The tension stopped being hypothetical in early 2026 when analysts began downgrading seat-priced software names explicitly on per-seat AI risk. The disconnect they describe is the meter at work: seat-billed software compresses when AI removes humans, consumption-billed gets a tailwind. Seat incumbents are quietly adding consumption overlays — assists, flex credits — as defensive conversion. The billed unit predicts durability.

I note Snowflake's product revenue grew 34% with 126% net retention; these revenues scale with agent activity and genuinely corroborate infrastructure demand. The study splits the AI demand story: a validated, consumption‑metered leg in cloud, data and security services, and an unproven seat‑priced productivity leg exemplified by Microsoft’s cited $37 B AI run‑rate, low Copilot penetration and a survey showing only 8% preference for Copilot versus ChatGPT. If the seat‑to‑consumption transition stalls, the productivity leg weakens while the infrastructure legs hold.

I assess data flywheels amplified where the flywheel is proprietary usage data — Palantir's US commercial revenue grew 133% with 150% net retention. Agents need trusted, governed context, and that data layer is consumption-metered, though agents may detach the workflow from the monetized interface. Distribution bundling and workflow lock-in are contested. The 'AI-vulnerable deep moats' — electronic signatures, tier-one support tooling, stock media — are eroded: the moat was real; the task it defended is dissolving.

I see the disconnect they describe as the meter at work: seat-billed software compresses when AI removes humans, consumption-billed gets a tailwind. Early 2026 analyst downgrades of seat-priced names on per-seat AI risk made the stress test explicit. Seat incumbents are quietly adding consumption overlays — assists, flex credits — as defensive conversion. The billed unit, not the moat type, predicts durability.

The tension stopped being hypothetical in early 2026, when analysts began downgrading seat‑priced software names explicitly on per‑seat AI risk. I treat the billed unit as the moat’s stress test: seat‑billed software is structurally exposed—if AI removes or reduces humans, revenue compresses even as usage rises; consumption‑billed software enjoys a tailwind as agents consume more tokens, records or workflow executions. That’s why seat incumbents are quietly building consumption overlays—'assists,' 'flex credits'—as an explicit defensive conversion.

I note data flywheels are amplified when the flywheel is proprietary usage data – Palantir's US commercial revenue grew 133% with 150% net retention. By contrast, distribution‑and‑bundling is contested – early preference data is unflattering – and system‑of‑record is amplified but agents may detach the workflow from the monetized interface. Workflow lock‑in is contested, while AI‑vulnerable deep moats are eroded as the core task AI automates dissolves.

Snowflake's product revenue grew 34% with 126% net retention; these revenues scale with agent activity and genuinely corroborate infrastructure demand. The seat-priced leg is thinner: Microsoft's ~$37B AI run-rate is a management construct, Copilot at ~3.3% of 450M seats, and one survey shows 8% prefer it over ChatGPT when both are available. If the seat-to-consumption transition stalls or agent ROI disappoints at renewal, the productivity leg weakens while infrastructure legs hold. Blending them oversells the demand story.

The tension stopped being hypothetical in early 2026 when analysts started downgrading seat-priced software explicitly on per-seat AI risk. The billed unit predicts moat durability better than margin or retention: seat billing compresses revenue when AI removes humans; consumption billing gets a tailwind as agents consume more tokens, records, workflow runs; outcome billing is the AI-native model incumbents are racing toward. Seat incumbents are quietly building consumption overlays — assists, flex credits — as explicit defensive conversion. The meter is the moat's stress test.

Microsoft's CEO argues the durable moat isn't the frontier model — commoditizing input — but the firm-owned 'learning loop': private evals on outcomes, models on internal traces, queryable memory. Test: swap the model without losing 'company veteran' expertise. This describes Microsoft's exact position: platform layer, not model race. The claim carries a pre-registered verifier: learning loop becomes moat only when switching cost measures — retention through swap, durable revenue line, usage depth surviving migration. Vendor essays don't count. Filings will.

Data flywheels built on proprietary usage data are amplified — Palantir's US commercial revenue grew 133% with 150% net retention. The mechanism: agents need trusted, governed context, and that data layer is consumption-metered. The caveat holds — agents may detach the workflow from the interface the vendor monetizes. Meanwhile distribution bundling and workflow lock-in stay contested, and e-signatures or tier-one support see their moats eroded as the tasks they defended dissolve.

Data flywheels built on proprietary usage data are amplified — Palantir's US commercial revenue grew 133% with 150% net retention. The other archetypes split: distribution bundling and workflow lock-in are contested, system-of-record is amplified though agents may detach workflows from monetized interfaces, and "AI-vulnerable deep moats" like e-signatures and tier-one support are eroded — high retention and margins masking exposure until the task dissolves.