Silicon Valley, 2026:
“You don’t use a Ferrari to go to the grocery store.”
Soviet cinema, 1968: already solved this.
More precisely.
The Ferrari analogy has been circulating in AI circles for more than a year: do not route trivial tasks to the most expensive frontier model.
Perplexity’s CEO used it again this week.
Three weeks ago, I wrote my own version:
“Many companies bought a Porsche and now drive it across the yard to buy bread.”
But we all chose the wrong vehicle.
In The Diamond Arm, a Soviet comedy classic, a building superintendent delivers the line:
“Наши люди в булочную на такси не ездят.”
“Our people don’t take taxis to the bakery.”
Not a Ferrari.
A taxi.
And that changes the economics entirely.
A Ferrari is capital expenditure.
You bought it. It sits in your yard. It may be absurdly overpowered for a bread run, but the journey is not being billed by an external meter.
The problem is mainly mismatch: too much machine for too little task.
A taxi is operating expenditure.
You own nothing.
You pay for every kilometre, every detour and every minute spent waiting.
That is much closer to a frontier API.
Pay per token.
Meter by meter.
And when an agent repeatedly re-reads the same bloated context just to make one small step, that is a taxi idling in traffic while the meter keeps running.
Most companies did not buy a Porsche.
They are sitting in a taxi, convinced they own one.
The taxi metaphor also reveals two things the Ferrari never could.
First, one neglected axis in AI architecture is not simply:
smart model versus cheap model.
It is:
owned versus metered.
Open weights running on your own hardware resemble owning a vehicle: higher entry cost, maintenance and responsibility — but far greater control over marginal use.
Frontier APIs resemble living on taxis: low entry cost, immediate convenience and a meter that never disappears.
Which one is rational depends on your mileage.
The same as in ordinary life.
Second, the taxi driver knows the route.
Addresses. Stops. Timing. Habits. Repeated journeys.
In AI terms:
prompts, workflows, proprietary context, organisational behaviour and the internal structure of your business.
With local infrastructure, you can decide whether that route ever leaves your control.
With an external API, someone else is necessarily present in the transport layer — even when the contractual and technical protections are strong.
So a Soviet comedy from 1968 quietly compressed several questions that enterprise AI is now rediscovering at conference-keynote prices:
proportion,
the meter,
ownership,
and privacy.
The Ferrari analogy explains excessive power.
The taxi explains the business model.
Some truths are grounded enough to be reinvented in every economy.
Scarcity. Bakery. Tokens.
The bread is close.
Walk.