May 28, Tianjin. Lenovo's exhibit list at the 2026 World Intelligence Industry Expo is quite interesting: 1,000 agents running persistently, a cost of less than 1 yuan per million tokens, DeepSeek inference throughput doubled, and Geely Automobile's wind resistance simulation efficiency increased 28-fold.
This is not the kind of empty rhetoric common in press releases—phrases like “AI comprehensively empowers enterprise digital transformation.” Lenovo this time provided concrete details—what architecture it used, how much it spent, and what results it achieved. For people working in enterprise digitalization, these numbers are far more useful than descriptions such as “strategic layout.”
What is Lobster Lake talking about
Lenovo’s enterprise solution this time is branded “Lobster Lake,” essentially a privately deployed agent cluster solution that can support up to 1,000 “lobsters” (agents) online simultaneously.
What caught my attention most was not the volume, but the cost structure: the operating cost per million high-quality tokens is less than 1 RMB. A year ago, this figure was almost impossible, but now it is a price that can be promised at the product level.
Agent cluster scale: 1,000 private deployments, ongoing operating costs
Private deployment deserves a separate discussion. Over the past two years, most enterprises have used AI through cloud APIs and paid by usage. As they continued to use it, they found that bills were higher than expected, and even more concerning are data security and compliance issues—in industries such as export processing, semiconductors, and finance, a large amount of data simply cannot leave the internal network.
The positioning of Lobster Lake is to lower this threshold: private deployment, data kept within enterprise boundaries, and self-owned computing power. This logic runs counter to the practice of encouraging the use of public cloud in previous years, but it is more in line with the actual procurement preferences of Chinese manufacturing and leading enterprises.
Hybrid AI is not a new term
In his opening speech, Yang Yuanqing used the term “hybrid artificial intelligence,” a phrase Lenovo has been using for at least two years, but this time the wording was somewhat different. He said the essence of global economic competition is “a comprehensive contest of AI + real economy ecosystem capabilities,” and the keyword is “ecosystem”—not technology, not computing power, but ecosystem.
In the age of artificial intelligence, innovation is no longer a breakthrough in a single product or technology, but systemic innovation—technology, application scenarios, ecosystems, standards, and governance are all indispensable.
I think this judgment is correct. In the past few years, the failure rate of enterprise AI projects has been high, and one core reason is that they focused only on technology while ignoring scenarios and governance. Technology itself is actually no longer the biggest pitfall; the biggest pitfalls are the two questions of 'which scenario to start with' and 'who is responsible when problems arise'.
Lenovo also announced two specific investments in Tianjin: a general-purpose server production line will begin mass production in September 2026, and a next-generation AI computing product R&D and manufacturing center will begin mass production in autumn 2027, both located in the Tianjin Lenovo Smart Innovation Service Industrial Park. These two timelines are quite specific, not vague promises of "sometime in the future."
What did Qingtian Claw run out?
The Lenovo Wanquan Heterogeneous Intelligent Computing Platform 4.0 is the underlying computing power support for the Qingtian AI Platform, and two specific cases were showcased at this exhibition:
| Scenario | Result | Explanation |
|---|---|---|
| DeepSeek inference throughput | Double | Compared with the standard configuration, the processing capability is doubled under the same computing power. |
| Geely Automobile Wind Resistance Simulation | Efficiency increased by 28 times | HPC high-performance computing scenarios, typical manufacturing requirements |
The wind resistance simulation case is quite representative. In automotive development, CFD (computational fluid dynamics) simulation has always been a time-consuming process; a complete simulation may take several days. An efficiency improvement of 28 times means that work that originally required a week can now be completed within six hours. This is not just 'a little faster'; it changes the rhythm of the entire design iteration.
In the vertical scenarios of intelligent manufacturing, Lenovo also showcased three industry-specific intelligent agents: design assistance, intelligent quality inspection, and Magic Cube intelligent customer service, as well as a City Super Intelligent Agent (a 1+N architecture that has been deployed in multiple cities such as Shanghai and Chongqing).
How should companies view these matters
There are a few questions worth careful consideration.
First, whether privately deployed agent clusters are suitable for the organization itself. The Lobster Lake logic applies to enterprises with high data sensitivity and existing IT infrastructure, such as leading manufacturing enterprises, financial institutions, and state-owned enterprises. For small and medium-sized enterprises, maintenance costs and configuration complexity may instead become a burden.
Second, the premise behind the figure of less than 1 yuan per million tokens is that Lenovo provided self-test data; factors such as token usage, call frequency, and operations and maintenance costs in actual deployments will all affect the real TCO (total cost of ownership). This figure is a reference benchmark, not the ultimate goal.
Third, the reproducibility of results such as the 28x wind resistance simulation. Different industries have vastly different HPC requirements; the CFD needs of the automotive industry are not on the same scale as semiconductor EDA simulations. The range of industries in which Lenovo can reproduce this result is narrower than what was stated in the exhibition PPT.
However, none of this affects a basic judgment: the technical threshold for deploying enterprise-grade intelligent agents is rapidly decreasing, and private solutions are becoming increasingly viable in terms of price and capability. In 2026, the question will no longer be "Can it be done?" but "Who will lead it, and from which scenario should it start?"
For enterprises deploying ERP systems such as Odoo, the path to introducing intelligent agents is actually clearer than outsiders imagine—ERP itself is the most data-intensive system in the enterprise. From procurement assistants and inventory anomaly alerts to customer service automation, running privatized small models on the internal network precisely bypasses data security concerns.
