# 中国"Token 工厂"：AI 推理的新型基础设施

- 来源：X.PIN (@thexpin)
- 发布时间：2026-08-07 23:06
- AIHOT 分数：57
- AIHOT 链接：https://aihot.virxact.com/items/cmsj32yw3238eronk36q2vud6
- 原文链接：https://x.com/thexpin/status/2085744351456735676

## AI 摘要

中国无锡等地正兴起“Token 工厂”这一新产业，不训练模型或造芯片，而是将算力转化为按 token 计量的推理能力，像公用事业一样交付。2025 年中国大模型采购合同超 7,500 份、金额近 300 亿元，需求集中在部署、安全、可靠性与可控性，而非模型智能度。与英伟达“AI 工厂”概念不同，中国已将其发展为独立行业。

## 正文

http://x.com/i/article/2085666288148623360

Token Factories: This Is How China Built the Missing Layer of AI

As foundation models become commodities, a new Chinese industry, the "Token Factory", is burgeoning to turn them into usable intelligence.

A Token Factory doesn't sell compute capacity. It sells inference capacity: manufactured intelligence, priced and delivered like a utility.

In Wuxi, China, a new type of AI infrastructure facility has appeared. It does not train frontier models. It does not manufacture chips. It manufactures something else: tokens.

That may sound like a strange distinction. After all, most of the AI race today is still measured by a familiar scoreboard: which model is smarter, which benchmark it wins, which company reaches the next generation of reasoning capability.

But inside the enterprises actually trying to deploy AI, the questions are becoming much less glamorous.

What happens if the network goes down?

Who catches it when the model is wrong?

Who is responsible when an AI system leaks information or makes a decision that cannot be explained?

China's enterprise AI market is already moving beyond the model race. In 2025, industry tracking data from China Large Model Procurement Monitoring Report 2025, recorded more than 7,500 large-model-related procurement contracts, with disclosed contract values approaching 30 billion yuan. But the surprising part was what those contracts actually demanded: they rarely focused on which model was smartest. Instead, they focused on deployment, security, reliability, and control.

Read enough of these contracts and a pattern emerges. The recurring requirements are central deployment, edge execution, offline caching, resume after disconnect, tiered permissions, and audit trails. Acceptance criteria alone can run for dozens of pages.

These are not the questions people ask when they open ChatGPT. They are the questions organizations ask when AI moves from a demo into a production system.

That gap - between whether a model can answer well and whether an organization can actually run it - is where a new layer of the AI industry is emerging.

China is not the only place where this layer could exist.

NVIDIA CEO Jensen Huang has described the next generation of AI infrastructure as "AI factories" systems that transform compute and electricity into tokens. In the United States, however, this capability is largely absorbed inside vertically integrated platforms such as AWS, Azure, Google Cloud, and the major AI labs.

What makes China different is not that it invented the concept, but that it is among the first markets where this idea has begun to evolve into an independent industry with local governments, infrastructure providers, and startups building what China calls "Token Factories."

What a Token Factory Actually Is

Three misconceptions are worth clearing up first, because each one understates what's being built.

It is not an API relay. A basic API proxy just forwards a request to a model provider's endpoint; it doesn't touch the underlying efficiency of inference. A Token Factory does something structurally different: it runs its own inference stack, splitting the process into stages that can be optimized independently, and pooling chips from multiple vendors under a single scheduler. The goal is one number: how many tokens this hardware can produce per dollar, per watt, per second.

It is not simple compute rental. A cloud sells compute capacity: GPU hours, a fixed unit regardless of what you do with it. A Token Factory sells inference capacity instead, the output of that compute, metered in tokens. Its margin depends entirely on how efficiently it converts hardware and electricity into that output. This is a genuinely different business, where the core metric isn't how many GPUs a company owns but how many tokens each GPU can produce, and at what quality. Push that efficiency up, and the same hardware generates more revenue without buying a single additional chip.

It is not just a data center. A useful way to see the stack: at the bottom sit GPUs, electricity, and networking. In the middle sits the software layer that actually does the work: inference frameworks, model scheduling, cache optimization, cross-chip compatibility. On top sits the delivery layer: APIs, agents, industry-specific applications that enterprises actually touch. A data center is the bottom layer alone. A Token Factory is all three, sold as one product.

