New19 Sep 2026
Peak Water, Not Average Water.
Peak-water capacity screening for Northern Virginia.
Read publication →Critical Load / Place assessment
AI capacity is a property of place.
We measure how much compute a place can support, what that compute changes from manufacture to operation, and which conditions should govern development.
240 MWPlanned capacity
176 MWIllustrative threshold
WaterBinding constraint
Explore the research
Why Critical Load exists
Compute has physical limits. Critical Load measures local capacity, impacts and development conditions.
The research boundary runs from semiconductor manufacturing, hardware and workload energy through electricity, cooling, water, land, infrastructure and community effects.
New19 Sep 2026
The Physical Cost of Useful AI Work.
Task-normalized energy accounting for agentic AI workflows.
Read publication →Reference research & evidence
The external evidence base Critical Load builds on.
External publications, datasets and open tools remain attached to their original authors and institutions. They are evidence inputs to Critical Load research, not Critical Load publications.
CL-R03 / 2026
Where Do the Joules Go?
Jae-Won Chung, Ruofan Wu, Jeff J. Ma and Mosharaf Chowdhury / ML.ENERGY
Task choice can move inference energy by 25×.
Measurements cover 46 models, seven tasks and 1,858 configurations; video generation can exceed image generation by more than 100×.
CL-R05 / 2025
HGX H100 Product Carbon Footprint
NVIDIA / analysis by WSP
1,312 kgCO₂e cradle-to-gate per HGX H100 baseboard.
Materials and components account for 91% of the reported footprint, creating a concrete downstream anchor for embodied-compute accounting.
CL-R04 / 2023
Aqueduct 4.0
World Resources Institute
Twenty-five countries face extremely high annual water stress.
The basin context turns facility water demand into a place-specific constraint rather than a context-free efficiency ratio.
CL-R06 / 2025
GREEN
Kaiqiang Xu, Decang Sun, Han Tian, Junxue Zhang and Kai Chen / HKUST and USTC
Up to 41.2% lower cluster carbon footprint.
The NSDI evaluation also reports a 12% peak-power reduction, with a 3.6–5.9% job-completion-time trade-off.
CL-R06 / 2024
DC-CFR
Soumyendu Sarkar et al. / Hewlett Packard Enterprise
14.5% carbon, 14.4% energy and 13.7% cost reductions.
The multi-agent control study evaluates carbon-aware workload, cooling and battery decisions across multiple regions for one year.
CL-R06 / OPEN TOOL
DCRL-Green
Soumyendu Sarkar et al. / Hewlett Packard Enterprise
An open environment for joint data-centre control research.
The modular multi-agent environment supports reproducible experiments across workload scheduling, cooling and energy storage.
Reference records remain distinct from Critical Load publications. Authorship, affiliation, method, source and release status stay visible.
01 / Research programme
Six research programmes, one measurement problem.
The programme connects direct telemetry, lifecycle compute carbon, AI inference benchmarks, local water context, semiconductor manufacturing carbon and workload routing into a single place-based model of compute capacity, local impact and decision.
Reference telemetry
Workload energy
409.7 W direct GPU reference signal
Hardware energy counters can anchor workload accounting before allocation.
Working reference recordReference validation shows 34.4–49.5 W MAE for one estimator run; provenance remains part of the release gate.
Open research →
LifecycleCompute carbon
Lifecycle carbon does not end at electricity
Semiconductor, hardware, facility, electricity, cooling, network and replacement are allocated to useful compute.
Working model illustration1 MWh IT spans 90–1,080 kgCO₂e across the current reference electricity contexts before the full lifecycle stack is added.
Open research →
BenchmarkAI inference
/1k
25× energy / response
Reasoning averages 4,625 J vs 184 J for conversation in the current B200 benchmark set.
ML.ENERGY / 202646 models · 7 tasks · 1,858 configurations.
Open research →
ContextWater constraint
>80% = extreme stress
A litre only becomes decision-useful when placed inside basin availability and competing demand.
WRI AQUEDUCT 4.0WRI reports 25 countries facing extremely high annual water stress.
Open research →
Industrial methodSemiconductor carbon
FABWAFERDIEPKGSYS
From fab carbon to embodied compute
Process gases, abatement, electricity, yield and advanced packaging are carried through to useful compute.
Working methodSemiconductor method development + public product-carbon benchmark bridge.
Open research →
DecisionTime & region
NOW
SHIFT
Time & region shifting under constraints
Published scheduling research shows that flexible workloads can reduce carbon, but the feasible move depends on service and infrastructure constraints.
Research synthesisCritical Load extends carbon-aware routing to water, grid, reliability and cost.
Open research →
02 / Active research
What we are working on now.
Programmes stay broad; projects make the work concrete. Status labels distinguish active method development, evidence assembly and scoped future work.
ID
Project
Programme / responsibility
Status
Next public output
CL-P01
Compute Energy Measurement ProtocolFrom hardware telemetry to accepted-task energy evidence.
