# walker-onlycfo-ai-tokenomics-2026-10-22

## Veille

Guest article by **Peter Walker**, Head of Insights at **OpenRouter**, published on **October 22, 2026** in the Substack newsletter **OnlyCFO** (approximately **1,500 words**, OpenRouter charts). The piece addresses chief financial officers and draws its figures from the traffic of the author's inference gateway.

## Titre Article

The Current State of AI Tokenomics | What CFOs Should Know

## Date

2026-10-22

## URL

https://www.onlycfo.io/p/ai-tokenomics

## Keywords

tokenomics, inference economics, CFO, chief financial officer, AI cost, price per token, token volume, OpenRouter, AI gateway, model marketplace, open-weight models, Chinese labs, inference providers, endpoints, AI agents, agentic traffic, prompt caching, cache hit rate, landed cost, retry cost, routing, automatic failover, availability, zero data retention, ZDR, commodity purchasing, usage-spend decoupling, guardrails, FinOps

## Authors

Peter Walker — Head of Insights d'OpenRouter ; article invité dans la newsletter OnlyCFO.

## Ton

Profile: a financial-explainer guest article, in English, written in first person plural ("we" = OpenRouter), addressed to CFOs and more broadly to "capital allocators." The introduction is signed by the newsletter's editor; the article is preceded by an unrelated sponsored insert (Remote).

Structure: an opening figure (volume versus price), a thesis (inference behaves like a commodity, not like software), four numbered forces, a section on pricing, one on "landed cost," one on non-price criteria, four recommendations.

Register: pedagogical and assertive, anchored by one figure per paragraph. The analogies are financial ("landed cost," "commodity input," "the canary in the cost coal mine"). Humor stays understated ("no comments on any or all upcoming IPOs!").

Evidence: exclusively OpenRouter platform data, with no detailed methodology or sample outside the platform. The routing section describes OpenRouter's default algorithm, and the availability example compares providers to OpenRouter.

## Pense-betes

- **Source**: guest article by Peter Walker (OpenRouter) in OnlyCFO, October 22, 2026; OpenRouter data.
- **Thesis**: companies buy AI like software (vendor, annual contract) whereas inference behaves like a **commodity**: falling unit costs, rising volume, best provider changing every few months.
- **Volume / price**: **0.5 → 164 T tokens/week** (Jan 2025 → late Sept 2026); average price **$0.77 → $0.14/M tokens**.
- **Supply**: **862** models launched since 2024, of which **362** in 2026 before October; models above 1 T tokens/week: **3 → 27** in six months.
- **Open-weight**: **~60%** of tokens, **26%** of spend; two-thirds on models from Chinese labs, of which **44%** served by US-based providers. Model × provider endpoints: **~330 → 1,342** in 2026. See [[sfeir-kimi-k3-moonshot-frontier-open-weights-2026-07-16]].
- **Agents**: overtake humans on **February 6, 2026**; by late September, **131 T** tokens/week versus **18 T**.
- Cache: agents resend the same long context at every step. Agentic traffic cache hit rate **88%** for this quarter's models versus **55%** in spring 2025; providers above 90% cache: **2/35 → 11/54**.
- **Leader**: changed **5 times in 8 quarters**; top lab's share **~40% → 21%**. Anthropic + OpenAI spend: **~80%** Anthropic in early 2026, near parity by September.
- **Decoupling**: AI-native startups tokens **×18**, spend **×5.1**; large organizations **×5.1** and **×2.4**.
- **Causes of the decline**: cheaper, more capable models (Kimi K3, DeepSeek V4 Flash); optimized caching; frontier labs' economic models (Luna, Haiku). Prices actually paid **43% to 89%** below list price for three open-weight models.
- **Landed cost**: effective price after cache discount + latency cost + retry cost. A $1.00 request that succeeds 80% of the time costs **$1.25** per successful response. **77%** of requests go to models served by multiple providers.
- **Beyond price**: ZDR endpoints for **93%** of tokens (18 of the top 20 models); **92.8%** success with the first provider, **96.9%** after failover; during a saturation event, **18%** for the 15 most affected providers.
- **Recommendations**: more tokens ≠ more spend; an open-weight model varies by provider; guardrails (ZDR, regional routing, per-function caps, fallback); expect change.
- ⚠️ **Scope**: data from a single intermediary, which sells the routing and failover it recommends; no published methodology in the article.

## RésuméDe400mots

On October 22, 2026, the OnlyCFO newsletter publishes a guest article by Peter Walker, Head of Insights at OpenRouter, a gateway that provides access to several hundred language models through a single account. The publisher frames AI cost as on track to become companies' largest expense line, and the topic as a matter for the finance function.

