The host channel in Russia is the second bottleneck after electricity. A slow, asymmetric or unstable channel means low utilization, a failed bandwidth test and lost rentals.
Vast.ai requirements
- Bandwidth scales with VRAM: ≈2.6 Mbit/s per 1 GiB of TOTAL machine VRAM (all cards combined), but no less than 100 and no more than 500 Mbit/s — separately for download and upload
- Symmetry matters: clients upload results and models, so upload is as critical as download
- Low packet loss and a stable channel (high loss = dropped rentals)
- Latency is not critical for batch ML
Example: a machine with a single RTX 5090 (32 GB) → 2.6 × 32 ≈ 83 Mbit → the 100 Mbit floor kicks in, so such a node already meets vast.ai’s requirement at ~100/100. But the SUM across all cards is what counts: a 4×RTX 4090 rig (96 GB) → 96 × 2.6 = 250 Mbit, and the 500 Mbit ceiling is only reached from roughly 192 GB of total VRAM. Symmetric ≥500/500 is the ceiling of the requirement and good headroom, not a mandatory minimum.
The main problems in Russia
- Tariff asymmetry: often 1000 down / 100–300 up. Such an upload usually passes the formal requirement (the floor is 100 Mbit), but you lose to symmetric nodes in catalogue filters and when clients push results back
- Blocking: Docker Hub and the bandwidth test are throttled by DPI → a VPN tunnel via a foreign VPS is needed. Distinguish two things: protocol-level DPI is cured by tunnel settings (port, SNI, flow), while throttling by the VPS host's AS is cured only by moving to another, niche AS — changing SNI or port does nothing against it
- Network hardware or driver can underperform even on a good tariff
From the field: cheap gigabit Realtek chips (RTL8111/8168 on PCIe x1) cap upload at ~330 Mbit — but that is a property of the chip, not the vendor: in our fleet the rig on Realtek 2.5GbE does 706/625 Mbit while the rig on an Intel I225-V does 524/518, so judge by measurement. NIC offloads (GSO/TSO/GRO) cause loss on a link BEHIND A TUNNEL — disabling them there adds 40–55%; on a direct link they are useful instead. An ISP can silently cut the channel (one case went 680 → 19 Mbit); rentals push hundreds of TB per month — unlimited traffic is mandatory. Measured: a single rental moved 8.32 TB in 37 hours, about 7.5 TB per day.
To start, get a symmetric ≥ 500/500 plan (ideally gigabit symmetry) with unlimited traffic and a static IP, and plan from the outset for the channel to run through a VPN tunnel.
About traffic money: the host sets a $/TB price in both directions, but it only works on docker rentals. Vast does not bill traffic for virtual-machine (VM) rentals at all — this is the official support answer. Verified on our own machine: with $3/TB set, a client moved 8.32 TB while the report showed Upload $0.00 / Download $0.00, whereas a neighbouring docker machine did accrue traffic charges. The consequence: a VM node with a fast link attracts people who want free transit rather than GPUs; the defence is a bandwidth cap on the router or a docker-only setup.