Put simply: the product isn't compute, and it isn't an interface. It's manufactured intelligence, priced and delivered like a utility.

Why This Layer Is Emerging in China

No single explanation carries the whole story. Five conditions are converging at the same time, and it's the combination, not any one of them, that matters.

1. Open-weight models supply the raw material.

DeepSeek and Alibaba's Qwen release full-size open weights. Moonshot's Kimi and Zhipu's GLM have released open weights too, though open-sourcing plays a smaller role in Moonshot's own business than it does for DeepSeek. The exact mix varies company to company, but the pattern holds: several competitive Chinese models are simultaneously available for any team to download, deploy, and optimize, something that doesn't exist among the closed U.S. frontier of OpenAI, Anthropic, and Google. If a model's weights are public, the model builder trains, and someone else can manufacture and distribute.

Meta's Llama is the obvious counterexample, and it's worth addressing directly. An open U.S. model exists, but it has not produced the same independent inference ecosystem that is emerging around China's model ecosystem. Most of the serious inference infrastructure built around U.S. models sits inside AWS Bedrock, Azure OpenAI Service, and Google Vertex AI, each of which already does real multi-model inference and routing across providers like Anthropic, Meta, and Mistral. That's genuine infrastructure, just not an independent layer: an integrated version of the same function, embedded inside a hyperscaler ecosystem rather than sold as a public utility that works across any model or chip.

Unlike the U.S. frontier AI ecosystem, where companies like OpenAI are tightly integrated with hyperscale cloud platforms such as Microsoft Azure, China's AI landscape is more fragmented. Some major technology companies, including Alibaba, ByteDance, Tencent, and Baidu, operate large cloud platforms of their own, but no single AI stack dominates the market from frontier models to global-scale cloud infrastructure.

That fragmentation, combined with open-weight models, creates room for a coordination layer to emerge.

2. A fragmented ecosystem needs someone to hold it together.

China's AI stack is fragmented in a way the U.S. stack isn't. On the model side: DeepSeek, Qwen, Kimi, GLM, and more, each with different architectures, strengths, and pricing. On the chip side: a shrinking pool of legacy Nvidia hardware as export controls tighten, alongside a fast-growing lineup of domestic accelerators such as Huawei Ascend, Cambricon, and Moore Threads, each running its own software stack, none of them interchangeable without real engineering work. On the infrastructure side: local data centers, telecom operators, and third-party clouds, each built for a different region and a different regulatory environment.

None of these pieces talk to each other natively. Someone has to translate a request into which model, on which chip, deployed where, and return a clean token stream on the other end. That's the role SiliconFlow founder Jeffrey Yuan (Yuan Jinhui) has described his company playing: connecting models, chips, and applications. SiliconFlow has built exactly that. Its self-developed inference engine has adapted to more than 150 models, serving close to 10 million users and over 10,000 enterprise customers. Third-party data puts it among the top four providers in China's public-cloud model-serving market, the only startup in that group, and on OpenRouter, an independent U.S. model-routing platform. It has emerged as one of the most-used Chinese model-serving platforms, frequently ranking among the top providers by daily token volume. A smaller, more technical example is Vectron, an optimization platform built by the Chinese startup Shishi Technology, which advertises compatibility with more than ten domestic chip types and twenty mainstream models: a narrower operator making the same underlying bet, that the money is in squeezing more usable tokens out of the same hardware, not in owning more of it.

3. Enterprise demand requires the model to leave the cloud.

Ask any enterprise buyer what stopped their AI rollout, and the answer is rarely that the model wasn't smart enough. It's usually that the data can't leave the building. This isn't unique to China. Samsung banned external generative AI tools company-wide in 2023 after employees pasted proprietary source code into ChatGPT, and similar bans have rippled through banks, chipmakers, and law firms worldwide since. What started as an emergency reaction has hardened into standard compliance practice across those industries. In China's government and enterprise market, this requirement has its own name: data never leaves the domain.

That constraint reframes the entire buying decision. It's no longer which API is cheapest per million tokens. It's who can install this behind our firewall, with permission systems that track which employee can see which document, with audit logs that hold up if something goes wrong, and with a service contract that says who's accountable when it doesn't. That's a systems-integration problem, not a subscription. A Token Factory is built to sell exactly that: not intelligence as a feature, but intelligence with a chain of custody attached.

Read Full Article