Phase A — First server runbook (to live listing)¶
Goal: One 4× A100 (40GB or 80GB) server live in colocation, financed per term sheet, listed on Vast.ai (verified path), with insurance and ≥6 months fixed-load reserve.
Canonical strategy: BUSINESS-PLAN.md §9.0, §10, §14.
Before you spend money (gates)¶
| # | Gate | Pass criteria |
|---|---|---|
| G1 | Entity + tax | UG (or agreed structure) + Regelbesteuerung if Steuerberater recommends Vorsteuer path |
| G2 | Colo quote | Signed or firm: space, kW, remote hands, IPv4, SLA, exit terms |
| G3 | Finance quote | Payment, term, early buyout, personal guarantee scope (minimize / cap) |
| G4 | Hardware quote | 4× GPU + server BOM; native PCIe confirmation (see 01-GPU-RFQ.md) |
| G5 | Unit economics | Rebuild §9.0 with your colo + your monthly payment; confirm go thresholds |
| G6 | Liquidity | Bank balance plan: ≥6 months of (colo + insurance + maintenance + financing) after any down payment |
Week-by-week sequence (typical)¶
Weeks 1–2 — Structure and RFQs¶
- Steuerberater brief (entity, VAT, depreciation intent).
- Send GPU RFQ (DE/EN per supplier table in RFQ doc).
- Send colo RFQ (02-COLO-RFQ.md).
- Parallel (low intensity): Christine shortlists 2–3 RE sites for Phase C (no binding capex).
Outputs: Comparable colo + finance + hardware offers; updated §9.0 spreadsheet.
Weeks 3–4 — Contracting¶
- Pick colo + finance + supplier; legal skim on default, termination, buyout, SLA credits.
- Order hardware (align delivery with colo ready date).
- Elektronikversicherung + Betriebshaftpflicht bound before rack power-on.
Outputs: Signed contracts; insured asset schedule; install date.
Weeks 5–6 — Build and ship¶
- Assemble or receive integrated server; burn-in (stress +
nvidia-smi, ECC/Xid watch). - Ship / deliver to colo; label assets; IPMI + OS baseline documented.
Outputs: Burn-in log; access checklist (IPMI, SSH, monitoring).
Weeks 7–8 — Rack and network¶
- Install per provider; PDU / dual feed check; link up; latency smoke test.
- Backup access path (provider console / KVM).
Outputs: Rack photos (if required); network diagram; monitoring alerts.
Weeks 9–10 — Marketplace¶
- Vast host onboarding; verified checklist (drivers, persistence, templates).
- Listing copy + initial price band from latest snapshot methodology (
tooling/marketplace/). - Soft launch pricing → tune on realized net EUR/GPU-h and occupancy.
Outputs: Live listing; daily dashboard: util, $/h gross, fees, incidents.
Weeks 11–12 — Decision checkpoint¶
- Compare 90-day (or minimum 60-day) rolling metrics to BUSINESS-PLAN.md Phase A goals.
- Hold / optimize / exit per stop-loss (§9.0): e.g. 2-month rolling net < -€400/mo → freeze expansion.
Documents and tools (repo)¶
| Item | Path |
|---|---|
| BOM + price snapshot | 03-BUYING-LIST.md |
| GPU first-offer RFQ | 01-GPU-RFQ.md |
| Colo first-offer RFQ | 02-COLO-RFQ.md |
| Unit economics snapshot | 04-UNIT-ECONOMICS.md |
| Vast snapshots / reports | tmp/market-snapshots/, build/vast_market_report.md (regenerate via tooling) |
KPIs to track from day one¶
- Occupancy (rented GPU-hours / available GPU-hours).
- Realized net EUR/GPU-h (after platform fee + FX — define formula and stick to it).
- Incidents (tickets, downtime minutes, provider faults).
- Cash: colo + finance + insurance + maintenance vs revenue.
Post-live gates: 30 / 60 / 90 days¶
Use calendar days from first paid rental hour (or from listing verified live if you need a fixed start — pick one and keep it consistent).
| Horizon | Review focus | Pass (typical) | If fail |
|---|---|---|---|
| 30 days | Listing health, initial util, incident rate, net EUR/GPU-h vs model | No extended 0% / dead listing; burn-in issues resolved; cash not breaching stop-loss alone | Fix template, pricing, verification tier, colo network; pause any expansion talk |
| 60 days | Rolling 30d occupancy stabilizing; support backlog; provider SLA | Occupancy trending toward ≥60% (or documented ramp reason); 2-month rolling net above stop-loss floor in §9.0 | Root-cause: demand vs price vs reliability; adjust ops budget |
| 90 days | Rolling 60–90d economics; reserve vs 6-mo fixed load; Phase B readiness | Meets Phase A goals in BUSINESS-PLAN.md §10; reserve still ≥6 months fixed load | Hold scaling; execute remediation or exit per §9.0 |
These gates are in addition to the week-11–12 checkpoint above (whichever calendar matches your ramp).
When Phase A is “go” for server #2¶
See BUSINESS-PLAN.md §10 Phase B: reserve floor, DSCR-style stress, no new uncapped personal guarantees.