Lenovo AI PC Sales Cross 50% in China, but Software Must Prove the Shift
Lenovo says AI PCs now represent more than 50% of its notebook sales in China, up from roughly one-third several quarters earlier. The Lenovo AI PC claim marks a major change in product mix, yet it does not prove that buyers are purchasing machines for their AI features.
Liu Jun, Lenovo’s executive vice president and China president, also said global Lenovo AI PC sales grew 86% year over year. His reported remarks connect that growth to Tianxi AI, Lenovo’s personal agent for devices sold in China.
The real contest is no longer Lenovo against another PC manufacturer. It is Lenovo’s agent-led computing model against the familiar hardware replacement cycle. Microsoft, Intel, AMD, Qualcomm, HP, Dell, and other vendors are putting neural processors into new computers, so capable hardware is becoming standard.
Lenovo wants to differentiate itself through software that understands personal context, works across devices, and executes multi-step tasks. That promise will depend on local model performance, developer participation, privacy controls, and whether people keep using Tianxi after setting up a new machine.
What Lenovo’s 50% AI PC Figure Actually Changes
The reported milestone shows that AI hardware has entered Lenovo’s Chinese mainstream, but its meaning depends on the denominator.
Liu’s figure refers to AI PCs as a share of Lenovo’s notebook sales in China. It does not mean Lenovo controls more than half of China’s entire notebook market. It also does not establish that more than half of all notebooks sold by every brand qualify as AI PCs.
That distinction matters because vendors use different AI PC definitions. Lenovo’s financial reporting defines AI devices as PCs and smartphones equipped with a neural processing unit, or NPU. An NPU is a processor designed to perform machine learning calculations with less energy than a general-purpose CPU.
Microsoft applies a narrower standard to Copilot+ PCs. Those machines require an NPU capable of more than 40 trillion operations per second, alongside other hardware and software requirements. A computer can therefore appear within a vendor’s AI PC category without supporting every Copilot+ feature.
Lenovo has not published a model-level breakdown behind Liu’s latest China percentage. The company has not disclosed how much of the reported growth came from consumer notebooks, business fleets, or channel inventory. It also has not shown whether customers regularly use Tianxi after purchasing those systems.
Even with those gaps, the direction is consistent with Lenovo’s earlier disclosures. In November 2025, the company said AI PCs accounted for 33% of its worldwide PC shipments. It also reported a 31.1% share of the global Windows AI PC segment.
Crossing 50% of Chinese notebook sales suggests a faster transition in Lenovo’s home market. China is also where Tianxi provides Lenovo with a more controlled software layer than the company has in markets centered on Microsoft Copilot.
The number arrives alongside broader financial momentum. Lenovo’s fiscal 2025/26 annual results showed that AI-related revenue grew 105% and reached 33% of group revenue. Lenovo defined that category broadly, including NPU-equipped devices, GPU servers, and AI services.
Lenovo reported full-year revenue of $83.1 billion, a 20% increase. Its Intelligent Devices Group produced $58.9 billion in revenue, up 17%. During the fourth quarter, Lenovo held 24.4% of the worldwide PC market, according to the company.
Those results establish that Lenovo has scale, distribution, and a large installed base. They do not isolate how much incremental revenue came from AI software rather than new processors, premium designs, or routine upgrades.
That is the tension created by Liu’s announcement. The company has moved AI-branded notebooks into its core sales mix. It must now show that the software experience creates durable value beyond the purchase itself.
Lenovo AI PC Growth Is Part Hardware Cycle, Part Agent Bet
AI PC adoption is accelerating during a replacement cycle that would have moved many buyers to new hardware anyway.
The PC industry entered 2026 with several overlapping demand drivers. Businesses continued replacing older computers, while the end of Windows 10 support pushed organizations toward Windows 11. New processors also improved battery life and added dedicated AI acceleration.
Component shortages complicate that picture. IDC reported that global PC shipments increased 3% year over year during the first quarter of 2026, reaching 65.6 million units. However, it attributed part of that growth to buyers and channels purchasing early before expected shortages and price increases.
IDC’s later PC market outlook forecast an 11.3% decline in worldwide PC shipments for the full year. It also projected average selling prices would rise 18.3% as memory and storage constraints affected supply.
That environment can make premium AI PCs appear stronger for reasons unrelated to their software. When entry-level availability tightens, manufacturers prioritize higher-margin configurations. Buyers may also choose newer processors because those are the models available within current enterprise and consumer product lines.
Lenovo’s own fiscal fourth-quarter results provide a clue. Premium PCs reached 50% of the company’s shipment mix, while premium shipments grew 29% year over year. That performance can support AI PC growth because newer premium machines commonly include NPUs.
