We Are Not in the Cheap Uber Era of AI

Somebody forwards me the “AI is just Uber” thread about once a week now. The subsidy’s ending - fine, true. Nobody ever gets to the part where I’m actually stuck paying whatever they decide to ask.

Everyone Read the Same Doom Thread

You know the story. The AI you use every day is cheap because someone else is paying for it, and the day that stops, you’re done. It’s the Uber story. You got the cheap rides, sold your car, fired your taxi guy, and then the promo codes dried up and the same trip across town quietly tripled.

The people telling it aren’t fools, and the receipts are real. Ed Zitron has built an audience on the line that “subscriptions were always a subsidy scam.” The numbers back the diagnosis. Hacker News threads this year peg the inference everyone relies on as subsidized by ninety-plus percent of its true cost. Microsoft was reportedly losing somewhere between twenty and eighty dollars a month on each Copilot user while charging ten. GitHub has already flipped to usage-based pricing, which is exactly what you do right before the meter starts spinning toward the customer.

Then there’s the anecdote that was too perfect not to spread. Uber - the actual company in the analogy - torched its entire 2026 AI budget by April. Per-engineer Claude Code spend ran five hundred to two thousand dollars a month, the COO went on record doubting it was worth it, and Zitron got his punchline: you think anyone goes from two hundred a month to two thousand and says “I love this”?

I’m going to concede all of that. The subsidy is real. The era of someone else covering your inference is ending. The diagnosis is correct. It’s the prognosis that’s wrong - and it’s wrong because they lifted it straight off the rideshare wars without checking whether the one thing that made rideshare gouging work actually carries over. It doesn’t.

The Subsidy Is Real. The Trap Doesn’t Follow.

Give the doomers everything. Inference is sold below cost, propped up by compute-credit deals and equity swaps that book today’s losses against tomorrow’s imaginary margins. Flat per-seat coding plans are on borrowed time; they’ll reprice to usage within a year because the math doesn’t close. The labs hold real pricing power right now. All true.

I’ll even hand them their best card. The deepest cost problem in 2026 isn’t the price per token - it’s what Artefact called the “token cost illusion.” Per-token prices are falling off a cliff and total bills climb anyway, because an agentic loop chews through more tokens solving one ticket than a 2024 chatbot burned in a week. That’s what happened to Uber’s engineers. Nobody gouged them. The tooling made consumption frictionless and they lit the budget on fire themselves.

So yes, your premium frontier bill is going up. I’m not promising AI gets free.

But here’s the line, and the whole doom thread survives by smudging it. This is a defense against gouging - rent extracted from a buyer with nowhere else to go. It is not a promise that AI stays cheap in absolute terms. Jevons applies; as the bar rises, everyone runs bigger loops and aggregate spend climbs no matter what a token costs. That’s an arms race. An arms race is a fight you walk into. A captor setting your price is a fight you can’t leave. The doom thread needs those to be the same event. They aren’t.

The Word the Analogy Is Hiding

Captive.

Every rideshare comparison rests on one fact nobody says out loud, because saying it kills the analogy: Uber riders were captive. When the promo pricing ended you couldn’t conjure a competing global driver network from your laptop. You couldn’t download last year’s drivers and run them on your own hardware. Your options were a taxi market you’d spent three years helping to strangle, or your feet. That missing alternative isn’t a footnote to the gouging. It’s the precondition for it.

This isn’t even a hostile reading. Zitron’s own sympathizers admit Uber is a strange comparison, because Uber’s costs don’t scale with use the way inference does - one more mile is gas the driver already bought. Uber’s pricing power never came from rides being expensive to produce. It came from there being nowhere else to stand.

So here’s the actual rule, and once it’s stated plainly the doom thread has to argue with it instead of pointing at a logo. Gouging requires captivity. A seller extracts rent only when the buyer has no honest floor to retreat to - no good-enough substitute at a knowable, lower price. Where a real floor exists, the most anyone can charge is the price of that floor plus your switching cost, and not a cent more.

That’s the test. The only question that decides whether the rideshare ending happens to AI is whether AI buyers have a floor. Forget the slogans and the screenshots of someone’s terrifying invoice. Uber riders didn’t have one. You do.

