We measure how much compute a place can support, what that compute changes from manufacture to operation, and which conditions should govern development.
Explore the research
The research boundary runs from semiconductor manufacturing, hardware and workload energy through electricity, cooling, water, land, infrastructure and community effects.
These publications, datasets and open tools form part of the institute's active research base. Each record keeps its authors, institutional affiliation, method boundary and source link visible so the work can be inspected, cited and extended.
Jae-Won Chung, Ruofan Wu, Jeff J. Ma and Mosharaf Chowdhury / ML.ENERGY
Measurements cover 46 models, seven tasks and 1,858 configurations; video generation can exceed image generation by more than 100×.
NVIDIA / analysis by WSP
Materials and components account for 91% of the reported footprint, creating a concrete downstream anchor for embodied-compute accounting.
World Resources Institute
The basin context turns facility water demand into a place-specific constraint rather than a context-free efficiency ratio.
Kaiqiang Xu, Decang Sun, Han Tian, Junxue Zhang and Kai Chen / HKUST and USTC
The NSDI evaluation also reports a 12% peak-power reduction, with a 3.6–5.9% job-completion-time trade-off.
Soumyendu Sarkar et al. / Hewlett Packard Enterprise
The multi-agent control study evaluates carbon-aware workload, cooling and battery decisions across multiple regions for one year.
Soumyendu Sarkar et al. / Hewlett Packard Enterprise
The modular multi-agent environment supports reproducible experiments across workload scheduling, cooling and energy storage.
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.
Hardware energy counters can anchor workload accounting before allocation.
Semiconductor, hardware, facility, electricity, cooling, network and replacement are allocated to useful compute.
Reasoning averages 4,625 J vs 184 J for conversation in the current B200 benchmark set.
A litre only becomes decision-useful when placed inside basin availability and competing demand.
Process gases, abatement, electricity, yield and advanced packaging are carried through to useful compute.
Published scheduling research shows that flexible workloads can reduce carbon, but the feasible move depends on service and infrastructure constraints.
Programmes stay broad; projects make the work concrete. Status labels distinguish active method development, evidence assembly and scoped future work.
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.
Proposed load: 240 MW. The first local constraint is reached at 176 MW, leaving the proposal 64 MW outside the illustrated supportable envelope.
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.
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.
Reasoning tasks averaged 4,625 J per response versus 184 J for conversation tasks across the benchmark synthesis.
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.
System boundaries, constraint logic, evidence classes and confidence.
Six research programmes assembled into the first working measurement architecture.
A common record structure for measured, modelled and derived environmental evidence.
A 2027–2030 research agenda turns the six programmes into a multi-year public research mission rather than a collection of one-off studies.
Reproducible measurement from chips and hardware to workloads and useful compute.
Energy, lifecycle carbon, water, land, infrastructure and community exposure.
Build comparable place assessments and a longitudinal evidence base for major compute regions.
Capacity envelopes, scenarios, conditions, alternatives and public decision records.
The institute's job is to make compute capacity, lifecycle impact and local consequences legible, comparable and reviewable.
Definitions and boundaries are published before ratings.
Measured, modelled and derived values are never collapsed into the same label.
Grid and watershed context are part of the unit of analysis.
Versioning, review, limitations and data provenance are part of every publication.
Six long-term programmes connect semiconductor manufacturing, lifecycle carbon, physical telemetry, AI workloads and place constraints into one public measurement architecture.
Research questions lead to active projects; projects release methods, datasets, working papers and assessments.
A place is constrained by the first physical infrastructure limit or environmental budget that reaches its limit.
| Class | Meaning | Example |
|---|---|---|
| Measured | Direct physical observation | GPU cumulative energy counter |
| Observed / reported | Operational or provider record | Cloud usage, facility PUE |
| Modelled | Estimated using versioned coefficients | Instance energy, facility embodied allocation |
| Contextual | Location/time environmental factor | Grid factor, basin stress |
| Derived | Critical Load calculation carrying upstream IDs | lifecycle gCO₂e / GPU-hour, water-limited MW |
Physical telemetry, usage, manufacturing and site records.
Grid, watershed, climate, procurement and regulatory boundary.
Move shared manufacturing and infrastructure impact to workload units.
Estimate supportable compute under each physical or environmental budget.
Compare the place baseline with the proposed compute scenario and preserve the local deltas.
State the supportable envelope, binding limit, project conditions and alternatives.
