Four comparisons of the two substrates — energy, code, scale, lifetime — each one sourced, bounded, and stating what it counts before it states the number.
a seriespart of w3ai.org
Both substrates compute. One has been doing it for about four billion years on sunlight;
the other for eighty, on electricity. Read the columns down: how much of each there is, how
much code each carries, how long each part lasts, and what each pays for a unit of work.
The third line in each block is the one that matters. Where the two substrates differ
most, there is usually something on biology's side with no counterpart on ours — and
that absence is a description of something to build.
Lifefour billion years
Computingeighty
01 Energy — what a unit of work costs
18× the floorATP hydrolysisWithin about eighteen times the thermodynamic minimum for an elementary operation.
2.5×10⁸× the floorOne GPU multiply-addSeven orders of magnitude further from the same limit.
Share of energy: photosynthesis captures 130 TW · all data centres draw 47 GW — 0.036% of the biosphere, 1.5% of human electricity
What is the equivalent of a mitochondrion?Biology puts one to two thousand independent power converters inside every cell. We put one power supply in each machine and accept it as a single point of failure. The software question is sharper than the hardware one: if every agent run were metered and charged at the point of work, rather than billed to a facility at the end of a month, what would change about which runs happen at all?The energy page ›
02 Code — how modules are shared
1 codon tableUniversal interfaceAny gene from any organism can, in principle, be expressed in any other.
thousandsRuntimes, ABIs, formatsMutually unintelligible. Our registries are comparable in size and far less composable.
Library size: both substrates hold on the order of 10⁸ distinct modules — biology then runs 10³⁰ copies
What is the equivalent of a chromosome?A chromosome is a packaging unit that replicates atomically and segregates whole. Our nearest thing is a lockfile beside a container image — but those do not recombine. Biology treats horizontal transfer as a supported feature, which is why resistance crosses species in months. What would a unit of agent capability look like if it could be handed to another organisation and expressed there without a migration project?The code page ›
03 Scale — how many of each there are
4–6 × 10³⁰Prokaryotic cellsCounted by a deliberate census, with published error bars.
≈ 10²²Transistors ever madeInferred from sales. No census exists.
The one row that matches:8.2 billion people and ≈7 billion smartphones — within a factor of two, reached in fifteen years
What is the equivalent of a species?A species is a boundary within which designs recombine freely and outside which they do not. The nearest analogue on the other side is the firm — three to four hundred million of them — and the fit is poor, because firms exchange nothing genetic. This is the gap an AI Nation is meant to fill: a boundary loose enough that skills and workloads cross it, tight enough that it means something to be inside.The scale page ›
04 Lifetime — what outlives what
parts disposableThe whole persistsA body replaces its gut lining every few days and lasts decades.
wholes disposableThe parts outlive themServers last ~5 years; the company around them, ~15.
The tell: the only cells biology does not replace are the ones holding state that cannot be reconstructed
What is the equivalent of turnover?Biology replaces anything it can regenerate from a specification and preserves only what it cannot. Applied to an agent system that is a sorting question, not a metaphor: which parts are rebuildable from a written spec, and which hold state that would be lost? Everything in the first group can be treated as consumable — which is the argument for keeping the spec in a document and the state in a ledger.The lifetime page ›
What this series is not
None of this argues that computers are becoming alive, or that biology is secretly a
computer. The comparisons are structural. Where the two are close it is interesting;
where they are many orders of magnitude apart it is more interesting. Every figure carries
a box saying what is counted, what is excluded and how wide the error bars are, because at
these scales the definition decides the answer more often than the measurement does.