The goal is a build around ~1M rub that maximizes income and passes vast.ai verification without hitting the known traps.
Key criteria
- GPU — demand and $/W: bet on RTX 4090 / 5090
- RAM — 32–64 GB per card (low RAM = fewer rentals)
- Disk — NVMe SSD: from 500 GB per server + 300 GB per additional card
- PSU with 20–30% headroom, 80+ Gold/Platinum
- Cooling — open-air frame with directed airflow
- NIC — judge by measurement, not by vendor: cheap gigabit Realtek parts on PCIe x1 cap upload at ~330 Mbit, while modern Realtek 2.5GbE and Intel deliver an honest link
- CPU — many-core (EPYC / Threadripper / Ryzen 9)
Three builds for about 1M rub
- A — one powerful 4×4090 rig (~950k rub): max income and simplicity, but concentration risk; disable VMs due to the reset bug
- B — two 2×(2×4090) rigs (~1M rub): diversification, but pricier and more complex
- C — 2×5090 (~820k rub): top demand, VMs possible (2 cards — mitigations needed), headroom to grow
| Build | GPU | Cost | VM | Profile |
|---|---|---|---|---|
| A | 4×4090 | ~950,000 rub | no | max income, simple |
| B | 2×2×4090 | ~1,000,000 rub | no | diversification |
| C | 2×5090 | ~820,000 rub | yes* | top demand, headroom |
Pitfalls: choose the NIC by measured upload, not by vendor. On VMs: failures returning a card from a VM only happen on builds with TWO OR MORE cards — on a single-card server VM works normally on any generation. * On multi-GPU builds: RTX 4090 + VM = reset bug, the card dies on exiting a VM and only a cold reboot brings it back — keep VMs off permanently and rent containers; on 5080/5090 the form is milder (sticking in P0, cured by a host reboot), and the 50 series needs -open drivers only, the whole nvidia stack on apt-mark hold and nvidia_drm.modeset=0 pcie_aspm=off in GRUB. Also note: vast does not bill traffic for VM rentals at all, and the docker audience is wider than the VM one. Do not skimp on RAM; size the PSU from the whole-rig peak: the cards draw 4×450 W but the rig peaks at 2.2–2.4 kW.