CL-LC defines a compute lifecycle boundary spanning semiconductor manufacturing, hardware integration, facility construction, electricity, cooling, shared services, replacement and end of life.
| Layer | Included carbon sources | Typical denominator |
|---|---|---|
| Semiconductor | Fab electricity, process gases, abatement, chemicals, UPW, wafer/die yield, packaging | good die / packaged device |
| Hardware | Boards, servers, racks, memory, storage and supporting electronics | lifetime useful compute |
| Facility | Building, steel/concrete, UPS, transformers, switchgear, batteries, cooling plant | capacity × time × utilization share |
| Operation | IT electricity, facility overhead, backup generation and location-specific electricity context | workload energy / GPU-hour / job |
| Cooling | Cooling electricity when measured bottom-up, refrigerants, embodied cooling equipment | facility or workload allocation |
| Shared services | Network and storage attributable to the compute service | GB transfer / GB-hour / service share |
| Replacement / EOL | Component replacement and end-of-life treatment within the declared boundary | lifetime useful compute |
Critical Load keeps three electricity statements separate: physical location-based electricity, procurement or market-based claims, and lifecycle emissions of the generation assets themselves.
What the local electricity system supplied in the stated time and place boundary.
PPA, certificates and other contractual attributes remain a distinct claim rather than replacing the physical record.
Manufacturing, construction, maintenance and end-of-life emissions of electricity infrastructure where the study boundary requires them.
Lifetime, utilization, GPU-hours, delivered FLOPs or another declared useful-work denominator.
Facility embodied carbon is assigned by capacity, service life, IT load share and utilization.
Traffic, storage-hours or a documented service allocation connects shared infrastructure to the workload.
Each term is first normalized to the same useful-compute denominator and time boundary.
If facility electricity is estimated as IT energy × PUE, cooling electricity is already inside the facility total. A bottom-up cooling-energy record replaces that portion; it does not add to it.
If a lifecycle electricity factor already includes generation-asset construction, those assets are not added again as a separate term.
Semiconductor, package and system records are reconciled before aggregation.
Fab process inventory → wafer basis → good-die allocation → advanced packaging → product carbon → lifetime useful compute. CL-SCM is the upstream manufacturing method feeding the broader CL-LC lifecycle carbon model.
Only after the full lifecycle intensity is normalized to the same time and capacity boundary does carbon enter the Critical Load minimum operator.
The data model preserves the chain from physical observation and manufacturing evidence to lifecycle compute carbon and Critical Load.
Telemetry, manufacturing inventory, reported usage, model coefficient or environmental context with source, time, location and uncertainty.
Place baseline, scenario impacts, lifecycle intensity and constraint calculation for a defined site, region and workload boundary.
Selected optimization option plus rejected alternatives and the constraints that governed the choice.
| Field | Purpose |
|---|---|
| Place boundary | Site, grid zone, watershed and administrative region |
| Time boundary | Observation, manufacturing and validity interval |
| Compute unit | MW, GPU-hour, job, request, token or useful-compute denominator |
| Lifecycle boundary | Included semiconductor, hardware, facility, operation and EOL stages |
| Evidence class | Measured, reported, modelled, contextual or derived |
| Method version | Calculation, allocation and coefficient release |
| Uncertainty | Confidence, error and missing-data state |
The lifecycle record does not collapse manufacturing and operation into one unexplained number.
| Record group | Representative fields | Evidence rule |
|---|---|---|
| Semiconductor | Fab, node, gases, abatement, wafer, yield, package, HBM | Carry CL-SCM method and primary/secondary evidence state |
| Hardware | Accelerator, board, server, rack, lifetime, utilization | Declare product boundary and useful-compute allocation |
| Facility embodied | Building, UPS, transformer, switchgear, battery, cooling plant | Declare service life, capacity and workload share |
| Electricity | IT kWh, facility kWh, PUE or bottom-up overhead, grid factor | Keep location-based, procurement-based and lifecycle factors distinct |
| Cooling | Cooling electricity, refrigerant type/leakage, equipment embodied carbon | Record whether energy is already captured by PUE |
| Shared services | Network GB, storage GB-hour, service allocation | Version model coefficients and service boundary |
| Lifecycle result | kgCO₂e/GPU-hour, kgCO₂e/job, gCO₂e/1k tokens or useful compute | Carry every upstream evidence ID and allocation rule |
The semiconductor record preserves the variables that determine whether a chip-carbon result can actually be reproduced.
