of GPUs operating with near-zero tolerance for network latency. The hyperscalers — AWS, Google, Azure — have structural advantages in procurement, power infrastructure, and data-center geography that are not easily replicated.
Decentralized networks address a different part of the market: inference workloads that are less sensitive to inter-GPU latency, smaller training runs that do not require thousand-card clusters, and geographies underserved by major cloud providers.
Economic models under pressure
Most decentralized GPU networks operate a two-sided marketplace: GPU owners supply compute; AI developers demand it. The marketplace earns a take rate on transactions. This model works if the supply side is disciplined and the demand side grows faster than hyperscaler capacity.
The failure mode is familiar from other two-sided markets: a race to the bottom on pricing, undifferentiated supply, and a hyperscaler pricing response that makes the decentralized option structurally uncompetitive for workloads that could run on either.
Where we see durable differentiation
Networks that build differentiation through privacy-preserving computation, geographic compliance, or specialised hardware (inference chips rather than training GPUs) have a more defensible position. The commodity compute market is not the right fight.