Skip to content

GPU compute
for research, agents and humans.

Cloud-agnostic GPU compute for your existing code. Deploy Bash scripts, Docker containers or Python environments through your agent or the CLI.

What is this?

This illustrates a radio pulse from a pulsar: a spinning neutron star whose radio beam sweeps past Earth like a lighthouse. The first pulsar was discovered in Cambridge in 1967.

As a pulse travels through plasma in space, lower radio frequencies arrive later than higher ones. This spreads the signal out in time: dispersion. Astronomers search for these pulses by dedispersing the data, aligning the frequency channels just as you do when you scroll.

Many possible alignments can be checked in parallel, making the search a natural fit for GPUs. C3’s first scientific workload was a search for radio signals like these.

Connect Claude Code
claude mcp add --transport http c3 https://api.cthree.cloud/mcp

Then run /mcp in Claude Code and choose Authenticate for C3.

Used by researchers at
From a request to a result

You set the work.
C3 runs it.

Like Slurm, but your laptop is the login node and the world is your GPU cluster.

Run scripts through your agent or the CLI, with logs and a shareable page for every job.

Example run · H100
youRun run.sh on an H100. Bring back the loss curve.
agentI'll deploy your workspace through C3.
toolc3 deploy run.sh --json
c3SCHEDULING → RUNNING
c3saved outputs/loss.png
agentDone. Loss 0.93 → 0.11.
Status
Succeeded
Run time
4m 12s
Example cost
£0.17
Output
loss.png

The infrastructure layer for automated research

We’re building the scheduling layer between research agents and GPU compute. Our neural scheduler coordinates jobs across clouds, matching changing workloads to available capacity so your agents can keep experimenting.

Agent swarmNeural schedulerGPU time slotsCapacity stateGPU AGPU BGPU CNowUpcoming
Ultra-malleable GPU capacity for agent swarms.
From agent requests to GPU time
The AIDDA result

18.5× lower GPU cost

25 agents · 2,038 jobs

Measured on one discovery loop against an always-on GPU per agent, before the neural solver. Savings vary by workload.

Read the AIDDA 2026 write-up
Built around your work

Your code. No custom functions.

Docker, Bash or Python.

Deploy Docker containers, run Bash scripts or use Python environments. Keep your code and dependencies without platform-specific function wrappers.

  • Docker
  • Bash
  • Python

Choose your cloud.

Let C3 find available compute, or pick a provider yourself. Keep the same workflow when you switch clouds.

  • GCP
  • Crusoe
  • Nebius
  • NexGen Cloud

Pay for the work.

Scale independent GPU jobs with demand. Pay per second, with logs and results you can inspect.

Run your first job.

Free credits to start. Per-second billing.

Connect your agent
claude mcp add --transport http c3 https://api.cthree.cloud/mcp

In Claude Code, run /mcp and choose Authenticate for C3.

Install the CLI
curl -fsSL https://cthree.cloud/install.sh | sh

Then run c3 login to sign in.