| Field group | Representative fields | Why it matters |
|---|---|---|
| Fab identity | Fab location, reporting period, process node, wafer diameter | Sets electricity, technology and manufacturing boundary |
| Process inventory | Electricity, process gases, abatement, chemicals, UPW | Captures direct and energy-related manufacturing impact |
| Yield allocation | Wafer starts, die area, gross dies, good-die yield | Determines embodied impact per usable die |
| Packaging | Substrate, backend process, HBM / memory, package type | Adds post-fab manufacturing impact |
| Evidence quality | Primary / secondary, confidentiality, uncertainty, source date | Prevents confidential gaps from disappearing into a single factor |
| Method record | Allocation rule, method version, evidence IDs | Allows the final lifecycle result to be challenged and reproduced |
The working release can export the research registry and schema. A public source-data repository and code repository are not yet claimed as released.
Working publication IDs, versions, types, status and programme links.
Working CL-E / CL-A / CL-D field structure. This is a specification, not a research dataset.
Source-level licensing, provenance and place evidence will be released after evidence lock.
No public code repository is claimed in this working release.
Critical Load is built by one interdisciplinary internal team spanning research leadership, carbon systems, compute and data engineering, verification, sustainability, regulation, policy and research partnerships.
Roles below describe current responsibilities inside the institute. Publication authorship and review responsibility are attached only when a specific research output is released.
The team works across the six research programmes and the shared measurement system that connects compute, place, impact and decision.
Names, roles and backgrounds reflect current internal responsibilities. Programme ownership, publication authorship and review roles remain visible at the level of each specific output.
Critical Load is an independent research initiative focused on how much compute a place can carry, what additional compute changes locally, and how those results should inform infrastructure decisions.
Our objective is not to label compute “green.” It is to make compute capacity, lifecycle impact and local consequences measurable enough to be challenged and used.
Funding must not buy a research result or rating.
Method, boundaries, evidence class and limitations accompany published results.
Common units and versioned records allow places and workloads to be compared.
Published methods are designed to be reproduced, tested, corrected and revised.
The working release does not claim a board, peer-review system, charity status or completed place ratings that have not yet been established. Formal governance and external review are built alongside the research.
Programme structure, working methods, named team responsibilities and versioned research outputs are public.
Board and scientific-advisory structures will be published only after appointments are accepted and conflicts are disclosed.
Critical Load ratings are not issued until the constraint model, evidence classes and review procedure are stable enough to support a public claim.
A future record will state capacity, binding constraint, local impact deltas, confidence, boundary, version and decision conditions — not just a letter grade.
Funding is tied to public research assets: methods, data, place assessments, reproducible analysis and independent review. A funder does not purchase a conclusion, rating or preferred policy outcome.
Each workstream is prepared as a 12-month research package with defined milestones, publication requirements and independent review.
Develop the lifecycle carbon framework from semiconductor manufacturing and hardware through facility and operation to useful compute.
Build public, evidence-traceable place assessments that connect physical capacity, project scenarios, local impacts and decision conditions.
Build the versioned data infrastructure needed to compare major compute regions over time without collapsing local constraints into one opaque score.
Support expert review, disclosed conflicts, responses to review, uncertainty work and a public corrections log.
No funder control over research conclusions or ratings.
Project support and material conflicts are disclosed with the output.
Publication terms are agreed before funding; negative findings are not suppressed.
Donation or tax treatment is stated by the actual contracting entity; this site makes no unverified charity-status claim.
Priority uses are researcher time, source-data access where lawful, compute and data infrastructure, field or place evidence, independent technical review and public publication.
How physical GPU and host energy becomes a workload-level environmental record.
Abstract. The programme connects direct board power and cumulative energy counters with activity-based attribution and model validation. The objective is not merely to report watts, but to establish an auditable path from physical hardware to a workload, job or GPU-hour.
A reference A100 telemetry record exposes 409.692 W board power and 319,159,391 mJ of cumulative energy — 88.655 Wh at the observation point.
Instantaneous GPU board power in the reference record.
Hardware energy counter used as the integration anchor.
Cumulative energy converted to a more usable unit.
Energy counter differences provide an independent power signal.
Evidence status: the 409.692 W telemetry example is a working reference record. Source provenance must be attached before it is promoted to a formal Critical Load publication result.
Node and GPU energy is shared. A useful compute metric requires an explicit attribution rule.
Read GPU and host energy counters plus utilization and process activity.
Estimate idle and dynamic power over the same observation window.
Allocate dynamic energy using workload activity and shared resource time.
Convert power samples into energy for the job, pod, process or GPU-hour.
Preserve source, estimator version, window and uncertainty in the evidence record.
