Computing's next constraint is electricity.
For fifty years the limit on computing was how small a transistor could get. That is no longer what binds. Data centres drew roughly 485 TWh in 2025 and the IEA's base case puts them near 945 TWh by 2030; one US state, Virginia, was already spending a quarter of its electricity on them in 2023. Capacity, not capital or talent, is what schedules the next model. This topic maps the escape: the physics that says a far cheaper computer is permitted, the machines being built to claim it, and the honest question of what would count as proof rather than a press release.
The Energy Lab has two exercises: derive the memory tax from first principles, then price the transition from stated assumptions and find which one carries the answer.
Eighty-five recorded claims with their verification status, arranged so that reading them is practice in telling a measured result from an announced one.
PAI-270 builds the same material into five lessons with runnable benches, from Landauer's floor to a receipt that refuses to launder a simulation.
Arithmetic is nearly free. Moving the data is the bill.
Landauer proved in 1961 that the only step of computing which must cost energy is erasing a bit, and at room temperature that floor is about three zeptojoules. Everything above it is a design decision. Stack today's costs against that floor and the structure of the problem is unmissable: the arithmetic is not the expense, the data movement is.
| Operation | Energy | Above the floor |
|---|---|---|
| Erase one bit (physics floor, 300 K) | ~3 zJ | 1× — the wall itself |
| One 8-bit multiply-accumulate | ~0.1–1 pJ | ~10⁸ |
| Fetch operands from on-chip SRAM | ~1–10 pJ | ~10⁹ |
| Fetch from off-chip DRAM | ~100+ pJ | ~10¹⁰ — 100–1,000× the arithmetic it feeds |
| One frontier-model query (serving median) | 0.31 Wh ≈ 1,100 J | ~10²³ |
A DRAM fetch costs a hundred to a thousand times the multiply it feeds. That single ratio is why the escape route is not a faster multiplier: it is an architecture that stops moving operands. Note also where the bill actually lands — between sixty and ninety percent of AI energy goes to inference, not training. The cost is per answer, forever.
Settle, don't shuttle
Today's stack was selected by the hardware that happened to exist, not by energy. GPUs reward dense synchronous matrix multiplication, so the field built models made of it. The alternatives share one move: treat computation as relaxation. An energy-based model answers by settling into a low-energy state rather than executing a fixed sequence of multiplies. Equilibrium propagation learns from two relaxations instead of a backward pass. Ternary models remove the multiply entirely. Each is a way of letting physics do the arithmetic instead of paying to shuttle numbers to a separate unit and back.
Nature already filed the existence proof. A human brain runs perception, planning and continual learning on roughly twenty watts, and does it while learning, not in a separate training phase. That is the benchmark, and the gap between it and a data centre is the size of the opportunity.
One record is already held. The important one is still open.
The paradigm is not waiting on new silicon to prove itself. Toshiba's Simulated Bifurcation Machine performs constrained decisions at microsecond latency in production, finding optimal eight-currency arbitrage with better than ninety percent success. That is a capability record, not an efficiency claim, and it belongs to physical relaxation today. Elsewhere the substrates keep landing: a $75M CHIPS letter of intent for thermodynamic sampling, a record $475M seed for unconventional AI, Zhejiang's Darwin Monkey past two billion spiking neurons, and Samsung and SK hynix — direct rivals — jointly standardising processing-in-memory through JEDEC.
The record still open is the one that matters most: adapting on the fly. No system yet demonstrates learning while inferring, under drift, at a decisively better cost per decision. That is the prize this lab is organised around, and it is why the benchmark question is not a footnote to the hardware story but the centre of it.
Why previous waves died, and what differs now
Every earlier alternative-computing wave died of the same three things: unverifiable claims, a missing toolchain, and no model class that needed the hardware. Four things differ this time. The energy constraint now binds the incumbent, so for the first time the alternative's fitness function is also the industry's bottleneck. The algorithms arrived. The capital arrived. And the software layer is being fought over in the open, which is the tell that participants believe the hardware question is real: Qualcomm paid $3.9B for a write-once-run-anywhere compiler stack.
That history is also the warning. Portability layers get bought as containment, which is why the pieces this lab contributes are deliberately capture-proof: an open receipt schema, an open relaxation dialect, and a reference runtime that anyone can run in a browser tab.
What the Institute contributes
Joules, not tokens
A decision priced in joules and carrying a signed receipt, so that a simulated number and a metered one can never be quietly confused.
see MathGround · TR-2026-06
A browser cannot read power
So browser runs emit modelled receipts and metered hardware emits measured ones, permanently distinguishable by provenance rather than blended into one flattering figure.
provenance is a field, not a footnote
Seven years, lab bench to industrial policy
Worth studying as a case in how a technology transition actually moves, because it is faster than the usual telling and the order is not the one people expect. The physics prize arrives after the engineering, and the state money arrives last.
