C3 - GPU Compute for Researchers and Agents

C3 Logo

NO QUEUES

Pay per second billing

Submit GPU code in <1 second. Researchers and agents can edit locally and deploy globally with no queues via a SLURM-like interface where any machine can be the login node.

Quick installLive
$curl -fsSL .../install.sh | shCopy
Documentation|Free credits for researchers
terminal
$

Trusted by researchers worldwide

University of Cambridge
University of Oxford
Kings E Lab
Cambridge Center for Frontier Technologies
MIT Media Lab
Imperial College London
University of Cambridge
University of Oxford
Kings E Lab
Cambridge Center for Frontier Technologies
MIT Media Lab
Imperial College London

How It Works

1
2
3
4
5
Step 1 of 5

EDIT LOCALLY

Your laptop. Your editor. Your workflow.

Write your research code locally. No remote setup, no SSH, no cluster configuration.

train.py
import jax
import jax.numpy as jnp
from flax import linen as nn
# Your research code here
model = TransformerLM()
state = train_step(state, batch)
CLI-native acceleration

GPU Compute for Agents

Agents running on small machines can use C3 as their GPU layer: fully interpretable commands, global capacity, and pay-per-second bursts of compute.

Single control machine
x1
$ for agent in agents
agent.launch("c3 deploy job.sbatch")
local CPU stays small; GPU work bursts out
c3 deploy
pay per second
Global GPU network
burst
N agents
launch jobs
pay/sec
GPU bursts

CLI-first by design

c3 deploy job.sbatch

C3 is fully legible to software agents: commands, files, logs, job IDs, and outputs are all inspectable text instead of hidden UI state.

Fan out from small machines

for agent in agents

A cheap control box can run many agents. Each one launches GPU work only when it needs acceleration.

Burst to global GPUs

c3 squeue --watch

Agents can deploy into the C3 network and pay by the second, turning one machine into access to unfiltered GPU power when the workload demands it.

From Code to Results

Write your science script

1
2
3
4
5
6
Visual Studio Code
🐍py
📄sbatch
⬛Terminal
1import jax
2import jax.numpy as jnp
3from jax_md import space, energy, simulate
4from galax import MilkyWayModel, NFWHalo
5
6# Initialize 10M particle galaxy
7galaxy = MilkyWayModel(n_particles=10_000_000)
8halo = NFWHalo(mass=1e12) # solar masses
9
10# Configure GPU-accelerated N-body solver
11sim = simulate.NBodySimulator(
12 galaxy, halo,
13 dt=1e6, # 1 Myr timestep
14 softening=100 # parsecs
15)
16
17# Evolve for 10 Gyr and render animation
18sim.run(t_final=10e9, checkpoint_every=100)
19sim.animate("plots/galaxy_evolution.gif")

No queues. No setup. Pay only while your code runs.

Live
$curl -fsSL .../install.sh | shCopy