Do the unit economics of AI products actually work?

◇ Neutral question

Conflates lab economics, app-layer economics, and customer ROI: different answers each.

◷ Share view, as of June 2026: confidence, rationales and priors are hidden. The questions, stances, claims, tests and evidence stay visible.

Stances · candidate answers

Q3-S3A

Margins are structurally improving; the model is viable; we're early.

Held by lab CFOs, growth investors

Inference cost/token has fallen orders of magnitude; serving gross margins positive and improving; losses are training capex and S-curve acquisition. Revenue growing faster than prior software categories.

would move me offRevenue growth decelerates sharply while training costs keep rising, with no path to amortizing training over a growing base.

Strongest opposing: Q3-S3B

Evidence behind its claims: 328 support · 69 contradict · 301 neutral

Q3-S3B

Labs subsidize usage; true unit economics are hidden.

Held by skeptical software analysts, VCs who passed

Headline margins exclude the cost of the model rebuilt every 12–18 months; 'cost/token falling' is true but irrelevant if you must keep spending more to reset the clock. Revenue per dollar of compute is poor.

would move me offA lab shows a full cycle (train → serve → retire) where a generation's lifetime revenue exceeds fully-loaded training + serving cost, disclosed cleanly.

Strongest opposing: Q3-S3A

Evidence behind its claims: 218 support · 45 contradict · 194 neutral

Q3-S3C

Lab vs application economics are different questions; the app layer is where value accrues (or doesn’t).

Held by app-layer founders, enterprise SaaS investors

Bear version: labs commoditize apps from below. Bull version: workflow integration + proprietary data create defensible margin above the model.

would move me off(bear) Durable high-retention app businesses with margins that survive the next model release commoditizing their core feature.

Evidence behind its claims: 93 support · 51 contradict · 143 neutral

The map: how claims bear on stances

supports contradictshover a node to highlight its links

Claims

3.1

Inference cost per unit of capability is falling rapidly.

Q3-S3A

3.2

Frontier training costs rise fast enough to outpace per-model revenue.

Q3-S3BQ3-S3A

3.3

Enterprises achieve measurable ROI and renew / expand, not just pilot.

Q3-S3AQ3-S3BQ3-S3C

3.4

App-layer companies retain margin after the model commoditizes their feature.

Q3-S3C

3.5

Recognized revenue (not bookings / ARR run-rate) grows at claimed rates.

Q3-S3AQ3-S3B

3.6

Training-data IP liability is now a quantifiable balance-sheet cost for AI vendors, and vendors unable to document clean training-data provenance face incremental procurement exclusion risk in enterprise and government channels.

Q3-S3AQ3-S3B

3.7

AI coding tools reduce enterprise software headcount or external vendor spend faster than new AI-related roles or tooling costs offset the savings, producing net cost compression.

Q3-S3AQ3-S3C

3.8

Enterprise contract renewal and expansion rates for AI software are high enough, and contract durations long enough, to generate a revenue backlog that justifies current data center capex levels.

Q3-S3A

7.4

Measured productivity gains from AI in knowledge work are large enough to change staffing plans.

Q3-S3A

Stances

Q3-S3A

Margins are structurally improving; the model is viable; we're early.

held by lab CFOs, growth investors

Q3-S3B

Labs subsidize usage; true unit economics are hidden.

held by skeptical software analysts, VCs who passed

Q3-S3C

Lab vs application economics are different questions; the app layer is where value accrues (or doesn’t).

held by app-layer founders, enterprise SaaS investors

Claims, each with its falsifiable test

3.1 · Inference cost per unit of capability is falling rapidly.

would falsifyCost-per-token-at-fixed-quality stops declining.

EconomicsClean to resolve

3.2 · Frontier training costs rise fast enough to outpace per-model revenue.

would falsifyA generation's lifetime revenue exceeds fully-loaded cost.

EconomicsSlow to resolve

3.3 · Enterprises achieve measurable ROI and renew / expand, not just pilot.

would falsifyNet revenue retention + pilot-to-production stay low across the base.

EconomicsSlow to resolve

3.4 · App-layer companies retain margin after the model commoditizes their feature.

would falsifyApp-layer margins compress sharply on each major model release.

EconomicsSlow to resolve

3.5 · Recognized revenue (not bookings / ARR run-rate) grows at claimed rates.

would falsifyAudited recognized revenue diverges materially from announced ARR.

EconomicsSlow to resolve

3.6 · Training-data IP liability is now a quantifiable balance-sheet cost for AI vendors, and vendors unable to document clean training-data provenance face incremental procurement exclusion risk in enterprise and government channels.

would falsifyAt least two enterprise or government procurement RFPs within 12 months include mandatory training-data lineage documentation as a pass/fail vendor-screening criterion, or at least one additional settlement exceeding $500M is reached within 18 months of the $1.5B ruling.

Bridge-claims connecting out

⤧ B1Agentic reliability gains will translate into enterprise ROI.Capability → Economics ⤧ B2Enterprise ROI will translate into durable recognized revenue (not bookings).Economics → Market ⤧ B5Hyperscaler capex now outpacing operating cash flow forces external debt financing, which raises the cost of AI infrastructure build-out and creates credit-market feedback that can slow deployment timelines independent of physical power or chip constraints.Economics → Build-out

Q3 / 9 claims / 3 stances / 3 bridges·confidence moves are made manually; AI proposes only