A single GPU, fully used
Even one card — local, on-prem, fully private. TorqHUB schedules the queue and loads/unloads models automatically, so a single GPU is reused evenly instead of idling between tasks.
Stop paying for idle metal
Your GPUs are booked around the clock and working a fraction of it. TorqHUB streams workloads onto the machines that can actually take them.
Your capacity is mostly air
Reserved 24/7. Working, say, 18% of the time. Billed for 100% of it.
Billed 100%
Idle spend / month $33,696 Illustrative — 10 racks reserved, 18% utilised.
Bar chart: ten GPU racks. Two are actively computing, shown in blue. Eight sit idle, shown in grey. A coral line across all ten represents the full cost you pay. Illustrative figures — real utilisation varies.
Occupied Used
A job holds the card for eight hours and computes for forty minutes. The card is busy. Nothing is happening.
Occupancy
How much of each reserved GPU is actually doing work. Drag to feel the gap.
18%
Duty cycle slider. Current value 18 percent of reserved capacity. Range 5 to 60 percent. Use the left and right arrow keys to adjust by one percent.
What it costs you
Move the sliders. This is the money that turns into heat and nothing else.
The cost of waiting
Reserved GPUs burn money every hour they sit idle. Change the fleet — then drag the occupancy slider to watch the gap close.
Duty cycle 18% Set by the occupancy slider above.
Idle spend / month
Idle spend / year less than one engineer / yr
Projected idle spend: $9,446 per month, $114,931 per year, at 18 percent duty cycle. less than one engineer / yr. Illustrative figures.
Every job finds a GPU that fits
Different jobs, different shapes, different machines. Flip the switch and watch the gaps disappear.
OFF — pinned to home GPUs
- DONE
- 24
- UTIL
- 58%
- QUEUE
- 34
- GPUS
- 3
ON — TorqHUB brokered + elastic
- DONE
- 79
- UTIL
- 21%
- QUEUE
- 0
- GPUS
- 3
One SDK token. Infinite scale.
You write the logic. We route the compute.
- 01
Install
One package, one token. No cluster to provision, no broker to configure.
- 02
Submit
Send a job with its type and payload. TorqHUB routes it to the GPU that fits.
- 03
Scale
Batch to hundreds, poll or await results. The router fills the gaps for you.
import { TorqClient } from '@torq/sdk'
const torq = new TorqClient({
token: process.env.TORQ_TOKEN,
})
// submit a job — the router picks the GPU that fits
const job = await torq.jobs.submit('llm-70b', {
model: 'meta-llama/Llama-3-70B',
prompt: 'Summarize this transcript…',
})
// wait for the result (or poll, or batch to hundreds)
const result = await torq.jobs.wait(job.id)A mesh that's already running
Capacity where the work is. The router sees every node and every gap.
One GPU, used fully
A single card — local, on your desk, fully private. TorqHUB schedules the queue and loads/unloads models automatically, so even one GPU stays busy instead of idling between tasks.
12 Regions
8.4K Workers
24 Job types
Private corporate mesh
Pool the GPUs you already own into a closed, private mesh. TorqHUB redistributes compute across the fleet inside your walls — your jobs, your machines, nothing leaves the network.
Public crowd mesh
Or tap a crowdsourced pool: rent capacity on demand, or put your idle cards on the mesh and earn. The router streams every job to wherever the right GPU is free.
Network mesh: a central TorqHUB router connected to GPU worker nodes across multiple regions. Illustrative scale figures. TorqHUB runs two ways: as a public crowdsourced compute mesh, or deployed privately inside your own corporate systems — contact us for deployment details.
STOP RENTING AIR
Tell us what you need. We'll come back with access, capacity or a deck.