For hosts · Updated October 13, 2026

What you need to rent out your GPU: hardware, internet and safety

KernelMachine is preparing an invite-only beta; no launch date is set. Where this guide describes KernelMachine, it's how it works at launch. Join the waitlist

Short answer: a desktop PC or server with an NVIDIA graphics card (RTX 30-series or newer, or a data-center card such as the H100, with 16 GB of memory or more), Ubuntu 24.04 or 26.04, at least 250 GB of free disk and a steady internet connection, wired if you can. Laptops, Macs, ARM computers, AMD cards, cards with less than 16 GB of memory and RTX 20-series or older cards can't host on KernelMachine.

This guide covers what a PC needs to rent out its GPU reliably, what doesn't work and why, and the safety checks worth doing before your card runs at full load for hours. To check your own PC in a few seconds, use the system check.

The checklist

Setting up

About 20 minutes, mostly waiting. Paste one line on Ubuntu:

Copy your install line from My machines and the PC joins your account by itself (with the public line, scan the code it shows; it works for a day). Setup restarts once for the graphics driver; the program then tests its cards and measures its internet by itself. The installer tells you what it's doing at each step, and it's safe to run again. During the beta a person approves each new host once, usually within a day. More on hosting.

What doesn't work yet, and why

• Macs and ARM computers: the host program needs a normal 64-bit (x86) PC running Ubuntu.
• AMD cards: renters' software (PyTorch, ComfyUI, Ollama builds) expects NVIDIA's CUDA, so we start with NVIDIA.
• RTX 20-series, GTX 16-series and older: they lack bfloat16 and FlashAttention-2, which renters' current software expects, so jobs that work elsewhere would fail on them. Older cards still have a driver problem too: RTX 50-series cards only work with NVIDIA's "open" driver (NVIDIA), and running one driver on every machine is what keeps updates safe. NVIDIA's 580 driver line is also the last for GTX 10-series cards (HotHardware).
• Laptops: the installer turns them away. They overheat under hours of full load and sleep when the lid closes.

Several cards, and old mining rigs

One machine can rent out several cards, even different models, and each card gets at least its size's minimum of the machine's processor, memory and disk, with the rest shared out between them. Mining rigs often connect cards through x1 risers (one-lane extension cables). That's fine for mining but slow for AI work, which moves a lot of data to the card, so renters' jobs run slower on them. Rigs also often have little system memory and a power supply that can't run every card at full load at once.

Disk space

Renters' software comes as container images that can reach tens of GB each, and each renter needs working space. Docker's newer storage keeps compressed and uncompressed copies of each image unless it's set up otherwise (Docker docs). That's why the minimum is 250 GB, and why other marketplaces ask for 200 GB (Vast.ai), 256 GB (Nosana) or recommend 500 GB (Salad).

Your internet connection

• Upload speed matters most. Renters download their results through your connection. The average US home uploads at about 30 megabits a second (Pong, a secondary source), so a 10 GB result takes about 45 minutes.
• Shared addresses. 5G home internet and Starlink share one public address between many homes, which makes direct connections harder and can slow rentals.
• Your plan's terms. Many home plans forbid running servers or business use, and renters' downloads count toward any data allowance. Xfinity's policy reportedly bans home servers used by people outside the home (Xfinity; we read a summary, as the page refused access). Check yours. Your dashboard shows how much data rentals used, and a business plan avoids the question.
• Wired beats Wi-Fi. Wi-Fi works, but drops more, and drops count against your machine.

Staying listed

Renters come first, so reliability decides. A new machine goes live after its automatic checks and 48 hours with no problems; one with 8 or more cards goes live as soon as they all pass the checks. A problem is a rental that fails on it, or going offline for more than 10 minutes without pressing Pause; each one starts the 48 hours again (on an 8-card machine it only counts toward removal), and we email you why. Three in 90 days remove the machine. Restarts for updates we run don't count. Staying online also pays: top hosts are offered rentals first. And if your machine drops mid-rental, the renter can recover from their last complete save on another available card of the same model, and you aren't paid for the few minutes of work lost since then, so press Pause before planned downtime.

Safety checks before you start

RTX 4090 and 5090 power plugs

The single high-power plug on these cards (called 12VHPWR or 12V-2x6) can still melt. In 2026 a reviewer's own RTX 5090 melted its plug while properly seated on a 1000-watt power supply (Club386), and another melted with the owner's power limit set to 500 watts (Wccftech). Before hosting:

• Push the plug fully in, until no gap shows.
• Don't bend the cable sharply close to the plug.
• Consider a lower power limit. One test found an RTX 4090 at 70% power keeps about 94% of its speed (RunAIHome, a secondary source). Limits reset on every restart unless something re-applies them.

RTX 3090 memory heat

The RTX 3090's memory chips slow themselves down at about 110 °C, and Linux tools can't read that temperature (NVIDIA forums). A 3090 that runs slow is often just hot. Better airflow helps, and one owner cut memory temperature from 110 °C to 82 °C by replacing the card's thermal pads (forum post).

Cards that drop out at idle

If your system log shows "Xid 79" (the card "fell off the bus") while idle, the usual fix is to turn off PCIe power saving in your PC's startup settings. On riser cables, setting the slot to an older link speed often helps (NVIDIA forums).

Your home network stays yours

On some marketplaces, hosts open router ports and their addresses show up in public data. In September 2026, researchers found attackers who pulled 297 host addresses from Vast.ai's public feed and scanned them (CloudSEK). On KernelMachine, your machine connects out to us, so there are no ports to open, and rentals run in isolated containers that can't reach the other devices on your network.

Using your PC while it hosts

Use it as usual. Cards you're using are skipped automatically, and one Pause button stops new rentals; rentals already running keep going until they end. Updates, card tests and speed checks run by themselves.

Read next

• How much can you earn renting out your GPU? Honest 2026 numbers
• How to rent a GPU in 2026: what to check before you pay
• Renting a GPU vs buying one: a simple way to decide