The Carbon:Silicon Firm
Every business is an information processing machine. It has human components — carbon — and automated components — silicon. The silicon needs electricity. The carbon needs HR, management, offices, and politics. As AI and decades of IT investment shift the balance toward silicon, the support infrastructure built for a high-carbon machine loses a large part of its purpose. Most organisations know something is wrong. Few have a language for what it is. This framework provides one.
Something structural is happening. Anyone who is using Claude knows. The market knows too: the SaaSpocalypse is a repricing of assumptions about where value is created and by whom. It is not entirely irrational.
The technology disruption is compounding with geopolitical disruption. Supply chains are fragmenting. Defence budgets are being rebuilt from scratch. The energy transition is reshaping industrial cost structures. Housing has become a political emergency across most of the developed world. Each of these crises is, at its core, a crisis of productive capacity. They are forcing a reckoning with a question that was easy to defer during the long period of relative stability: how much can our organisations actually make, and how fast?
Amidst this uncertainty, one thing is not in doubt: the direction of travel is more digitalisation, more automation. The impact of that automation will be profound and is difficult to fully understand.
This automation will disrupt productivity, for decades in stately decline,1 at its foundations. AI looks and feels — viscerally — like it is automating work that was previously assumed to require human judgment. This author rebuilt the Giraffe website in a few hours. If SaaS — where businesses compete against massively capitalised global titans and have to relentlessly innovate or die — is being disrupted, there is no chance that other businesses will remain safe.
We have been trying to work out the meaning of this, and working with our customers to do the same. One analytic framework has been very fruitful: to think of a business as an information processing machine. Not a culture, not a team, not a mission statement. A machine in which data enters, decisions come out, and everything between those two points is either processing or overhead.
The humans in the machine are data processing components made of carbon. The computers are components made of silicon.2 The framework that follows is deliberately inhuman — it brackets questions of talent, motivation, and organisational culture not because those things are irrelevant, but because they obscure the more fundamental question that the current moment poses: what is the machine actually doing at each layer, how much of it requires human involvement, and how much of that human involvement is a structural legacy of conditions that no longer exist?
The framework proposes three layers — M0, M1, and M2 — and a single diagnostic ratio: carbon to silicon. The argument is that in most organisations, the ratio has shifted dramatically and structurally, while the machine itself has not been rebuilt to reflect it.
Three Layers
M0
M0 is the work itself. The physical, productive act that the business ultimately exists to perform. In manufacturing, M0 happens on the factory floor. In logistics, it is the movement of goods. In property development, M0 is what happens on site — the act of construction, the transformation of materials into built form.
M0 also has a symbolic or cognitive dimension: architectural drawings, contract documents, websites, feasibility models. These are not physical products, but they are still acts of production — outputs that are directly directed by the business toward M0 ends. For a long time, this cognitive dimension of M0 was necessarily human. It required judgment, craft, and specialised knowledge that could not be codified into a repeatable process.
The physical dimension of M0 was largely automated across the twentieth century through industrialisation. We are now in the era of the dark factory: where production happens without people, or the light their eyes need. The white collar work is now being automated by AI. Websites, feasibility studies, contract drafts, design iterations — outputs that required skilled human time a decade ago are now produced in seconds. M0, across both its components, is becoming almost entirely silicon.
M1
M1 is the information processing layer. It takes reality as input and produces instructions for M0 as output. In practice, M1 encompasses market research, pricing decisions, procurement strategy, financial modelling, product design, and the sequencing of decisions that determines how a business moves from opportunity to execution. M1 is the machine’s intelligence — the layer that decides what M0 should do and when.
A useful distinction: people working in the business are working at M0, or as M0-adjacent functionaries within M1. People working on the business are working at the M1 layer — designing the machine to understand the world and coordinate supply to meet demand.
M2
M2 is the support infrastructure for the humans needed to run M1. Human resources, remuneration committees, management layers, performance review cycles, office leases, travel policies, incentive structures, and the accumulated administrative apparatus that exists to coordinate, motivate, and retain people as information processing components inside the machine. M2 does not process information in service of M0. It exists to keep the carbon components of M1 productive.
