Two questions a new machine has to answer at once.
Quantum computing has spent a decade learning a hard lesson in public, and every novel computing paradigm should study it before repeating it. A quantum result appears. It is impressive. Then a classical algorithm is found that reaches the same place on ordinary hardware, and the advantage evaporates. The technique has a name — dequantization — and it has been unusually productive: in April 2026 the short-path algorithms for constraint problems were dequantized, and in August a quantum method for short-time dynamics of local systems went the same way. What survives is real but narrow, and often turns out to be an advantage in memory rather than in time.
Why this keeps happening
The pattern is not bad luck. It follows from what kind of claim is being made. A speedup claim is a statement about computational complexity — this machine reaches the answer in fewer steps — and complexity claims are exactly the kind that a cleverer algorithm can overturn, because the comparison is between two ways of computing rather than between two physical situations. Anyone can attack it without building anything.
So the first discipline for a new paradigm is to know which of your claims are of that kind. If your advantage is that you run the same computation with fewer operations, expect it to be dequantized, and expect the person who does it to need nothing but a laptop and a good idea.
The claim that does not dequantize
The energy paradigm is making a different kind of claim, and the difference matters more than it first appears. Its advantage is thermodynamic rather than algorithmic: not that fewer operations are performed, but that the physical transport those operations normally require does not happen at all. No algorithm makes an off-chip memory fetch cheaper. If the arithmetic occurs where the data already sits, the fetch is not optimised away — it never occurs.
This has a clean test attached. Ask of any energy advantage: does it come from doing fewer operations, or from where the physics happens? The first is dequantizable and should be treated as provisional. The second is a statement about the arrangement of matter, and an algorithm cannot argue with it.
Be honest about the boundary, though. Better algorithms do reduce energy, because fewer operations mean fewer joules, so an energy claim built on operation count inherits every weakness of a speedup claim. The paradigm only escapes the trap where the saving is structural.
What the old machine cannot reach
This is the second question, and it is the one most often dodged. It is also the one most often stated wrongly. The usual phrasing is that the new machine does work that is impossible on traditional hardware, and taken literally that is false: a classical computer can compute anything computable, so any claim of the form only our machine can compute this is either mistaken or is hiding a resource bound without saying so.
The version that is both true and checkable is about resources rather than computability. There is a task, an energy budget and a latency bound, the budget and the bound coming from a situation that genuinely has them, and the old machine cannot finish inside both while the new one can. Stated that way the claim can be attacked, because someone can attempt the task on ordinary hardware and report what happened. Each of the four below is a case of that shape rather than an impossibility.
Randomness is free here and expensive there.
A thermodynamic sampler draws its noise from the thermal bath it sits in — the fluctuations are the power source, not a cost. A deterministic machine has no access to that. It must synthesise randomness by computing it, and pay for every bit. For workloads that are mostly sampling, this is not a speed difference; it is a difference in what the machine is made of.
Settling is one physical event, not a search.
A circuit relaxing into a low-energy state performs, in one continuous motion, something the digital machine must approximate by iterating. The answer arrives at the speed of the physics rather than the clock, which is why microsecond constrained decisions were demonstrated on this hardware years before the surrounding field noticed.
Learning without a separate training phase.
Where the weights are the physical device, adaptation happens as a side effect of operating. There is no gradient copied back through a machine that is otherwise idle. This is the open capability, and it is the one nobody has yet demonstrated convincingly.
A power envelope that cannot be met by scheduling.
Some work has a hard budget — a device in a field, on a battery, with no link home. If the joules per decision exceed the envelope, no amount of scheduling or batching rescues it. This is a constraint the datacentre framing does not have, and it selects for different hardware.
The bar this sets for our own work
A paradigm that only answers the second question is a curiosity: interesting physics with no way to compare it to anything anyone runs. A paradigm that only answers the first is competing on the incumbent's terms, on hardware built for those terms, and will lose. Both answers are required, which is why the measurement discipline and the impossibility argument belong in the same programme rather than in separate papers.
Concretely: report joules per task on workloads people actually run, with provenance attached so a modelled figure cannot pass as a measured one — and separately, state plainly which capability is being claimed as unreachable and what evidence would settle it. Anything that cannot survive both tests should be described as promising rather than proven.
The paper
The argument above is set out in full as a technical report: the dequantization record, the separation of complexity claims from thermodynamic ones, why an advantage located in an algorithm is unstable while one located in the arrangement of matter is not, and the three honest forms a what the incumbent cannot reach claim can take. Every reference carries a verification grade, and the section on what remains open is written so the argument can be attacked at its weak points.
↓ Building the Energy Compute Future · TR-2026-26 · PDF
Work the physics yourself →Run the attack yourself (PAI-270)Work to be doneThe map and the evidence