Cory Doctorow thinks AI is a bubble, and he argues it better than almost anyone who cheers him on. In his December speech at the University of Washington, the case runs like this: the tech giants stopped growing years ago, and a company priced as a growth stock faces catastrophe the moment the market stops believing. So they pump whatever keeps the multiple alive. Pivot to video, crypto, NFTs, Metaverse, now AI, each pitched with the same conviction, each abandoned when the next vehicle arrives. "Superintelligence" is the pitch this cycle because science fiction makes a better P/E story than enterprise software ever did.
His sharpest tool is the reverse centaur. A centaur is a human assisted by a machine: you drive the car, the car amplifies you. A reverse centaur is a human bolted onto a machine as its peripheral, hired to absorb the failures the machine can't, at machine pace, under machine supervision. Doctorow's thesis is that the money behind AI is betting on reverse centaurs: workers demoted into error-cleanup for systems sold to their bosses as replacements.
Take him seriously. The losses are real, the incentives he names are real, and I have sat through enough board-mandated AI pilots to know the reverse centaur is not a hypothetical. And then notice that his argument, stated carefully, is about two different balance sheets that the doom discourse keeps smearing into one.
Two Balance Sheets, One Shouting Match
Seller economics asks whether the labs make money. Company-level losses are enormous, but look at where they live: training runs, data-center capex, sales, and the free tier. The serving unit underneath is a different story. SemiAnalysis puts Anthropic's gross margins in the mid-60s — an independent analyst number, though The Information reported the same company projecting 40% when inference costs spiked, so treat the exact figure as contested. The direction is not. Serving paid tokens is gross-margin-positive; what bleeds is everything wrapped around it.
Even the famous counterexample proves the shape. When Sam Altman admitted OpenAI loses money on the $200 Pro plan, the reason was that heavy users outran the flat price. Sell unbounded agent workloads at a fixed monthly fee and the top of the usage tail eats you. That is a pricing-model problem, and pricing models get fixed. Physics problems don't. This one is being fixed right now, mostly at the expense of people like me: usage caps, tier splits, metered agents.
Buyer economics asks a different question: does a scoped deployment pay for itself? That is where operators live, and none of the training-run losses show up on this side of the table. I've written down what my own stack costs per month, and why "$20K in tokens against a $250K engineer" is the wrong math even when it flatters the tools. A deployment either produces measured value over measured cost or it doesn't. The lab's income statement has nothing to do with it.
Doctorow is right about the first balance sheet. The mistake is letting the first one answer for the second.
The Question That Does All the Work
Here is my actual thesis, earned from reps rather than from a valuation model: the economics work themselves out, but only for operators disciplined enough to ask "what problem am I solving" before pointing the tool at anything, and rigorous enough to measure the answer afterward.
An efficiency you don't understand is not an efficiency. It is an unmeasured cost wearing a demo. The team that deploys a coding agent with no baseline, no success metric, and no owner has not automated anything; it has hired a very fast intern nobody supervises and booked the salary as savings. That is blind deployment, and blind deployment is exactly the reverse-centaur failure: the humans end up serving the machine's output because nobody defined what the machine was for.
Which means my thesis and Doctorow's frame are the same observation read from opposite ends. Scoping rigor is what makes you a centaur instead of a reverse one. The human who decides what the machine is for stays the head. The human who cleans up after an unscoped machine becomes the peripheral.
The scoping questions I use, written down (writing them down matters, and the last section says why):
1. What is the problem, stated without naming a tool?
2. What does it cost today, in hours or dollars I can point at?
3. What does "working" look like, measured how, by whom?
4. What is the failure mode, and who catches it?
5. What is the kill criterion, decided before the pilot starts?
The Payoff and the Pin Are the Same Force
Now the part I have not seen anyone say plainly: the rigor that makes AI pay off is the same force that pops the bubble.
Walk through what happens if scoping discipline actually spreads. Every FOMO seat license gets audited against question 2, and half of them fail. Every board-mandated pilot meets question 3 and dies for lack of a metric. The all-you-can-eat experimentation budgets get metered. A real slice of today's AI revenue is hype-spend, and disciplined buyers stop paying for hype. Revenue evaporates precisely as deployments get better.
That is not bearish on AI. It is bearish on the bubble, and those are different positions. It is Doctorow's own "some bubbles leave behind something productive" scenario made concrete. He reaches for Worldcom, a fraud whose CEO died in prison while the dark fiber it buried still carries Doctorow's 2-gigabit home connection. Swap the fiber for scoped deployments quietly compounding inside businesses while the FOMO spend burns off. The economics working out and the bubble deflating are the same event, narrated by an optimist and a pessimist.
The Craftsman Clause
There is an investor-side implication buried in this that is more uncomfortable than any crash prediction.
If every deployment that pays needs a human who understands both the domain and the tool's failure modes, then AI is not the frictionless labor replacement the trillion-dollar valuations are priced on. It is a power tool that still needs a craftsman. A nail gun does not fire the carpenter; it makes the carpenter's judgment the binding constraint on every house. Attention was always the bottleneck; the tools just moved it.
Great news for buyer ROI. Quietly fatal for the "human exits the loop" story being sold upstream, because the wholesale-replacement bet only pays if the craftsman requirement goes away, and every honest deployment I have run or reviewed says it doesn't. The augmentation bet keeps cashing small checks. The replacement bet keeps pre-spending checks nobody has figured out how to write.
Two Half-Lives, One Fork
So if the scarce input is the human who can scope, the personal question is what that scarcity is worth and how long it lasts.
The scoping skill is teachable in principle and bottlenecked in practice, and the bottleneck is structural, not a fad. The skill is tacit. It lives on the seam between a domain and a tool, where few people sit. And organizations promote people for shipping, not for the scoping that made the shipping cheap; nobody's OKRs reward the pilot that didn't happen.
But the edge has two halves with opposite half-lives, and confusing them is the trap. Tool mechanics — prompt-craft, knowing this month's model's failure modes, the tricks — depreciates fast, because every model release is the labs folding exactly that knowledge into the product. The agent-loop mechanics I wrote up are already aging out from under that post. Problem formulation — translating a business's real problem out from under its stated one — depreciates slowly, because it is organizational translation, and organizations stay human. The trap is that the fast-depreciating half is the legible, demoable half. It feels like the edge because you can show it off. The durable half looks like "just asking questions."
Which leaves the fork, and I am going to leave it as a fork rather than resolve it for you. Stay the indispensable bottleneck and you extract rent: real money, now, capped at your personal throughput, gone the day the scarcity ends. Or codify the tacit half into artifacts a colleague can run without you — a written scoping method like the checklist above, reference patterns from the deployments that worked — and you dissolve your own bottleneck while capturing the value of having systematized it. Rent pays this quarter. The moat is owning the codified version when everyone else finally needs it.
The bubble being real is what creates the opening; nobody pays a premium for discipline in a calm market. The economics work out for whoever is on the right side of the rent-to-moat conversion. And the thing to be dismantling, deliberately, starting now, is the bottleneck being you personally — before the labs dismantle the cheap half of your edge for free and leave you holding only the part you never bothered to write down.
Top comments (0)