The overlap does not invalidate Liu’s number. It changes the question readers should ask. The relevant test is not whether customers bought NPU-equipped computers, but whether agent features affected their purchasing and continued use.
Gartner previously forecast that AI PCs would account for 54.7% of worldwide PC shipments in 2026. Its AI PC forecast projected 143.1 million units for the year, compared with 77.8 million in 2025.
If that forecast holds, passing 50% will soon stop distinguishing any major PC vendor. AI acceleration will become a standard feature, much like webcams, fingerprint readers, and dedicated graphics options became familiar parts of notebook configurations.
That shift pressures Lenovo to compete above the processor layer. Intel, AMD, and Qualcomm supply capable AI silicon to multiple manufacturers. Microsoft provides Windows AI components across supported hardware from Lenovo, Dell, HP, Acer, Asus, Samsung, and its own Surface line.
Lenovo’s answer is Tianxi in China and Qira in international markets. The company describes these systems as personal agents that retain context and coordinate actions across PCs, phones, tablets, wearables, and other devices.
This “one AI, multiple devices” strategy seeks to turn a notebook into the center of a personal computing environment. The agent, rather than any single screen, becomes the interface connecting a user’s files, preferences, applications, and active tasks.
That is a more ambitious claim than offering local image effects or transcription. It also creates a more difficult product challenge. Cross-device agents need reliable identity, permission management, application integration, context retrieval, and predictable execution.
Lenovo’s sales milestone indicates that compatible hardware is reaching users. It says much less about whether the agent layer has earned their trust or become part of their daily work.
Local Tokens Give Lenovo Its Strongest Technical Argument
Keeping more inference on the device can lower cloud dependence, reduce latency, and limit unnecessary data transfers.
Liu said Lenovo is upgrading computing capacity at the device and edge levels. The goal is to keep more token production local while balancing privacy and operating costs. Tokens are the small text or data units an AI model processes when interpreting input and generating output.
Cloud models remain useful for complex reasoning and tasks requiring large amounts of computing power. However, routing every request to a remote service introduces network delay, server expense, and questions about where personal information travels.
A hybrid design chooses among the PC, a nearby edge system, and the cloud. Simple or sensitive tasks can stay local, while demanding workloads move to larger remote models. The practical challenge is deciding where each task should run without confusing users.
Microsoft’s Windows AI guidance shows why PC manufacturers see an opening. Copilot+ PCs use NPUs with at least 40 trillion operations per second, allowing supported models to run without occupying the CPU or relying entirely on a network connection.
Windows ML can identify available hardware accelerators and select an appropriate execution provider. If the preferred NPU route is unavailable, the system can fall back to a GPU or CPU. This makes local inference more accessible to application developers across different processor platforms.
Lenovo can build on that common foundation, but Tianxi must provide additional value. A local summarization model alone does not create a personal agent. The system needs relevant context, useful integrations, and authority to complete actions safely.
Consider a worker preparing for a customer meeting. A valuable agent might locate local notes, summarize recent documents, identify unresolved commitments, and draft a briefing. It would need to distinguish the correct customer, respect file permissions, and cite the material supporting its summary.
This type of workflow explains the appeal of a personal knowledge base. The model itself is only one component. Reliable retrieval and transparent source connections determine whether an answer helps someone make a decision.
Local execution can protect some data by avoiding cloud transfers, but “local” is not a complete privacy policy. Users still need to know what the agent indexes, how long it retains context, which applications can access that context, and how information moves between devices.
Cross-device operation adds another layer. A phone, PC, and tablet may hold different accounts or confidential materials. Synchronizing context can improve continuity, yet it can also expand the consequences of an incorrect permission or compromised account.
Local processing also has performance limits. Small language models use less memory and power than large cloud systems, but they generally handle narrower tasks. Complex planning, broad factual questions, and long documents can still require remote processing.
Developers must optimize models for several combinations of NPU hardware, memory, and software runtime. Quantization, which reduces numerical precision to make a model smaller and faster, can improve efficiency but sometimes reduces output quality.
Lenovo’s hybrid approach is therefore a routing problem as much as a hardware strategy. A useful system must choose between local and cloud models according to latency, sensitivity, capability, and cost.
If Lenovo handles that routing well, users may experience faster private tasks without thinking about infrastructure. If it handles the process poorly, they may see inconsistent answers, repeated permission prompts, missing context, or unexplained network activity.
The strongest part of Liu’s argument is economic. Moving frequent inference away from centralized servers can reduce the marginal cloud expense attached to each interaction. That matters if personal agents run continuously across millions of devices.
Yet local hardware shifts some cost and complexity to buyers, software teams, and device manufacturers. More memory and capable NPUs increase system requirements. Maintaining models and runtimes across many configurations requires sustained engineering after the sale.
The Lenovo AI PC thesis works only when those costs produce applications people value. Hardware availability opens the door, but useful software determines whether local tokens become more than a technical metric.
