More compute per watt
on your GPUs

We find and lock the best operating point your stack can't hold

Apply for early access Video demo

Runs alongside vLLM, SGLang and TensorRT. Under your power envelope, with no rewrite of your stack.

Demo

Inference is where AI
hits its cost ceiling

Every GPU has a fixed limit on what it can serve, and adding more is slow and expensive.

You're left with capacity you can't grow and thin margins on every token. The limit isn't the hardware. It's the operating point your stack settles into and never leaves.

Solutions

How the CarbonForge loop works

1 · Measure

Power Telemetry

Sub-millisecond power and latency, with kernel-level attribution.

2 · Optimize

Optimization Engine

Searches the operating point that captures what your monitoring misses.

3 · Re-lock

Runtime Controller

Re-locks the operating point when models, traffic, or hardware change.

Re-locks continuously as models, traffic, and hardware change

Run the complete Loop on your fleet. More tokens per GPU under the same power envelope.

Get Early Access
Team

Built at Mila

Leadership

Jean-Maxime Larouche Jean-Maxime Larouche, Co-Founder & Co-CEO · Technology
Jean-Maxime Larouche
Co-Founder · Tech
Serial AI Entrepreneur
Laurent Maisonnave Laurent Maisonnave, Co-Founder & Co-CEO · GTM
Laurent Maisonnave
Co-Founder · GTM
Serial AI Entrepreneur
Pierre-Luc Bacon Pierre-Luc Bacon, Co-Founder, CarbonForge
Pierre-Luc Bacon
Co-founder & Chief Scientist
Reinforcement learning, Mila
Sacha Lepretre Sacha Leprêtre, CTO, CarbonForge
Sacha Lepretre
CTO · AI platforms
Ex-LuxCarta, CAE, Mila

Advisors

Christophe Dubach Christophe Dubach, Advisor, CarbonForge
Christophe Dubach
Scientific Advisor
Compilers and GPU codegen
McGill, CIFAR AI Chair
Richard Reiner Richard Reiner, Advisor, CarbonForge
Richard Reiner
Strategic Advisor
Enterprise and infrastructure
Former Data Center CTO, Intel

Become an early adopter partner

A limited number of slots in 2026 for teams serving, operating, or running inference at scale.

Locked pricing on first deployments

Locked pricing on first deployments

Priority input on the roadmap

Priority input on the roadmap

Direct collaboration with the team

Direct collaboration with the team
Apply Now