AI & Infrastructure

The value of this framing is that it pulls the discussion away from "whose model is stronger" back to physical constraints. Compute depends on electricity; electricity depends on generation capacity, grids and transformer supply. When chips iterate yearly while grid infrastructure takes five to ten years, the bottleneck shows up somewhere nobody is watching.

Counter-intuitively, hardware scarcity can force a lower-power route that leans less on top-end silicon. Biological intelligence is a harsh benchmark: a human brain runs on tens of watts at close to the theoretical efficiency limit, while the best silicon is orders of magnitude away. No algorithm closes that gap — only architectural change does.

Several signals point the same way: open models now lead on some download measures; fine-tunes derived from them make up a large share of newly released models; and per-unit inference cost keeps falling by orders of magnitude. "Having a model" is no longer a moat.

For adopters the conclusion is not to stop investing in technology but to move the investment upstream or downstream. Upstream means proprietary data and scenarios; downstream means deployment and process. The middle layer — selling access to a general model — gets flattened first.

Separating the four explains why early conversational tools could not carry a whole task: no memory, so every session started from zero; no execution authority, so they could only advise; and no self-checking, so they could not catch their own errors. With all four in place, the unit of work becomes a person plus several dispatchable capabilities.

That raises the bar for managers in two tiers: getting real work out of these tools is a scarce execution skill, while designing and supervising several such units is the strategic one — and the second requires the first, because you cannot spot the sticking points without having operated one.

Cost and barriers are real. Heavy use can produce substantial usage bills; setup requires some technical environment; and once execution permissions are granted, exposure and audit become mandatory concerns. Official security bodies have already flagged risks in default configurations, so public exposure, unnecessary access, authentication and operation logging all need review.

The progression is visible: first how to phrase a request, then what context to supply, and now what working environment makes output reliably usable — tasks decomposed into verifiable steps, explicit acceptance criteria, primary sources available for lookup, and human confirmation at the points that matter.

What this means for an organisation is that the durable asset is the environment and the workflow, not someone's private collection of prompt templates. The work worth doing is building the setup that makes a given task produce dependable output — and building it so others can use it.

A recurring observation: firms using the same model get wildly different results. The cause is rarely the model. It is whether business data has been made usable, whether it is clear which decisions are assisted and which stay human, and whether the tool is wired into daily process rather than idling beside it.

This is why smaller organisations often have an opening. Once the capability threshold drops, the deciding variables become depth of understanding of a specific niche and how short the decision chain is. Large firms hold data and talent advantages alongside process inertia and slower iteration.

So the question for any AI investment is not whether you use the strongest model, but in how many real workflows the capability is actually running, how many hours it saves per month, and whether error rates went down.

Emergence: simple agents following simple rules interact and, at the system level, produce novel properties unpredictable from any single agent — the whole exceeds the sum of its parts. Three mirrors: ants follow pheromone rules, yet the colony optimises foraging paths; no single neuron thinks, yet billions of connections yield consciousness; in a large language model every parameter performs basic arithmetic, yet past a hundred-billion-parameter threshold, in-context learning and chain-of-thought reasoning appear abruptly — a phase change, like water freezing at 0°C. Weak emergence (bird flocks) is computationally reducible; strong emergence (consciousness) is not.

For managers the lesson is structural: do not attempt to command the macro outcomes of complex systems — markets, organisations and innovation are all emergent. Design the micro-rules and incentives, safeguard the density and quality of interaction, and tolerate unpredictability. Directives produce compliance; emergence produces surprise — and excellence nearly always arrives by the latter route.

The human equivalent measures the ratio between AI's marginal cost of producing a unit of intelligence and the human cost. Its brutal lesson: replacement is driven by economics, not capability ceilings — in customer service, translation and routine paperwork the ratio has reached hundreds to thousandsfold, and substitution follows. The moat test is therefore clear: shallow water (standardised, structured, well-specified tasks) drains fast; deep water — judgement under unstructured risk, ethical decisions, meaning-making, relationships built on empathy and trust — remains human-led.

The response for individuals and organisations is migration, not confrontation: move from competing with AI on efficiency to acting as its principal and editor — framing good questions, making value judgements, owning outcomes. This matches the core claim of the cognition discipline: after the democratisation of knowledge, human value lives precisely on the layer AI cannot reach.