
The Decade of the $6 Ride
For roughly a decade, a small miracle recurred in American cities. A stranger would arrive in a clean car, drive you clear across town, and charge you less than it cost to provide the ride. The reason was strategy, and an expensive one. Between 2014 and the early pandemic, Uber passengers were covering only about 41% of the true cost of their trips, according to the transport economist Hubert Horan; the missing 59% was wired in by venture capitalists, and later by public-market investors, who had decided that buying the market outright was cheaper than competing for it. Uber lost money on a scale that beggars belief; cumulative operating losses well north of $25 billion before it posted its first annual profit in 2023, and it did so entirely on purpose.
The point of the exercise was dominance. Lyft had helped pioneer the peer-to-peer model, and for a while it ran a real race; in some cities the two traded punches, subsidizing so aggressively that in 2014 Uber was reportedly handing Berlin drivers around $10 an hour in incentives against a gross fare of roughly $2. But Uber had raised more money than most central banks, and in a subsidy war the deepest pocket wins by default. Riders enjoyed a decade of artificially cheap travel, competitors and public transit were hollowed out, and once the field lay clear the coupons expired. Fares have since climbed something like 40 to 80% depending on the city, surge pricing hardened from an occasional bug into a permanent feature, and the ride that once felt cheaper than owning a car now routinely costs more than the taxi it replaced.
I rehearse this history because we are living through the sequel.
Your $200 Coupon
Consider the going rate for frontier artificial intelligence. Two hundred dollars a month buys a Claude Max or ChatGPT Pro seat: near-unlimited access to a system that can draft, code, analyze, and argue at a level that would have read as science fiction in 2022. The naive interpretation is that raw intelligence has simply become absurdly cheap. Look closer and you are holding a coupon.
Sam Altman said as much, out loud, in early 2025: OpenAI was “currently losing money” on its $200 Pro tier because, in his words, “people use it much more than we expected.” Anthropic has been blunter still. When it clamped weekly rate limits onto its Max plan in 2025, it explained that a small number of subscribers were consuming “tens of thousands of dollars” in model usage on that same $200 subscription, largely by running coding agents around the clock. One analysis reckoned a Pro user who maxed the thing out could cost the provider up to $14,000 a month.
How large is “the rest”? I built a simple model from published 2026 API prices — the wholesale rate the labs charge for the very same tokens they serve you at retail (like I said, a simple model, so take it with a grain of salt).

A casual user who fires off a few chats a day costs perhaps five dollars a month; that person is a rounding error, and reliably profitable. But a “typical” heavy Max user, working maybe three focused hours a day, already burns through roughly $350 of tokens. A serious engineer leaning on Claude Code for six hours costs about $1,400. And a true power user, running the priciest model eight to twelve hours a day — the exact person for whom these tools are transformative, the person the entire marketing apparatus is aimed at — costs somewhere between $3,000 and $4,600 a month to serve. He hands over two hundred dollars. Silicon Valley covers the other ninety-odd percent.

In both cases the customer pays a slice of the true cost; in both cases the gap is financed by investors wagering that dominance bought today can be milked tomorrow; in both cases the losses are colossal and deliberate. OpenAI booked around $13 billion of revenue in 2025 and lost somewhere between $21 and $38 billion earning it, depending on whose leaked figures you believe. Its inference bill alone (the raw cost of answering your questions) ran to roughly $5 billion on Microsoft’s cloud in the first half of 2025. Across the industry, the hyperscalers are shoveling three to four hundred billion dollars a year into the data centers that make your $200 feel reasonable. You are the rider in 2014, marveling at how little the trip costs and declining to ask who is feeding the meter.
The optimist’s rebuttal is that inference keeps getting cheaper, and on the narrow point they are correct: the price of a million tokens has been falling by roughly an order of magnitude a year. What the rebuttal underweights is that consumption is exploding faster than price is dropping. The new “reasoning” models think in long, expensive internal monologues before they answer, and agents loop over a single task dozens of times, so the token count of one job has ballooned ten- to a hundredfold. When the meter spins a hundred times faster, a meter that is ten times cheaper still leaves a larger bill at the end of the ride.
Cheaper Than Humans?
The reigning assumption — in boardrooms, in layoff memos, in ten thousand breathless threads — is that artificial intelligence is now decisively cheaper than human labor. A knowledge worker costs $80,000 a year, fully loaded; a Claude seat costs $2,400. Thirty-three times cheaper. The case looks closed.
The case is wide open, for two reasons the headline number is designed to hide.
The first is the subsidy we have just anatomized. That $2,400 is a 2014-Uber price, not a 2024-Uber price. Re-run the comparison at the true cost of serving a productivity-heavy user — the only sort that actually displaces a salary — and the annual figure climbs toward $37,000. The seat that looked thirty-three times cheaper than an American employee turns out, once the coupon is peeled away, to be merely half the price, and already dearer than a capable human in Manila or Bangalore. The discount does the persuading. The unsubsidized economics are a far closer contest.
The second reason is subtler and, I suspect, more lasting. The comparison assumes that AI output is usable unsupervised — that a model’s draft is a finished good rather than a promising lump of raw material. Anyone who has shipped real work with these systems knows better. The output must be read, checked, corrected, and now and then binned, and that reviewing is performed by a human who draws a wage. Fold in even a modest supervision tax — a person spending a quarter of their time verifying the machine — and the “unsubsidized AI plus the human who babysits it” figure drifts up toward $57,000, breathing on the neck of the $80,000 employee it was meant to abolish.

None of this makes the machines a fraud; they are extraordinary, and for narrow, high-volume, error-tolerant work — translation, boilerplate code, document triage — they are already and unarguably cheaper than people. The mirage is the universality of the claim. When the task is judgment-heavy and the tolerance for error is thin, the honest verdict is a shrug.
And there is a final twist that ought to keep the strategists awake, because it is exactly where the Uber analogy snaps in AI’s disfavor. Uber’s subsidy had a clean exit ramp: the marginal cost of a ride is trivial, so once the company dominated it could raise fares, throttle driver pay, and watch the profit pour in. The frontier labs enjoy no such luxury. The marginal cost of inference is real, large, and — this is the sharp bit — it rises with engagement. The more indispensable a customer finds the product, the more he costs to serve, which is precisely why Anthropic reached for rate limits rather than simply lifting the price. Uber grew cheaper to run as it grew larger; the AI labs grow more expensive to run as their best customers fall in love. A subsidy cannot be switched off cleanly when success itself is the thing doing the bleeding.
So enjoy the discount, wholeheartedly. It is one of the great bargains in the history of commerce, a miraculous technology sold for a sliver of what it costs to produce. Simply remember what the last decade taught anyone who bothered to read the fine print. The coupon and the price are different numbers, and the second one is on its way. When it lands, the tidy assumption that declares the machine cheaper than the man will have to be redone at full cost — with the man who checks the machine’s work still, stubbornly, on the payroll.
See ya, folks. And enjoy that discount (;
—J&E