What flows in and out?
Usable energy, conversion losses, waste heat and available work. Include the entire system and state its operating conditions.
Explore the assumptions
What would have to be true?
Inspect the proposed connections, follow costs across the system, and test how demand can offset efficiency.
The proposed model
Which connections need testing?
This proposed model has not been tested. Choose a factor to inspect the assumptions behind its connections.
Choose a factor
Each factor has a list of proposed connections. On wider screens, lines show those connections; their width and position do not indicate strength or importance.
Question to investigate
Storage density
Does the storage system’s weight or size limit the outcome?
Storage density → Useful outcomes
Assumption: Mass or volume is a binding constraint; installed-system overhead does not erase the gain.
How to test it: Compare delivered service at fixed system mass or volume and duty cycle.
Storage density → Cost and access
Assumption: Changes in materials, manufacturing and lifetime can outweigh storage gains.
How to test it: Compare full lifecycle costs at identical service and reliability.
Storage density → Costs and harms
Assumption: A denser chemistry or carrier changes extraction, containment and end-of-life burdens.
How to test it: Publish a comparative lifecycle inventory with material quantities and uncertainty.
Research computing → Storage density
Assumption: Extra search produces replicable, manufacturable storage improvements.
How to test it: Audit discoveries against prior art and independent physical tests.
Read all 13 proposed connections
Storage density → Useful outcomes
Assumption: Mass or volume is a binding constraint; installed-system overhead does not erase the gain.
How to test it: Compare delivered service at fixed system mass or volume and duty cycle.
Storage density → Cost and access
Assumption: Changes in materials, manufacturing and lifetime can outweigh storage gains.
How to test it: Compare full lifecycle costs at identical service and reliability.
Storage density → Costs and harms
Assumption: A denser chemistry or carrier changes extraction, containment and end-of-life burdens.
How to test it: Publish a comparative lifecycle inventory with material quantities and uncertainty.
Energy efficiency → Cost and access
Assumption: Operating energy costs matter and savings are not fully offset by capital or maintenance.
How to test it: Compare total costs using a declared duty cycle and lifetime.
Energy efficiency → Costs and harms
Assumption: The same service is delivered and manufacturing overhead does not erase operating savings.
How to test it: Compare whole-system totals across lifecycle and demand scenarios.
Infrastructure → Useful outcomes
Assumption: Delivery or availability constrains the service.
How to test it: Compare outage, queue or curtailment data before and after an intervention.
Infrastructure → Costs and harms
Assumption: Infrastructure adds embodied impacts but can reduce other losses.
How to test it: Include construction, replacement and displaced equipment in the inventory.
Cost and access → Total demand
Assumption: The cost variable is service price and demand responds to it.
How to test it: Estimate a price response using a credible counterfactual; separate unrelated growth.
Total demand → Costs and harms
Assumption: Marginal service has nonzero resource requirements.
How to test it: Measure aggregate consumption and marginal supply effects.
Cost and access → Research computing
Assumption: Energy costs materially constrain research compute and capacity can expand.
How to test it: Compare energy share of compute cost and actual capacity constraints.
Research computing → Energy efficiency
Assumption: Additional research compute may produce valid, deployable improvements.
How to test it: Compare independently validated research yield per total resource cost.
Research computing → Storage density
Assumption: Extra search produces replicable, manufacturable storage improvements.
How to test it: Audit discoveries against prior art and independent physical tests.
Useful outcomes → Total demand
Assumption: Changes in service quality, access and substitution alter total use.
How to test it: Track uptake, displaced services, quality and absolute consumption.
The model guide explains its scope and limits. No causal ranking has been established.
Follow the whole system
Where did the entropy go?
Check where the costs and harms go, too. Energy losses, pollution, expense and risk need separate measures.
Include the full life cycle: extraction, manufacturing, delivery, use, maintenance and end of life.
What grows elsewhere?
Extraction, water, land, emissions, toxicity and replacement needs. Report the absolute totals as well as each unit of service.
Who carries the cost?
Affordability, safety, access, reliability, supply dependence and construction time. Name affected groups and missing evidence.
See the full set of candidate metrics
Choose relevant objectives and hard constraints for a specific service. Keep tradeoffs visible; no automatic combined score. Missing measurements are unknown.