Gasoline Never Got Ten Times Cheaper

Uber’s marginal cost had a floor made of physics - gas, a human’s time, a car that wears out. Those never got dramatically cheaper, which is precisely why the pricing power held once the competition was dead. Inference runs the other way. The cost of intelligence is in free-fall, on improvements no boardroom can revoke.

Look at the mechanisms, not the slogan, because the mechanisms check out independent of anyone’s investment thesis. Quantization drops a model to four-bit and cuts compute two to four times with little quality loss - that’s arithmetic. Distillation squeezes a giant’s behavior into something small and cheap. Caching reclaims repeated context. Routing sends the easy ninety percent to a cheap model and saves the expensive one for the hard slice. And the silicon keeps moving; one widely-shared 2026 result had providers cutting costs tenfold by running open models on Blackwell. None of these are promotions.

Add it up and you get the number that should end the conversation: roughly tenfold cheaper per year, something near a thousandfold over three. GPT-4-class output that ran about twenty dollars a million tokens in late 2022 sits near forty cents now. a16z named the trend “LLMflation,” and I’ll use the word while noting they’re long AI to the eyeballs and would obviously have a cheerful name for this. So don’t take it on their say-so. Take it on the silicon.

A promotion gets yanked overnight. You can’t un-discover quantization. The would-be captor isn’t waiting for the subsidy to lapse so it can raise prices - it’s watching its own pricing power erode by an order of magnitude a year.

The Floor Is a Curve

A cheap floor only matters if it’s good enough to stand on, and here I have to be more careful than the doomers, because the lazy version of my argument dies just as fast as theirs. “Open weights are free, so you’re safe” is a slogan a sharp skeptic guts in one breath: free weights aren’t a free model. Running a frontier-class one at real throughput means enterprise GPUs, power, and an ops team that doesn’t show up from a pip install.

So concede it. There are two floors. The cheap one is already remarkable - Devstral Small 2 scores 68% on SWE-bench Verified on a single RTX 4090 or a 32GB Mac. A real coding agent, owned outright, that nobody can switch off or reprice. It is not the frontier. The capable floor - DeepSeek V4 Pro, Kimi K2.6, GLM 5.1 - gets you close to the closed leaders but wants serious iron. Cheap per token, not cheap to stand up. A doomer points at that spread and calls it your captivity.

Except the spread is a curve, and it’s the curve from the last section. Today’s rack-scale model is next year’s workstation model is the year after’s laptop, because the same tenfold-a-year collapse drags the whole capability ladder down with it. You’re not betting the cheap floor is already frontier-class. You’re betting it keeps climbing while the price keeps falling, and that bet has paid every year this technology has existed.

Then the smartest objection: maybe the last five percent is the whole point - the genuinely hard work only the frontier can do, so for what actually pays you’re captive anyway. Answer: split the spend. Rent the frontier for the hard slice, and you should. That’s premium spend you choose with a budget, not ransom. Meanwhile the gap is closing fast enough to watch - the open-versus-closed MMLU spread fell from 17.5 points to 0.3 in a single year. And the receipt that ends “good enough means weak”: Kimi K2.6 and Composer 2.5 are genuinely excellent at about a tenth of frontier pricing. A tenth. That’s not a fallback you tolerate when the rug pulls. It’s a tool you’d reach for on the merits.

The question was never “is open the best?” The big three hold the real frontier; concede it all day. The question is whether the alternative is good enough that nobody can hold you hostage. It is, and it’s one ollama pull or one cheaper API key away.

There’s a deeper thing here, and it’s the part I actually believe most. The model is commoditizing, the same way the act of typing code already did. For years we mistook raw capability for the moat. It never was. What’s scarce is what you build with the capability - and once that capability is a swappable input you can get from a dozen vendors and a hard drive, nobody gets to charge you rent for it.

Dozens of Players, and Governments Funding the Free Ones

Rideshare settled into a two-player truce because the prize was modest and the IP was thin. There was no secret to Uber, just capital and a head start. Two players reached an equilibrium, put the knives away, and started extracting. That’s the future the doom thread quietly assumes for AI: the survivors consolidate and the pricing begins.

