A European initiative organised by Kraftcompute
How contested capacity is allocated
Three independent members rank requests only when more than one team wants the same capacity. The provider always has the final say on its own donation.
How it works
- Kraftcompute sends the competing requests to the panel. The panel replies in writing within 5 working days.
- Each member scores each request against the published criteria. The ranking and a one-line reason are published on the ledger.
- No team receives more than 40% of allocated GPU-hours in a quarter unless nobody else applied.
- Members serve 12 months, unpaid. Kraftcompute acts as secretary and has no vote.
Criteria and weights
Weights are a starting proposal.
| Criterion | Weight | What scores well |
|---|---|---|
| Openness of outputs | 30% | Weights, code and results released under open licences, with a date |
| Readiness | 25% | Can start within the provider's window; software stack already runs on that GPU type |
| Research and public value | 25% | Fills a gap public allocations don't cover; benefits more than one group |
| Efficient use | 20% | Realistic GPU-hour estimate; can pause and resume; will report utilisation |
Conflicts of interest
A member with a link to an applicant team or to the provider steps aside for that decision.
The three seats
Members are named here once they have agreed. The aim is three countries across the three seats, and no member employed by a compute provider.
| Seat | Profile |
|---|---|
| Model builder | An academic who has trained large models and is not on an applicant team |
| Allocation practitioner | Someone who has run or reviewed compute allocations at a national centre |
| Open-source voice | Someone from a non-profit in open-source AI or open science |
To suggest a panel member, use the contact page.