Vultr Joins National Compute as Founding Member of Shared AI Compute Pool

Cloud provider Vultr has joined National Compute as a founding consortium member, contributing capacity to a shared pool meant to make frontier-class AI compute easier to reach for researchers, startups, universities and government labs.
- Vultr has joined National Compute as a founding consortium member and will contribute capacity to a shared pool.
- The pool, called the Grid Exchange, lets research teams reserve compute on demand instead of relying only on long-term commitments.
- The consortium targets 2 GW of pooled capacity by 2030 across NVIDIA, AMD and TPU architectures.
- The Grid prices around "goodput", meaning completed, uninterrupted work, rather than GPU-hours.
- According to National Compute, 10 diverse participants can cut waste per FLOP by roughly 3x and 100 participants by roughly 10x.
A shared pool called the Grid Exchange
National Compute is bringing together compute capacity from different clouds, sites and chip architectures and linking it through an orchestration layer. According to Vultr, the result is the Grid Exchange, a unified pool in which research teams can reserve compute when they need it instead of relying only on long-term infrastructure commitments.
Vultr says Grid Exchange is designed to become the world's largest burst-compute pool for AI. It would give startups, universities and government labs access to capacity that would traditionally require hyperscaler-scale contracts.
The consortium's long-term goal is to reach 2 GW of pooled capacity by 2030. The pool is to span a diversified silicon base that includes NVIDIA, AMD and TPU architectures.
Pricing around goodput instead of GPU-hours
Vultr points out that traditional cloud infrastructure is often priced by the amount of infrastructure consumed, such as GPU-hours. The Grid takes a different approach and prices around "goodput", which Vultr's post describes as completed, uninterrupted work.
The reasoning is that a GPU which is allocated but idle, waiting on another resource, or interrupted mid-job does not deliver the same value as one that keeps a workload moving toward completion. According to Vultr, the Grid's model rewards providers that operate reliably, keep availability high and engineer for efficient utilization.
The post also says the Grid's architecture builds on systems developed to improve utilization across Google's fleet, where similar approaches helped raise occupancy from approximately 65% to 99% at a gigawatt scale.
Why pooling could lower barriers
Vultr argues that the current model for frontier AI research creates significant barriers to entry. Organizations often have to secure large amounts of compute through long-term commitments before they know how their research will develop. Startups, academic researchers and government laboratories may lack the scale or resources to negotiate hyperscaler-sized contracts years ahead.
Pooling relies on statistical multiplexing, which lets diverse workloads share a larger set of resources and reduces the capacity that must sit idle to cover unpredictable demand. According to National Compute, 10 diverse participants can reduce waste per FLOP by roughly 3x, while 100 participants can reduce it by roughly 10x. Vultr notes that the value of the pool grows as more providers, workloads, modalities and regions join.
The Grid is also built to be heterogeneous, pooling capacity across NVIDIA, AMD and TPU architectures and across multiple cloud and infrastructure providers. Vultr says this matters as organizations optimize around performance, availability, cost and workload-specific needs rather than defaulting to a single provider.
Policy context and what it means
Vultr ties the initiative to national priorities, including the U.S. AI Action Plan's emphasis on a healthy and competitive AI compute market and the Genesis Mission's objective of accelerating scientific discovery and doubling the productivity of American science.
For researchers, the pitch is flexibility: reserving capacity when needed rather than committing years in advance. For infrastructure providers such as Vultr, it offers a way to move beyond static capacity commitments, with available compute flowing to the workloads that need it. Vultr describes its own role as contributing high-performance infrastructure to the network. The source does not give details on pricing, availability dates or how much capacity Vultr will contribute.


