Reading an Energy Claim
Almost every energy figure published about a machine is a property of something other than the machine. Watt-hours describe a pack; hours of runtime describe a duty cycle somebody chose. Neither answers what a buyer, an operator or an inspector needs, which is how much work comes out per joule in. This course builds the competency for reading such a claim adversarially: what the number is per, what grade of evidence it carries, which conditions were selected, what sits outside the boundary, and who is answerable for it. Every bench runs on published figures, and the closing lab produces a refusal that names its own acceptance conditions.
▶ Start the course ← All coursesWhere this sits, and what moves it.
Binding constraint · Energy per unit of work, at a stated grade of evidence. A figure without a denominator cannot be compared between machines, and a figure without a provenance grade cannot be composed with another one, because a composed claim takes the weakest grade among its parts rather than the average.
Energy claims for machines were read as numbers rather than as claims. Capacity and runtime were treated as efficiency, provenance travelled in prose if at all, and a reader had no method beyond scepticism, which loses to a schedule.
The measurement exists and the reporting does not. Humanoid cost of transport is published at 0.41 to 0.42, a physics-based power model for a commercial humanoid arm was identified from 897 measured trajectories, and an ISO Technical Specification for industrial robot energy consumption was due in August 2026, while no machine on the market states energy per unit of work at a declared grade.
A reader who can name which of five elements is missing, and the remedy for each, is what turns a reporting format into a practice. What would move it further is the obligation itself: data-centre operators already report energy indicators annually in the European Union and an AI energy label is under exploration, and no equivalent exists for a machine that carries its own supply.
Every hard thing was impossible until the constraint that made it impossible was named. How we read a frontier →
What a number has to carry
Separate a property of a pack from a property of a machine, and attach a provenance grade that composes by its weakest part.
- L2What the number is perA vendor publishes capacity and runtime but not duty cycle. Who decides the efficiency number?Separate a property of a battery from a property of a machine, and identify the unit of work a claim is denominated in.→
- L2The grade, not the numberLocomotion is metered and is 68% of the joules. Sensing is a stand-in. What grade is the total?Attach a provenance grade to every figure in a composed claim, and apply the weakest-part rule rather than averaging.→
How a claim flatters
Show that selected conditions move a figure further than engineering does, and test whether an omitted term can change the conclusion.
- L3Choose the conditions, choose the answerA vendor publishes one cost-of-transport figure for a robot. What has been withheld?Show that a single machine yields a wide range of defensible energy figures depending on conditions the publisher selects, and identify which conditions move it most.→
- L3What was left outDetect an incomplete boundary in an energy claim by testing whether the omitted term can change the conclusion.→
- L4Worth how much compute?One table gives eleven architectures' gains from a 5.7x step in training compute, ranging from 1.76 to 6.20 points. A new model leads the baseline by 4.66 points. Converting that lead into compute with each architecture's gain in turn, how far apart are the largest and smallest answers?Convert an architecture's lead into the training compute it would replace, and show that the answer is fixed by the scaling curve chosen to convert it and by how far that curve is carried.→
Judging it
Price a conformance level against the decision it protects, and write a refusal that names its own acceptance conditions.
- L3What a signature buysWhy does the correct conformance level change with fleet size?Price the difference between a self-declared, a metered and a witnessed claim, and identify what each level actually removes.→
- L4Refusing a claimWhat separates an engineering refusal from an objection?Assemble the five questions into a single review, apply them to a real published claim, and write a refusal that names what would make it acceptable.→
- L4Bounding what you cannot measureThe claim omits the cost of reading the answer out of the machine: 2,592 values per trajectory. For the advantage of 10^11 to survive, what must reading ONE of those values cost?Bound an omitted term by inverting the claim it was left out of, and judge the claim by whether that bound is physically reachable.→
- L4What the table would not sayA published table gives a device's inference latency as 102.3 ms and its energy as 2.398 kJ, and does not say what the energy is per. What can you conclude?Recover the energy denominator of a real, published cross-vendor table from physics, and then show what demanding a control rate costs in joules.→
- L4Energy at one length, accuracy at anotherAn analog system is 3.6x more energy efficient but loses accuracy to noise; four sampled answers buy the accuracy back. What is left of the advantage?Combine a published energy advantage with the accuracy it costs, find that the two were measured on different workloads, and locate the single quantity that decides whether the advantage survives.→
- L4Where a switch spends its joulesAn optical network draws under 3% of a supercomputer's power. Swap it for a packet-switched network and training slows by 10%. Before the new switches draw a single watt, how much more energy does a training run take?Recover the share of training time a network governs from two published slowdowns, and show that its energy is decided by the seconds it saves every other chip rather than by its own share of the power.→