Aura Digital
Sofia · 42.69°N 23.32°E
All articles
Guides · September 2, 2026 · 8 min read

What a GPU Server Actually Earns on Vast.ai: The Real Math

The income side of renting out a GPU server: six weeks of real median rates, the three costs that eat the margin, and why nobody promises you utilization.

A row of closed server cabinets down a data centre aisle

TL;DR

The income side of a GPU server is arithmetic, and every term in it is knowable except one. Market rates are real and we track them weekly. Power draw is measurable. The platform's commission is published. Utilization is not knowable in advance, and anyone who quotes you a payback period without saying what utilization they assumed is selling, not calculating.

This article gives you the arithmetic, six weeks of real median rates from our own archive, and three scenarios at different utilization levels. It does not tell you what your machine will earn, because we cannot know that.

The formula

gross   = market rate ($/h) × GPU count × utilization × 730 h
yours   = gross × your share after commission
power   = kW × 1.3 × 730 h × utilization × electricity price
net     = yours − power

The 1.3 is the system factor: CPUs, RAM, NVMe, fans and PSU losses draw power alongside the cards. Vast.ai keeps a commission and typically leaves the host 80-85%. Our GPU ROI calculator runs exactly this formula with live rates and your own numbers.

What rates actually did: six weeks from our own archive

We refresh the rate table on this site weekly from Vast.ai's public offers, and git keeps every version. That makes our repository an accidental archive of the rental market. Here is the median rate per GPU-hour:

GPU 07 Aug 20 Aug 31 Aug Change
B200 5.96 5.88 6.00 +1%
H200 4.21 3.95 4.18 −1%
H100 SXM 2.10 2.07 2.37 +13%
RTX PRO 6000 WS 1.05 1.09 1.20 +14%
A100 SXM4 0.93 1.00 0.99 +6%
RTX 5090 0.34 0.38 0.45 +32%
RTX 4090 0.35 0.38 0.36 +3%
RTX 3090 0.12 0.12 0.12 0%

The one interesting thing in that table

The RTX 5090 rose 32% in 24 days. Nothing else moved like it - not its own sibling the 4090 (+3%), not the previous generation 3090 (flat), and not the flagship datacenter cards, which went nowhere at all.

The pattern is not "consumer versus datacenter". It is that the mid tier is where demand is moving: H100 +13% and RTX PRO 6000 +14%, while B200 and H200 sit still at the top. The 5090 is the outlier in its own class.

Two honest caveats. Twenty-four days is a short window - six samples cannot separate a trend from a fluctuation. And a median across listed offers can move because the mix of hosts changed, not because anyone raised a price: a batch of cheap new listings pulls it down without a single price cut.

The number nobody can promise you

Utilization decides everything, and it is the one term we deliberately do not publish a figure for. Not because we are hiding it, but because our number would tell you nothing about yours. Occupancy depends on your region, your network, your reliability score, your price relative to everyone else's that week, and how many machines like yours came online.

So instead of a single answer, here is the same machine at three utilizations. 8× RTX 5090, median rate $0.446/h - the unrounded figure behind the 0.45 above - 730 hours, 82.5% host share, electricity at $0.15/kWh, system draw 5.98 kW:

Utilization Gross Your share Power Net
100% $2,605 $2,149 $655 $1,494/mo
60% $1,563 $1,289 $393 $896/mo
30% $781 $645 $196 $448/mo

Note what the table does not do: it does not divide by a hardware price to produce a payback period. Your build cost is yours - put it into the calculator with the utilization you consider realistic, and treat any number that arrives without an assumption attached as marketing.

What eats the margin

  • The commission. 15-20% off the top, before any cost.
  • Electricity, and it does not scale down as fast as you hope. An idle machine still draws standby power, and the 1.3 system factor applies whether the GPUs are rented or not.
  • Being unremarkable. The reliability score is what gets you rented over an identical machine. It is earned over months and lost in one bad week.

When it does not pay

If you need the machine yourself more than half the time, the rental question is the wrong one - see renting versus buying. If your electricity is above ~$0.25/kWh, the bottom half of the table stops working. And if you are buying hardware purely to rent it out, understand that you are taking a position on a market whose prices moved 32% in one direction in three weeks, and can move the other way just as fast.

None of this is a forecast, and none of it is advice about what you should buy. It is the arithmetic, with the real inputs we can measure and an honest gap where the unknowable one sits.

We build the machine, you decide the business

We design and build GPU servers for clients - 8-GPU platforms, tested under full load before they ship, with EU warranty and RMA. What the machine earns afterwards is your market and your call. The configurator returns a real quote within 3 working days, and the ROI calculator runs your own numbers against live rates.

#vast.ai#gpu server#roi#gpu hosting#ai инфраструктура#наем на gpu

We build custom AI servers

From a single GPU workstation to a warranted server - we design, assemble and maintain. Including SEV-SNP builds with RTX PRO 6000 / H200.

Get in touch
Call Us