Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
InfrastructureInsight · On RadarHabr AI→ original

Four Hours on an RTX 5090 Cost 348 ₽, or From 1 730 ₽ With a One-Day Minimum

Need an RTX 5090 for one evening, say to run a model for four hours? The bill depends on the minimum rental period. At GPUGO's rate of 87 ₽/h, four hours cost 348 ₽. Where the minimum is one day, you pay from 1 730 ₽, roughly five times more. Before ordering, check whether time is billed per second and what keeps being charged after you stop the machine.

AI-processed from Habr AI; edited by Hamidun News
Four Hours on an RTX 5090 Cost 348 ₽, or From 1 730 ₽ With a One-Day Minimum
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1GPUGO checked the listings of 8 services as of 20 November 2026
One day as a paid minimumEKACOD 2 604,80 ₽, FirstVDS 1 768 ₽ (out of stock)
Key metric #3At immers.cloud, a regular shutdown does not stop GPU charges; you need Shelve

Imagine you need an RTX 5090 one evening to run model inference for four hours, and by morning you won't need it anymore. It seems enough to compare the hourly price. In practice, three things determine the bill: the minimum rental period, the resources around the card, and what keeps being charged after you stop.

GPUGO, a Russian service that rents out GPUs itself, compared the public listings of eight providers as of 20 November 2026. At GPUGO's rate of 87 ₽/h, four hours cost 348 ₽ (4 × 87). At providers where the minimum paid period is one day, the same run costs 1 730–1 750 ₽, even if you release the card in the evening.

The difference is 1 382–1 402 ₽, roughly five times. A caveat: the comparison was published by a market participant, the prices come from catalogs, and no real orders were placed with competitors.

Pod or VPS: what exactly are you renting

In both formats, GPUGO provides a whole physical RTX 5090 with 32 GB of video memory. Everything else differs. A pod starts from a container image and uses the host machine's kernel. For a one-off run this is convenient: you take a ready-made image with PyTorch or ComfyUI and start working. A VPS boots a separate guest OS with its own kernel. You install system packages yourself, configure services through systemd, and set up an API and workers that survive a reboot.

A VPS does not add video memory. It does, however, come with more CPU in GPUGO's configuration: 14 vCPU, 56 GB RAM and a 300 GB disk for 87 ₽/h. A pod with 8 vCPU and 32 GB RAM costs 57 ₽/h, with disks billed separately. The extra cores help with video decoding, tokenization and preparing several batches of data. If a run is bottlenecked by the card itself, they won't speed it up.

Who charges how much for an RTX 5090

Demand is growing: according to Apple Hills Digital and Selectel, GPU infrastructure consumption in Russian public clouds reached 8,2 billion ₽ in 2025, 68,7% more than a year earlier. Of that, 3,9 billion ₽ went to dedicated servers.

What the listings show:

  • GPUGO: 1 GPU, 14 vCPU, 56 GB RAM, 300 GB disk. 87 ₽/h, 2 088 ₽ for 24 hours in the "For a Term" mode, 62 640 ₽ per month.
  • IQHost: the same configuration at 2 209 ₽ per day or 47 130 ₽ per month. The cheapest VPS on the product page costs 1 730 ₽ per day, with 72,08 ₽/h listed next to it.
  • FirstVDS: a whole card, 4 CPU cores, 16 GB RAM, 300 GB disk. 1 768 ₽ per day, but out of stock at the time of the check.
  • immers.cloud: 8 vCPU, 32 GB RAM, 60 GB disk. 130,76 ₽/h, per-second billing.
  • EKACOD: 14 vCPU, 56 GB RAM, 300 GB disk. 2 604,80 ₽ per day, with one day as the paid minimum.
  • Cloud4U: 66,18 ₽/h excluding VAT, but CPU and RAM are not listed in that line, and the final offer is available on request. This line was not included in the full-machine comparison.

The same card name does not make the machines the same. First normalize CPU, RAM, disk, network and rental period, then calculate the difference.

Where else you can lose time and money

You can't always get a card the same evening. At Yandex Cloud, the starting GPU quota is zero and must be requested separately; compute is billed per second, while disk and public IP are charged under their own rules. For GPU access in Cloud.ru Evolution, the documentation specifies a top-up of at least 15 000 ₽ with no bonuses, and a free card must be available.

Stopping doesn't always mean the charges end, either. Stopping a Python process, shutting down the machine and releasing the GPU in the control panel are three different actions. At Selectel, paid resources such as disks remain after a freeze. At immers.cloud, a regular shutdown keeps the GPU charge, while the Shelve operation releases compute resources, but storage and a saved IP may still be billed. At GPUGO, persistent storage is billed separately and survives a pod stop. For a VPS, there are restart, stop (which keeps the resources and the charge for them), and delete, which ends the rental.

In the "For Rental Time" mode, a 3,5-hour run at GPUGO costs 304,50 ₽. In the company's tests, the longest VPS launch took 1 minute 5 seconds, but that is the maximum across their runs, not a newcomer's path from registration to a working application.

What to do in practice

For a short experiment, look at per-second or hourly billing and at exactly how the card is released. For a permanent service, compare monthly rates: here IQHost looks cheaper than GPUGO. After your first short rental, check the billing log. For a repeatable measurement, the authors recommend two timers: from first visit until you can order a VM, and from order confirmation until a successful nvidia-smi over SSH. Next to them, record the image, the GPU model and the actual charge.

Story

Key Questions Answered

What are the key details on Pod or VPS: what exactly are you renting?

In both formats, GPUGO provides a whole physical RTX 5090 with 32 GB of video memory. Everything else differs. A pod starts from a container image and uses the host machine's kernel. For a one-off run this is convenient: you take a ready-made image with PyTorch or ComfyUI and start working. A VPS boots a separate guest OS with its own kernel. You install system packages yourself, configure services through systemd, and set up an API and workers that survive a reboot. A VPS does not add video memory.

Who charges how much for an RTX 5090?

Demand is growing: according to Apple Hills Digital and Selectel, GPU infrastructure consumption in Russian public clouds reached 8,2 billion ₽ in 2025, 68,7% more than a year earlier. Of that, 3,9 billion ₽ went to dedicated servers. What the listings show: GPUGO: 1 GPU, 14 vCPU, 56 GB RAM, 300 GB disk. 87 ₽/h, 2 088 ₽ for 24 hours in the "For a Term" mode, 62 640 ₽ per month. IQHost: the same configuration at 2 209 ₽ per day or 47 130 ₽ per month.

Where else you can lose time and money?

You can't always get a card the same evening. At Yandex Cloud, the starting GPU quota is zero and must be requested separately; compute is billed per second, while disk and public IP are charged under their own rules. For GPU access in Cloud.ru Evolution, the documentation specifies a top-up of at least 15 000 ₽ with no bonuses, and a free card must be available. Stopping doesn't always mean the charges end, either.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…