01 Trigger Event
On November 10, Lenovo launched the “Wanquan Hybrid Token Factory” brand and partnered with the Zhangjiakou municipal government to build its first hybrid Token mother factory.
The initial plan includes a 15,000P heterogeneous compute cluster, an AI computing center of approximately 100MW, annual Token production exceeding 2 quadrillion, and peak output of 200 million Tokens per second. By 2030, the company expects the relevant core metrics to increase more than tenfold.
The critical development is not the larger “200 million Tokens per second” figure but a change in the unit of pricing: Lenovo is no longer selling only servers and data-center capacity. It is attempting to repackage infrastructure around AI output—Tokens.
Data from the National Data Administration show that China’s average daily Token consumption surpassed 500 trillion as of June 2026. Agents performing searches, analyses, decisions, and tool calls break a single interaction into more model calls. Compute demand is therefore expanding from training toward continuous inference.
One boundary must be established first: I have not operated this cluster inside Lenovo and cannot verify the methodology behind its “annual Token production,” model mix, or effective utilization. The figure of 2 quadrillion represents planned capacity, not necessarily saleable Tokens, and certainly not revenue.
02 The Real Significance
The issue is not whether Lenovo has built another AI computing center, but whether it can occupy a position closer to cash flow within the industry chain.
Over the past three years, the most certain AI profits came from selling picks and shovels: GPUs, servers, switches, data centers, and power. Customers make these purchases as one-time capital expenditures, and revenue is largely recognized after delivery. The risk shifts to the customer: being able to buy equipment does not mean being able to afford the models running on it.
The Token factory seeks to bridge that gap by packaging heterogeneous hardware, models, and customer workloads into continuously operating AI production services. Lenovo’s “three-layer hybrid power” model is straightforward:
- The bottom layer is the Wentian AI infrastructure, providing heterogeneous compute.
- The middle layer manages compute, schedules models, and produces Tokens.
- The top layer consists of the Qingtian and Baiying Agent solutions, connecting Tokens to business processes.
In essence, this is an aggregation strategy. Rather than merely controlling compute supply, Lenovo wants to aggregate model scheduling, inference services, and application distribution, becoming customers’ default procurement entry point for Tokens.
What will ultimately be priced is not bare compute, but the difference among unit Token cost, effective output rates, and business outcomes.
Compute rental is static revenue. Token operations generate operating revenue. The latter sits closer to profit, but it also requires continuous optimization: model routing, KV cache, batching, prefill/decode ratios, fault recovery, and capacity planning can all change unit economics.
This is why “15,000P of heterogeneous compute” establishes qualification to enter the market but not commercial quality. P measures compute scale rather than effective Token throughput. A peak of 200 million Tokens per second is not average utilization. Annual production of 2 quadrillion Tokens represents supply capability rather than customer demand.
Lenovo’s real bet is that selling servers yields only a one-time delivery profit, while controlling Token production and distribution creates an opportunity to earn recurring service fees in the AI era.
03 Historical Analogy / Structural Comparison
The model resembles the migration of the industry chain following the release of the iPhone in 2007.
Hardware assembly, displays, and supply-chain production were initially the most profitable segments. Value subsequently shifted toward operating systems, app stores, and developer ecosystems. Hardware did not disappear, but bargaining power gradually moved toward platforms controlling distribution and access to usage.
Cloud computing followed a similar path. AWS initially sold EC2 instances and storage before expanding its revenue boundaries through S3, managed databases, and usage-based PaaS. Infrastructure providers eventually sold not only resources, but also the scheduling systems that enabled those resources to be consumed reliably.
Lenovo’s Token factory can be viewed as an extension of the IDC into the “New Cloud”: land, power, servers, and networks remain at the bottom, while software scheduling transforms heterogeneous resources into AI capabilities produced on demand.