The model must expose the allocation of shared idle power and the activity signal used for dynamic power.
In the reference validation, estimator-side mean absolute error was materially lower than the local model.
Compute does not become zero-carbon when its electricity becomes green.
Abstract. This programme defines a full lifecycle carbon boundary for compute: semiconductor manufacturing, hardware integration, data-center infrastructure, electricity, cooling, shared network/storage, replacement and end of life. The result is allocated to useful compute and can then be translated into a carbon-limited capacity.
The method separates semiconductor, hardware, facility, electricity, cooling, shared services and replacement/end-of-life so that upstream evidence remains inspectable instead of disappearing inside one factor.
All terms must first share the same system boundary and allocation denominator.
The visual below is a qualitative scenario, not a measured percentage dataset. It illustrates the accounting consequence of sharply reducing operational electricity carbon while manufacturing and infrastructure remain.
Normalized schematic only. The actual shares are calculated from the declared hardware, facility, electricity, lifetime and manufacturing evidence for each assessment.
At PUE 1.135, 1,000 kWh of IT energy becomes 1,135 kWh of facility energy. Across the six reference electricity contexts, the operational result spans 90.1–1,080.3 kgCO₂e per MWh IT.
Location-based, procurement-based and lifecycle electricity factors remain separate evidence statements.
Included through PUE when using a top-down facility model, or measured directly in a bottom-up model.
Leakage is a separate greenhouse-gas source and is not electricity carbon.
Chillers, pumps, heat exchangers and other cooling assets contribute manufacturing and construction carbon.
Bottom-up cooling energy replaces the relevant facility-overhead estimate; it is not added a second time.
Facility embodied carbon also includes building structure, transformers, UPS, switchgear, batteries and power distribution. Those assets are allocated by service life, capacity and workload share.
Allocate manufacturing carbon by lifetime utilization, GPU-hours, delivered FLOPs or another declared useful-work denominator.
Allocate building and equipment embodied carbon by capacity × time × utilization share.
Allocate shared service carbon by traffic, storage-hours or a documented service allocation.
Output units can be kgCO₂e/GPU-hour, kgCO₂e/job, gCO₂e/1k tokens or another reproducible compute denominator.
Once full lifecycle carbon is expressed on the same compute and time boundary, a stated emissions budget can be converted into maximum supportable compute.
The result is normalized to the same time and capacity basis as power, grid, water, cooling and other Critical Load constraints.
| Layer | Evidence class | Example record | Release rule |
|---|---|---|---|
| Semiconductor | Primary / modelled | fab inventory, yield, package | carry CL-SCM version |
| Hardware | Reported / modelled | product PCF, lifetime | declare product boundary |
| Facility | Reported / modelled | construction and equipment carbon | declare capacity-time allocation |
| Electricity | Measured / contextual | kWh, PUE, grid factor | state electricity claim type |
| Cooling | Measured / reported | energy, refrigerant, cooling assets | state PUE overlap treatment |
| Lifecycle compute carbon | Derived | kgCO₂e/GPU-hour or job | carry all upstream evidence IDs |
Reasoning, batching, hardware and precision change the energy cost of a response.
Abstract. The programme synthesizes measured inference benchmarks into an environmental intensity framework that separates response-level energy, token efficiency, hardware choice and methodological uncertainty.
Model families included in the reference benchmark synthesis.
Conversation, reasoning and multimodal workloads.
Hardware / batch / precision combinations.
H100 and B200 in matched comparisons.
Chung et al. (2026), Where Do the Joules Go? Diagnosing Inference Energy Consumption, ML.ENERGY, arXiv:2601.22076 ↗. The published measurement boundary and configuration-level results are retained in this research record.
At the reference settings, conversation energy was 95 J per response; reasoning was 2,192 J.
| Qwen 3 32B / B200 | Conversation | Reasoning |
|---|---|---|
| Mean output | 627 tokens | 7,035 tokens |
| Max batch | 512 | 128 |
| Energy / token @ BS128 | 0.209 J | 0.312 J |
| Energy / token @ max batch | 0.151 J | 0.312 J |
| Energy / response | 95 J | 2,192 J |
Higher concurrency can cut energy per token by roughly three to five times in the reference analysis.
B200 won 63 of 72 matched-latency LLM comparisons, with a median ~35% energy reduction.
FP8 won none of seven small-batch comparisons; median energy was about 30% higher.
A 400-token Mistral Small 3.2 estimate changed from 1.43 Wh to 0.022 Wh after the methodology incorporated more production-like batching and hardware assumptions.
