Question 1

What does “billed per second” actually mean?

Accepted Answer

The meter starts when your machine starts running and stops when you release it, rounded to the second. A 90-second job costs 90 seconds. There’s no one-hour minimum and no rounding up to the next minute. If a machine never gets as far as letting you log in, it costs nothing.

Question 2

What about the disk?

Accepted Answer

Every machine comes with one, at $0.08 per GB-month, billed by the second from when the machine starts until the disk is deleted. That’s about $0.03 an hour for the 279 GB on a 1-GPU machine, and $0.11 an hour for the 1,023 GB on an 8-GPU machine.

Question 3

Is my disk safe?

Accepted Answer

A standard disk is fast but has no redundancy: if its hardware fails, the data is lost. For $0.15 per GB-month, a redundant disk is just as fast and mirrored to three drives, so it survives that. You choose when you launch. Either kind is deleted with the machine, so keep anything you can’t lose somewhere else.

Question 4

Is there really no egress fee?

Accepted Answer

Really. Moving data out of Koi, to the internet or between regions, costs nothing. We think egress fees exist to keep you from leaving, and we would rather you stay because it’s good.

Question 5

Do prices change with demand?

Accepted Answer

No. On-demand and Drift prices are fixed and published here. With Drift, what varies is whether spare capacity is free, not what it costs. When we lower them, every customer gets the new price automatically.

Question 6

What’s the catch with Drift?

Accepted Answer

It can be taken back. When the capacity is needed, your machine is reclaimed at once, without notice, and released along with its disk. Billing stops at that moment. It suits work that checkpoints often and resumes: batch jobs, sweeps, most training. Or ask for it back when you launch: the GPU stops billing, the disk is kept and billed as usual, and your container restarts when capacity returns, or the machine and disk are deleted after a day. If your checkpoints live on that disk, a redundant one protects them from a hardware failure too.

Question 7

Can I reserve capacity?

Accepted Answer

Yes, for a month or longer, in the region you choose. Reserved prices depend on the GPU, how many, the term and the region, so we quote them per request instead of listing them. Email us what you need and an engineer replies within a few hours.

Question 8

How do I get more than the self-serve limit?

Accepted Answer

Email us the GPU type, count and dates. An engineer, not an account executive, replies with what we can place and a confirmed price from this page.

Question 9

Can I set a hard spending limit?

Accepted Answer

You can’t spend more than you’ve added: Koi is prepaid. We email you when your running GPUs have about two hours of credit left, and stop them if it runs out, so there’s never a surprise invoice. You can also set a machine to stop by itself after a while, and give an API key a limit on what its machines may cost.

GPU	Memory	Bandwidth	Interconnect	Availability	Per GPU-hour
B300	288 GB HBM3e	8 TB/s	NVLink 5 · 1.8 TB/s	Up to 8 self-serve	$9.99 per hour
B200	180 GB HBM3e	8 TB/s	NVLink 5 · 1.8 TB/s	Up to 8 self-serve	$8.99 per hour
H200	141 GB HBM3e	4.8 TB/s	NVLink 4 · 900 GB/s	Up to 8 self-serve	$4.99 per hour
H100 SXM	80 GB HBM3	3.35 TB/s	NVLink 4 · 900 GB/s	Up to 8 self-serve	$3.99 per hour
RTX PRO 6000	96 GB GDDR7	1.6 TB/s	PCIe Gen5	Up to 8 self-serve	$2.29 per hour

Every price.
One page.

GPUs

Storage, and the
number zero.

What does
idle cost you?

The idle-tax calculator

The fine print,
which is short.

Get in
the water.

Every price.One page.