National Compute's Genesis credits: what $100 million buys in AI compute
The consortium says single-tenant data centres average under 15% utilisation. The largest listed neocloud gets 98% of its revenue from take-or-pay contracts, leaving owners as the likely source of idle hours. OpenAI and Anthropic are expected to keep their capacity for themselves.
Expensive chips sit idle in AI data centres while smaller labs and startups struggle to rent compute time. National Compute, a new company that has pooled computing power from major technology firms, plans to donate $100 million in computing credits to the Trump administration's Genesis Mission, Politico reported, citing two people familiar with the matter and a draft paper on the project. The credits are expected to be announced Thursday at an event with Michael Kratsios, director of the White House Office of Science and Technology Policy. The White House declined to comment. Cofounder Anjney Midha says the aim is to "reduce the barrier to entry, increase compute access".
The capacity National Compute wants to pool sits in owner-operated clusters that run in bursts around training sessions. GPU clouds sell reserved blocks instead. The grid is therefore a market for interruptible hours, and $100 million is a small anchor order: about 15 megawatt-years at reserved rates, or roughly 2% of a year of the 760 megawatts the consortium says are connected or in sight. Companies that rent out compute have a reason to supply those hours. OpenAI and Anthropic, whose clusters train their own next models, do not. Politico reports that major providers often prioritise long-term contracts with the largest AI labs.
The likely supply comes from equipment owners. AI companies often use nearly all their capacity during training runs, while many chips sit idle between runs, according to Axios. The consortium's paper says independent single-tenant data centres average less than 15% net utilisation. Some idle time is headroom a lab keeps for its next training run. The grid lets members contribute unused capacity while reserving larger clusters for planned training, so supply depends on what each owner decides it can spare before its next run.
CoreWeave's annual report says committed, take-or-pay contracts accounted for 88% of revenue in 2023, 96% in 2024 and more than 98% in 2025. Its latest quarterly filing puts the share at 98% for the three months to 30 June. The largest listed neocloud has sold nearly all its capacity in long blocks, making it an unlikely source for the pool. Sam Sinha, head of AI at 1X, says larger players can pay far more and sign long-term contracts that smaller operators cannot.
CoreWeave's second-quarter results provide a price comparison: revenue of $2,575 million and 1.5 gigawatts of active power. Annualised, that is about $10.3 billion across 1,500 megawatts, or $6.9 million per megawatt-year. At that rate, $100 million buys about 15 megawatt-years. The consortium says 760 megawatts are connected or in sight and aims for 2 gigawatts by 2030. Politico's account of the draft paper puts capacity allocated so far at 750 megawatts, making the gift about 2% of a year of the 760 megawatts. Interruptible hours should cost less than reserved capacity, so the same money would buy more, but not enough to change the order of magnitude.
The gift is small beside the programme it supports. On 22 July, the White House announced more than $5 billion in federal commitments for the Genesis Mission, including 278 selected projects and a Department of Energy platform connecting researchers with data, compute and AI tools. National Compute's $100 million would equal 2% of that sum. The release does not mention the company, though it promised further announcements about industry and philanthropy. Politico reports that each public-sector user is also due $100 in National Compute credits on top of the donation, spreading the money across many small accounts.
The incentives point to who will sell. An owner with an uneven training schedule has idle hours but no easy way to sell them in short blocks without exposing weights or data, or risking interruption when its own run starts. A scheduler listing available capacity, chip type, location, price and utilisation could lower the search and credit costs. A government buyer would pay by the hour, while users would pay only for the power they use. Small clouds with unsold capacity have a reason to join; the named backers include Vultr, Crusoe and universities. The draft paper also lists labs and clouds running Nvidia and AMD chips, plus Google's tensor processing units, as possible sources. Google capacity could therefore come through cloud partners rather than Google's own training clusters. Genesis researchers holding credits bear the cost: interruptible hours are the only ones the grid can plausibly sell below reserved rates, so their jobs must save state and resume. The engineering time for that comes out of their project budgets.
The frontier labs still have a role in the plan. Politico reports that National Compute is launching Grid Defense, a programme in which government agencies test AI models through Marshall, an agent embedded in the grid. Labs then decide voluntarily whether to release a model. Agencies set their own procedures; Midha says they remain confidential in part so frontier models cannot learn to evade the evaluations. The programme entered invite-only early access with some agencies in June. Politico says it will likely depend on both labs and the federal government opting in. That would put OpenAI and Anthropic on the grid as subjects of testing, not suppliers of compute. A lab can submit a model for evaluation without putting its training cluster on someone else's scheduler. A cluster built for the next frontier run has no idle hours worth selling below what they are worth to the lab. By 31 March 2027, neither OpenAI nor Anthropic will be named as contributing its own cluster; the listed capacity will still come mostly from independent clouds, universities and partners selling Google's chips.
CoreWeave's third-quarter 10-Q, expected around mid-November, will show the share of revenue from committed contracts. Last year's filing came on 13 November. If idle hours were accumulating inside reserved neoclouds, the 98% share would start to fall as uncommitted capacity grew. It will not. The filing also shows $103.7 billion in remaining performance obligations, 41% due for recognition in the 24 months through June 2028. That is about $5.3 billion in contracted revenue per quarter, roughly twice the $2,575 million the company booked in the second quarter. CoreWeave's new capacity is already spoken for. Its third-quarter committed share will be 98% or higher, leaving about $50 million a quarter uncommitted, half the size of the gift itself.