A litre of water does not mean the same thing everywhere.
Abstract. The programme connects data-centre water use with basin-level scarcity and seasonal conditions, then translates that context into a maximum supportable compute load.
The reference framework classifies annual water stress from low (<10%) to extremely high (>80%) based on withdrawals relative to available supply.
| Stress band | Withdrawals / available flow | Interpretation |
|---|---|---|
| Low | <10% | Relatively low competition for supply |
| Low–medium | 10–20% | Increasing competition |
| Medium–high | 20–40% | Material constraint signal |
| High | 40–80% | High competition for supply |
| Extremely high | >80% | Strong potential binding constraint |
Reference finding: extremely high annual water stress.
Exposed in the annual reference finding.
Face high water stress at least one month per year.
Threshold used for the strongest stress class.
World Resources Institute, Aqueduct 4.0 / 2023 water-stress findings ↗. Basin stress, facility demand and competing uses are kept distinct in the place assessment.
At PUE 1.20 and WUE 0.50 L/kWh, a continuously loaded 100 MW IT facility implies roughly 525,600 m³ of annual site water use.
The next step is not to label this number “good” or “bad”; it is to test it against basin, utility, seasonal and permit constraints.
From fab process inventory to wafer, die, package and lifetime useful compute.
Abstract. Critical Load's semiconductor programme treats one of carbon accounting's most data-constrained areas as a first-class measurement problem. The working method traces fab electricity, process gases, abatement, chemicals, ultrapure water, yield and advanced packaging through to product-level embodied carbon and, finally, a compute-intensity denominator.
The hardest variables are often inside the fab: process gas use and destruction, cleanroom electricity, chemical and water inventories, node-specific process intensity, equipment allocation, wafer yield and confidential supplier data.
Gas-specific inventory and destruction efficiency can materially change direct process emissions.
Nameplate abatement is not automatically equivalent to realized destruction performance.
The same wafer-level impact is allocated across the number of usable dies that leave the process.
Water and chemical treatment belong inside the manufacturing inventory, not outside the boundary.
Electricity, process gases, abatement, chemicals, UPW, equipment and facility allocation.
Process node, wafer diameter, wafer starts and manufacturing period.
Die area, gross dies, yield and usable output determine the allocation denominator.
Substrate, advanced packaging, memory/HBM and backend manufacturing are added.
Product carbon is normalized by lifetime utilization and delivered compute.
Yield, allocation rule and data-quality class remain attached to the result.
Our semiconductor work is designed around sensitivity rather than a single universal chip factor. The method tests which input actually controls the result under the stated manufacturing configuration.
| Variable | What changes | Expected result effect | Evidence requirement |
|---|---|---|---|
| Fab electricity mix | kgCO₂e / kWh | Operational manufacturing carbon shifts with local electricity | Fab / supplier period |
| Process gas + abatement | Gas use and destruction | Direct process emissions can become a major driver | Gas-specific inventory |
| Good-die yield | Usable dies per wafer | Lower yield raises impact allocated to each good die | Node / product yield |
| Advanced packaging | Substrate, HBM, backend | Packaging can shift the balance after die fabrication | Package bill / process data |
Directional result structure. Proprietary fab-level primary data is not exposed in the public working site.
A public HGX H100 baseboard PCF reports 1,312 kgCO₂e cradle-to-gate, with 91% attributed to materials and components. Critical Load uses public product-level PCFs as benchmark checks, not as a substitute for upstream semiconductor process research.
In the existing public reference comparison, embodied compute intensity falls from 0.66 to 0.50 gCO₂e/exaflop — about 24% — illustrating why product carbon alone is not enough for hardware comparison.
Utilization, lifetime and delivered work are first-class parameters.
CL-SCM ends at an embodied compute intensity. CL-R02 then combines that upstream result with hardware integration, facility, electricity, cooling, shared services and replacement/end-of-life to produce full lifecycle compute carbon.
The semiconductor result never substitutes for the rest of the lifecycle; it feeds it.
The lowest-carbon location is not always the correct place to run compute.
Abstract. This programme treats workload placement as a constrained optimization problem: first define the feasible set from service and infrastructure conditions, then compare carbon, water, grid stress and cost without hiding the trade-offs.
Published studies report material reductions under specific workloads, regions and control assumptions. Critical Load treats those results as evidence about opportunity, not as a transferable constant.
Cluster-wide carbon-footprint reduction reported in the GREEN evaluation.
