Articles and guides

• Linux commands for a GPU host: a searchable cheat sheet 12 min Over a hundred commands you need on a Vast.ai server every week: GPUs, disks, network, services, Docker, logs and the Vast daemon itself. A one-line explanation for each and keyword search. Free.
• Hashcats on an RTX 4090: how long a cat takes and what it really costs 7 min Live numbers from hashcats.fun: a 42–45-bit target, 4–35 trillion hashes per cat, 15 minutes to 2 hours on one RTX 4090 in CUDA. And why the bottleneck is not the card but the $206 entry price.
• How to connect to a server without a public IP over a reverse SSH tunnel 12 min No public IP, port forwarding impossible — and we still get in. A reverse SSH tunnel to a hub, a browser terminal, our own latency numbers, and the four mistakes that got us here.
• Clore or Vast.ai: where to rent out your GPU — a 7-point comparison 10 min Demand, pricing, payouts, entry barrier, tooling, support and the score — measured on our own six-machine fleet: $1,737 in August 2026 at 90% utilization, 100% of the rate to the host and a +33% markup added for the client. Including what broke on our side and where we got it wrong.
• Renting out your GPU on Vast.ai: the pre-launch checklist 8 min Five items, ten checks with commands and thresholds, our 2026 incident log, and a section on where we got it wrong ourselves.
• What hardware to buy to rent out GPUs for AI workloads 11 min Three builds for about 1M rub with an income forecast, NIC measurements from our own fleet (330 vs 706 Mbit), RAM, disk and PSU in concrete numbers — plus a «where we got it wrong» section: from a dead memory module memtest never saw to $0.00 paid for 8.32 TB of traffic.
• How much power a GPU server draws: fleet measurements, a 16 A line and cooling 10 min A 2.2–2.4 kW peak per 4×4090 server against a measured 26–360 W per card, what a single 16 A line really covers, the monthly bill at 4/6/8 rub per kWh, and six mistakes we made on the site itself.
• What internet you need to rent out a GPU: bandwidth, symmetry, DPI 13 min The vast.ai formula: 2.6 Mbit/s per 1 GiB of VRAM; the minimum is 100 Mbit/s, the cap is 500. Measurements across six fleet machines, throttling by the VPS host's AS (OVH and M247 choked, Private Layer clean), 8.32 TB in 37 hours and why it earned $0.00.
• How much a GPU server and GPU rental make in 2026: our $1,820 month, card by card 23 min Over 31 days, from 29 August to 28 September 2026, our six machines with ten GPUs brought in $1,820 at a share of paid hours of about 82% — each machine's income divided by its own card-hours and its own asking rate. We break the sum down card by card, subtract downtime and costs, convert to roubles and show the method for estimating payback. Every machine's income is published openly on the "Fleet" page of our site, vast.sale, and refreshes every hour.
• Installing Ubuntu 22.04 Server for Vast.ai: step-by-step guide 17 min From boot USB to a ready system: BIOS, every installer screen and disk partitioning — a picture for each step. Free.

Your GPUs, your power tariff: see what they would earn.

• From idle to income: the complete Vast.ai server migration 16 000 ₽ From mining, rentals or Clore — 9 chapters from hardware audit to first orders: commands, install, RKN-bypass, pricing and monitoring. Everything to switch without losing income. 56 min. Buy
• Server setup for Vast.ai: from hardware to first order 8 000 ₽ Drivers, Docker, host agent and tests — a stable node from the first launch. 30 min. Buy
• Pricing and maximizing GPU utilization 8 000 ₽ How to price so your cards stay busy and earn the most. 25 min. Buy
• Security, uptime and host reputation 8 000 ₽ How to secure your node, keep uptime and grow rating for steady orders. 25 min. Buy
• Backup tunnel: how to survive a VPS outage without losing clients 8 000 ₽ Every VPS at one provider rebooted at once — and the rigs went down with them. The solution is a backup VPS in another country, which the machines jump to automatically. 9 chapters: what happens to the score, why the usual advice fails, and a ready failover script with rollback. 32 min. Buy
• The machine is “inactive” yet everything works: a stuck vast controller 5 000 ₽ A rare failure no routine check catches: the link is perfect, the tunnel is up, clients are running — yet the marketplace considers the machine dead. The diagnosis and a watchdog that fixes it on its own. 11 min. Buy
• VM mode on a multi-GPU server: from BIOS to the first client 5 000 ₽ Virtual-machine mode opens a client segment that container rentals never see. 9 chapters on setting up a multi-GPU server: the three admission conditions, ACS in the BIOS, hardening the cards against the power-saving trap, a controlled ladder of 1 → 2 → 3 → 4 cards in virtual machines before the first client, three failures of the enabling script that report the wrong thing, and the main trap of a multi-GPU build — the marketplace's hourly job that wipes the VM flag. Working scripts and watchdogs included. 63 min. Buy