Dedicated GPUs for training your own models

Full access to the hardware, no shared servers. We source capacity across providers at market price or below, and keep every server healthy with the software we build for GPU clouds.

talk to us

C

talk to us

C

talk to us

C

Compute / request
Pick how many GPUs you need and for how long. We check our partner providers and come back with capacity that’s actually available.
ready
GPU
H100 SXM80 GB
Count
8
16
64
256
Duration
2 weeks
3 months
6 months
Access
bare metal · rootno VM
Find capacity↵
osl-1
fra-2
ams-1
sjc-1
sin-1
lon-2
Cluster / —0 / 64 GPUs
node-01
node-02
node-03
node-04
node-05
node-06
node-07
node-08
provisioned8 nodes
nccl baselinepassed
harbor attachedmonitoring
awaiting request

Buying compute shouldn’t take a week of calls

The listed price is marketing. The capacity isn’t there when you get on the call. And when something finally fits, the minimum is three years.

Pricing

Prices that move

The advertised price is for a setup nobody actually uses. The real price only shows up after the second sales call.

Capacity

Capacity that isn’t there

GPUs listed as available often aren’t there when you ask. Training a model needs dozens or hundreds of GPUs at the same time.

Terms

Years-long lock-in

Direct access to the hardware is reserved for the biggest buyers. Everyone else gets multi-year contracts and a virtual machine between their code and the GPU.

You tell us the workload. We find the hardware

We’re not a marketplace. We source on demand across the providers we work with, which is why what we offer is actually available.

02

We source it

We check capacity across the providers we work with and come back with what’s actually available, at a price that holds. Market rate or below, without the back and forth.

03

You get bare metal

Full control of the hardware, with nothing virtualized in between. Your cluster is ready for your training code from day one, with Harbor watching every server from the start.

01

Brief us

Tell us the GPU type, cluster size, duration and what you’re training. That’s everything we need to start looking. One conversation, not five calls with five providers.

02

We source it

We check capacity across the providers we work with and come back with what’s actually available, at a price that holds. Market rate or below, without the back and forth.

03

You get bare metal

Full control of the hardware, with nothing virtualized in between. Your cluster is ready for your training code from day one, with Harbor watching every server from the start.

02

We source it

We check capacity across the providers we work with and come back with what’s actually available, at a price that holds. Market rate or below, without the back and forth.

03

You get bare metal

Full control of the hardware, with nothing virtualized in between. Your cluster is ready for your training code from day one, with Harbor watching every server from the start.

01

Brief us

Tell us the GPU type, cluster size, duration and what you’re training. That’s everything we need to start looking. One conversation, not five calls with five providers.

02

We source it

We check capacity across the providers we work with and come back with what’s actually available, at a price that holds. Market rate or below, without the back and forth.

03

You get bare metal

Full control of the hardware, with nothing virtualized in between. Your cluster is ready for your training code from day one, with Harbor watching every server from the start.

Cheap compute, without the usual trade-off

Compute / routing
live
allocated GPUs · requests · avg ready in [Xm]3 providers · bare metal only
Foundation models/ large clusters
Custom kernels/ root access
Domain models/ own data
Fine-tuning/ short terms
Inference/ dedicated
Research sprints/ on demand
harbor
osl-1 · bare metal · H10032/108
fra-2 · bare metal · H10051/108
ams-1 · bare metal · H10066/108
allocation log10:28:07
foundation models
50× H100 · fra-2$1.96/h · 10:26:42
inference
20× H100 · fra-2$1.96/h · 10:26:59
fine-tuning
8× H100 · fra-2$1.96/h · 10:27:16
inference
17× H100 · osl-1$1.74/h · 10:27:33
custom kernels
15× H100 · fra-2$1.96/h · 10:27:50

Sourcing

Routed to the best price

Every request goes to the provider with real capacity at the lowest rate. Market price or below, without a week of calls.

Bare metal

Direct access to the GPU

No virtual machines, no sharing servers with other customers. Full control of the hardware, ready for your training code.

Same layer

Kept healthy by Harbor

The same software we build for GPU clouds watches every server in your cluster. Problems are caught and fixed before they slow your training down.

For teams that need the whole GPU

Built for teams that need full control of their GPUs but can’t sign a three-year contract.

Large clusters

Foundation models

Pre-training needs hundreds of GPUs for a few months, not a few years. We source the capacity and hand you the full cluster.

Root access

Custom kernels

Write and tune your own low-level GPU code directly on the hardware. Nothing virtualized, full admin access on every server.

Own data

Domain models

Voice, radiology or anything where general models fall short. Train on your own data, on your own cluster.

Short terms

Fine-tuning

Serious hardware for runs that take weeks, not years. Start when you’re ready, hand it back when you’re done.