The shape is the lesson. Seven years separate a research curiosity from a national funding line, and for most of that span the work looked unpromising to anyone counting revenue. The last row is the one to watch, because it is a capability nobody has demonstrated rather than a product nobody has shipped.
The same standard, applied here
The map grades 85 organisations on how well their claims are evidenced. It would be indefensible to publish that and exempt ourselves, so the table below records what our own stack does and does not do, checked against the source rather than the roadmap. Two entries are absent that a reader might reasonably assume are present.
| Capability | State | How this was established |
|---|---|---|
| Runs on the three native graphics backends | Present | Vulkan, Metal and DX12 paths all appear in the source; checked by inspection, August 2026. |
| Imports models from a standard interchange format | Present, partial | A pure-Rust ONNX importer exists with a starter operator set. Enough for linear and small networks; the transformer operators are not yet written. |
| Counts the work a decision costs | Present | Operations per decision are derived analytically from the architecture rather than sampled, so the count is exact rather than estimated. |
| Converts that count into joules | Absent | This needs a calibration table per device class, and no such table has been measured. Counting operations is not the same as knowing what they cost. |
| Emits a signed energy receipt | Absent | Specified in the draft standard on this site; not implemented in any library crate. |
| Refuses work that would exceed an energy budget | Absent | Searched for and not found. The only budgets in the code are memory and scheduling, which are unrelated. |
Publishing the absent rows is the point. A programme arguing that the field should record what it can and cannot demonstrate has no standing unless it does so first, and the gap between counting operations and knowing their cost is exactly the kind of distinction that gets quietly elided elsewhere.
A proposed build order, and why it is ordered that way
The review argues that an open runtime is needed for any of this to be checkable, and sketches how one could be built. This is a proposal, not work in progress: none of the stages below has been started, and the table above records honestly which pieces exist today. It is published because the reasoning about sequence is the transferable part.
Costs weeks rather than years and creates the adoption surface. Anyone can reproduce a result without installing anything, which is what converts a sceptic where a paper does not.
Identical outputs, receipts differing only in how the energy figure was obtained. This is what makes the browser numbers honest rather than merely convenient.
Part of the work on a settling substrate, the rest on a GPU, and the composed receipt taking the weakest provenance. This is the credibility event, and notably it is about plumbing rather than a performance win.
Entrants must beat their own no-adaptation ablation, with sealed seeds. Otherwise the result measures the problem, not the method.
The standards event. It comes last because it is the only stage that depends on someone outside the project agreeing to watch.
The ordering principle is worth extracting from the specifics: cheap and reproducible first to earn attention, the awkward integration second because that is what nobody believes until they see it, and the externally witnessed measurement last because it is the only step you cannot take alone. Nothing in the sequence waits on new silicon.
Open questions
Four questions are unresolved, and the answers are not known. They are recorded here with their current status because a map that only lists supporting evidence is not a map.
Does computing by energy descent retain its advantage at scale?
Unreplicated. The original scaling claims have not been independently reproduced, and no substrate has yet demonstrated a measured tenfold gain on a real workload.
Will a portability layer for relaxation substrates come from inside the incumbent stacks or outside them?
Undecided. No incumbent compiler currently targets these substrates natively; several open specifications exist in draft.
Is adaptation at inference necessary, or is context-as-adaptation sufficient?
Untested. No benchmark currently scores regret under drift at controlled latency, so the comparison has not been run.
Does electricity remain the binding constraint on computing?
Currently yes on published forecasts. Two consecutive downward revisions from the major forecasters would change that.
This is a snapshot, compiled in August 2026 from public sources, and it should be read as one. The field moves faster than a review of it can, so entries will go stale and some already have. The conventions used in compiling it are stated so the work can be checked and extended: every figure carries a verification grade, letters of intent are not counted as awards, and entries that failed are kept with the lesson rather than deleted.
The papers
The full review and the two open specifications it proposes. All three are drafts released for comment; the receipt schema and the relaxation dialect are CC0 by construction, because a portability layer that can be captured is worth nothing.
↓ The Energy-First Turn · TR-2026-25 · PDF↓ OER/1 energy receipts · RFC · PDF↓ RELAX/1 dialect · RFC · PDF
Open the Energy Lab → The global database (85 entries) The open specifications How the map is kept The bar a new machine must clear → Take the course (PAI-270)
The forecast, priced honestly with the amenable share bounded and efficiency gains set below vendor claims, is modest through 2030 and compounds afterwards. That is the unglamorous shape of a real transition, and stating it plainly is the point: this map is kept with verification statuses attached to every figure, and with the conditions that would falsify the thesis written down beside the ones that would confirm it.