The Biggest Motor in the County
A factory in 1895 ran on one steam engine and a ceiling full of spinning steel. Line shafts down every bay, leather belts dropping to each machine, and the whole building drawn around the engine's reach instead of the work's flow. When central electric power arrived, most owners made the obvious move: pull the steam engine, bolt one giant electric motor in its place, and keep every shaft, belt, and floor plan exactly where it stood. Cleaner, quieter, somewhat cheaper to run. Output didn't much notice.
The economic historians wrote this postmortem a long while back. Warren Devine's shafts-to-wires accounting puts electric motors at under five percent of American factory drive in 1899, near eighty by the late twenties. The curve everybody cites is that adoption. The one that stings is productivity: Paul David's dynamo paper traced it sitting stubbornly flat for the better part of four decades after the dynamo showed up, because the first wave of electrification changed the power source and nothing else. Owners laid the new system over the old architecture, and the old architecture kept paying old dividends.
The revolution came when owners quit swapping engines and rewired the premise instead. A motor on every single machine, unit drive in the trade's term, and the layout finally answered to the work. Machines lined up by process flow instead of shaft proximity; single-story plants with overhead cranes; light and air where the belt forest used to hang, and any one machine running alone without spinning the whole ceiling. Ford's Highland Park line started moving in 1913, and it couldn't have moved an inch under line-shaft drive. The gains went to the institutions that redesigned how they worked, not to the ones that owned the biggest motors.
The Giant-Motor Play, Rebooted
That's the year AI adoption is living through right now, and I'd put most of the market squarely at the giant-motor stage. The standing play, 44 months into the chatbot era: buy seat licenses, lay them over an unchanged workflow, announce an AI strategy. It's the 1895 move in a software skin, and it's collecting 1895's marginal results.
MIT's NANDA group measured the wreckage last summer: 95 percent of enterprise generative-AI pilots returned nothing measurable to the P&L. Not thin returns; nothing the ledger could find. And their diagnosis wasn't the models. It was the approach: generic tools draped over workflows that didn't change shape. The five percent getting paid did the opposite, embedding the system inside one specific workflow and letting the workflow reorganize around it. The unit-drive crowd, in other words. Same divide, new century, and the teams learning fastest are the ones pulling away.
Electricity's Product Was Never Lamps
Underneath the giant-motor play sits a category error I run into most weeks: the notion that AI is a chatbot interface you subscribe to. That's meeting electricity and concluding the product is lamps. Lighting was the first thing electricity was obviously for, and lamps sold like mad, and anybody who stopped the analysis there missed streetcars, elevators, refrigeration, the electrified factory floor, and the reorganization of the working day itself. Electricity didn't stay a product. It became the substrate everything else got redesigned on.
The chatbot's a fine lamp; I burn one all day. It drafts, it summarizes, it answers what I ask. Honest tooling, honestly useful. But a lamp doesn't reorganize an institution, and a subscription seat can't hold your workflow, your data, or your audit trail. It just glows.
Unit drive in 2026 means a motor on the machine itself. A governed model inside the claims operation, reading the demand package as it lands, drafting the file note, teeing the subrogation trigger before the adjuster's diary comes due. An agent working the clinic's denial queue, decoding CARC and RARC pairs and assembling the appeal with the right chart excerpts attached. The privilege-log grind at a firm, that per-document judgment against the cast-of-characters list, running inside the firm's own walls with every line of the log accountable. Different machines, same rewiring: the model goes where the work is, and the floor plan changes to suit.
Every one of those workflows carries data that isn't allowed to leave the building, which is why the rewiring question and the sovereignty question turn out to be the same question. A motor bolts to the machine; it doesn't phone the utility for permission per revolution. The AI equivalent is open weights running on hardware you own, governed by controls that sit in your paddock rather than a vendor's.
You Don't Get Four Decades
The analogy breaks in exactly one place, and it breaks against you. Electricity gave institutions time. David dates the productivity payoff to the 1920s, four decades on, and a factory owner who slept through the first twenty years could still catch the wave with money to spare. AI's capability curve extends no such courtesy. METR's task-length work has frontier agents doubling the length of work they can finish roughly every seven months, and across 2024 and 2025 the doubling ran closer to four. Months, compounding, while budgets move annually, hiring moves quarterly, and trust moves at whatever speed the last burned hand allows.
So the flat spot won't be a forty-year grace period everybody shares this time. It'll be a sorting. The redesigners pull ahead on a curve that compounds monthly, and the giant-motor shops discover their marginal gains were the whole prize.
What the Analogy Doesn't Buy You
Honesty about the analogy's edges, because analogies flatter whoever's selling with them. A 1905 motor did what its nameplate promised every hour of its service life. A model is probabilistic; it can be confidently, fluently wrong, and an agent with hands needs controls a belt never did: identity, egress, metering, audit, a kill switch that means it. Redesigning a workflow around a system you can't govern isn't unit drive, it's a loose flywheel. The governance plane comes first or the rewiring shouldn't happen at all.
The rewiring also never happened as one heroic demolition. Devine's own stages ran line shaft to group drive to unit drive across years, machine by machine, and that's the honest cadence here too: one workflow rebuilt, measured, owned, then the next. Anybody promising the whole floor in a quarter is selling you a bigger motor.
And the work was never self-service. Unit drive minted a trade: the plant electrician who'd walked a hundred floors and knew which machine earned its motor first. That's the trade I practice now. I sit next to the person who's done the job for fifteen or twenty years until the workflow's real cadence is legible, and the recommendation doesn't exist until that Discovery has run. Island Mountain carries no catalog and no standing configuration, by design, so nobody here profits from selling you the biggest motor in the county. I'll add that I write from a stretch of California the grid never quite finished reaching, where late electrification isn't a metaphor; it's a neighbor's generator bill. Arriving late to a foundational technology is a bill somebody real ends up paying.
Put a Motor on One Machine
So skip the bigger-motor bid this budget cycle. Pick the one workflow your people understand cold, the intake queue, the denial worklist, the quarterly report somebody rebuilds by hand every single time, and put a motor on that machine: a governed model, owned weights, the floor rearranged just enough that the work flows through it instead of around it. Measure what changed. Then rewire the next machine. The owners who worked that way from 1913 on set the terms of their industries for half a century. They got decades to be wrong first. You'll get quarters.