China Has Found Its Next Commodity to Export: The Token

The industrial playbook that commoditised steel, solar and electric vehicles is now rewriting the economics of artificial intelligence.

The Pattern of Chinese Industrial Commoditisation

There is a pattern that has repeated for twenty years. An industry is born in the West with high margins and barriers that look unassailable. China enters late, builds capacity at a scale nobody considered rational, and saturates its own market. Dozens of producers tear each other apart in a domestic price war, margins collapse — and the product then crosses the border at a price Western competitors cannot match. Not because they are less capable, but because their cost structure belongs to a different world.

Steel. Solar, where European manufacturers vanished inside one investment cycle. Batteries. Electric vehicles: per China Leadership Monitor, over 400 Chinese EV producers have ceased operations since 2018, and average sector margins fell from 7.8% in 2017 to 4.3% in 2024 — the two worst years on record. Brutal at home, devastating abroad.

In 2026 the product is the token — not a blockchain token, but the inference token: the unit by which a language model’s work is measured, billed and sold. Roughly three quarters of a word. The unit of account of the AI economy.

2026 AI Token Pricing: Chinese vs. Western Models

Price per million tokens, July 2026, input / output in USD.

Model

Company

Input

Output

Context

Qwen3-8B

Alibaba

0.01

0.01

—

Hy3 Preview

Tencent

0.063

—

—

DeepSeek V4 Flash

DeepSeek

0.14

0.28

1M

MiniMax M2.7

MiniMax

0.30

1.20

—

DeepSeek V4 Pro

DeepSeek

0.435

0.87

1M

Kimi K2.6

Moonshot AI

0.68

3.42

262K

GLM-5

Zhipu / Z.ai

—

3.20

200K

MiMo V2.5 Pro

Xiaomi

—

3.00 (flat)

1M

Qwen3 Max

Alibaba

—

3.90

262K

Against the Western frontier: Gemini 3.1 Pro at $2/$12, Claude Opus 4.8 at $5/$25, GPT-5.6 Sol at $5/$30, Claude Fable 5 at $10/$50.

DeepSeek V4 Pro costs roughly 34x less than GPT-5.6 Sol on output. V4 Flash undercuts OpenAI’s budget tier by about 21x.

The trajectory matters more than the level: Chinese labs cut API prices six times in the first half of 2026, three of those cuts declared permanent. Not launch promotions — structural repositioning. And one line item goes unnoticed but decides whether agent products work at all: Moonshot holds its cache-hit price at $0.07 per million, DeepSeek at $0.0036. When the same system prompt is resent thousands of times daily, that is where the economics live.

Market Share Shift: Chinese Open-Weight Model Adoption

Benchmarks are a narrative battlefield. Traffic flows are not.

OpenRouter is the largest neutral model router in the world, processing over 20 trillion tokens a week with no vendor lock-in. It is one of the few places you can see what developers actually choose when nobody is watching the brand.

• Per a Bloomberg chart built on OpenRouter and Exponential View data, US models — Google, OpenAI and Anthropic combined — fell from roughly 70% of token volume in June 2025 to roughly 30% in June 2026.

• A joint OpenRouter–a16z study covering 100+ trillion anonymised tokens puts Chinese open-weight share at 1.2% to nearly 30% within a year.

• DeepSeek is now the largest single provider by volume, at 16.3% — ahead of Google, Anthropic and OpenAI.

• The crossover came in the week of 9–15 February 2026, when Chinese systems processed more tokens than American ones for the first time. The ratio now runs about three to one.

The signal is not isolated. Vercel AI Gateway’s Production Index shows open-weight models rising from 11% in April 2026 to 29% in June. a16z’s Martin Casado estimates roughly 80% of startups on open-source stacks run Chinese models. Inside China, National Data Administration figures show daily token calls going from 100 billion in early 2024 to 140 trillion by March 2026.

Why Chinese LLM Tokens Cost So Little: Architecture & Energy

Architecture. The leading Chinese models are Mixture of Experts pushed to the extreme, activating a fraction of total parameters per inference — MiniMax M2.5 holds 229 billion parameters and activates 10 billion. Independent tests cited by The China Academy put DeepSeek V3’s inference cost at roughly 36x below GPT-4o’s. This is unit-cost engineering, not dumping.

Energy. Total electricity costs in China run about 40% below US levels. And here is the most interesting mechanism in the whole story: the cluster runs in China, the electricity comes from the Chinese grid, and the result returns to a user in Europe or America.

