§ 01 — CEH™ Standardv2.1 · Published April 21, 2026

Compute,
denominated in energy.

Compute Energy Hour (CEH™) is a proposed standard unit for the energy intensity of compute — the facility energy it takes to produce one unit of useful output. One denominator for comparing chips, sites, and claims on energy fundamentals, not GPU-hour pricing alone.

Operating principle — measurement before management

Before we engage —

Which lens are you reading the standard through? It argues in your language.

§ 02Why CEH™ exists

Three numbers. None of them the answer.

AI infrastructure is scaling fast, but no single unit connects hardware performance, facility efficiency, energy price, and carbon. Each common proxy describes only a fraction — buy on one and the others move against you.

kWhEnergy

Counts the power. Blind to the work.

Tells you what a system draws — nothing about the useful compute it returns for it.

FLOPSThroughput

Counts the work. Blind to the cost.

A spec-sheet peak. Says nothing about the energy a real workload burns to reach it.

$/GPU-hrPrice

Counts the rent. Blind to both.

A rate that floats with the market — not a measure of efficiency, and never of carbon.

→

One denominator that holds all three accountable.

CEH™ resolves energy, work, and outcome into a single comparable unit: the facility energy it takes to produce one unit of useful compute.

§ 03The unit

Compute, expressed in the energy it costs.

Horsepower let buyers compare engines across makers. The BTU made heat a tradable quantity. Each took something everyone argued about and gave it one honest number.

CEH™ does the same for compute — a hardware-neutral unit for the energy a system spends to deliver useful work. It is a measure of energy intensity, not a performance benchmark: lower CEH means more useful compute per unit of energy. Originated and maintained by Oak Ridge Management.

Base definition

CEH = kWh ÷ output

facility energy per unit of useful compute · per hour

Total facility energy drawn over one hour — inclusive of PUE overhead — divided by the useful compute delivered (tokens, TFLOPS, frames, or a declared output unit), at real utilization. Lower is better.

§ 04What CEH™ measures

Four inputs. Three reported figures.

CEH integrates the four variables that materially shape compute economics — hardware power draw, throughput under a declared workload, facility overhead (PUE), and the price and carbon of delivered energy. From one reading it reports three figures.

Figure 01 · Intensity

CEH

kWh / output

The core energy-intensity figure: facility energy per unit of useful compute. Lower is better.

Figure 02 · Economics

CEH Cost

$ / output

CEH multiplied by the delivered price of electricity — the energy cost of a unit of useful compute.

Figure 03 · Emissions

CEH Carbon

gCO₂e / output

CEH multiplied by the carbon intensity of delivered electricity — emissions on the same basis as cost.

§ 05Benchmark methodology

A controlled basis, fully disclosed.

The v2.1 benchmark applies one controlled comparison basis so hardware is assessed like-for-like. A CEH benchmark is reproducible only if it discloses every assumption below.

Define

Fixed workload — LLM inference — with output tokens as the declared compute output unit.

Meter

Energy measured at the node on an 8-unit basis, then PUE-adjusted to the operating site.

Normalize

Across utilization, configuration, and software stack so two readings actually compare.

Publish

Versioned, reproducible from disclosed inputs, and open to challenge from any party.

FormulaCEH=Facility energy · kWh, PUE-adjustedUseful compute output→ lower is better

Disclosure requirements — to be reproducible & comparable

Hardware configuration — chip model, unit count, TDP
Measured utilization during the benchmark window
Declared or measured PUE
Workload type and compute-output-unit definition
Throughput method — software stack and workload parameters
Energy rate used for CEH Cost
Carbon-intensity source and vintage used for CEH Carbon

§ 06Hardware Reference Registry

What can be compared, and how well it's evidenced.

The Registry is the reference layer — which accelerators can be compared under CEH today, and how strong the public evidence is for each. It is not a certified ranking. It defines eligibility and evidence, not results.

