Willison argues the eye-watering bills landing in engineering org inboxes are the clearest signal yet of PMF. Coding agents like Claude Code, Cursor, Aider, and Codex have become daily drivers for highly-compensated professionals, running hot tool-call loops for hours — and Anthropic's rumored first profitable quarter validates that this is a real, paying market, not hype.
The editorial highlights the developer who burned $2,180 of on-demand token equivalent on a $200/month Max plan — a 10x subsidy on the flagship tier. That ratio historically signals either an imminent market explosion or an imminent repricing, and conflating heavy usage with sustainable PMF papers over a much messier picture.
Working backwards from the $5-10 trillion in AI infrastructure capex (Stargate, Microsoft's $80B, Meta's $60B, Google's $75B) that needs recouping over five years, token spend would need to hit ~$1 trillion per year sustained. With only 30 million developers globally, every single one would have to spend several thousand dollars annually on inference indefinitely for the math to work — otherwise the data centers get written down.
Simon Willison published an essay arguing that Anthropic and OpenAI have, finally and unambiguously, found product-market fit — and the evidence is the eye-watering bills landing in engineering org inboxes. Anthropic is strongly rumored to be heading into its first profitable quarter. The mechanism isn't chatbots or enterprise seats. It's coding agents: Claude Code, Cursor's Claude integration, Aider, Codex, and the long tail of agentic IDE wrappers that now run hot loops of tool calls against frontier models for hours at a time.
The specific data point lighting up Hacker News (1,049 points, hundreds of comments): a developer reporting that their $200/month Claude Max subscription metered out to $2,180.16 in equivalent on-demand token spend in a single month. That's a 10x subsidy on a flagship plan, and it's the kind of unit economics that historically signals either a market about to explode or a market about to be repriced.
Willison's framing is that this is the iPhone moment for LLMs as professional tools — "daily drivers for the work carried out by extremely well-compensated professionals." The counter-framing, which dominated the HN thread, is that PMF and profitability are different things, and that conflating them papers over a much messier picture.
The most cited comment in the thread came from `trjordan`, doing the back-of-envelope math: the AI hardware buildout — Stargate, Microsoft's $80B capex, Meta's $60B, Google's $75B — represents somewhere between $5 trillion and $10 trillion that needs to be recouped over the next five years, or the data centers start getting written down. That implies token spend needs to hit roughly $1 trillion per year, sustained. There are about 200 million knowledge workers globally and 30 million developers. The arithmetic only closes if every developer in the world is spending several thousand dollars a year on inference, indefinitely.
That's plausible if you believe coding agents become the default IDE. It's much less plausible if you've watched the open-weights curve. Commenter `binary0010` made the point bluntly: "how do OpenAI and Anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?" GLM-5.1, DeepSeek V4, Qwen 3.5 Coder, and Kimi K2 have collapsed the capability gap on coding tasks specifically. The frontier labs are charging premium prices for a capability lead that has historically been measured in months, not years.
There's a third critique worth taking seriously, from `prepend`: "Tokens don't have an intrinsic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I'm getting a billion dollars worth of pots and pans for $19.99." The on-demand API price is set by Anthropic. Of course the subsidized plan looks like a steal when you benchmark it against a number Anthropic itself controls. The real question is what the marginal cost of serving those tokens is — and nobody outside the labs actually knows.
What we do know: inference costs per token have dropped roughly 10x per year for equivalent capability since GPT-3.5. Hopper-class GPUs are being depreciated over five years. Blackwell is shipping. If the gap between "price charged" and "cost to serve" is currently 50%, and inference costs drop another 10x while prices drop 3x, the labs absolutely do hit profitability. The bear case isn't that frontier labs can't be profitable — it's that they can't be profitable at valuations that assume they own the entire market.
The Uber comparison Willison invokes cuts both ways. Yes, Uber eventually found profitability after a decade of subsidies. Uber also destroyed roughly $30 billion in investor capital getting there, and the end state is a business worth a fraction of its peak valuation, in a market where rideshare prices are now substantially higher than taxi rates were in 2014. "PMF" describes whether people want the product. It says nothing about what they'll pay once the subsidy ends.
If you're a senior engineer making architecture decisions in 2026, the practical implications fork sharply depending on what you believe:
If you believe the frontier labs win: keep building on Claude/GPT APIs, lean into structured tool use, and assume capabilities continue to ratchet up faster than prices ratchet down. Your bet is that the cognitive overhead of switching models exceeds the savings of running open weights. Budget for your inference spend to roughly triple over the next 18 months as the subsidies on power-user plans get rationalized.
If you believe open weights catch up: invest in inference infrastructure now. Stand up vLLM or SGLang behind a router. Build a thin abstraction layer that lets you swap GLM-5.1 in for Claude on the 80% of tasks where it's good enough, and reserve Claude for the hard 20%. Your bet is that the marginal capability of frontier models is worth 10x on the long tail of work, but not on the median PR review or refactor.
The pragmatic middle, which is where most shops will end up: run a cheap open-weights model as the first-pass on every agent loop, and escalate to a frontier model only when the cheap model fails a confidence check. This is the same pattern that emerged in search (cheap lexical → expensive semantic) and it's already showing up in production agent frameworks. If your stack doesn't have a tier-and-escalate pattern by Q3, you're either overpaying or you've made a deliberate bet on a single vendor.
One quieter implication: the $2,180-of-tokens-for-$200 stories are great copy, but they're also the most actionable telemetry your finance team has. If you're an engineering lead, audit your team's agent usage before procurement does it for you. The labs will eventually meter these plans more aggressively; better to know which engineers are running which loops now.
The HN thread's center of gravity wasn't "is this PMF" — it was "who pays for the buildout." That's the right question. Anthropic posting a profitable quarter on coding agents is a real milestone, but it's profitability against current opex, not against the trillions in capex that need to be serviced. The next twelve months will tell us whether the frontier labs can hold pricing power as open weights close the gap on coding specifically — the one domain where they've found PMF. If they can't, the writedowns start in 2027, and the architecture decisions you make this quarter will look either prescient or expensive.
I feel like there's a bit of AI psychosis in this particular post.>"These are tools which burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals.">"Somehow this fragment turned into headlines
I find this analysis confusing. PMF for coding was likely reached some time last year. Profitability, which is different, we don’t know. The article kind of confuses both without making a strong economic case or using numbers in a compelling way. I don’t understand what the Uber case has to do with
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?I don't see the business model working. My closest friend actually does automation software for large companies.He does not use Claude or openai at all. He primarily uses gpt 120
> $2,180.16 worth of tokens for $200“Tokens” don’t have an intrisic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I’m getting a billion dollars worth of pots and pans for $19.99.I think it’s funny how we are throwing critical thinkin
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a worl