← The energy-first turn · Research
Work it out for yourself.
Teaching apparatus for a claim you should not take on trust: that electricity, not transistors, is what now limits computing. Two simulations and two calculations, in that order. You will watch a computation pay for itself and find that the arithmetic is not the expense; then put numbers on it. You will watch a hundred coupled variables fall into an answer without searching for one, which is what the substrates in the map are built to do. Then you will price the transition yourself, including the settings where it disappoints. Everything runs on your own device, and every number is labelled with what kind of number it is.
1. Derive why moving data costs more than computing with it, and say how much more.
2. Price a technology transition from stated assumptions, and identify which assumption is doing the work.
3. Explain how a system computes by settling, and why that suits some problems and not others.
4. Distinguish a modelled figure from a measured one, and refuse to average them.
Watch a computation pay for itself.
Below is a small calculation running one operation at a time. Each dot is an operand being fetched so the arithmetic can happen, and the meter counts what that costs. Predict first: if you move the data closer, or stop moving it at all, how much do you expect the bill to fall? Then change the architecture and watch.
The arithmetic is identical in all three cases. The only thing that changes is how far the numbers travel to reach it, and that is what moves the meter. Most people expect a smaller difference than they see, which is the reason this constraint went unaddressed for as long as it did.
Where does the energy actually go?
Before you move the dial, predict the answer. A thousand multiply-accumulates, the atom of machine learning, and one variable: how far the operands travel. What fraction of the energy do you expect the arithmetic to account for? Now set the dial to zero, so nothing is fetched from off-chip memory at all, and read the ledger. Most people are wrong about this, and being wrong about it is why the memory wall was underestimated for a decade.
ARITHMETIC
0.30 nJ
MOVEMENT
24.0 nJ
MOVEMENT SHARE
ABOVE LANDAUER FLOOR
8×10^9
Movement dominates.
modelled · order-of-magnitude figures: MAC ≈ 0.3 pJ, SRAM ≈ 5 pJ, DRAM ≈ 100 pJ, Landauer floor ≈ 3 zJ at 300 K
Turn the dial to zero and the arithmetic is still nearly ten billion times the physics floor. That is the headroom the field has never claimed. Turn it up and the arithmetic stops mattering at all, which is why the honest response to the energy wall is not a faster multiplier but an architecture that stops fetching — settling in place, computing in memory, or spending no multiply at all.
Let it fall into the answer.
A hundred variables, coupled so that satisfying one tends to frustrate its neighbours. Somewhere among the 2¹⁰⁰ possible arrangements are good ones. Predict first: to find one, how many arrangements do you think a machine has to look at? Then let it settle and count the flips. Nothing here searches. Each variable simply responds to the ones it touches, and the whole system slides downhill into an answer.
left: the variables, gold and blue for the two states · right: the system's energy as it falls
The count is the point. A few dozen flips against a space of 2¹⁰⁰, and the machine never enumerated anything: it followed the gradient it was already sitting in. This is what is meant by computing as relaxation, and it is why the substrates in the map look so unlike a processor. If falling downhill is the computation, then any physics that falls downhill can compute — a circuit settling, spins annealing, currents summing in a memory array. The arithmetic stops being something you fetch operands to perform, and becomes something the material does.
Run it a few times. Notice that you rarely get the same final energy twice, and that it always stops somewhere rather than continuing to improve. Both facts matter: settling finds a good answer quickly and an optimal one rarely, which is exactly the trade that makes it useful for decisions under time pressure and unsuitable for problems where only the exact best will do.
What is the transition worth, and which assumption decides?
Run each scenario and watch which input moves the answer most. This is the research workbook's own model, not a simplification of it. AI data-centre electricity starts at 190 TWh in 2026 and grows 22% a year to 2030, then 12% a year to 2035. Three scenarios set how much of that workload is genuinely amenable to relaxation-style computing, how much better the new substrate is, and how fast anyone adopts it. Every input is an assumption and is labelled as one.