R01 / R03 · Yan Zhang
Methods release published
CL-P02
Carbon of ComputeLifecycle carbon from semiconductor manufacturing to useful delivered compute.
R02 / R05 · Runwen Jin
Working research
Working paper + open schema
CL-P03
Place Water Constraint ModelTranslate facility water demand and basin context into a supportable compute envelope.
R04 · Research Group · hydrology reviewer sought
Method development
Model note; independent hydrology review planned
CL-P04
CL-001 Place AssessmentFirst place-based evidence assembly linking capacity, impact and decision.
Cross-programme · Xuguang Ma / Yan Zhang
Evidence assembly
Baseline assessment; no rating issued yet
CL-P05
Constraint-aware Compute RoutingTime and region shifting after service, grid, water and sovereignty constraints.
R06 · Yan Zhang
Working research
Research note + decision-record format
03 / Capacity → impact → decision
Capacity is the first answer. Impact makes it useful.
Critical Load first identifies how much compute a place can support. It then tests what a proposed project changes locally and turns the result into a planning decision. The working example below keeps the existing V8.5 illustrative 240 MW / 176 MW case — it is not a rating of a real place.
Illustrative assessment / same V8.5 reference case
176MW
Water is the binding constraint.
Proposed load: 240 MW. The first local constraint is reached at 176 MW, leaving the proposal 64 MW outside the illustrated supportable envelope.
Capacity
240 MW proposed → 176 MW supportable
−64 MW margin
Water impact
The local water threshold is crossed before the full project load is reached.
Binding
Grid / power
Still tested independently; not the first binding limit in this reference case.
Constraint record
Carbon / environment
Impact remains visible as a separate result rather than disappearing into one ESG score.
Impact record
Decision outputDo not approve the 240 MW proposal unchanged. Resize to the supportable envelope or change the condition that makes water binding, then reassess.
04 / Scenario comparison
The same place produces a different decision as compute grows.
A Critical Load assessment is not a single score. It compares proposed scales against the same place boundary so the point at which a constraint becomes binding is visible.
| Assessment line | +100 MW | +176 MW | +240 MW |
|---|---|---|---|
| Capacity state | Inside envelopeBelow the first binding threshold. | Critical LoadThe first local constraint reaches its limit. | 64 MW overProposal exceeds the illustrated supportable capacity. |
| Binding state | No binding limit yetContinue monitoring all constraints. | Water bindsWater defines the maximum supportable load. | Water exceededMitigation or resizing is required. |
| Decision | Feasible in envelopeSubject to the full evidence record. | Threshold conditionDo not treat the boundary as spare capacity. | Redesign / mitigateDo not proceed unchanged. |
Illustrative decision structure only. A real place assessment replaces these reference states with measured or sourced baseline data, project assumptions, impact deltas, uncertainty and a versioned evidence chain.
05 / Featured research
Reasoning changes the environmental unit of AI.
A response is not a fixed unit of compute. Output length, batching, model behaviour and hardware can change energy by an order of magnitude or more.
25×
Mean energy per response in the reference benchmark.
Reasoning tasks averaged 4,625 J per response versus 184 J for conversation tasks across the benchmark synthesis.
FIGURE 03.1 · CL-E03-B01 · SOURCE / ML.ENERGY · arXiv:2601.22076
06 / Publications
Every result carries its evidence trail.
Methods, data sources, assumptions, uncertainty and version history remain visible in every publication. Working outputs are labelled as such and are not presented as peer reviewed.
CL-MN-001
Critical Load methodology
System boundaries, constraint logic, evidence classes and confidence.
CL-RS-001
Compute constraint synthesis
Six research programmes assembled into the first working measurement architecture.
CL-DS-001
Evidence & data schema
A common record structure for measured, modelled and derived environmental evidence.
07 / Research agenda
The questions we intend to answer next.
A 2027–2030 research agenda turns the six programmes into a multi-year public research mission rather than a collection of one-off studies.
MISSION 01
Measure compute
Reproducible measurement from chips and hardware to workloads and useful compute.
MISSION 02
Measure the full footprint
Energy, lifecycle carbon, water, land, infrastructure and community exposure.
MISSION 03
Measure places
Build comparable place assessments and a longitudinal evidence base for major compute regions.
MISSION 04
Turn evidence into decisions
Capacity envelopes, scenarios, conditions, alternatives and public decision records.
08 / Institute
Designed to behave like a measurement institution, not a dashboard.
The institute's job is to make compute capacity, lifecycle impact and local consequences legible, comparable and reviewable.
01
Methods before claims
Definitions and boundaries are published before ratings.
02
Evidence classes stay visible
Measured, modelled and derived values are never collapsed into the same label.
03
Place matters
Grid and watershed context are part of the unit of analysis.
04
Research remains challengeable
Versioning, review, limitations and data provenance are part of every publication.