Slurm with a wallet: metered compute for offices, same pattern as WiFi · content/office-compute-hub.md
The Office Compute Hub — metered compute on the local network
Date: 2026-09-13 · Founder's vision: "data exchange similar to WiFi should also work for transferring compute power from one computer to another — very helpful for local compute hubs in offices and seeing user activity." · Companions: qalhotspot (the WiFi twin — the same escrow+meter+settle pattern) · compute market · the Fleet
The insight: WiFi and compute are the SAME pattern
| qalhotspot (WiFi) | Office Compute Hub | |
|---|---|---|
| The scarce resource | Starlink backhaul | GPU-hours on the office machines |
| Who has it | the operator (one payer) | the office (or one beefy workstation) |
| Who wants it | village guests | colleagues + their AI agents |
| Metered by | MB | GPU-seconds · tokens · job-count |
| Escrow | hold at connect | hold at job submit (already live in qmarket!) |
| Trust | receipts anchored | receipts anchored (already live!) |
| Zero-rated tier | messaging always free | agent identity/heartbeat always free |
| Cash onboarding | voucher keys | voucher keys (same system) |
| Auto-billing | auto-pay keys | auto-pay keys against a daily compute budget |
The office hub IS the compute market (qmarket), pointed inward, with the hotspot's billing UX. The primitives already exist — what's new is the packaging.
The office architecture
┌─── The Office (Tailscale or LAN) ────┐
│ │
Colleague A ──────┤ "run this training job" │
(AI agent, capped │ │ │
AU$20/day) │ ▼ │
│ OFFICE HUB (qalhotspot pattern) │
Colleague B ──────┤ "render this video" │
(human, voucher │ │ │
key from admin) │ ├── Machine 1 (RTX 4090) ──► provider
│ ├── Machine 2 (A100) ──────► provider │
The Admin ────────┤ └── Machine 3 (idle laptop) ► provider (CPU)
(sets caps, │ │
sees activity) │ ▼
│ every job: escrow → run → attest → receipt → settle │
│ activity feed visible to the admin (anchored) │
└───────────────────────────────────────┘
What the admin sees (the "user activity" ask)
- Live board: every machine in the hub — who's using it, what job, how many GPU-seconds, cost
- Per-colleague daily spend: their agent's cap, spent today, remaining
- Job receipts: every completed job has an anchored receipt — the audit trail is public to the office, tamper-evident
- The trust model: escrow before compute (nobody freeloads), attested outputs (nobody fakes), slashing if a machine lies about what it ran
The three onboarding paths (same as WiFi)
| Path | Who | How |
|---|---|---|
| QALS account (auto-pay) | regulars, their AI agents | authorize a daily compute budget → every job auto-escrowed from it |
| Voucher key (prepaid) | visitors, cash-payers | admin creates a "50 GPU-hour" voucher → the visitor redeems it → compute up to that limit |
| Free tier | identity-only tasks | agent heartbeat, identity proofs, small scheduled jobs — zero-rated (the "messaging always free" equivalent) |
What's already built vs what's the build
| Piece | Status |
|---|---|
| Compute escrow (hold → job → settle at actual) | LIVE (qmarket, 62+ assertions) |
| Provider registration + benchmarking | LIVE (qmarket provider.py) |
| Receipts + anchoring | LIVE (ComputeReceipt pattern) |
| Per-agent caps | LIVE (loopd — owner-set, optional; uncapped by default) |
| Voucher keys | BUILT (qalhotspot voucher system) |
| Auto-pay authorization | BUILT (qalhotspot autopay) |
| The office packaging (the hub UI, the admin board, LAN discovery) | the build — ~1-2 weeks |
The build order
- v0 (days): a config file on one beefy machine lists "office providers" (colleague machines that opted in); a simple "office-hub" HTML page (like the phone bridge) shows the live board: jobs, machines, spend. The voucher/auto-pay systems from qalhotspot are re-pointed at compute instead of MB.
- v1 (+2 weeks): LAN/mDNS discovery of providers; the admin dashboard in the Tauri wallet's Network tab; job submission from any colleague's phone (bridge).
- v2: office-to-office federation (two offices share compute across Tailscale — the same escrow, the same receipts).
Why this matters commercially
The office hub is the enterprise wedge: every company with a GPU workstation has idle compute and colleagues who want it. The current alternatives (Slurm, manual queues, "just SSH in") have no billing, no receipts, no caps, no agent-safety. The office hub is Slurm with a wallet — and the wallet is what makes it safe to say "yes, your agent can use my GPU."