The inference economy: sizing the prize

Start with the only demand data that can't be argued with, disclosed token volumes. Google processed ~9.7 trillion tokens a month across its products in 2024, ~480 trillion by May 2025, and over 3.2 quadrillion by May 2026, roughly 7x in the last year alone. Microsoft processed over 100 trillion tokens in a single quarter of early 2025, up 5x year-on-year. The largest independent routing layer went from negligible volume to 25 trillion tokens a week. And the composition is shifting as fast as the volume: reasoning models went from a rounding error to more than half of all routed tokens in about a year, and average tokens-per-request has roughly quadrupled since early 2024. Agents compound this, credible estimates put agentic workloads at 5-30x the tokens of a chat interaction, with complex multi-tool tasks far above that.

Phase 1, the training-capex era (2023-2025). Value pooled at silicon and scarce compute; the neocloud tier was born in the gap between hyperscaler capacity and frontier-lab demand. This phase built the tier's revenue, and its cost structure.

Phase 2, the token price war (2025-2027, now). Inference overtakes training as the dominant workload, by 2026 this is industry consensus, and the chip vendor now describes the entire buildout as "token manufacturing." Raw compute commoditises, the terrain of our companion deep-dive, Neoclouds & Inference Platforms; token prices collapse; and the collapse itself drives adoption: enterprise generative-AI spend went $1.7B → $11.5B → $37B across 2023-25, tripling in the last year, with ~$12.5B of that flowing through model APIs. Consensus forecasts put the inference market at ~$255B by 2030, with the hosted-platform slice around $105B, but note that these forecasts assume ~17-19% annual growth in a market whose observed spend is tripling annually. Every credible forecast of cloud in 2013 was wrong in the same direction.

Phase 3, the consumption economy (2027-2030). Agents turn inference from a product feature into a metered operating expense of ordinary business, closer to electricity or payroll than to software licensing. When token consumption is a board-level cost line, cost-per-token stops being a nice-to-have and becomes procurement's problem: workloads migrate toward whoever serves equivalent quality cheapest at enterprise grade. That is the moment the open-model platform layer either consolidates into two or three AWS-shaped winners, efficiency edge, workflow gravity, owned capacity, or dissolves into the hyperscalers' bundles. Everything in our subsector stance is about telling those two endings apart early.

The rejoinder to every commodity critique of this layer is a single company. In 2013, AWS did $3.1B of revenue renting a commodity, compute, at margins everyone called structurally thin, in a market everyone knew the incumbents would eventually dominate. It ended 2025 at $128.7B of revenue with ~35% operating margins, generating more operating income than the rest of Amazon combined in some years. The mechanism was precisely the one the platform archetype is attempting: win developers first with price and ergonomics, layer sticky services on the commodity, let enterprises follow the developers, and convert scale into both cost leadership and margin. The leading open-model platforms today, at roughly $1B of revenue, sit where AWS sat around 2012-13, with demand growing faster than cloud's ever did.

What is the winning open-model platform worth in 2030? We refuse to answer with a point estimate, but the arithmetic of three coherent scenarios is knowable, and it is the honest way to hold both our discipline and the upside in one frame. All three use the same skeleton: market size × independent-platform share × leader's share of that layer × margin-appropriate multiple. Every input is stated so you can attack it.