The analogy cannot be applied directly, however. The iPhone in 2007 established a new interaction entry point, while the entry point for Agent workloads remains unsettled. Models, protocols, billing systems, and developer tools are still migrating, and customers are likely to retain multicloud and multimodel architectures. Lenovo has an opportunity to become an aggregation layer, but not necessarily the sole entry point.
The gap with AWS in 2014 is equally clear. AWS had years of accumulated cloud control-plane capabilities, developer habits, and a global capacity network. Lenovo’s mother factory is only beginning, while its brand, developer distribution, and stable customer workloads remain unproven.
Accordingly, this is not “rebuilding AWS.” It is a server vendor attempting to cross into an operating layer with lower margins and more intense competition.
04 What This Means for AI Builders
There is no need to migrate architectures immediately simply because of the “Token factory” name. This month, however, the economics of inference should be recalculated.
First, decompose Token prices into procurement costs and effective task costs. A low-priced model may become more expensive after longer chain-of-thought, repeated context, and failed retries. The relevant measure is not billable Tokens, but how many Tokens, units of waiting time, and tool calls are required to complete one successful task.
Second, evaluate whether building model routing in-house is worthwhile. The value of provider aggregation lies in dynamic routing across differences in price, latency, capacity, and quality rather than sending every request to the cheapest model. Poor routing can make the abstraction layer an additional source of failures, a concern that may receive insufficient attention when each business has distinct stability requirements.
Third, reconsider KV cache and prompt caching. If a Token factory is to generate gross profit, it must reduce marginal costs through prefix reuse, continuous batching, and cache locality. For builders, these are more important provider-selection questions than context window: can long context be reused repeatedly at an acceptable cost?
Fourth, preserve model portability in advance. Do not bind customer data, workflows, and evaluation systems to a single Token factory. Interface compatibility can reduce migration costs, but model behavior, concurrency limits, billing definitions, and SLAs still require independent validation.
Fifth, place Token costs and business metrics on the same table for Agent products. If the core metrics are completed work orders, qualified code generated, or revenue recovered, gross margin cannot be discussed separately from task success rates. For AI startups, this matters far more than promoting an annual production capacity of several quadrillion Tokens.
05 Counterargument / Risks
The assessment above may be too optimistic: Lenovo may not be able to convert its hardware advantages into an advantage in Token operations.
First, capital expenditures may become depreciation pressure faster than expected. A 100MW AI computing center requires continuous power, cooling, networking, and equipment maintenance. When demand falls short, idle compute is harder to operate than a conventional IDC. Peak Token throughput, in particular, can conceal long-term average utilization.
Second, model suppliers hold pricing power. Anthropic, OpenAI, and Google provide both models and their own infrastructure and Agent products. As an aggregator, Lenovo may gain access to demand, but upstream providers may also compress its margins. There is no clear evidence that Lenovo possesses differentiated model capabilities; it is more likely to compete through cost, delivery, and localized services.
Third, Token is not a stable unit of value. Improvements in model efficiency will reduce the Token demand required for the same task, while Agent frameworks will reduce redundant calls. If Tokens per unit task continue to decline, the “Token factory” may grow while the price per Token falls even faster. This mirrors cloud storage becoming progressively cheaper while more valuable computing and data services benefit, producing the same structural pattern.
Finally, the upper-layer Agent solutions may be only a transitional measure. The real control point may lie not in racks or scheduling layers, but in enterprise workflows, the MCP ecosystem, developer tooling, and customer distribution. Lenovo’s ability to produce Tokens does not mean customers will continue buying them from Lenovo.
Consequently, Lenovo’s launch deserves attention, but that does not mean a “New Cloud” has already been established. I would reduce the criteria for success to three metrics: the share of paid Tokens from unrelated customers, average cluster utilization over several consecutive quarters, and contribution margin per Token.
If those numbers fail to materialize, the Wanquan Hybrid Token Factory may remain a heavier version of the server business rather than a new infrastructure platform.