The Subsidy Era Is Ending
Anthropic is projecting its first operating profit and OpenAI is reportedly prepping an S-1. When your suppliers start caring about margins, your architecture inherits the problem.
Two pieces of news this month deserve to be read as one. Anthropic is projecting around $10.9 billion in second-quarter revenue and, more interestingly, its first quarterly operating profit. And multiple reports say OpenAI is preparing a confidential S-1, aiming at a public listing as early as the end of the year.
Set aside the horse race. The structural fact is this: the two most important model vendors are switching objectives, from “grow at any cost” to “demonstrate margins.” Every builder downstream of them should understand what that objective function change does to prices, because for three years we have all been building on top of someone else’s venture-funded generosity, and generosity is not a load-bearing architecture.
You have been subsidized
Be honest about the era we are leaving. Token prices that fell 9x to 900x a year did not fall purely because serving got cheaper (though it did, spectacularly). They fell because share mattered more than margin, and every lab priced accordingly. Free tiers that would bankrupt a normal SaaS. Context windows sold below the cost of the attention they compute. Reasoning tokens billed like commodity completions. This was customer acquisition spend wearing a price tag.
A company preparing an S-1 has to show the opposite story: unit economics that improve, revenue that outruns compute commitments measured in hundreds of billions. A company hitting operating profitability wants to keep it. Neither reprices the menu overnight; what changes first is the direction of surprise. In the subsidy era, price surprises were pleasant. From here, expect floors under the headline numbers and margin recovered in the fine print: rate limits, tier boundaries, cache pricing, “priority” lanes, enterprise minimums. The list price falls; the invoice does not.
The floor below and the ceiling above
Why can’t they just raise prices outright? Because the open-weight floor I wrote about in January holds firm: strong downloadable models cap what anyone can charge for mid-tier capability. So closed labs are squeezed between a floor they cannot price below and investors who now demand margin. There is only one direction to escape: upward, into differentiated capability, reliability guarantees, and integration depth that a file of weights cannot match, sold at premium tiers to buyers who can justify them.
Which means the market bifurcates. Commodity intelligence keeps getting cheaper, genuinely. Frontier capability under an SLA gets more expensive, or at least stops being quietly discounted. The middle, where most 2024-era app architectures live, “we call the second-best model for everything and eat the bill,” is exactly the stratum that stops making sense.
What I would do about it
I spent last year measuring what LLM serving actually costs under latency constraints, and the honest summary is that most systems buy far more intelligence than their traffic needs, then leak the surplus as margin to their vendor. In the subsidy era that waste was somebody else’s money. From here it is yours. Four moves follow:
Put a price on your own queries. Instrument cost per successful task now, so that when list prices shift you know your exposure by workload instead of by invoice-shock. Teams that metered nothing will renegotiate blind.
Build the routing muscle before you need it. Difficulty-based routing between tiers (the SAGE argument, one level up) converts vendor price changes from emergencies into config updates. The option to shift traffic is negotiating leverage even if you never move a token.
Treat efficiency as a durable skill, not a phase. A generation of engineers learned to ignore token counts the way pre-2008 web developers ignored memory. The ones who learn budgets, caching, truncation discipline, and early termination will carry an advantage that outlasts any particular price sheet.
Read your contracts for asymmetry. Committed spend with uncommitted pricing is the standard shape right now. Notice who bears the risk of the menu changing.
The healthy version of this story
None of this is doom. An industry that cannot show margins eventually shows layoffs, and an ecosystem built entirely on subsidy is a bad place to store your company. Profitable suppliers are stable suppliers, and stability is what you want from something you have wired into every workflow.
But the transition rewards a specific posture: assume the free lunch is a loan, and that the balloon payment arrives right after the IPO window. The teams that treated cheap tokens as a temporary tailwind will barely notice. The teams that treated them as physics are about to discover economics.