Home / Compute & infra

Compute & infrastructure

A pooled GPU cluster and cloud-credit pool for active members. Enough compute to test an idea — not enough to brute-force a problem, by design. Allocations scale with contribution level, but every active member gets a baseline and project leads always get more.

8
H100 GPUs
alphabell cluster
16
A100 GPUs
+ idle-time donations
$420,000
Cloud credit pool
AWS, GCP, Lambda
94%
Utilisation · 30d
healthy, not pinned

What's in the pool

ResourceQuantityWhereNotes
H100 80GB SXM8Helsinki + SingaporeShared via Slurm; max 7-day reservations
A100 80GB16Cluster + 4 community donorsIdle-time priority for project leads
A100 40GB / L40s~30Cloud passthroughAllocated against $ credit pool
Cloud credit$420,000AWS, GCP, LambdaQuarterly burndown
Open datasets~3 TBS3 mirror, eu-north-1Free for all members
Eval gradersSelf-hostedUsed by competitions; free to run for any project

How allocation works

  1. Every active member at L4+ gets $50/month in cloud credit by default.
  2. Project leads on accepted projects get up to 1 H100-week + $500/month by default.
  3. Fellows get $25,000 in cluster + cloud across their fellowship.
  4. Special-request bursts (3+ GPU-weeks, or $5k+ in cloud) get reviewed by the compute committee. Decisions in 7 days.
  5. Quarterly burndown: unused allocation rolls 50% into the next quarter; the rest goes back to the pool.

The compute committee

Compute is allocated by a 3-member committee, rotating quarterly, drawn from L8+ members. They are paid for the work. Their goal is to maximise research output per GPU-hour across the lab — not to maximise individual fairness. They publish a quarterly report showing where compute went, what came out of it, and what they would do differently.

Donating compute

If you have spare GPU time at home or in a lab, alphabell can use it. We have a turnkey contributor stack — Docker container, sandboxed, opt-in workloads only, audit logs. Donors earn contribution points for every GPU-hour donated.

Read the contributor docs