Liquidtraining, intelligence, compute
Turn globally distributed, heterogeneous and unreliable compute into a cluster indistinguishable from a centralised one.
Scattered GPUs aren't a cluster.
We make them one.
Run large training jobs on compute you don't own, reserve or wait for. You pay less and never lock into a fixed cluster.
Pool the compute
GPUs from datacentres, clouds and spare capacity. Different hardware, different places, different owners.
Run it as one machine
IOTA runs one training job across every machine. When machines drop out or rejoin, the run carries on. The systems work is ours, not yours.
No compromise
Same model. Same monitoring. Same code. You keep the workflow you know, and the quality you expect.
Train anywhere,
train everywhere.
Liquid Compute supplies the GPUs. The IOTA SDK runs your training code across them.
Compute platform
Liquid Compute
Scaled compute in seconds. Preemptible GPUs across multiple clouds, provisioned as one fleet. Idle GPUs at 0.2× the reserved hourly rate.
$lc auth login$lc run start train.py$lc run listPut idle GPUs to work on high-value training runs. Utilisation rises without buying more hardware.
Contribute computeTraining SDK
IOTA SDK
A drop-in replacement for your distributed backend. An elastic mesh takes care of fault tolerance and machines joining and leaving.
-dist.init_process_group(backend="nccl")+iota.use_mesh(mesh, group="all")+dist.init_process_group(backend="iota")Follow live training runs on the IOTA dashboard.
See IOTA in actionH100, per GPU-hour, September 2026. Reserved price: 1-year median, AIMultiple GPU index. Idle compute: spot prices from FastGPU and Vast.ai.
Stay in control.
Or leave the training to us.
Integrate IOTA into your own stack, or have our researchers and engineers run training for you.
Self directed
Make IOTA part of your stack
Add distributed training to your own stack. You keep control of the code, infrastructure and models.
- Choose Liquid Compute or your own GPUs
- Keep your existing training code
- Fault tolerance built in
With our team
A managed training run
Tell us what you want to train. We plan the run, source the compute and manage training through to delivery.
- Plan the run with our research team
- We handle orchestration and monitoring
- You own the trained model
More from Macrocosmos
Explore our ecosystem of intelligence.
The General Intelligence Platform
Apex
Turn your hardest optimisation problem into an open competition. Developers worldwide build AI agents to solve it, and you watch the best climb the leaderboard in real time.
Social Data
Data Universe
Over 55 billion rows of structured social data from X, Reddit and YouTube, with up to 80 million new rows a day. Built for monitoring, research and model training.
