The Naysayers Are Reading the Bill Wrong

The Naysayers Are Reading the Bill Wrong

Agent bills are exploding, with teams reporting monthly spend that dwarfs what any chatbot subscription used to cost. Most enterprise pilots never make it to production. And to the skeptics, every new model release feels less dramatic than the last one, which they read as proof the technology is topping out. That is the case against AI, and none of it is invented. The naysayers are still wrong, because they are reading their own evidence backwards.

Britain had this argument in 1865

Britain went through the same argument in 1865. The country was convinced it was burning through its coal, and the comforting theory was that more efficient steam engines would fix the problem, since a better engine burns less coal per unit of work. William Stanley Jevons wrote The Coal Question to show the theory had it backwards. As engines gained in efficiency, coal consumption rose anyway, because cheaper effective power made coal worth burning in places where it had never made economic sense before. Efficiency did not shrink demand, it created it.

That is the signature AI is showing right now, and you can see it most clearly at the bottom of the price list. The top end is trending upward. Fable launched at twice the token price of Opus, OpenAI's current flagship costs double what last year's did, and the best reasoning on the market charges a premium because it can. The floor is a different story. In late 2022 the best model money could buy was GPT-3.5 at $20 per million tokens. Today that level of capability is the budget tier: Haiku beats it outright and costs $1 per million input tokens. The ceiling rose while the floor collapsed. Yet total spending keeps climbing, because the number of tasks worth pointing a model at grows faster than the floor falls. The naysayers look at rising bills and see an industry that cannot pay for itself. Jevons would look at the same bills and see a general-purpose technology doing exactly what coal did: becoming cheap enough to be used everywhere.

Efficiency did not shrink demand, it created it.

The pilots are failing the old layout

The failed-pilot statistics deserve the same historical patience. Factories saw no productivity boost the day they swapped steam for electricity. The economist Paul David showed the gains took decades, because they only arrived once factories were redesigned around the new power source instead of bolting it onto layouts built for a central steam shaft. Many AI pilots fail the same way. They bolt a model onto a workflow designed for people carrying every handoff, measure the result, and call the technology overhyped. I have written before about building at the speed of AI and removing the human glue from workflows, and this is why. That redesign is where the productivity boost actually comes from.

The revolution rides the cost curve

The plateau argument is the one I buy least. I do not accept the premise, because the jump to this generation of frontier models did not feel incremental from where I sit. But even if capability froze today, the argument misses where revolutions actually come from. The steam engine spent a century pumping water out of mines before anyone built a railroad with it, and the railroad became an industry nobody had imagined once the economics of steam caught up with the idea. The revolutionary phase of a general-purpose technology arrives on the cost curve, not the capability curve. That is why I spend as much of my attention on smaller, cheaper models as on frontier launches.

The revolutionary phase of a general-purpose technology arrives on the cost curve, not the capability curve.

The elevator operator

AI does bring one thing coal and electricity never had. It can start to act as its own boss, making sure the work gets done without a person driving every step, and the closest precedent I can find is the elevator. I noticed it watching Mad Men for the first time recently: the office elevator has an operator whose whole job seems to be standing there and pressing the button for people. I looked into whether that was historical accuracy or a plot device, and it was accurate. Automatic elevators existed for decades while people refused to ride them, and buildings kept paying operators because passengers wanted a human in the car.

When New York's elevator operators walked out in September 1945 and shut down 2,100 buildings in under a week, the math changed. Buildings went automatic and spent the next decade winning back trust with recorded voices, emergency controls, and a human reachable at the critical moment instead of riding along on every trip. Mad Men opens in 1960, fifteen years after that strike, and the operator is still in the car. Trust moves slower than technology. That is the right model for AI supervision. The systems I run check their own work, retry their own failures, and keep the flow moving, but a human still owns the quality sign-off at the points where a bad output creates real damage.

Trust moves slower than technology.

Models do not originate

What no model has done yet is originate. AI has not invented a new use for itself, and humans have found every killer app so far. Its most celebrated outputs, the famous Go move, the predicted protein structures, the discovered algorithms, all happened because a human posed the problem, defined what winning looked like, and recognized that the output mattered. The steam engine did not decide to become a railroad. Electricity did not propose replacing streetlights. People did that, and people will find the uses for cheap intelligence that nobody has thought of yet, including the ones that become industries.

The steam engine did not decide to become a railroad. Electricity did not propose replacing streetlights.

Drive your own cost curve down

You also do not have to wait for the labs to make this cheap. In my own systems, the frontier model is the architect, not the laborer. I develop ideas with Fable or Sol, have them adversarially review each other's plans when the stakes justify it, save the plan, and hand execution to a fresh headless agent with one instruction: pick the cheapest model you believe can complete each step, and here is what you have access to, from Sonnet, Terra, Haiku, and Luna down to DeepSeek V4 Flash and a local Qwen. My YouTube transcript pipeline runs this way. Cleaning 229 raw transcripts, more than five million characters of captions, was bulk token work that Haiku handled for pennies per video and the local Gemma box handles for nothing but electricity when I am done for the day. Sonnet only touches the small judgment pass at the end. The routing is never settled either. I keep retesting whether a cheaper model can hold the routine work, because every step down the ladder saves money and token spend, and every handoff keeps my own context window clear for the work that deserves it.

That is Jevons applied at home. I made my own AI cheaper, so I found more uses for it, so I use more of it. The naysayers will keep reading rising bills as proof of failure, and while they wait for the collapse, the people driving their own costs down are finding the new uses. Somebody is going to stumble onto this era's version of the railroad. It will not be a model, because models do not originate. It will be a person who was using the technology enough to recognize what it could become.

I made my own AI cheaper, so I found more uses for it, so I use more of it.

The next railroad

As intelligence becomes cheaper, everything downstream of it becomes cheaper too, and I have not even mentioned what it is doing to the cost of writing software. The space of things worth trying grows every month, and nobody has envisioned the future that produces yet. That is the part I am excited about. You cannot trip over the next railroad from the sidelines, so I will keep building and finding out.

Sources: