a legacy company coordinates people with people: meetings, status chains, approval layers. what that costs is well measured — bain: the average company loses more than 20% of its productive power to organizational drag; a quarter of a typical supervisor's week is unnecessary meetings and email.
an ai-native company is a small team directing agents that hold real jobs — releases, defects, spend — with coordination written to a ledger by machines. the leverage precedent is old: whatsapp served 420 million users with 55 employees in 2014, before agents. agents raise that ceiling.
>20%productive power lost to organizational drag (bain)
30%success rate across 900 digital transformations (bcg)
95%of enterprise genai pilots: no measurable p&l impact (mit, 2025)
10–50xyearly deflation of llm inference at constant capability (a16z, epoch ai)
takeaway — the two companies pay to know things in different currencies: payroll vs compute. everything below follows from that.
02 · the inflection
the inflection
cost to serve one user per month, $ — log scale: each gridline step is 10x · solid: current · dashed: projection · ai-native line models an ai-rich product (far more inference per user than a legacy app does)
basis: the ai-native line is a modeled ai-rich consumer product at current vendor prices, payroll-realistic headcount included, deflating with a decelerating blend of measured inference-price trends (bearish-to-base dials, not the historical extreme). at product-parity workloads the entrant is already below the incumbent line today; at ai-rich depth the projected crossing lands 2026–2028. the legacy line anchors to the leanest public incumbent's ~$1.36/user-month, drifting with payroll inflation.
the fact pattern, side by side — public filings for the incumbents, our model (payroll included) placed at the leanest incumbent's scale and monetization for comparison:
category leader
leanest incumbent
ai-native (modeled, same scale)
employees
~2,200
~151
~15–30
operating cost / yr
$2.6B
~$244M
~$80–130M, falling
cost per user-month
~$16 per paying user
~$1.36
tens of cents
revenue per employee
$1.6M
$2.9M
~$15–30M(illustrative)
modeled column: illustrative — our archetype at the leanest incumbent's scale and monetization; assumptions and bases in the full paper. per-user cost is comparable to the incumbent's today at ai-rich product depth, already below at product parity.
takeaway — the level was never the story. the slope is.
03 · the deflation divide
the deflation divide
composition of the cost to build and serve the product, % of total — growth marketing excluded on both sides (both pay market rates for attention) · illustrative shares
basis: published measurements put compute at 57–70% of total spending at leading ai companies (epoch ai) and 40–50% of revenue at ai-first startups — companies with staffs many times larger than this archetype. our leaner model lands at the top of the range (70–90% of serve-cost); the bar is drawn at the middle. the legacy mix approximates people-dominated operating cost from public filings, marketing excluded.
inference gets ~10x cheaper per year (measured up to 50x). published cost mixes: compute is 57–70% of total spend at leading ai companies, 40–50% of revenue at ai-first startups; our leaner model runs higher. even at the floor — a 50% inference share — half the cost base deflates every year while the legacy base, ~0% inference, inflates with payroll.
takeaway — two years out, today's small gap becomes several-fold to an order of magnitude. no decision inside the legacy company changes which curve it's on.
04 · same token, different returns
enterprises: 95% of genai pilots show no measurable p&l impact (mit); 42% of companies abandoned most ai initiatives in 2025, up from 17% (s&p global, n>1,000). the rational response is already visible: caps, rationing, deferred spend. when tokens don't return, throttling is correct.
ai-native companies: inference is the product, so every token faces a margin test. spend concentrates where tokens earn; waste-hunting (caching, routing, lean context) is a survival discipline, not an initiative.
takeaway — the divide isn't who pays less for tokens. it's who gets paid back for them. this is why "just adopt ai" doesn't close it.
05 · the incumbent's dilemma
the arithmetic, on the leanest incumbent: staying margin-whole at half today's monetization requires cutting ~half its cost base; meeting the entrant's projected two-year floor requires ~90%. and cutting isn't transforming — remove a third of the people and the coordination model remains. the base rate for far shallower change: 30% success across 900 transformations (bcg).
the transformation window
conceptual — not to scale · y: $ · x: time → · the window is the region where transforming still costs less than the capacity to pay for it
schematic. the mechanism: the capability gap to close grows with the deflation curve while margin and cash shrink with ceded growth — so delay raises the price and shrinks the payer.
transformation is cheapest today, when the incumbent is strongest — and most painful today, because it cannibalizes a working business. waiting feels rational every quarter: defend margin, return capital, trim single digits. the category leader put over 100% of free cash flow into buybacks and dividends last year while paying users declined. margin defense spends the window.
takeaway — we believe most coordination-built companies will not survive the repricing: managed decline, not bankruptcy. what transplants: knowledge bases, decision capture, machine escalation. what doesn't: the cost structure. that has to be founded.
06 · what would prove us wrong
deflation stalls below ~2x/yr for two straight years.
no ai-native entrant achieves real market liquidity by mid-2028.
an incumbent proves the reduction — >1,000 employees, >50% cost cut in 24 months, function preserved.
ai-native unit economics break upward at production scale — per-user ai costs are easy to underestimate.
07 · notes & sources
this post summarizes a longer research paper (archetypes, scorecard, full cost model with every assumption's basis). all named statistics are published: sec filings, bain (mankins & garton), bcg (n=900), mit project nanda, s&p global (n>1,000), epoch ai / visual capitalist, cisq, a16z. modeled figures are labeled as modeling and stated as magnitudes. not investment advice; the authors hold no positions in any company named.