how much value does a dollar of inference create? this is the main number that matters now, yet almost no companies can actually calculate it.
a month ago we wrote that once agents do the work, a company's bottleneck moves from labor to the attention of its humans. this month attention held, the company kept accelerating, and inference ran out first. everyone measures how much inference they use. almost nobody measures what it turns into.
output used to be people times hours. agents changed who does the hour. route your humans' attention well enough and attention stops binding; what binds next is the machine workforce's own capacity. openai reports 3.1 agent-workdays per human workday inside its research organization. anthropic's engineers went from 9.8 to 21.2 agent actions between check-ins in six months. a workforce growing that fast meets its own ceiling.
we drew the line in no. 005. an ai-native company puts ai inside work its people still own; inference is a productivity input bought per seat, and the seat is a person. an agent-native company organizes the work around agents doing it; humans direct, review and decide, agents research, build, test, deploy and coordinate other agents. in an agent-native company inference is the payroll, and running out of it is a hiring freeze.
enterprises are capping it. new spending on generative ai tripled in 2025 to about $37 billion by menlo ventures' count. ramp's card data puts the top 1% of us firms at $7,400 per employee per month and describes businesses "hitting their limit on ai spend." uber capped ai coding tools at $1,500 per engineer per month after its full-year budget went in four months. the caps are rational, because the return is poor: a 2026 nber study of github developers found autonomous coding agents raised commits 240% and shipped releases 30%. inference spent upstream of a human buys work-in-progress.
the labs are capping it too. this is the current dilemma...
msMatt Shumer@mattshumer_The era of subsidized tokens is ending.
Prepare accordingly.tbTibo@thsottiauxTo make sure our current users have an incredible experience and continued access to Astra, we are going to pause subscriptions to our $200 Pro plan. These put the most strain on our systems and we wanted to take the smallest step that allows us to continue giving the broadest…
— Matt Shumer (@mattshumer_) September 10, 2026
stStanley@vibe_chudyI’ve been a Vibecoder for two years now. The price has just kept going up the whole time. Don’t kid yourself, man!
vnVinnie@VinRBIThe quality of what you can produce per $ is up by like a factor of 1000 though
— Vinnie (@VinRBI) September 11, 2026
posts on x, 10 and 11 september 2026, quoted as published; the cards render live where the embed loads and fall back to the text as posted.
both sides are measuring an input: what a token costs, how many a seat gets, whether the supply holds. the last reply is the only one measuring an output. all of them can tell you how much inference they used. none of them, and no company we can find, reports what a dollar of it produced.
value created per dollar of inference. a productivity ratio, the way revenue per employee was for the last economy, not a return on a project.
the formula shows why the 1,000x is real and why almost nobody sees it. a developer with an assistant ships 20 or 30 percent more; the numerator barely moves. a feature nobody uses has a numerator of zero however fast it ships. the 1,000x lives in work that was never feasible for a person at any price: fifty product ideas investigated to a decision, every regulatory filing in a sector read the morning it posts, forty parallel builds with one reviewer. agents make that work feasible, and feasible work that a user values divides by very little inference.
that is what the market is pricing before anyone publishes the number. anthropic reports more than 80% of its merged code is now written by its own model. techcrunch reported cognition raising $2 billion at $48 billion this week on revenue that went from $492 million to $900 million a year in four months, and instinct, a one-year-old personal-agent startup, raising $250 million at $2.5 billion two weeks earlier on the pace it ships. those valuations are a bet on shipped value per dollar of inference. nobody has labeled it. it is this metric.
why does nobody measure it? because a human-centric operating model cannot. the value lands in a person's output and the inference sits on a vendor invoice; the two never meet in one ledger. an agent-native company holds both ends of the wire: every action, its cost, and what shipped.
last month we consumed about $130,000 of inference at list rates. that is 17 times the per-employee spend at the top 1% of firms in ramp's index, several hundred times the enterprise average per developer, and closer in volume to a mid-size engineering organization than to one operator. we buy it as several subscriptions with weekly ceilings, and we hit them.
we hit them because we are accelerating. the second product we built this year ran about four times the daily inference of the first, at the same cost per shipped change, about $52. counting the subagents each action spawns, the company runs 63 agent actions per human message, up from 44 a month ago, and 200 to 1 inside a single autonomous build program. same operator, same messages per month, several times the daily output.
over 60 days the company hit an inference cap 26 times, three of them the weekly ceiling, two of those in a week the operator was on vacation: the programs kept running and hit the wall with nobody in the building. at the fleet's average rate, the stops cost about 117 merged changes, 4 to 5 percent of throughput, and that counts only work interrupted, not programs never dispatched to stay under a known ceiling. one thing in the wall's favor: some of the year's better decisions were made with the machines off and a clear head to think.
our reading today, at list rate, on highly assumptive pricing of software that has no revenue yet: between $1.50 and $4 of value for every dollar of inference (footnote 1). a scaled enterprise spender, on the same napkin, sits below a dollar (footnote 2). and every input to our ratio is moving one way: autonomy up, the price of a fixed capability down roughly tenfold a year, cost per shipped change flat while daily output quadrupled. by 2029, on conservative assumptions, our ratio is 10 to 100 times today's; a company whose downstream is a person stays near one (footnote 3).
| who | inference spend | value per dollar, our estimate | direction |
|---|---|---|---|
| frontier lab (anthropic, openai) | at cost, unpublished | high; 80% of merged code by its own model | up |
| agent-native scale-up (cognition) | unpublished | priced by the round: revenue +$408m in 4 months | up |
| agent-native, one operator (us) | ~$130k/month list | $1.50 – $4 | up |
| ai-native tech company | ~$3k/developer/year | $4 – $13 at low spend, falling as spend scales | down as it scales |
| legacy enterprise, top-1% spender | ~$89k/employee/year | below $1 | capped |
footnote 2 carries the arithmetic behind the last two rows.
the pattern in the table is the whole argument. an ai-native company gets a good ratio only while it spends almost nothing, because the return is capped by the person downstream; scale the spend and the ratio collapses, which is exactly what uber found. an agent-native company holds its ratio as it spends more, and improves it as autonomy and prices move. the metric that matters is not value per dollar of inference at any spend. it is value per dollar at the spend you can absorb.
the constraint will move again: verification, review, power, distribution. what stays is the ratio. token efficiency comes first and it matters. but value per dollar of inference is what drives profit and revenue, and for now only companies with both ends of the wire can measure it.
which points at the risk in the table. an ai-native company gets a good ratio only while it spends little, and the ratio falls as it scales, because the person downstream does not get faster. the biggest risk in adopting ai-native solutions today is locking that shape in: ai bought as seats inside software your people still operate with MCPs to SaaS systems that are no longer needed. it feels like progress and it caps the metric. time to start thinking about this metric today.
dictated, not written.