Silicon components require electricity and cooling. They do not require annual leave, career development, or conflict resolution. M2, in this framing, is correctly sized for the carbon component it supports. The question the current moment forces is whether that carbon component still reflects what the work requires — or whether the machine has changed and M2 has not.
The Carbon:Silicon Ratio
The carbon:silicon ratio describes, for any given organisation or layer within it, the proportion of work being performed by humans versus automated or algorithmic processes. A machine that is 90% silicon and 10% carbon has a fundamentally different cost structure, management overhead, and competitive vulnerability to one that is 70% carbon and 30% silicon.
The ratio is not static. It shifts with technology, with process design, and with deliberate decisions about what to automate and what to retain as human work. AI is currently accelerating this shift, but it is more accurate to see AI as an incremental step — significant, but built on decades of improving IT systems that have been steadily converting carbon to silicon across M0 and M1 since the 1980s.
The central observation of this framework is that across most industries, the underlying carbon:silicon ratio has shifted dramatically toward silicon, while M2 — sized for the old ratio — has not moved. M2 persists not because the work requires it, but because the machine has not been redesigned.
To get a sense of what existing ratios look like, architecture firms are a useful reference point. The taxonomy is not identical, but the underlying structure is recognisable: firms track an overhead multiplier — the ratio of total indirect costs to total direct labour — and the industry target sits at 1.5 to 1.75. For every dollar of productive work, between
The Transmission
A transmission exists because an internal combustion engine only outputs power efficiently within a narrow rev range. Below and above that power band, the engine performs poorly. The transmission solves this problem mechanically — it translates engine output across a wider operating range, allowing the vehicle to function at speeds the engine alone could not efficiently serve. The transmission is not an elegant addition to the machine. It is a compensatory mechanism for a fundamental inefficiency in the engine. It adds weight, complexity, and a significant source of mechanical failure. But given the engine, it is necessary.
When the internal combustion engine is replaced by an electric motor — which outputs power efficiently across its entire rev range — the transmission becomes dead weight. The inefficiency it existed to compensate for is gone. It is structurally redundant.
M2 is the transmission. It exists to compensate for the cost and complexity of running a machine with a high carbon component. As the carbon:silicon ratio shifts — as M0 automates and M1 becomes leaner — M2 does not gradually shrink. It sits, unchanged, as a legacy of the machine it was built for. The engine has been replaced. The transmission remains bolted on.
How the Machine Loses Contact With Itself
The Engineering Problem
Engineering is, by its nature, proximate to reality. If the bridge is designed badly, the bridge falls down. If the code is wrong, the program fails. The feedback loop between action and consequence is short, and the discipline that develops within that loop tends to be rigorous.3
The feedback loop for M2, by contrast, is nowhere near as short. A poor HR policy may not manifest in measurable organisational damage for years. A badly designed incentive structure may produce perverse outcomes across an entire business cycle before the connection is made. The learning cycle at M2 is structurally slower than at M0 or M1, which means M2 accumulates errors in ways that M0 cannot. The machine at the top of the organisation is operating with a much longer lag than the machine at the bottom — and is rarely aware of it.
The Outsourcing of M0
Understanding why M1 drifts from M0 reality requires understanding a structural feature of how most modern businesses are organised: M0 is contracted out.
In property development, the developer does not build. Construction is engaged through a head contractor, who manages a supply chain of subcontractors. The developer’s staff interact with the built form primarily through documentation — programmes, specifications, RFIs, payment claims. In manufacturing, the factory may be offshore, run by a third party under a supply agreement. In software, development may be handled by an external team or agency. In each case, the organisation’s internal staff occupy M1, while M0 is performed by parties who are structurally external to the machine.
This is efficient in many respects — it allows organisations to access specialist M0 capability without carrying it on the balance sheet. But it has a significant structural consequence: the people designing and running M1 have less direct experience of M0. Their model of the work is increasingly abstract, built from reports, drawings, site visits, and secondhand accounts rather than direct physical engagement with what the machine actually does.
The Drift
The critical property of M1 is its proximity to M0 reality. When M1 is designed close to the actual work, it tends to be lean. When M1 drifts from M0 reality, it accumulates processes that serve the abstraction rather than the work.
