-->

OpenRouter is the platform where millions of developers pick and call multiple LLMs through a single API. Because it reflects real usage by cost-sensitive developers, it gets cited often as a leading indicator for the broader market. On that platform, US model token share dropped from roughly 70% to roughly 30% in a single year.

Conceptual image showing token flows being redistributed across multiple model nodes

This post verifies that data, surfaces the second layer the headline tends to obscure, and draws out what it means for ThakiCloud and Paxis strategy.

What Happened

Start with the numbers. US model token share (OpenAI, Anthropic, Google) was around 70% in mid-2025. By mid-2026 it had fallen to about 30%. Over the same period, Chinese models (DeepSeek, Qwen, MiniMax, Moonshot, Tencent, and others) climbed from under 2% a year earlier to roughly 46%. The inflection point was the week of February 9-15, 2026, when Chinese models processed 4.12 trillion tokens against 2.94 trillion for US models – the first time they crossed.

Different analysts land at slightly different digits. The analyst whose work spread this data most widely counted US models at 72% dropping to 33%, with Chinese models at 47%. A more precise count that separates unidentified traffic puts Chinese models at about 46% and US models at about 36%. Either way, the direction is the same. The one thing to watch is that comparisons like “US 33% vs China 47%” obscure the unidentified bucket, so treat the exact numbers as directional rather than definitive.

Why It Happened So Fast

The driver is a sharp shift toward open-weight releases from leading Chinese labs. DeepSeek released R1 and V3 at effectively no cost and achieved near-top reasoning quality. Alibaba’s Qwen series delivered strong results on multilingual and coding tasks. Both families carry MIT or Apache licenses, which made commercial adoption straightforward and pulled developer uptake up fast. Qwen now leads Hugging Face download charts ahead of Llama.

There is also an argument that Nvidia export controls – the restrictions on H100, H200, and B200 sales to China – created a paradoxical incentive. Constrained on compute, Chinese labs had stronger motivation to squeeze more performance from fewer resources. Add the fact that a large portion of OpenRouter’s user base is cost-sensitive startups and individual developers, and a structural tilt toward better price-to-license ratios becomes predictable.

But Tokens Are Not Revenue

The headline reading is “China beat the US.” One layer deeper, a different picture appears. On the same OpenRouter platform, one analysis found that Anthropic holds about 12% of token share but captures roughly 46% of revenue. That is a signal of market bifurcation.

One segment is a commodity market where the cheapest model wins. High token volumes flow through it, but margins are thin. The other is a premium segment – coding, autonomous agents, tasks where failure is expensive – where the model that actually performs commands revenue even at higher prices. Token share reflects the first segment. Revenue share reflects the second. They are not the same metric.

There is one more thing to add. OpenRouter’s user base skews heavily toward cost-sensitive developers, which means it does not represent the full enterprise market. Reading the share reversal directly as “US defeat” means leaving out both of those points and jumping to a conclusion that overstates what the data shows. The real event is not a win or a loss – it is a market splitting into two tracks: tokens and revenue.

The ThakiCloud and Paxis Angle

That split is exactly where ThakiCloud and Paxis have a favorable position. Two points.

First, model neutrality. In a market where commodity and premium tracks diverge and model rankings flip every quarter, the most resilient structure is one that is not locked to any single vendor. Paxis aims at model-neutral routing – letting customers choose their own price-performance tradeoffs the way OpenRouter does, but inside an enterprise-grade layer. The rise of Chinese open-weight models is not a threat to this strategy; it reinforces it. Whatever model rises to the top can be treated as a first-class citizen, which means market volatility becomes opportunity rather than disruption.

Second, the compliance layer. When enterprises start running DeepSeek or Qwen families to control costs, the questions that follow immediately are about commercial license terms and data governance. ThakiCloud’s Keycloak-based multi-tenancy and ArgoCD GitOps pipeline are technically well-suited to support diverse model backends. But to be honest, the layer that automatically validates per-model commercial licenses and per-customer data compliance is still homework we need to build. That is both a gap and the clearest opportunity in front of us. The provider that combines a first-class inference pipeline for Chinese open-weight models with license and data validation tooling will be the one that wins regulated-industry customers.

Where This Logic Weakens

For balance, here are the scenarios where this reasoning breaks down.

First, enterprise purchasing decisions can move differently from developer usage patterns. There is no guarantee the token share curve translates directly into enterprise adoption. Second, if data risk concerns around Chinese models block uptake in heavily regulated industries like finance and government, the share curve may fragment by sector. Third, the unidentified traffic bucket is large enough that precise numbers vary meaningfully across sources.

Three indicators are worth watching: whether the revenue share curve follows the token share curve in the same direction; whether Chinese model adoption actually rises in regulated industries; and whether the same trend appears in enterprise AI gateways outside OpenRouter.

Summary

The share reversal on OpenRouter is real. But it does not mean the US lost. It is one facet of a market restructuring in which tokens and revenue are pulling apart. The winners will be those who are not dependent on whichever model rises next – and who can provide the validation layer that makes running that model legal and safe. That is the position ThakiCloud and Paxis are building toward.

Related reading: The Real Logic Behind Big Tech’s GPU Overinvestment: Asymmetric Insurance and the Next Generation Toll Gates

Sources

참고

빅테크 GPU 과투자의 진짜 논리: 비대칭 보험과 다음 세대 톨게이트

 

2026년 하이퍼스케일러 캐펙스가 약 7,250억 달러로 전년 대비 77% 늘었습니다. 거품처럼 보이는 이 지출을 비대칭 보험과 의도 라우터 톨게이트라는 두 구조적 논리로 읽고, METR의 태스크 호라이즌 데이터와 신뢰도 임계선 수학으로 검증한 뒤, 톨게이트에 종속되기 싫은 엔터프...