the gains arrive when the floor plan changes

in 1987 the economist robert solow made the most quoted complaint in tech history: "you can see the computer age everywhere but in the productivity statistics."

he was late to the feeling. factory owners could have said the same about electricity in 1900. electric motors had existed for decades. factories bought them. productivity barely moved for nearly forty years. the best explanation of that gap, the economic historian paul david's, is uncomfortably relevant right now.

steam factories were built vertically around one giant engine, power delivered through belts and shafts, machines arranged by their hunger for power rather than by workflow. when electricity arrived, owners did the obvious thing: swapped the steam engine for one big electric motor and kept the building. same layout, same process, new power source. the gains were tiny.

the payoff came a generation later, when engineers finally redesigned the whole thing: a small motor on every machine, single-story plants, machines arranged in the order work actually flows. output per worker jumped. the technology was never the unlock. the floor plan was.

we are in the bolted-on years

swap "electric motor" for "language model" and the story reads like this quarter's earnings.

the world is spending on the order of $400 billion a year on ai infrastructure against roughly $100 billion of enterprise ai revenue. the famous mit study found the overwhelming majority of corporate genai pilots produced no measurable p&l impact. gartner expects over 40% of agentic projects to be canceled within two years. the commentary looks at all this and concludes the technology is overhyped.

wrong diagnosis. this is what bolting a new power source onto an old floor plan looks like. a copilot subscription dropped into an unchanged workflow is the big electric motor wired into the steam building: technically present, structurally ignored.

meanwhile, the companies with no old building are posting revenue per employee an order of magnitude above the software norm, with teams of ten doing what took a thousand. they aren't smarter, and they don't have better models. everyone has the same models. they have a different floor plan: nothing in their structure assumes cognition is expensive, so nothing has to be unlearned.

same tools. different building. that's the whole gap.

why incumbents mostly can't

the cruel part of the electricity story: the redesign took forty years not because it was hard to imagine, but because of who would have had to approve it.

a floor plan is never just logistics. it's the power map. in 1910, the men who ran factories were the men who understood steam. their status was load-bearing in the old layout. redesigning around electricity didn't just move machines; it demoted the people who decided whether to move machines. so it waited for a generation to retire.

an org chart works the same way. reorganizing a company around ai means collapsing the layers whose function is moving information between humans, and the people who staff those layers are the ones running the reorg. asking an organisation to redesign itself around ai is asking the floor plan to vote on its own demolition.

some will manage it. microsoft rebuilt itself around cloud by treating its own cash cow as legacy, so it's possible. but notice what that took: it was a floor-plan change imposed from the top, not an adoption program. adoption programs produce pilots. floor-plan changes produce companies that look structurally weird before they look inevitable.

what the new floor plan looks like

honest speculation, so hold it loosely:

• teams get small and spans get wide. the coordination layers go first. a manager with agents needs far fewer humans between decision and execution.
• compute becomes the new headcount. budget lines that were salaries become inference. some companies will report revenue per employee the way they used to report headcount growth: as the flex, inverted.
• judgment moves to the edge. when execution is instant, the expensive thing is deciding what's worth executing. the winning structure pushes that decision to whoever touches reality, not up a chain built for a slower century.
• the org chart becomes a runtime, not a hierarchy. which agents, which humans, which checkpoints, decided per task and redrawn daily. this is where my own bets are.

maybe patience solves it

maybe this is just early, and patience solves it. the electricity gap closed eventually. incumbents adapted. banks survived the internet. why should this time be a massacre?

the gains will diffuse eventually, agreed, and "eventually" made the dynamo story a footnote instead of an extinction event. but two things are different. speed: electricity gave incumbents forty years of grace, while model capability compounds monthly and an ai-native competitor gets founded in a weekend, not a generation. and softness: a factory redesign needed capital and construction. an org redesign needs the will to demote yourself. capital was always easier to find than that.

what to watch instead

solow's paradox eventually resolved: us productivity finally jumped in the late 1990s, a full generation after the complaint, once businesses had rebuilt their processes around computing instead of placing a pc on every desk of an unchanged office.

so ignore the model announcements when judging where we are in this transition. watch for floor plans changing. the signal that the payoff decade has started is big companies bragging on earnings calls that revenue grew while headcount shrank, on purpose, as strategy. that talk has already started.