How much power a GPU server draws: fleet measurements, a 16 A line and cooling

Where the server sits, and whether the site has enough power, cooling and stable electricity, is the first question that breaks a profitability calculation. Cards can be bought in a week; a dedicated electrical line in an old building cannot. In practice the site, not the hardware, most often limits hosting on vast.ai. Below: a power measurement across our fleet, the line calculations, and a section on our own mistakes.

Peak, average and the number on the bill are three different things

A 4×RTX 4090 machine on a server platform: peak ≈ 2.2–2.4 kW, steady 1.8–2.0 kW under ML load. The average depends on how much the cards actually compute: about 1.0 kW at a realistic 50%, about 1.4 kW at 70%. Two servers already peak at ~4.5 kW. Do not confuse three numbers: the peak sizes the breaker, the cable and the power supply, the average produces the electricity bill, and the sum of the cards' rated power (4×450 W = 1800 W) is neither.

A breaker does not trip on the average: a client launches training and four cards hit their limit within seconds, along with the CPU and the drives. That is why the line and the power supply are sized from the whole server's peak with 20–30% headroom: the cards account for 1800 W while the server draws 2.2–2.4 kW at the wall — the difference goes to the EPYC, the memory, the NVMe drives and power-supply losses. The opposite mistake costs as much: multiply 2.4 kW by 730 hours and your electricity line is overstated roughly threefold.

What we measured on our own fleet

A single snapshot proves nothing, so we took a series: fifteen samples 20 seconds apart, five minutes of observation per machine; the table shows averages and spread. Some of the cards were under a paid rental.

Two rows explain why a spec-sheet calculation fails. An RTX 5090 with a 600 W limit averaged 43 W: the server was rented and paid for, 30 of its 32 GB of memory occupied, and the fans never moved once — the card earns money at seven percent of its rated maximum. At the other edge of the table an RTX 5080 sat at exactly 360 W for all five minutes. The gap between paid idling and real work is roughly tenfold.

Two conclusions follow. Size the line, the breaker and the power supply from the peak — it arrives suddenly. Calculate the electricity bill from a measurement, or at least from the range between paid idling and a pinned card: plug your own cards, tariff and utilization into the calculator at the income calculator, and how paid hours diverge from consumed kilowatt-hours is worked through in our income breakdown card by card.

Electrical: what to check before you buy

• The capacity granted by your supply contract, not the thickness of the wire: a flat usually has 5–10 kW in total (one server at the limit), a house ~15 kW, and three-phase 380 V removes the question almost entirely.
• A dedicated line from the board with its own breaker, at least 2.5 mm² copper and sockets rated for the current — not an extension cord and not the kitchen circuit. One 16 A line (≈3.5 kW) covers exactly one 4×4090 server; a second server needs its own.
• Continuous load no higher than 80% of the breaker rating: a C16 realistically means ~2.9 kW, and a 2.2–2.4 kW peak only just fits. Come summer an air conditioner joins the circuit and the breaker starts tripping while you are away.
• An RCD or RCBO on the line: a metal chassis, two kilowatts, and a person who will reach inside without cutting the power.
• A stabilizer for sags: in rural areas an evening 190 V is not unusual, and power supplies tolerate it worse than a short outage.
• The tariff: across 730 hours a month, the difference between 4 and 8 rub per kWh is the difference between a working business and one that never pays back.

What it costs per month

We calculate for one 4×4090 server over 730 hours, as "average power × hours × tariff". The average comes from the share of time the cards actually compute, not the share of paid hours. The first row of the table is not a hypothesis but our measurement: four cards idling under a paid rental, plus the platform, is about 0.35 kW.

A reality check: our actual bill for the whole fleet — six machines, ten cards — is about 6,000 rub a month. The model gives the same figure for one server at 70% and 6 rub per kWh. Two reasons: the cards are not pinned most of the time, and our tariff is below the typical city rate. The conclusion is not that electricity is nearly free, but that it has to be calculated for your own site. The full cost structure and payback — our income breakdown card by card.