OpenRouter's data show a rare combination: the volume processed rises from half a trillion tokens per week in January 2025 to 164 trillion by late September 2026, while the average price paid falls from $0.77 to $0.14 per million tokens. The author concludes that inference behaves like a commodity, not like software purchased from a vendor under an annual contract.

Four forces explain this evolution. Supply is exploding: 862 models launched since 2024, of which 27 exceed one trillion tokens per week. Open-weight models carry about 60% of tokens for 26% of spend; two-thirds come from Chinese labs, but hosting is chosen separately from the model, and 44% of those tokens are served by US-based providers. Agents became the top source of tokens on February 6, 2026; because they resend the same context at every step, prompt caching becomes the decisive price variable, and the cache hit rate for recent models reaches 88%. Finally, the leader keeps changing: five times in eight quarters.

Usage and spend decouple: at AI-native startups, tokens multiply by 18 and spend by 5.1. The decline stems from cheaper, more capable models, cache optimization, and frontier labs' economic models. For the same open-weight model, the price actually paid sits between 43% and 89% below list price, and the cheapest provider on the price list is often not the cheapest in practice.

The author therefore proposes buying at "landed cost": the price after cache discount, plus the cost of latency and the cost of retries. A one-dollar request that fails one time in five costs $1.25 per successful response. Add to that zero data retention, available for 93% of tokens, and availability: 96.9% success after automatic failover.

He concludes with four pieces of advice for CFOs: consuming more does not mean spending more, an open model varies by provider, guardrails are essential, and plans must remain revisable.

## GrapheDeConnaissance

- Peter Walker —travaille_chez→ OpenRouter (ORGANISATION, 0.97)
- Peter Walker —publie→ article The Current State of AI Tokenomics (DOCUMENT, 0.98)
- OnlyCFO —publie→ article The Current State of AI Tokenomics (DOCUMENT, 0.95)
- article The Current State of AI Tokenomics —est_basé_sur→ données de trafic d'OpenRouter (AFFIRMATION, 0.95)
- Peter Walker —mesure→ volume OpenRouter passé de 0,5 à 164 000 milliards de tokens par semaine entre janvier 2025 et septembre 2026 (MESURE, 0.95)
- Peter Walker —mesure→ prix moyen payé passé de 0,77 $ à 0,14 $ par million de tokens (MESURE, 0.95)
- Peter Walker —affirme_que→ l'inférence se comporte comme une matière première et non comme un logiciel acheté sur engagement annuel (AFFIRMATION, 0.92)
- open-weights —mesure→ environ 60 % des tokens hebdomadaires pour 26 % de la dépense sur OpenRouter (MESURE, 0.92)
- Peter Walker —mesure→ les agents dépassent les humains comme source de tokens le 6 février 2026 ; 131 T contre 18 T par semaine fin septembre (MESURE, 0.92)
- Prompt caching —réduit→ coût par token du trafic agentique (CONCEPT, 0.9)
- Peter Walker —mesure→ taux de cache du trafic agentique de 88 % pour les modèles lancés ce trimestre contre 55 % au printemps 2025 (MESURE, 0.9)
- Peter Walker —mesure→ part du premier laboratoire en tokens tombée de 40 % à 21 % ; leader changé cinq fois en huit trimestres (MESURE, 0.9)
- Anthropic —concurrence→ OpenAI (ORGANISATION, 0.9)
- Peter Walker —mesure→ dépense cumulée Anthropic + OpenAI passée d'environ 80 % Anthropic début 2026 à une quasi-parité en septembre (MESURE, 0.88)
- Peter Walker —mesure→ startups IA-natives : tokens ×18 et dépense ×5,1 ; grandes organisations : ×5,1 et ×2,4 (MESURE, 0.9)
- Kimi K3 —concurrence→ Claude Opus (TECHNOLOGIE, 0.75)
- Peter Walker —recommande→ Landed cost (CONCEPT, 0.92)
- OpenRouter —utilise→ Landed cost (CONCEPT, 0.9)
- OpenRouter —améliore→ taux de succès des requêtes de 92,8 % à 96,9 % par bascule automatique (MESURE, 0.88)
- Peter Walker —recommande→ Zero Data Retention (CONCEPT, 0.88)
- Peter Walker —prédit→ les dépenses d'IA deviendront le premier poste de coût de la plupart des entreprises (AFFIRMATION, 0.85)
- Peter Walker —référence→ FinOps (METHODOLOGIE, 0.72)

---
Canonical: https://www.thekb.eu/en/fiches/walker-onlycfo-ai-tokenomics-2026-10-22/