Reference accelerators — CEH™ v2.1

Vendor	Accelerator	Architecture	Deployment	Role	Evidence
NVIDIA	H200 SXM	Hopper	Shipping / deployed	Frontier inference & training	Auditable
AMD	Instinct MI300X	CDNA3	Deployed	Inference, fine-tuning, HPC/AI	Auditable
NVIDIA	H100 SXM	Hopper	Deployed	Production training & inference	Auditable
NVIDIA	L40S	Ada Lovelace	Deployed	Cost-sensitive inference	Auditable
NVIDIA	A100 80GB	Ampere	Deployed / legacy base	Existing-fleet baseline	Auditable
Google	TPU v4	TPU	Deployed cloud	Non-GPU reference	Modelable
NVIDIA	B200	Blackwell	Frontier / early ramp	Emerging frontier platform	Modelable*
NVIDIA	RTX 4090	Ada Lovelace	Deployed (consumer/edge)	Non-datacenter comparison	Modelable
NVIDIA	V100	Volta	Legacy	Historical baseline	Modelable
CPU cluster	EPYC-class	x86	Deployed baseline	Non-accelerated floor	Modelable

Evidence tiers

Auditable: Vendor specs plus public benchmark or deployment data support reproducible CEH modeling.
Modelable: Public hardware data exists, but throughput normalization still depends on declared assumptions.
Modelable*: Relevant and material, but public deployment or benchmark evidence remains incomplete.

§ 07Modeled CEH™ outputs

The benchmark, applied.

Modeled comparative outputs under the v2.1 methodology — not a universal market ranking. Each figure carries the evidence tier of its source hardware and should be read with the assumptions below. CEH is shown indexed to NVIDIA H100 = 100 so it reads at a glance; lower is better.

Accelerator ranking — CEH™ v2.1Provisional · modeled

Rank	Accelerator	Efficiency Index (H100 = 100)	CEH Cost	CEH Carbon	Evidence
01	NVIDIA B200 (Blackwell)	S39	0.02	94	Modelable*
02	Google TPU v4	A69	0.04	165	Modelable
03	NVIDIA H200 SXM	A78	0.04	187	Auditable
04	NVIDIA L40S	B83	0.04	200	Auditable
05	NVIDIA H100 SXM	B100	0.05	240	Auditable
06	AMD MI300X	B118	0.06	283	Auditable
07	NVIDIA A100 80GB	C123	0.06	294	Auditable
08	NVIDIA RTX 4090	C193	0.10	462	Modelable
09	NVIDIA V100	D345	0.18	825	Modelable
10	CPU baseline (EPYC)	F986	0.52	2358	Modelable

Grading — by efficiency index (lower is better)S< 50A50–79B80–120C121–200D / F> 200

Benchmark basis

WorkloadLLM inferenceOutput unitoutput tokensNode8 acceleratorsElectricity$0.085 / kWhCarbon0.386 kgCO₂e / kWhUtilization · PUEdeclared per benchmark

CEH indexed to NVIDIA H100 = 100 (lower = better) · CEH Cost in $ / 1M tokens · CEH Carbon in gCO₂e / 1M tokens. Modeled from published specs and public throughput references; third-party-audited public index in development.

§ 08Why it matters

Energy is the largest lever you control.

Under the v2.1 framework, delivered energy price is one of the largest operator-controlled levers in compute economics. Moving from U.S. commercial grid pricing near $0.085/kWh to behind-the-meter generation in the $0.030–0.045/kWh range can materially cut CEH Cost for the same hardware and the same workload.

Hardware choice, facility design, and energy procurement are no longer separable decisions. CEH puts them in one equation.

$0.085→$0.030–0.045

per kWh, grid → behind-the-meter

§ 09Governance

A versioned standard, validated in stages.

Versioned

Dated, versioned releases

v2.1 published April 21, 2026. Revisions are versioned — triggered by new workload classes, methodology changes, or new disclosure requirements.

Reproducible

Public-source recalculation

Every figure is reconstructable from disclosed inputs and public specs, so an independent reviewer can re-run it.

Neutral

Energy-source & vendor agnostic

No pay-to-rank. Hardware ranks by the measured numbers, regardless of generation source or commercial relationship.

Validation pathway

Provisional

Modeled from public data

→

Verified

Independently recalculated

→

Certified

Third-party audited

§ 10Who it's for

One unit, three rooms.

Operators

Publish efficiency on a standard basis.

Infrastructure operators and cloud providers get a common way to state and compare compute efficiency — across vendors, sites, and hardware generations.

Buyers

Procure on energy economics.

Enterprise buyers evaluate hardware and hosting on cost, efficiency, and carbon at once — not dollars per GPU-hour alone.

Capital

Underwrite on a real denominator.

Capital allocators and infrastructure investors underwrite AI infrastructure on energy fundamentals — efficiency as a line item diligence can defend.

Engage the standard. Or argue against it.

CEH™ is being developed as a measurement and certification framework for compute infrastructure. Bring the hard question — it is built to answer it.

Read the methodology

Compute,denominated in energy.