M1 begins to drift as follows. A process is added to manage a risk that arose from not understanding M0 well enough. Another is added to verify outputs that were misspecified because M1 was not close enough to M0 reality. A reporting layer is added to give M1 visibility over an M0 process that a more grounded M1 would not have needed to monitor separately. None of these additions are irrational in isolation. Each responds to a real problem. But collectively they represent M1 growing to compensate for its own distance from the work — accumulating complexity not because M0 is complex, but because M1 has lost the ability to read M0 directly.
This drift is self-reinforcing, and its relationship to M2 is non-linear. The number of coordination relationships required within an organisation grows as n(n−1)/2 — meaning that each additional human component added to M1 requires a disproportionately larger increase in M2 to manage the resulting relationships. Adding ten people to a team of ten does not double the coordination overhead — it increases it by a factor of roughly three. This is the mechanical basis of what C. Northcote Parkinson observed empirically in 1955: that work expands to fill the time available, and that administrative overhead grows independently of the productive work it nominally supports.4 Parkinson’s insight was sociological. The carbon:silicon framework makes it structural. M2 does not grow because of human laziness or political manoeuvring — though those are real enough. It grows because the mathematics of human coordination are punishing, and every incremental increase in the carbon component of M1 triggers a disproportionate expansion of M2.
The machine develops an immune system that protects its own complexity.
The Carbon:Silicon Firm
The analytical exercise this framework proposes is straightforward, though not easy to execute.
Map the machine. Trace every process, every role, every system back to the M0 act it ultimately serves. Identify whether each component belongs to M0, M1, or M2. For M0 and M1 components, ask honestly what proportion of the work is currently performed by humans and what proportion is automated or could be. Calculate, at least approximately, the carbon:silicon ratio at each layer.
What this exercise tends to reveal is that M0 is far more silicon than it is given credit for, and that M0’s remaining carbon components are on a clear automation trajectory. It reveals that M1, when stripped back to what it genuinely needs to do to serve M0, is a fraction of its current size. Small M1 means small M2 — and given the non-linear relationship between human components and coordination overhead, even modest reductions in M1 carbon produce significant reductions in M2.
The limitation of most organisational analysis is that it begins with people — with roles, relationships, and accumulated institutional knowledge — and asks what to optimise. This framework asks a prior question: what is the machine actually? Begin with that, establish what it needs to do, and ask what form that requires. The human components that survive that question are genuinely necessary. The ones that do not are transmission weight.
The organisations pulling ahead have learned to think inhumanly about what they actually are — and to rebuild the machine accordingly.5
1 As Robert Solow observed in 1987: “You can see the computer age everywhere but in the productivity statistics.” The paradox he identified — that transformative technology does not immediately register in aggregate productivity — appears to be resolving, rapidly and unevenly, in the current moment.
2 The framing of humans as carbon-based and machines as silicon-based has a lineage in science fiction. Arthur C. Clarke titled a 1992 essay “Greetings, Carbon-Based Bipeds!” — later the title of his collected essays (St. Martin’s Press, 1999). Isaac Asimov proposed the shorthand C/Fe — carbon and iron — to describe a culture combining human and robot life. The carbon:silicon formulation used here updates that contrast for the current technological moment.
3 The observation that engineering cultures produce faster learning cycles and more grounded and ambitious decision-making than lawyerly cultures is not new. Thorstein Veblen argued in The Engineers and the Price System (1921) that engineers, by virtue of their proximity to production, were in structural conflict with businessmen whose orientation was toward managing the price system rather than the productive machine. It runs through debates about industrial organisation from the early twentieth century forward. For recent treatments see Breakneck by Dan Wang, and adjacently, the Abundance Agenda.
4 C. Northcote Parkinson, “Parkinson’s Law,” The Economist, 19 November 1955. Later expanded as Parkinson’s Law: The Pursuit of Progress (John Murray, 1958).
5 What a leaner machine does with its freed capacity is beyond the scope of this framework. Whether the productivity gains are absorbed by doing more work, doing it at higher quality, or whether — per Jevons’ paradox — cheaper production simply stimulates demand until the machine is full again, are open and consequential questions. This framework addresses the structure of the machine, not the use to which it is put.