# Joule

> Efficiency down to the metal.

Joule (https://joule.lat/) is a phase-aware inference power economics engine. See which watts in your inference fleet are safe to reclaim, and which ones would break your SLOs. Metrics include $/M tokens, tokens/W, TTFT, TPOT, and goodput per watt on your infrastructure.

Founded by Abi Aryan (https://abiaryan.com/).

## One-line summary

Efficiency down to the metal. See which watts in your inference fleet are safe to reclaim, and which ones would break your SLOs.

## What Joule does

- Observe: shadow mode on the serving runtime alongside device power (KV occupancy, queue depth, cache). No production writes. Telemetry on your infrastructure.
- Attribute: estimates power by inference phase (prefill vs decode attribution) tied to TTFT, TPOT and goodput, not raw utilization.
- Decide: safe power headroom where capacity can be reclaimed and where a cap would break SLOs. Trade-offs measured with controlled tests on a pool you choose, with automatic rollback, including on-demand load shedding and latency cost.
- Start with a fixed-fee Power & SLO assessment (six weeks): recoverable capacity, safe power headroom, SLO risk map. Shadow mode first, then controlled tests with rollback. Telemetry stays on your infrastructure.
- Then continuous measurement and control as the fleet changes (scoped after assessment).

## Who it is for

- Production inference leads running vLLM or SGLang who need to predict TTFT/TPOT impact before power, batching or caching changes ship.
- Neocloud and AI host operators running managed inference or opt-in tenants: sellable capacity within a fixed power envelope and SLO-safe curtailment.
- Capacity and finance leads optimizing $/M tokens, recoverable capacity, safe power headroom, and grid flexibility under latency SLOs.

## What Joule is not

- Not a generic self-serve hyperscaler platform. One front door: request an assessment, then ongoing programs scoped in proposal.
- Not a facility-only DCGM dashboard or utilization story without serving context.
- Not replacing your serving stack. It measures and models economics on the path you run.

## Canonical phrases

- Tagline: Efficiency down to the metal.
- Category: Inference power economics · phase-aware inference power economics.
- Metrics: $/M tokens, tokens/W, SLO goodput, TTFT, TPOT, safe power headroom, recoverable capacity.
- Measurement gap: facility telemetry misses serving state (head-of-line blocking, KV-cache preemption). Joule reads runtime state to judge whether power decisions are safe.

## Pages

- [Home](https://joule.lat/)
- [Request an assessment](https://joule.lat/scoping)

## Contact

- hello@joule.lat