energy avoided per year, 2026 → 2035
AVOIDED IN 2035
—
SHARE OF AI DEMAND
—
ENERGY COST AVOIDED
—
CUMULATIVE CAPEX AVOIDED
—
CO₂ AVOIDED IN 2035
—
modelled · every input below is an assumption, stated with its reason
Open the assumptions, and argue with them
| Input | Value | Why that number |
|---|---|---|
| AI data-centre electricity, 2026 | 190 TWh | Anchored to the IEA's energy-and-AI base case for the AI share of data-centre load. |
| Growth rate, 2027–2030 | 22%/yr | Implies roughly 421 TWh by 2030, consistent with published trajectories. |
| Growth rate, 2031–2035 | 12%/yr | A deceleration assumption. Nothing guarantees it; the curve could stay steep. |
| Industrial electricity price | $0.08/kWh | US industrial average. Large buyers negotiate below it, so this is conservative for cost savings. |
| Grid emissions factor | 0.35 kg CO₂/kWh | Global average intensity. A grid decarbonising faster makes the CO₂ column shrink independently of any of this. |
| Capacity factor | 0.85 | High-utilisation AI facilities run close to flat, unlike general-purpose data centres. |
| All-in capex per GW | $10B | Recent campuses land between $8B and $15B per GW including power and cooling. |
| Amenable share of the workload | 10 / 20 / 30% | The share of AI joules in sampling-like or settling-like work. The least defensible number here, and the one worth attacking first. |
| Efficiency gain on that share | 10 / 100 / 1000× | Anchored low against vendor claims: in-memory compute shows roughly threefold at system level today. |
Change any one of these in your head and re-run the scenarios. The amenable share is doing most of the work, which is why it is stated three ways rather than once, and why a forecast quoting a single figure for it should be read with suspicion.
Two things are visible in every scenario. The savings arrive late: through 2030 they are small in absolute terms, because adoption takes time and the amenable share is bounded. And no setting removes the underlying pressure, because cheaper computing has historically become more computing rather than less. The case rests on capability per joule, not on a promise that demand falls.
Substrates down the side, model classes across. The diagonal is the story.
Eighty organisations, labs and national programmes tracked across two coupled axes: the physics a machine is built from, and the kind of model it can natively run. A cell is only interesting where both sides meet — a substrate that executes relaxation and a model class that computes by relaxing. Entries marked with an asterisk learn at inference time, which is the capability still unclaimed.
| Substrate ↓ / Model class → | Energy-based / EBM | Hopfield / associative | Learns at inference * |
|---|---|---|---|
| Thermodynamic / p-bit | Extropic DTM; Normal Computing | sampling-native candidate | p-bit Boltzmann learning (UCSB, Tohoku) |
| Ising / annealer | OIM-as-sampler (2026) | Toshiba SBM; Fujitsu DA; D-Wave; NTT CIM | — |
| Neuromorphic SNN | Darwin Monkey attractor nets | Loihi 2 / Akida on-chip plasticity * | PKU-CAS memristor neuromorphic |
| Photonic | photonic EBM proposals | photonic Ising (NTT, Lightelligence) | optical matmul for ternary nets |
| In-memory / CIM / PIM | memristor Hopfield crossbars | Hebbian outer-product is native here | CIM ternary inference (Witmem, Houmo) |
| FPGA settling + accounting | Klere ternary settling fabric | settle ×2 * | EFA on Ferric → fabric target * |
| Digital GPU (the incumbent) | EBT on GPUs | AXIOM (VERSES) * | Titans / TTT / DeltaNet *; BitNet b1.58 |
12
neuromorphic & spiking
largest single cluster, and the one with real neuron counts
11
photonic & optical
capital-rich, integration-poor
17
model classes that learn at inference
the algorithms arrived before the silicon
6
in the graveyard
kept deliberately: every wave that died, and why
The graveyard is not decoration. Every prior alternative-computing wave died of one of three things: claims nobody could verify, a toolchain that never arrived, or no model class that actually needed the hardware. Those six rows are kept beside the eighty live ones so the same failure is recognisable the next time it is being repeated.
A modelled number is not a measured one.
Both instruments above are labelled modelled, and they stay that way. The rule the wider programme runs on is that a simulated figure, a stand-in measurement and a metered one carry different provenance and are never averaged together. It is the reason previous alternative-computing waves collapsed: not that the physics was wrong, but that nobody could tell which claims had been checked.
Read the science report → Download the review (PDF) The open specifications The sampling instrument Work to be done