Skills Are the Difference Between a Chatbot and an Agent
Lenovo’s planned Skills expansion is the clearest test of whether Tianxi can move from answering questions to completing work.
Liu said Lenovo is accelerating the introduction of Skills to strengthen Tianxi AI’s L3 task execution. In this context, a Skill is an integration or structured capability that lets an agent interact with an application, service, or defined workflow.
The L3 label is Lenovo’s description, not an industry-wide certification. Readers should not treat it like a standardized autonomy rating. The meaningful questions concern which tasks Tianxi can finish, how often it succeeds, and what controls surround consequential actions.
A chatbot can generate a suggested itinerary. An agent must compare options, read a calendar, request approval, create reservations, and preserve the relevant details. Each step introduces possible errors and demands a clear permission boundary.
Useful Skills could allow Tianxi to organize documents, prepare presentations, manage schedules, or transfer context between a Lenovo PC and phone. Enterprise Skills might connect approved internal systems while keeping selected data on managed devices or local infrastructure.
This creates Lenovo’s largest software opportunity. PC manufacturers traditionally ship utilities that many users ignore. A frequently used agent could create an ongoing relationship after the hardware sale and make switching devices less attractive.
It also creates platform risk. Microsoft controls Windows and its system-level AI interfaces. Application developers decide which external agents can access their software. Processor vendors supply optimization tools that work across competing PC brands.
Lenovo must convince developers that supporting Tianxi is worth the effort. A large installed base in China helps, especially if the same Skill can reach several Lenovo device categories. However, developers already face demands from mobile platforms, cloud agents, and Windows itself.
Gartner expects 40% of software vendors to prioritize on-device AI capabilities by the end of 2026, up from 2% in 2024. That would enlarge the available software base, but it would also reduce the uniqueness of any one manufacturer’s agent.
Quality matters more than the number of listed Skills. An integration that launches an application or fills a single field does not establish agentic execution. Buyers need evidence that multi-step workflows finish accurately under normal conditions.
Lenovo has not released public success-rate data for Tianxi’s L3 tasks. It has not shown how often users correct the agent, abandon an operation, or switch to manual work. It also has not published comparable latency figures for local and cloud routes.
Those omissions are understandable for an evolving product, but they limit what the sales milestone proves. Fifty percent hardware penetration provides a potential audience. It does not demonstrate active agent adoption within that audience.
The company also needs careful controls for actions that change data or create commitments. Deleting a file, sending a message, approving an expense, or placing an order carries more risk than summarizing text.
Good agent design should separate planning from execution. It should show users what will happen, request approval at meaningful points, record completed actions, and provide a recovery path when something goes wrong.
Privacy controls need equal attention. A personal agent becomes useful by learning from sensitive context. That same context can include private messages, financial records, work documents, health information, and other people’s data.
Keeping inference local can reduce exposure, but an agent may still call cloud services through a Skill. Users need a clear view of which information leaves the device, which provider receives it, and whether the transfer is necessary.
Enterprise buyers will demand administrative controls, audit logs, identity integration, retention settings, and predictable updates. They will also test whether Tianxi fits existing security systems rather than creating a separate policy layer.
Consumer expectations differ. Individual users tend to reject setup processes that require extensive configuration before showing value. Tianxi must deliver useful initial workflows while giving people understandable control over deeper access.
Lenovo’s hardware scale gives it time to improve these systems, but scale also raises the cost of mistakes. A faulty agent behavior deployed across a large device base can quickly become a support, security, and reputation problem.
The Skills strategy is therefore the central mechanism in Lenovo’s claim. More integrations can turn local computing capacity into real work. Weak or shallow integrations will leave AI PCs looking like conventional premium notebooks with extra silicon.
What the Sales Number Still Does Not Prove
Lenovo has reported strong distribution, but it has not yet demonstrated sustained demand for personal agent behavior.
The first uncertainty is measurement. Lenovo has not provided a detailed definition for the China notebook percentage in Liu’s remarks. Investors and buyers need to know which processors, operating requirements, and AI functions place a model inside the category.
They also need unit figures. A share of Lenovo’s own notebook sales can rise when AI PC volumes grow, when conventional notebook sales decline, or when both happen together. The reported 86% global increase offers direction but lacks a disclosed starting base.
The second uncertainty is sell-through. Manufacturer shipments can reflect inventory moving into retail and distribution channels rather than completed purchases by end users. The original remarks use a sales framing, but no independent market tracker has confirmed the exact 50% figure.
The third issue is product bundling. If AI-capable processors become standard across midrange and premium notebooks, customers may buy AI PCs without choosing them for AI. Features such as battery life, display quality, weight, and upgrade timing may drive the decision.