Cooling: where two kilowatts of heat go

Almost all the power turns into heat: a 2 kW server is two oil heaters running non-stop. To remove 2 kW with a 10 °C air temperature rise you need to move about 600 m³/h through the room. A fan in a window does not deliver that; a 7 kW air conditioner does — but draws about 2 kW itself.

• Separate intake from exhaust: cold air to the front panel, hot air out. A room with one fan blowing "somewhere" does not work: the hot air goes round in a circle.
• Target below 80–83 °C on the cards under load. Above that the card throttles itself, and the client sees a slow server rather than a hot room.
• Plan for July, not January: in winter the server's heat is your heating, in summer it is a second cost line.
• Dust: an open-air frame in a living room clogs up within a month — an intake filter is cheaper than stripping the server down.
• Noise: a server platform with 40 mm fans at full speed is impossible in a living space.

Site networking and power: so a new server does not take down the rented ones

• A dedicated router rather than the ISP box, plus a managed switch: a new server plugs into a free port. Reconfiguring the router for the eighth server means downtime on the other seven.
• A UPS on the platform, the network and the router. Keeping two kilowatts of cards on batteries is pointless — the rental is interrupted anyway; the UPS's job is to prevent a dirty shutdown.
• Auto-restart after a failure: set Restore on AC Power Loss to Power On in the BIOS. Without it a power cut while you are away means a day offline: about $29 on a 4×4090 machine, plus a reliability dip.
• Remote power control: IPMI/BMC on a server board gives a reboot and a console when the system stops responding. Consumer boards have none — there you need a controllable socket.
• Bandwidth is part of the site: the marketplace requires 2.6 Mbit/s per 1 GiB of VRAM, at least 100 Mbit/s, with no upper limit; symmetric 500/500 is our target with headroom, and no data cap is needed because of rental traffic volume. Measure your link — during a free consultation; requirements and ISP traps — what network a host needs.

How much a home network takes part in hosting is visible in the traffic: one virtual-machine rental pushed 8.32 TB through our router in 37 hours — about 5.4 TB a day. With our price set at $3 per TB, the amount credited was $0.00: the marketplace does not bill virtual-machine traffic at all, as support confirmed. The site has to withstand transit you are not paid for.

Where we got it wrong

None of the problems below was a problem with the graphics cards. Every one of them was about the site, and about what we ourselves treated as secondary.

• The home router turned out to be a single point of failure for the whole fleet: on 20 September the tunnel on one machine came up before DHCP, packets went into its own interface and 28 thousand sockets filled the router's connection table — and four machines falsely switched to the backup VPS.
• We relied on thresholds instead of link quality: in the same episode interference on the home link crossed the automatic-failover threshold and four machines changed their IP for no reason, and changing an address penalizes reliability.
• Not every outage is power. On 25 August four machines went offline, and the electrical supply had nothing to do with it: the marketplace controller was stuck, fixed by resetting /var/lib/vastai_kaalia/controller_connection (write-up: the stuck controller write-up). Separately we caught the agent hanging on the marketplace's own web API: a blocking HTTP request with no timeout.
• Instruments do not always diagnose: memtest never saw a dead memory module on the 4090 server; we proved the cause by moving modules between slots and watching swap activity.
• Silent losses are worse than failures: the marketplace's own hourly job wipes the VM support flag when all cards are passed through to guests — a server with two 5080s stood without VM mode for ten days, from 19:22 on 11 September to 17:04 on 21 September.
• We picked the tunnel host without checking its autonomous system: download speed is throttled precisely by AS — OVH and M247 were choked for us, while the niche Private Layer was clean; the list changes over time, so re-measure.

The overall conclusion: a site is not a socket. It is the supply, the board, cooling, the router, the switch, the UPS, the startup order of services and an external VPS. Each of them has cost us income at least once, and none appeared in the hardware budget. What the machines earn right now is at our live fleet page, refreshed hourly.

Related reading: hardware and builds — what hardware to buy, bandwidth and ISPs — what network a host needs, economics and payback — our income breakdown card by card, our month card by card — our income breakdown card by card, installing the OS — installing Ubuntu for Vast.ai, VM mode on a multi-GPU server — VM mode on a multi-GPU host. Work out your own site — the income calculator. We can run the numbers for your hardware — our services.