Use IOTA

IOTA

Two ways in. One platform.

However you come to IOTA, the platform underneath is the same. Pick the route that fits how you want to work, run it yourself, or hand it to us.

Get Started → Talk to our Founding Team →

Choose how you use IOTA

The only real question is how much you run yourself.

Integrate IOTA and drive your own workloads, or bring us the workload and we take it from start to result. Same platform, same capability, whichever you pick.

Two routes / one underlying IOTA platform

Self-directed technical route

SDK

Integrate IOTA and run your own workloads, with direct control over how they are configured and executed.

Who it’s for: Teams with the technical capability to run their own training, and the infrastructure people to integrate a platform directly.
What you do: Configure your workload, connect to IOTA, and run it. You keep control of the workload, IOTA coordinates the compute underneath it.
What IOTA handles: The disaggregated training underneath, from coordination and execution to monitoring, recovery, and reporting.

Get Started →

Managed route

Train with IOTA

Bring us a defined workload, and we take it from qualification through to a finished, reported result. The managed route suits model-training workloads and platform or fleet workloads alike.

Who it’s for: Teams and platforms that want the outcome without running the training themselves, or that want it qualified and managed end to end.
What you do: Bring the workload, the objective, and what a good result looks like. We shape it with you and run it.
What IOTA handles: Everything from qualification and configuration through to execution, monitoring, recovery, and a result you can act on.

Talk to our Founding Team →

Which route fits your situation?

Find the one that sounds like you.

The right route usually comes down to how much you want to run yourself. Here is where each one tends to fit.

Scenario 01

Experienced ML team with a defined training workload

You have the expertise and the workload ready. Integrate directly and run it yourself.

SDK →

Scenario 02

A company that wants to train its own model but lacks distributed-training expertise

You know what you want to build, but not how to run it across disaggregated compute. We take that on with you.

Talk to our Founding Team →

Scenario 03

An AI platform with workloads that are hard to serve profitably

You have demand you cannot serve profitably today. We help you turn capacity you could not use before into workloads you can take on.

Talk to our Founding Team →

Scenario 04

A GPU operator with an AI workload backlog or capacity sitting idle

You have the hardware and the demand, but not the layer that matches them. Integrate IOTA and put more of the fleet to work.

SDK →

Scenario 05

An advanced research team with a defined workload and strong internal expertise

You have the capability and a clear objective. Run it yourself, on flexible capacity, without a reserved cluster.

SDK →

Scenario 06

A commercially important workload that needs qualification and managed execution

The workload matters too much to leave to chance. We qualify it, run it, and report back on the result.

Talk to our Founding Team →

Fault-tolerant training

Built to keep going when the hardware doesn’t.

Disaggregated compute is never perfectly reliable. Machines slow down, drop out, and rejoin. IOTA is built so none of that stops the work, and so progress is preserved rather than lost.

WORKLOADRESULT

Fault-tolerance concept

Failure without lost work

Work is spread so no single machine holds the only copy of a result. A large share of the machines in a stage can fail, and the combined output still comes together correctly, with over 99% of the work accounted for.

Fault-tolerance concept

No waiting on the slowest machine

The system moves forward once enough machines have done their part, instead of waiting for every one. Faster and slower hardware work together, and the slowest never sets the pace.

Fault-tolerance concept

Progress preservation

The state of the work is saved regularly, so nothing already done is lost when a machine drops out or falls behind.

Fault-tolerance concept

Rejoin and recover

A machine that dropped out, or a new one joining, picks up from the last saved point and gets back to work. Capacity lost in one round is replaced in the next.

Built for real training workloads

Not a demo. A platform with a track record.

IOTA is proven on real workloads, at scale, in the open. You can see how it holds up before you commit.

Technical documentation

Research

Workload validation

Reporting

Next step