The Grand Energetic Pivot

1. The hominid precedent

The constraint on both evolutionary and technological complexity is the metabolic bill. A high-functioning organ is a costly receiver: its energy requirement is continuous and non-negotiable. A twenty-watt brain cannot be funded by chewing longer. It has to be funded by moving part of the work outside the body — metabolic offloading — so that internal complexity becomes affordable.

Two frameworks describe how the budget was managed. They are usually presented as rivals. They are better read as two settings on the same ledger.

^{1 2} The transition was made possible by offloading devices that raised net energy yield before the food ever reached a cell. ^{3 4}

• Fire — externalized digestion. Gelatinized starch and denatured protein reduce the internal cost of chewing and digestion. The stomach is pre-paid outside the body.
• Tools — borrowed force. A stone edge applies force where the body could not afford to grow it: a one-time energy payment substituting for a permanent metabolic subscription.
• Cooperation — pooled surplus. Division of labour relocates the bill from a single stomach to a network of bodies.

2. The gut-brain axis, held at arm's length

There is a further layer under the anatomy. The gut microbiome is not a passenger; it sits in the conversion chain between fuel and host, and it plausibly co-evolved with encephalization.

A reported 2026 study transplanted microbiota from large-brained primates and from macaques into germ-free mice, and found divergent host gene expression: energy-production and synaptic-plasticity pathways in one case, fat-storage patterns in the other. ⁵

I want to be careful here, because this is exactly where a good argument goes wrong. This is one recent study in mice. It has not been independently replicated. Correlations that have been drawn between microbial profiles and neurodevelopmental conditions are contested, confounded, and not causal claims — and nothing in this brief should be read as a health claim or as advice about any condition or treatment. What the work earns, at this stage, is a place on the list of open edges: a candidate mechanism for how the internal ledger might be managed, waiting for repeatability.

The standard is not novelty. The standard is repeatability — the number of independent copies of a fact the world happens to be holding.

3. The gigawatt ledger

As cognition is offloaded to machines, the bill does not vanish. It changes address — from calories to grids — and the new address is alarmingly concentrated. ⁶

• Data centres consumed roughly 485 TWh in 2025, about 1.5% of global electricity generation.
• Projections put that near 3.0% of global electricity by 2030.
• AI-specific facilities accounted for roughly 155 TWh, about 0.5% of world electricity.
• Concentration is the real exposure: over 20% of Ireland's electricity consumption and over 25% of Virginia's demand.

Underneath the aggregate sits an efficiency discrepancy of about a million to one. A twenty-watt brain does work that an exaflop machine approaches at twenty megawatts. ⁷ The tasks are not identical, and the comparison is not a verdict — but it is the right ledger. At the micro scale, a single heavy reasoning query on the order of 50 Wh spends what a brain spends in two and a half hours, and a measurable slice of one European citizen's ~17 kWh daily electricity average.

4. Toward hardware that pays its own bill

The von Neumann architecture spends most of its energy moving data between memory and processor. Neuromorphic and in-memory designs — including reported synaptic-resistor circuits using hafnium-zirconium oxide, with Hebbian and spike-timing-dependent learning in place of backpropagation — attempt to remove that migration entirely. ⁸

This is early laboratory work, and I am recording it as a direction of travel rather than a delivered capability. The reason it belongs in the brief is structural, not promotional: it is the first hardware family whose organizing principle is the same one evolution used. Put the memory where the computation happens, and stop paying to carry information across the room.

5. Exactness as an energy strategy — and what the Riemann Hypothesis actually buys

There is an algorithmic version of the same argument. Stochastic modelling buys answers by spending cycles: more samples, more drift, more watts for the same confidence. Analytical structure — spectral methods, invariant constants, the statistics the Riemann zeros keep handing us ⁹ — buys the answer once and does not pay again for the variance. That much is ordinary numerical analysis, and it is a real energy lever.