Peak-power reduction reported alongside the GREEN carbon result.
Carbon reduction reported by the DC-CFR multi-agent control study.
Energy reduction reported in the same one-year multi-region evaluation.
Xu et al., GREEN, USENIX NSDI 2025 ↗; Sarkar et al., DC-CFR, AAAI 2024 ↗. Results remain bounded by each study's workloads, regions and control assumptions.
DCRL-Green provides a modular, configurable multi-agent environment for testing how workload scheduling, cooling and battery control interact. It makes assumptions inspectable and gives the programme a concrete base for scenario replication and extension.
The environment can support controlled comparisons of carbon, energy and cost strategies before place-specific grid, water, capacity and service constraints are added.
Deadline, latency, sovereignty, model and GPU availability.
Power, grid headroom, cluster availability and reliability.
Carbon, water, cooling and local environmental state.
Keep trade-offs visible instead of compressing them into one opaque score.
Publish the chosen option and the rejected counterfactuals.
| Candidate | Carbon | Water | Grid | Service | Decision |
|---|---|---|---|---|---|
| Region A / now | High | Medium | Constrained | Pass | Reject |
| Region B / +4h | Low | High | Available | Pass | Review water |
| Region C / +2h | Low–medium | Low | Available | Pass | Selected |
Illustrative decision-record format. Environmental states are placeholders for the final live evidence inputs, not claims about specific real regions.
Programmes define long-term questions. Projects show the work currently being developed, assembled or scoped, who is responsible, and what becomes public next.
How can physical GPU and host energy be turned into reproducible workload-level evidence without hiding allocation error?
What is the full lifecycle carbon intensity of compute once semiconductor manufacturing, hardware, facility and operation share one boundary?
When does local water availability, seasonal stress or utility capacity become the binding limit on additional compute?
Can one place assessment carry grid, water, lifecycle carbon, land, permitting and community evidence into a defensible capacity and decision record?
How should flexible workloads move across time and place after latency, sovereignty, grid, water, reliability and cost constraints are applied?
How can every published Critical Load result preserve sources, assumptions, uncertainty, method version and correction history?
Critical Load separates working papers, research briefs, method notes and data specifications. Working outputs are explicitly labelled and are not presented as peer reviewed. Every release keeps authorship, affiliation, methods, source links and review status attached to the result.
Formal releases carry authors, type, version, date, evidence IDs, method version, limitations, citation text and change history. Peer review is stated only when it has actually occurred.
The agenda describes intended research directions and milestones. It is a plan, not a claim that the outputs or geographic coverage already exist.
Develop common measurement and allocation rules from telemetry and manufacturing evidence to workload energy and useful compute.
Measure lifecycle carbon, semiconductor manufacturing, cooling, water, land, infrastructure and community exposure rather than reducing compute impact to electricity alone.
Build public place assessments with explicit grid zones, watersheds, time boundaries, evidence quality and comparable scenarios.
Translate measurement into capacity envelopes, impact deltas, mitigation conditions, rejected alternatives and public decision records.
The intended public infrastructure is a common method, open evidence records, reproducible research, comparable place assessments and a body of decision evidence that can be used by governments, infrastructure operators, investors and communities.
These are working institutional policies for funding, conflicts, review and corrections. Formal legal governance structures are published only after they exist.
Funding may define a research question or geographic scope, but it cannot prescribe a rating, conclusion, preferred technology or policy outcome.
Material project support is disclosed with the output. Funding relationships that could reasonably affect interpretation are stated alongside the research.
Authors, researchers and reviewers disclose material financial, professional or institutional conflicts relevant to a release. Conflicted contributions are not represented as independent review.
Working research remains challengeable. Material corrections receive a version change, dated note and explanation; prior versions remain identifiable where practical.
The current working site identifies internal team responsibilities but does not claim a formal board, scientific council or completed peer review where those structures or reviews have not been established. Governance appointments and reviewer disclosures will be added as they become real.
CL-001 is the first place-based pilot intended to connect compute growth with power, grid, water, lifecycle carbon, land, permitting and community evidence. No Critical Load rating is published on this working page.
The release gate is evidence completeness, not a launch date. Proposed scenarios for method testing are +100 MW, +500 MW and +1 GW. Results remain unpublished until place boundaries, evidence quality, uncertainty and independent review requirements are met.
Grid, watershed, administrative, project and time boundaries stated.
Sources, missing data, modelled values and uncertainty made visible.
Comparable project increments run against the same place baseline.
Independent domain review completed before a public rating claim.