- Usable energy per system mass — Wh/kg
- Usable energy per system volume — Wh/L
- Delivered power at required duration — kW
- Useful work per input energy — J/J
- Quality-qualified tasks per energy — tasks/kWh
- Lifecycle cost per functional unit — declared currency, price year / functional unit
- Capital cost — declared currency and price year
- Operating cost — declared currency and price year / year
- Material availability — material-specific supply measure
- Critical mineral requirement — kg / functional unit
- Manufacturing yield and process readiness — declared yield and readiness basis
- Achievable annual deployment — functional units / year
- Land occupation — m²·year / functional unit
- Water consumption by location — m³ / functional unit
- Heat rejection at stated temperature — kWh thermal / functional unit
- Operational emissions intensity — kg CO2e / functional unit
- Lifecycle greenhouse gas emissions — kg CO2e / functional unit
- Hazard and exposure — declared assessment method
- Recovered material with processing burdens — percent by mass plus process inventory
- Service availability and failure rates — percent availability; failures / service-hour
- Performance under stated disruption — service retained; recovery hours
- Hazards, exposure and residual risk — qualified assessment; not a universal score
- Supply concentration — supplier/country shares and method
- Additional infrastructure — physical inventory plus costs
- Time until service delivered — years from decision
Pareto frontier: options for which no feasible comparator is at least as good on every chosen objective and better on one. This only works with comparable boundaries and sufficiently resolved uncertainty. It does not automatically select a preferred option.
Read the comparison methodTwo possible feedback loops
Does efficiency reduce total energy use?
An efficiency gain can support more research. It can also encourage more use. Both effects need to be measured.
Energy supports research
- Better energy
- Cheaper, more efficient computation
- More scientific computation
- Better models, materials and engineering
- Potentially better energy
Lower costs can increase demand
- Cheaper computation
- More computation
- Greater aggregate energy demand
- More heat, infrastructure and materials
- Possible loss of the original benefit
Illustration, not a forecast
When does growth erase the saving?
Assume every task meets the same quality standard. Adjust the inputs to compare energy use before and after the change.
Energy use, compared with today
100%
The saving is exactly offset by task growth.
100% means no change. Below 100% means less energy.
0.80 × 1.25 = 1.00
A 25% increase in tasks erases a 20% saving per task.
Assumptions and limits
New energy ÷ baseline energy = (1 − saving per task) × (1 + task growth). It excludes manufacturing, hardware replacement, cooling changes, prices and other sectors. It does not estimate how demand responds to price or establish a lifecycle benefit.
Rebound: extra use caused by an efficiency-related change. Jevons-type backfire: induced use more than offsets savings. Task growth from unrelated causes is not automatically rebound. The benefit loop also requires useful, reproducible discoveries; more compute does not guarantee them.
The research process
Test, compare and update.
Investigate independently before sharing conclusions. Then compare evidence and agree on the next useful test. Keep unresolved disagreements visible.
ExpandChallengeTestConvergeRepeat
Start with several possible problem framings. Seek the cheapest test that could rule one out. Different AI roles do not, by themselves, create independent evidence.
Human and AI roles, independence and safety
- Explorer: Find competing explanations and neglected interventions.
- Engineer: Check manufacture, operation, maintenance, deployment and scale.
- Physicist: Check mechanisms, units, conservation laws and physical limits.
- Experimentalist: Design the cheapest discriminating test.
- Systems analyst: Track the service and downstream system changes.
- Environmental auditor: Find lifecycle burdens and affected communities.
- Economist: Compare capital, operating costs, access and deployment timing.
- Red team: Challenge physics, assumptions, metrics, safety, demand and scale.
- Prior-art investigator: Find patents, papers, abandoned programs and negative results.
- Synthesis system: Preserve disagreement and identify the next decision.
Freeze the initial question and record each participant's sources, model or expertise, conflicts and exposure to earlier conclusions. Compare independent evidence summaries before sharing conclusions. Preserve dissent and source overlap. This static repository cannot enforce blind review or run a multi-AI system.
Humans are first-class researchers in every role. Hazardous physical work must remain theoretical until qualified professionals complete the appropriate review and authorize its scope.
Read the research protocol · Read the safety boundaryDoes density deserve priority in the first selected service?
An open framing claim. The service, baseline and decision thresholds must be registered before the comparison.
Research record (JSON)First assignment: a bounded framing audit
A proposed desk audit. Compare density with feasible alternatives before deciding whether a physical test is warranted.
Research record (JSON)Falsified hypotheses
None recorded yet. That means no project hypothesis has been falsified—not that any has survived a test.
Falsified-hypothesis index (JSON)Results and replication
No completed project study or independent replication is recorded. Background sources explain terms; they do not validate the challenge.
Research index (JSON)How evidence and confidence are labeled
Unresolved: the evidence cannot yet settle the bounded claim.
Plausible — experiment required: supporting evidence exists, but the needed test has not resolved the claim.
Falsified: evidence contradicts the registered claim within its stated scope.
Supported within tested conditions: documented independent replication supports the claim within named conditions.
Hypotheses, assumptions, simulations and experimental results are recorded separately. Confidence has a dated rationale; it is never a truth meter.