type
Post
status
Published
date
Nov 15, 2026
slug
summary
tags
category
行业动态
icon
password

Lenovo's Token Factory: Converting Green Energy into AI Efficiency

As artificial intelligence integrates deeper into enterprise operations, a paradox has emerged: companies can afford AI hardware but struggle to afford its operational consumption. According to the National Data Bureau, China's daily token consumption surged from 10 billion in early 2024 to 100 trillion by the end of 2025, breaking the 140 trillion mark in March 2026. This exponential growth, driven by complex agent tasks and model evolution, has forced businesses to re-evaluate their spending. For instance, a mid-sized bank reported spending nearly 9 million RMB on token consumption in the first half of the year, accounting for 8% of its entire annual IT budget. Consequently, organizations are becoming more cautious about "burning tokens," seeking ways to align computational supply with actual business needs.

The Mixed Token Factory in Zhangjiakou

In response to this challenge, Lenovo announced the "Lenovo Wànquán Mixed Token Factory" in Zhangjiakou, Hebei, on November 10. The concept is straightforward: green energy goes in, tokens come out. This facility is not just a server farm but an integrated system designed to produce economic and efficient token services by organizing diverse computing resources and models. The project represents a shift from brute-force chip stacking to an industrialized model of high-value token production.
The first phase of the project involves a 15,000P heterogeneous computing cluster and an approximately 100MW AI data center. Once operational, the facility aims to produce 200 trillion tokens annually, with a peak output capability of 200 million tokens per second. The initial construction is scheduled for completion in 2027, with plans to expand the scale by more than 10 times by 2030. Zhangjiakou was selected for its unique advantages: it is China's leading city for wind and solar green energy, with renewable energy capacity exceeding 45 million kilowatts as of the end of 2025. Additionally, as a key hub for the "East Data West Computing" project, it offers proximity to Beijing and supports over 1.07 million operational standard racks across more than 50 data centers.

Hybrid Supply and Intelligent Scheduling

The core innovation of this facility is its mixed supply model. Unlike traditional approaches that focus solely on scaling hardware, Lenovo’s system dynamically matches computing power to business demands. This hybrid approach combines three dimensions: supply source mixing, where private self-generated tokens complement authorized public services; device-edge-cloud collaboration, where lightweight tasks run locally while elastic demands are handled by the cloud; and model ecosystem mixing, selecting between open-weight and proprietary models based on cost and quality requirements.
To manage this complexity, Lenovo deployed the Token Engine, which uses dynamic hybrid deployment and intelligent model routing to optimize resource allocation. This is managed centrally through the xCloud Lenovo Smart Cloud, which unifies private and public cloud resources. Within this framework, the Token Hub serves as the core for AI inference, dynamically adjusting resources based on task needs, model quality, data boundaries, latency, and cost. This allows for precise scheduling down to the chip and computing node level, ensuring that lightweight models run on smaller nodes without occupying large clusters, thereby maximizing resource utilization.

Three-Layer Optimization for Cost and Security

To ensure efficiency without compromising security, the system employs a three-level optimization strategy. First, it enhances hardware efficiency through Lenovo’s WenTian heterogeneous AI computing solution, which combines super-nodes and multi-chip engines to improve computational efficiency by more than 30%. Second, it controls the total cost of ownership of the data center using modular prefabrication, integrated liquid cooling, and green power direct current control, targeting a PUE (Power Usage Effectiveness) of less than 1.1. Third, it optimizes token efficacy for business scenarios, ensuring that simple tasks do not consume high-cost models.
Security is integrated into this architecture through three barriers: model security via a whitelisting certification process; content security through the Token Trust capability that intercepts sensitive information; and privacy computing using Trusted Execution Environments (TEE) to ensure data is "usable but invisible" to the platform. By internalizing these practices, Lenovo has achieved a 70% reduction in comprehensive costs for its own AI operations.

Defining "Effective Tokens"

The ultimate goal of the Mixed Token Factory is not just lower costs, but the creation of effective tokens. Lenovo emphasizes that raw token throughput is meaningless if the AI fails to complete tasks accurately. A system that requires repeated inferences and modifications to get the right result is less valuable than one that delivers qualified outcomes in a single pass. By focusing on task completion precision, green sustainability, and end-to-end security, the facility aims to transform AI from a cost center into a reliable industrial input. The stated targets include optimizing the cost per one million tokens by 70%, increasing token output per second by 5 times, and maintaining a token yield rate of over 99.9%.
出处:量子位,https://www.qbitai.com/2026/11/510094.html
AI正在加速人类文明进程程序员沉迷折纸飞机:最长滞空31.2秒,一个网站火了
Loading...