It won’t happen, and the reason is structural. The prize here runs to trillions and the IP leaks by nature. Weights get published. Papers get posted. Talent changes labs like it’s changing lunch spots, carrying the methods in their heads. And - no rideshare equivalent for this - governments are subsidizing the open side on purpose, because an AI monopoly is a strategic threat they won’t permit. Chinese open models already make up around thirty percent of all open-model downloads worldwide. A cartel needs discipline. This market is built for defection.

Now the sharpest counter, the one that kept me honest: open models distill from the frontier, so when the subsidy dies and the labs stop shipping, doesn’t the open floor stop rising too? Meta drifting toward a closed line while Llama stalls gets cited as the canary. It doesn’t land, because distillation is a one-way ratchet. The capability already pulled into open weights doesn’t un-exist when the bonfire dims. Those models are downloaded, mirrored, forked onto tens of thousands of machines in every jurisdiction on Earth. You can’t recall them. Worst case - every current steward retreats tomorrow - the world still inherits a foundation good enough to build on, and the money chasing this is large enough that someone will. The stewards are replaceable. The foundation isn’t.

On geopolitics I’ll give ground, because pretending otherwise is dishonest. If the US bans Chinese weights on security grounds, that genuinely narrows the menu for a regulated bank - DeepSeek and Qwen come off the table. But a narrower menu isn’t an empty one. Mistral, Gemma, the Llama lineage, a growing stack of domestic open weights, all remain. Captivity needs every exit to hit zero at once. Five substitutes instead of fifteen is still a market with no captor.

And the most sophisticated objection of all: commoditization doesn’t guarantee a cheap bill. AWS commoditized compute and its margins are still fat; the cloud bill still balloons. True - and owning that completes the argument instead of dodging it. Value migrates up the stack. When the model layer commoditizes, the margin moves to the product and the agent layer above it. That’s not a hole in the thesis. That’s the thesis. The expensive, defensible part ends up where it belongs - in what you build, not in the raw intelligence you rent.

The Only Cage Is the One You Build Yourself

So where’s the real danger? Because there is one, and pretending there isn’t would make me the cheerleader I just accused a16z of being. It’s not the market. It’s your own architecture.

It shows up in the best argument the doomers have left, the one I’ve been circling: you don’t need a monopoly to extract from someone. You need switching cost. Lock-in is real even when captivity isn’t. Spend a year writing prompts and evals tuned to one model’s quirks and “just switch to the open one” becomes a lie you tell yourself. Migrating a production agent fleet is never ollama pull - it’s a re-tuning project with a regression-risk tail, and a vendor who knows you can’t leave cheaply has all the leverage it needs. Not monopoly rent - friction rent. And the friction is something you built into your own stack with your own hands.

Which is exactly why I run my own multi-agent harness instead of marrying one vendor’s SDK. The harness treats the model as a swappable input. Route a task to Claude today, to Kimi or Composer tomorrow at a tenth the cost, to a self-hosted Devstral the day someone gets greedy - and nothing above the model layer changes, because nothing above it was ever allowed to care which model answered. Build that abstraction once and you’re structurally un-gougeable. I didn’t build it because I’m disciplined. I built it after watching a side project’s bill behave like Uber’s, and deciding I never wanted to feel that again.

The other half of the bill is the part the doom thread got right and then misread. Uber’s engineers didn’t get gouged. They over-consumed, because the tooling made tokens feel free and nobody watched the meter. That’s a discipline problem. Measure tokens-to-value, not tokens-to-feature, and treat a runaway agent loop like the runaway cloud bill it actually is.

Do those two things - keep the exit wired, watch the meter - and neither of the rideshare weapons can touch you. The Uber rider had to take the price hike; standing in the rain with no alternative was the entire point of the trap. You’re not in the rain. You’ve got a floor that’s cheap, getting cheaper, good enough, and impossible for any one company to take away. The only way you end up captive in this story is if you wire your whole stack to one SDK, bet the company on a vendor’s goodwill, and let the meter run.

The frontier labs were never the thing to fear here. The only cage in this story is the one you build yourself - and that’s the one you get to refuse.