Luxi · Energy architecture for AI compute

Energy-aware AI compute architecture.

Luxi is building the energy architecture for AI compute. LuxiEdge is the first public product in a system that spans deterministic computation, GPU execution, request scheduling, load shaping, and facility power control.

Each layer targets a different source of wasted energy. Each has its own measured scope and maturity label. We do not present early-stage concepts as deployed products.

LuxiEdge → See the evidence

Why one layer is not enough

Modern AI infrastructure wastes energy at every level: in how arithmetic is executed, in how GPU time is filled, in how requests are scheduled, in how facility load is shaped, and in how hardware is selected and configured. Optimizing a single layer leaves the others untouched. The Luxi architecture assigns a separate, measurable product to each layer, with evidence matched to its actual scope.

The six layers

Layers are listed from computation outward to facility. Maturity is labeled honestly.

Working · Scope-limited evaluation

LuxiQuant

Deterministic numerical engine

Controls Arithmetic precision and determinism across CPU and GPU paths

Waste addressed Nondeterminism forces redundant computation, audit overhead, and result mismatch between environments

Maturity Working engine. TESTfort QA Lab independently evaluated the defined non-linear numerical workload (December 2025). That evaluation did not cover every function, backend, or platform.

Next milestone Extended expression coverage and independent reproduction of additional function classes

Primary product · Third-party measured

LuxiEdge

GPU inference and prefill energy

Controls Board energy per prefill position and faithful model execution on H100

Waste addressed Excess GPU board energy per prefill position versus reference serving stacks in the tested configuration

Maturity An internal matched comparison on the locked prefill path measured roughly 1.18x the tested vLLM throughput at roughly 12% lower board energy per prefill position at batch 16 on the same-GPU matched Qwen2-7B workload. A separate internal absolute prefill champion measured roughly 44,860 positions per second at roughly 0.01532 board joules per position at batch 72, with no matched vLLM arm; the two are never combined. Internal, prefill-stage evidence, reported separately from the independent TESTfort baseline; independent reproduction is the next step. Public brief →

Next milestone Full decode and serving throughput measurement, independent third-party reproduction of the internal matched prefill results

In development

LuxiPack

Request packing and batch strategy

Controls How requests are grouped to maximize hardware utilization per batch

Waste addressed Underutilized batches leave GPU compute idle while still drawing active board power

Maturity In development. No independent measurement yet.

Next milestone Internal batch-utilization measurement on H100, compared to reference packing baseline

Prototype · Local validation only

LuxiPhase

Execution scheduling and phase alignment

Controls When computation phases execute relative to power delivery cycles

Waste addressed Phase misalignment causes avoidable power spikes and reactive power overhead

Maturity Prototype. Results are from local validation only; not yet tested in a data-center environment.

Next milestone H100 integration test with NVML power-draw logging

Early concept

LuxiLoad

Facility load shaping

Controls Aggregate power draw profile at the rack and facility level

Waste addressed Uncontrolled load spikes generate peak demand charges and stress cooling infrastructure

Maturity Early concept. No prototype or measurement exists yet.

Next milestone Simulation model for peak shaving against a defined facility power profile

Early concept

LuxiSDG

Sustainable deployment guidance

Controls Configuration and hardware procurement guidance for energy-optimal deployment

Waste addressed Mismatched hardware selection and configuration choices that lock in avoidable long-term energy overhead

Maturity Early concept. No prototype or measurement exists yet.

Next milestone Pilot advisory engagement with a data-center operator

Maturity labels

Every layer carries a label that reflects its actual state. We do not conflate concepts with products.

Third-party measured Independent measurement on the stated configuration. Evidence is public.

Scope-limited evaluation Working engine; third-party evaluation covers a defined, not exhaustive, workload.

In development Active engineering work; no external measurement yet.

Prototype / local validation Code exists; results are internal only; not tested in a data-center environment.

Early concept Design and rationale exist; no prototype or measurement yet.

LuxiEdge is available now

LuxiEdge is the measured, public-facing layer of the Luxi architecture. All claims on this site apply to LuxiEdge on the tested configuration. Evidence is on the Proof page.

We publish what we have measured, label what we have not, and link to the evidence. Contact us to discuss evaluation or data-center deployment.

See the evidence AI & data centers Contact