GPUs · B300
Rent B300s,
by the second.
NVIDIA's Blackwell Ultra GPU, and the most memory on one GPU here, for models too large for anything smaller.
01Price
$9.99 /GPU-hour
Billed by the second, with no minimum, plus the machine’s disk at $0.08 per GB-month. Drift is the same GPU on spare capacity, and it can be reclaimed without notice. Every price
| Machine | On demand, per hour | Drift, per hour | A month left on |
|---|---|---|---|
| 1 × B300 | $9.99 | $5.49 | $7,293 |
| 8 × B300 | $79.92 | $43.92 | $58,342 |
02The GPU
Specs
- Memory
- 288 GB HBM3e
- Memory bandwidth
- 8 TB/s
- Interconnect
- NVLink 5 · 1.8 TB/s
- Regions
- US, United Kingdom, France, Spain
03What it suits
Good for
- Training and serving the largest models
- Workloads limited by memory rather than compute
- Multi-GPU jobs that lean on fast NVLink
04Launch one
From your terminal
Or from the console or the API, once you’ve created an account. Read the quickstart
# One, stopping by itself after two hours
koi spawn b300 --ttl 2h
# Eight on one machine, on spare capacity
koi spawn b300:8 --drift05Questions
About the B300
- How much does it cost to rent an B300?
- $9.99 per GPU-hour on demand, and $5.49 per GPU-hour on Drift, which is spare capacity that can be reclaimed without notice. You pay by the second, with no minimum. Every machine also has a disk, at $0.08 per GB-month.
- Where can I run an B300?
- US, United Kingdom, France, Spain. You can pick a region, or let Koi use the cheapest with capacity.
- How many B300s can I get on one machine?
- One or eight. An account can run 8 GPUs at once to start with, and you can write to us to raise that.
- How long does it take to start?
- Minutes. Koi shows the machine as starting until SSH answers, and then gives you the command to log in. A machine that never becomes ready costs nothing.
Other GPUs
- B200$8.99/h
- H200$4.99/h
- H100 SXM$3.99/h
- RTX PRO 6000$2.29/h
Ready when you are
Get in
the water.
Pay by the second, only while your GPUs run. No call, no waitlist. Your first H100 is minutes away.