Howard Yu's Post — American AI models went from 70% of usage to 30% in 12 months

Howard Yu · article · 2026-07-11 source ↗ Reader 🖥️

Agentic AI AI Unit Economics Frontier Model Convergence / Commoditization Moats & Defensibility Open-Weight Models Technology Diffusion & Transformation Curves Worse Is Better

The AI-model market is flipping from US to Chinese dominance in real production usage, and the decisive driver is per-task cost — decisive precisely because the agentic era multiplies calls per task, so distribution consolidates on the cheapest 'good-enough' models faster than any residual US capability lead can matter.

theoretical (expository market analysis) with a practical undertow · market-shift analysis / opinion essay (short LinkedIn post) — On its face descriptive — it explains a market reversal — but it carries an implicit practical directive: stop paying the US 'best model' premium, and assume that in the agentic era cost, not benchmark rank, decides which models get built on. The P.S.S. explicitly anchors the piece in a disruption framework ('good enough and cheap' beats 'best and gated').

Structure

Part	Role
Opening claim + evidence (paragraphs 1-3)	establishes the reversal
Cause (paragraph 4)	assigns the mechanism
Corroboration (paragraph 5)	third-party and adversary testimony
Positioning contrast + closing aphorism (paragraph 6)	dramatizes and generalizes
P.S. / P.S.S. (framework anchor)	hedges a specific and grounds the argument in a theory

Key terms

distribution
installed real-usage share across the developer ecosystem — who actually gets called in production — not marketing reach or channel access
routed volume
token traffic actually dispatched to a model through a marketplace/gateway (OpenRouter) — real production calls, as opposed to downloads, benchmark runs, or press
open-weight Open-Weight Models
models whose weights are downloadable and self-hostable (Qwen, DeepSeek, Zhipu GLM) — deployable on domestic chips and freely forkable
for equivalent work
normalizing cost to a common task/eval so a cross-family price ratio (the ~9x premium) is meaningful