A Lenovo executive has previously acknowledged this broader industry problem. Luca Rossi, who leads Lenovo’s Intelligent Devices Group, said in 2025 that only a modest share of users fully understood and embraced AI features. He argued that attractive hardware remained a major reason people selected the machines.
That admission makes the current milestone more interesting. Lenovo is winning the distribution phase while still needing to prove the behavior phase. The two stages should not be confused.
Active-use metrics would help. Lenovo could report how many AI PC owners activate Tianxi, how often they return, which Skills they use, and how many tasks reach completion without manual recovery.
Retention matters more than initial trials. Many bundled assistants experience high curiosity after setup and then disappear from daily workflows. A useful personal agent should show consistent weekly or monthly engagement across several task categories.
The competitive response creates another uncertainty. HP and Dell can combine similar processor platforms with Microsoft’s AI services. Apple controls both hardware and operating software across the Mac, iPhone, and iPad, giving it its own cross-device integration advantages.
Microsoft remains especially important. Its Windows AI APIs provide text summarization, image processing, optical character recognition, and local language features on compatible devices. As these capabilities improve, PC vendors must explain what their own agent adds.
Lenovo’s position in China provides some insulation because Tianxi can be adapted to local services and market requirements. That advantage becomes harder to reproduce internationally, where application ecosystems and customer expectations differ.
The company has introduced Qira as its broader personal intelligence layer. However, Lenovo has not yet shown that Tianxi and Qira can deliver equivalent agent behavior across markets, languages, applications, and regulatory environments.
Economic pressure could also affect adoption. IDC’s forecast for lower 2026 PC shipments and higher average prices creates a difficult setting for discretionary upgrades. AI PCs may gain share inside a shrinking market without producing overall industry growth.
That outcome would still help Lenovo’s premium mix. It would not necessarily create a new PC supercycle. A true expansion would require buyers to upgrade sooner, purchase additional device categories, or pay for continuing services because of AI capabilities.
Lenovo’s reported numbers support a narrower conclusion. AI-ready hardware has become normal across much of its Chinese notebook portfolio, and the company is converting that transition into sales. The evidence for software-led demand remains incomplete.
Three Signals Will Decide Whether Lenovo’s Agent Strategy Works
The next phase should be judged through usage, developer support, and verified financial contribution, not another hardware penetration milestone.
The first signal is sustained Tianxi engagement. Lenovo should disclose active users as a share of AI PC owners, retention over several months, and task completion rates. Growth in these metrics would show that the agent is influencing behavior after purchase.
Falling engagement would weaken the strategy even if AI PCs approach 100% of shipments. That pattern would suggest the category is expanding because NPUs became standard hardware, not because customers value Lenovo’s software.
The second signal is a credible Skills ecosystem. Lenovo needs integrations that complete meaningful workflows across business and consumer applications. The useful measure is not a large catalog count, but repeated execution with clear permissions and low correction rates.
Developer announcements will matter when they include specific functions, supported markets, and availability dates. Demonstrations should show what happens when an application is missing, a permission is denied, or a task changes halfway through execution.
A strong ecosystem would differentiate Tianxi from a conversational interface layered over Windows. Weak participation would leave Lenovo dependent on Microsoft’s APIs and a limited collection of first-party utilities.
The third signal is financial separation between AI hardware and AI services. Lenovo already reports broad AI-related revenue, but that category combines devices, servers, and services. More detailed reporting could show whether personal AI produces recurring revenue or improves customer retention.
Service revenue is not the only possible proof. Lenovo could demonstrate shorter replacement cycles, higher attach rates across phones and tablets, or stronger loyalty among active Tianxi users. Any of these would connect agent use to business performance.
Readers should also compare Lenovo’s progress with the wider market. If every major vendor crosses similar AI PC shipment thresholds, hardware penetration will reveal little about competitive strength. Software engagement and workflow reliability will become more informative.
For developers, the opportunity lies in local models and agent integrations that work under real hardware constraints. They should watch Lenovo’s APIs, Skill distribution terms, permission model, and support across different processors.
Enterprise buyers should demand measurable outcomes before standardizing on an agent layer. A pilot should test retrieval accuracy, completion rates, auditability, network dependence, and the handling of sensitive information.
Individual buyers should focus on tasks, not labels. An NPU adds useful capacity, but its value depends on the applications they use and whether those applications can run meaningful work locally.
Lenovo’s 50% claim captures a genuine market transition. AI processors are moving from specialist configurations into ordinary notebook purchases. The unresolved question is whether Tianxi can turn that installed base into trusted, repeated action.
Watch what people do after the new PC leaves the store. If Tianxi becomes a dependable layer across files, applications, and devices, Lenovo will have built more than another upgrade cycle. If usage fades, the Lenovo AI PC milestone will mainly document how quickly new silicon became standard.