But I have to correct my own framing here, because a careful reading of the literature does not support the stronger version of the claim. The Riemann Hypothesis is unproven, verified numerically to enormous height, and its established consequences live in analytic number theory: the sharpest error term for the prime counting function, bounds on prime gaps, and something on the order of a thousand conditional theorems that would go unconditional overnight. ¹⁰ Its computational leverage is real but narrow — deterministic primality testing under GRH, sharper estimates for primes in arithmetic progressions. It does not underwrite modern public-key cryptography, and a proof would not break it.

The physics connection is genuine and genuinely beautiful, and it is also still a bridge rather than a road. Montgomery and Dyson noticed, and Odlyzko confirmed numerically, that the spacing statistics of the zeros match the eigenvalue statistics of large random Hermitian matrices from the Gaussian Unitary Ensemble — the signature of a chaotic quantum system. ¹¹ The Hilbert–Pólya conjecture asks whether the zeros are the spectrum of some self-adjoint operator. Nobody has produced that operator.

So: there is no peer-reviewed method that turns the Riemann Hypothesis into exact, zero-error nuclear or subatomic simulation at millisecond latency on commodity x86 and ARM silicon. Claims of that shape are speculative or promotional, and I am not going to carry them in this brief on the strength of their appeal. What survives the audit is modest and still worth stating: where an exact spectral treatment of a system exists, it is often cheaper and stiffer than a Monte Carlo approximation of the same system, and the saving is energetic and not merely aesthetic. That is falsifiable on a benchmark, which is the only property that makes it worth saying. The Riemann material earns its place in this archive as a structural analogy about spectra and repeatability — not as a product claim.

6. The long horizon, with its risks attached

If the efficiency dividend does arrive, the second-order consequences are larger than the compute savings. Four of them seem coherent enough to name.

Compute decentralizes. Collapse the efficiency gap by orders of magnitude and inference and continuous learning stop requiring gigawatt campuses. They move to edge devices, vehicles, and distributed micro-grids. The twenty-watt brain becomes an engineering target rather than a rhetorical benchmark, and the electricity freed up can go to desalination, carbon removal, and industrial electrification instead of to cooling.

Uncertainty becomes a residual rather than the dominant term. Wherever exact treatments scale, real-time digital twins of grids, climate subsystems, and supply chains become tractable with negligible numerical drift, and the posture shifts from reactive optimization under noise to predictive control. This is the branch most contingent on section 5 holding up, and I am flagging it as such.

Cognition stays multi-layer. The gut-brain work, if it replicates, says the internal ledger is still part of the efficiency story — that intelligence is a distributed metabolic network spanning biology and silicon, not a property of either alone. Nothing here is a health claim or a protocol; it is a statement about where the energy is managed.

Chokepoints move. Concentrated demand in Ireland and Virginia is a strategic exposure today. Methods that run on commodity silicon lower the barrier to entry and reduce pressure on exotic fabrication — which democratizes advanced modelling and, in the same motion, makes planetary surplus management a governance function rather than a procurement one.

The risks are proportional to the upside, and they are the ordinary ones. Concentrated neuromorphic or spectral capability creates new single points of failure. Over-reliance on any one analytical framework hides model misspecification, which is exactly the failure the stochastic methods were guarding against. And the transition demands enormous capital and coordination before the dividend appears — which is the part of the story that historically kills good arguments.

7. Metabolic re-absorption

The arc is one long record of managing the metabolic bill. Fire gelatinized the starch. Tools borrowed the force. Cooperation pooled the surplus. Machines took the cognition. Each move bought complexity by relocating cost, and each relocation eventually presented an invoice somewhere new.

The invoice is now on the grid, and it is concentrated in a handful of jurisdictions. The next move, if there is one, is inward: metabolic re-absorption — bringing biological efficiency back to our machines rather than continuing to buy capability by the megawatt. Twenty watts of discipline and the rigor of exact structure. That is the target. Whether we reach it is an engineering question, and it is being answered, right now, by whoever is paying the electricity.