Propositions

explicitThe AI-model market has reversed in twelve months: American models fell from ~70% to ~30% of usage while Chinese models climbed from ~1% (2024) to more than 60% by May, and by June weekly routed volume was ~18 trillion tokens Chinese vs ~5.5 trillion American on OpenRouter.conf: medium
supportsSimon Wardley's Post — Industrialisation of ML: nine notes on what to expect · p12 — Reaches Wardley's China-ascendancy conclusion by a different mechanism (cost/distribution here vs. state supply-chain mastery there).
explicitThe cause of the flip is price: a standard evaluation suite costs ~$4,811 on Claude, $3,357 OpenAI, $1,071 DeepSeek, $544 Zhipu GLM — the premium American model runs nearly 9x the cheapest Chinese one for equivalent work.conf: medium
supportsDeepSeek v4, and the end of the OpenAI/Microsoft AGI clause · p2 — Corroborates the price premise: DeepSeek-V4 is the cheapest across the market (Flash cheapest small, Pro cheapest of the larger frontier models).
supportsThe Pulse: a new trend, smart model routing · p1 — The ~9x US-over-Chinese premium is a concrete instance of the documented 10-20x cross-model price spread.
explicitThe chat-to-agents shift made the price gap decisive because an agentic task calls a model thousands of times (an overnight coding run), so per-call cost differences multiply into architecture-deciding sums.conf: high
supportsBaseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models · p9 — The agentic multiplier is the Jevons mechanism — cheaper inference pulls in far more calls, which is exactly why per-call price becomes decisive.
explicitcontestedDistribution, once it sets, is sticky in a way benchmarks never are — so an established real-usage lead now matters more than capability leadership.conf: low
contradicts ←Andreessen Horowitz's Post · p6 — This post claims developers are 'fair-weather' and chase whichever model is momentarily best, implying low switching costs and no durable lead; the other source claims the opposite — that once real-usage distribution sets for a model/provider, it is sticky in a way benchmark leadership never is.
refines ←You Outsourced the AI—but You Still Own the Risk · p6 — Refines that piece's 'distribution is sticky' claim by locating enterprise AI stickiness in the integration layer (prompts, workflows, safety filters and evals calibrated to one provider) rather than the router — where a model-string swap is cheap. So capturing the cost advantage of cheap commoditized models requires deliberately engineering portability, which doubles as a liability hedge.
explicitThe models have reached rough capability parity: Anthropic itself conceded in a May policy paper that US models are only several months ahead of Chinese ones and that Beijing is winning global adoption on cost — so competition has moved from capability to price and distribution.conf: high
supportsDeepSeek v4, and the end of the OpenAI/Microsoft AGI clause · p4 — Independent confirmation the capability gap is months, not years: Anthropic's 'several months' ahead ~= DeepSeek trailing frontier by ~3-6 months.
supportsThe Pulse: a new trend, smart model routing · p10 — Capability parity moves 'which model to use' off benchmarks and toward a commoditized, cost-driven choice.
explicitOpen-weight Chinese models now anchor the open-source ecosystem: a16z's Martin Casado estimates ~80% of young AI companies building on open-source stacks reach for Chinese models, and Alibaba's Qwen has become the most-forked open model on earth.conf: medium
supportsBaseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models · p10 — US firms adopting Chinese open models: Baseten frames Chinese open-source subsidies as an effective subsidy to US enterprises — the same adoption surge Casado quantifies.
supportsThe Pulse: a new trend, smart model routing · p5 — Open models sufficient for ~60% of coding work with usage strictly increasing corroborates the open-weight adoption surge.
explicitcorroboratedUS incumbents are moving the opposite way from the market: OpenAI launched its strongest model, Sol, to only 20 government-approved partners, and Meta — which defined open weights with Llama — has fallen below 1% of routed volume, i.e. gating and closing exactly as the market rewards cheap-and-open distribution.conf: medium
extends ←How the Escalating AI Wars Benefit You · p1 — This piece's 'ecosystem control' frame broadens the prior doc's specific observation (US labs gating/closing exactly as the market rewards cheap-and-open distribution) into a wider multi-front picture spanning litigation, chip export policy, and pricing, not just open-vs-closed model strategy.
supports ←Is Open-Source AI Really the Dangerous Path? · p1 — Confirms from the strategy side what the usage data in the other extraction shows: US incumbents gating and closing access (OpenAI's 20-partner release, Meta's retreat from open weights) enacts the 'control' bet this article attributes to Washington, mirrored by Chinese labs' open releases enacting the 'diffusion' bet.
explicitdevelopedThe underlying regularity is a disruption pattern: 'good enough and cheap' beats 'best and gated' with predictable regularity (the framework the author cites from his Intel/Nvidia/DeepSeek piece).conf: medium
extends ←AI and the Looming Competition for Margin · p4 — The observed 'good enough and cheap beats best and gated' disruption pattern among AI model makers is a specific instance of this article's general claim that widely diffused technology cannot sustain price/moat advantages.
extends ←Ordinary Engineers, Not Heroic Inventors · p3 — Diffusion theory provides theoretical grounding for the empirically observed pattern that cheap-and-distributed beats best-and-gated.

Arguments

The AI-model market is flipping from US to Chinese dominance in real production usage, and per-task cost is the decisive driver. strength: medium

Explicit premises:

Usage share on OpenRouter reversed in 12 months (US 70%->30%; Chinese ~1%->60%+; ~18T vs ~5.5T weekly tokens).
Chinese models cost a fraction (~1/9) of the premium US model for equivalent work.
The chat-to-agents shift multiplies calls per task, so per-call price now dominates spend.
Anthropic conceded US models are only several months ahead and that Beijing is winning adoption on cost.
~80% of young open-source-stack AI companies (per a16z) now reach for Chinese models.

Implicit premises:

OpenRouter routed volume is a valid proxy for the whole AI-model market.
Cost dominates capability once models are 'good enough'.
Distribution, once set, is sticky in a way benchmarks never are.

Inferences:

If cheap models are good enough and agents multiply call volume, buyers route to the cheapest -> usage share follows price.
Once usage share consolidates it locks in via distribution stickiness, so it outruns any residual US capability lead.

US incumbents are strategically mispositioned for this shift — gating capability and abandoning open weights exactly when the market rewards cheap-and-open distribution. strength: low