AI & Robotics
Open Compute Networks
Decentralized AI training and inference across compute hardware, energy grids, and sovereign regions
AI compute infrastructure is concentrated within a handful of hyperscalers and large labs, limiting experimentation and participation. That concentration also lets usable compute sit idle across edge environments. Additionally, a growing set of regions and institutions look for AI infrastructure they control rather than rent from a few central providers. Decentralized compute markets could pool capacity across where power is available, harnessing compute at the edge, and giving regions sovereign alternatives.
Inflection Point
A decentralized compute network trains a competitive model—demonstrating that decentralized infrastructure can match centralized systems for meaningful AI workloads, breaking the assumption that only hyperscalers can develop frontier AI.
AI infrastructure becomes globally decentralized. Model training moves from centralized labs to open, coordinated compute markets.
Tipping Signals
The Opportunity
Decentralized compute networks become viable infrastructure for meaningful AI training and inference by pooling capacity across power-rich sites the grid can't serve, sovereign infrastructure that regions and institutions control rather than rent, and compute across edge environments. The shared problem across all three is coordination—matching demand to decentralized supply reliably. Open compute markets attract independent researchers, startups, and eventually mid-tier labs seeking alternatives to hyperscaler pricing and control. PL's role is to support the research and early builders coordinating this capacity.
Context
Concentration of AI infrastructure is the primary competitive moat. Whoever controls compute controls who can build frontier AI. Decentralized alternatives must reach performance and cost parity to meaningfully shift this dynamic.
Supply-side coordination is the hard problem. Building decentralized compute networks that maintain reliability, consistent performance, and economic sustainability requires solving hard mechanism design problems that pure technology alone does not address.
Multiple pressures are pushing compute out of centralized clusters. Power and grid interconnection are increasingly as binding as chips—queues run years and new generation sits stranded—but they are not the only force: sovereign demand for infrastructure independent of a few providers, and usable compute stranded across edge environments, all reward architectures that decentralize rather than concentrate. Compute is easier to relocate than power, politics, or hardware already in the field, so the strategic move is to coordinate capacity where it already exists rather than wait for the grid, or a single provider, to serve centralized clusters.
Open infrastructure enables open innovation. When AI training infrastructure is permissionless, independent researchers, small labs, and new entrants can experiment and innovate without depending on large-scale capital or hyperscaler access.
The window is open but may close quickly. As AI infrastructure consolidation accelerates, the opportunity to establish open alternatives narrows. Decentralized compute must demonstrate viability before centralized lock-in becomes irreversible.
Friction
Centralization accelerates as model scale grows. The compute requirements for frontier models grow faster than decentralized alternatives can scale, creating a widening gap between centralized and decentralized capabilities.
Reliability and performance variance remains a major barrier. Decentralized compute networks struggle to offer the consistent SLAs that production AI workloads require, limiting adoption to experimental and non-critical use cases.
Speculative token design undermines real utility. Many decentralized compute projects have been designed around token incentives rather than genuine infrastructure utility, eroding trust and slowing adoption.
No shared standards for compute coordination. Fragmented protocols, APIs, and orchestration layers prevent interoperability between decentralized compute networks.
Centralized compute faces compounding structural limits. Grid strain delays roughly a fifth of planned data center projects, and interconnection queues reach a decade in some markets; at the same time, reliance on a few providers leaves regions without a sovereign fallback, and large amounts of usable compute sit idle across edge environments. Decentralized coordination can turn power, sovereignty, and stranded capacity into supply.
Field Signals
# of competitive AI models trained on decentralized compute infrastructure
Sustained, non-speculative usage across decentralized compute networks
Price per GPU-hour on decentralized networks vs. hyperscaler equivalents
# of startups building AI products on decentralized compute rather than hyperscalers
# of decentralized compute networks using shared coordination protocols
GW of compute co-located with stranded or curtailed renewable generation, bypassing grid interconnection