The token bill comes due

Subsidized intelligence is the bubble. Q3 is when the invoice arrives.

The AI trade rests on a number almost nobody prices: the true cost of a token. Every demo that wows a boardroom, every "agent" that drafts a contract or reconciles a ledger, runs on inference that is sold below what it costs to produce. The closed labs have been buying market share with investor capital, the way ride-sharing once bought rides. That subsidy has an expiration date, and I think the expiration date is close enough to trade on this quarter.

The math of the full-time agent

Chat was cheap because chat is bursty — a few thousand tokens, a human reading the output, long pauses in between. The product everyone is now selling is different: agents that work all day. An agent that genuinely automates a workflow reasons in loops, re-reads its context, calls tools, checks its own work. Token consumption per unit of useful output is orders of magnitude higher than chat, and reasoning-heavy models burn more tokens per query, not fewer. Run the arithmetic on a "digital employee" working full-time at frontier-model prices and you get an annual token bill that competes with a human salary — before the integration costs. At unsubsidized prices, it's worse.

That is the gap between the story and the spreadsheet. The promise being capitalized in public markets is full workflow automation. The thing that is actually affordable is autocomplete. The labs know this, which is why pricing keeps falling on the surface while compute commitments balloon underneath — the difference is being eaten by somebody, and that somebody is a balance sheet that ends.

The IPOs are the top

The reported IPO preparations among the major private labs are read by the market as validation. The better reading is that they are the top. Going public converts narrative into liquidity for insiders at precisely the moment the narrative is most valuable and the unit economics are least examined. Public companies file financials. Quarterly disclosure will, for the first time, force the question the private markets never pressed: what does inference cost, what is gross margin on a token, and how much revenue is circular — labs paying clouds paying chipmakers funding labs. The IPO window doesn't end the subsidy; the scrutiny that follows it does.

The flight to good-enough

Meanwhile the substitute is free and improving. Open-weight models are now good enough for the majority of deployed workloads — classification, extraction, summarization, routine code, internal copilots. When the closed labs reprice toward true cost, the migration doesn't happen gradually; it happens at the speed of a config change, because the OpenAI-compatible serving stack made switching costs approximately zero. Cheap tokens from commodity hosts running open weights become the default, and frontier-model pricing power collapses to the narrow set of tasks where the capability gap is genuinely binding.

That gap is also narrowing from the supply side. The most powerful capabilities are exactly the ones the proprietary labs are most constrained in shipping — autonomy, cyber-relevant skills, self-improving agent loops sit behind safety review, government scrutiny, and (post-IPO) board-level liability. The closed frontier decelerates by obligation while the open ecosystem inherits last year's frontier for free. Paying ten times the token price for a model eighteen months ahead made sense; paying it for a model six months ahead does not.

How it unwinds

Sequence, over the evaluation window and the quarters that follow: enterprise pilots fail to convert at the rate the revenue projections require, because the agent math doesn't close. Token subsidies shrink as labs dress up margins for public listing. Open-weight migration accelerates on every repricing. Lab revenue growth stalls — and with it the single assumption holding up the capex super-cycle: that demand for premium inference is infinite. Compute commitments get renegotiated quietly, then publicly. The datacenter buildout chain — the GPU resellers, the neoclouds, the power-and-shell speculators — is priced for a demand curve that goes vertical. It reprices for one that flattens.

The basket

Tilt away from the subsidy chain: the pure-play AI infrastructure complex — neocloud GPU landlords, server assemblers, speculative datacenter developers, and the crypto-miners-turned-AI-hosts — which carries maximal exposure to a capex pause with minimal pricing power, along with the most AI-narrative-dependent megacap multiples. Tilt toward the beneficiaries of cheap intelligence: platforms with open-weight strategies and the distribution to monetize free models; commodity inference and edge-compute plays that win when tokens are a commodity rather than a franchise; and cash-generative software and services companies that consume AI to cut costs rather than sell it as a dream — the businesses for which collapsing token prices are a margin tailwind, not an existential event.

What would break this thesis

The thesis fails if a frontier model ships in Q3 that is so capable it resets willingness-to-pay; if agent token-efficiency improves faster than subsidies unwind (making automation affordable at honest prices); if the IPOs are delayed and the private subsidy machine keeps running another year; or if open-weight progress stalls and the capability gap re-widens. The numbers to watch: published per-token pricing versus disclosed compute commitments, enterprise pilot-to-production conversion rates, and the share of new deployments on open-weight models.

The market has correctly concluded that intelligence is valuable. It has not yet asked what intelligence costs. Q3 is when someone finally has to answer in a filing.