The electricity never left the grid, but its value was delivered abroad.

The token is a derivative of electricity. It is intangible, clears no customs, meets no tariff, appears in no trade statistic. China has coined a phrase for it: token 出海 — “tokens going overseas.” In March 2026 the National Data Administration formally standardised the Chinese word for token (词元, cí yuán), with its director describing tokens as measurable, priceable and tradable. When a regulator adopts the vocabulary of commodities, the positioning has already been decided.

Involution (内卷). The term Chinese authorities themselves use for aggressive price-cutting among domestic firms. The U.S.–China Economic and Security Review Commission defines it as subsidised industries expanding output despite insufficient demand, lowering profitability for everyone. Over a dozen labs share the same track; loss-leading is the norm. Behind it, Huawei is reportedly aiming to nearly double Ascend 910C production in 2026 to around 600,000 units with SMIC, while ByteDance has planned some $14 billion of compute spend including $5.7 billion in Ascend processors.

Commodity vs. Premium AI Markets: The Revenue Divergence

Here is where shallow analysis misses. Volume has moved. Revenue has not.

On OpenRouter, Anthropic holds around 12.3% of tokens processed but captures a far larger share of platform revenue — reported near 46% — because each premium token is worth many commodity tokens.

This is not a platform changing hands. It is two markets forming in parallel. The commodity lane — bulk classification, extraction, chat backends, retrieval augmentation — is work measured in billions of tokens where “good enough” wins and cost per token is the only variable that moves. Chinese open-weight models are running away with it. The premium lane — complex multi-file engineering, long-horizon agents, reasoning where being wrong is expensive — is where reliability, guarantees and predictability get paid for. A feature-sized engineering task computes to roughly $0.21 on DeepSeek V4 Pro against $3.80 on Claude Opus 4.8, and there are contexts where that gap is irrelevant next to the cost of failure.

BYD and Porsche do not compete on the same metric. BYD’s unit volume tells you nothing about Porsche’s P&L.

The open question is whether the premium lane stays wide enough. The capability gap is narrowing faster than incumbents would like.

Hidden Risks: Governance, Sustainability, and Data Sovereignty

Sustainability. Selling below cost to win share is a strategy, not an equilibrium. Solar and EV show what follows: consolidation, exits, stabilising prices. Stress-test the scenario where $0.28 per million triples.

Governance. The USCC bulletin notes distillation allegations against several Chinese labs. Merits aside, what matters for a regulated business is licence terms, version pinning and traceability — risk line items, not cost line items.

Sovereignty, on the upside. Open weights let a European organisation host GLM, Qwen or DeepSeek on-premise, keeping data inside the perimeter. Closed American models cannot match that argument, and it likely drives more European enterprise adoption than price does.

Leaderboards are not evaluations. A model tops OpenRouter because it is cheap, new, or used by one high-volume customer — not because it fits your case. Run your own eval.

The Shift in Global AI Token Supply

For two years the framing was a capability gap: US labs ahead, China trailing by six to eighteen months, export controls widening the lag. In 2026 that became the wrong question.

The gap that matters is no longer who trains the best model. It is who supplies the tokens the world actually consumes — and on that axis, the open layer has already been won. This is what commoditisation looks like: the frontier becomes a feature, while the cheap, good-enough tier eats the volume.

For decision-makers there is one operational consequence. “Which model” has stopped being a decision and become a routing policy, differing by task class and revisited quarterly. Anyone still sending everything to the most expensive model is making the most common and most expensive mistake in production AI.

And for anyone watching the macro picture: China has found a way to export deflation in a form that clears no customs, pays no tariff, and appears in no trade statistic.

Sources

Pricing: provider documentation (DeepSeek, Alibaba Cloud, Moonshot AI, Zhipu AI, Anthropic, OpenAI, Google Cloud), cross-checked against BenchLM, CostGoat and Apidog trackers.

Market Share: OpenRouter public data via Bloomberg / Exponential View; OpenRouter × Andreessen Horowitz study (a16z State of AI Report, 100T+ tokens); Vercel AI Gateway Production Index, July 2026.

Policy and Cost Structure: U.S.–China Economic and Security Review Commission (USCC); National Data Administration (March 2026); China Leadership Monitor on Economic Involution; The China Academy on inference and energy economics.

Note: OpenRouter share figures vary between 45% and 61% across coverage from April–June 2026 depending on measurement week and denominator; this article uses the conservative reading. Prices verified July 2026 and move quickly.