Analyst: Lenovo undervalued in AI infrastructure shift, ISG revenue up 98% YoY

After attending an AI investment industry summit this week, the most memorable takeaway wasn't any speaker's market call—it was a detail from a factory floor.

Lenovo Group Vice President Chen Zhenkuan shared on a panel that the company's supernode factory is undergoing upgrades. Traditional server production lines could be placed on the second or third floor, requiring floor load capacity of just one ton per square meter. For supernodes, the minimum requirement is three tons, designed with a margin of five tons. This forces production lines to the ground floor, with experts demanding solid earth foundations—factories with underground parking garages simply cannot handle the job. Further down the chain, each supernode production base requires over 30 megawatts of power, necessitating a dedicated 110 kV substation.

While capital markets debate model ceilings and valuation fractiles, the industry is recalculating concrete load-bearing capacity. For secondary market investors, the key insight is this: AI infrastructure is shifting from "buying equipment" to "building systems." Systemic changes ultimately manifest in tangible, weighty realities.

01 The New Narrative Standard for AI

Professor Peng Wensheng from Shanghai Advanced Institute of Finance offered a framework more useful than most AI narratives: Large language models represent "scaled production of cognition," but unlike past asset-light internet platforms, every step forward requires real capital investment in computing power, data, and electricity. AI is pushing the digital economy into an industrialization phase with significantly front-loaded capital expenditure.

The numbers tell the story. In Q1 2026, Amazon, Microsoft, Google, and Meta—the four North American cloud giants—posted combined quarterly capital expenditure of $131.6 billion, up 70% year-over-year, with full-year estimates around $710 billion. Wang Qing of Chongyang Investment pushed further: U.S. leading firms' AI capex was approximately $400 billion last year, could double to $800 billion this year, and potentially reach $1 trillion next year. China, lagging slightly, spent RMB 500 billion last year and is expected to reach RMB 1 trillion this year.

At this scale, the rules change. Wang highlighted an often-overlooked variable: starting this year, these giants' free cash flow has turned negative, prompting large-scale bond issuance and off-balance-sheet financing. Projects with major-company credit backing carry financing rates of 5.3% to 6.6%; those without, 8.5% to 9%. For the first time, the technology cycle and financial cycle are so tightly intertwined. Liu Yuhui's assessment was blunter: AI is approaching the "intelligence singularity," which is precisely the most capital-intensive phase.

Thus, the core question in this AI trade has shifted from "does the story make sense?" to "can the money be recovered?" Orders prove demand, profit margins prove business models, and cash flow determines how much industrial prosperity translates into corporate value. After three-plus years of aggressive AI investment, the next phase's narrative logic must follow this standard.

02 Lenovo as a Case Study

By this standard, Lenovo provides a ready-made sample.

IDC's latest data: In Q2 2026, Lenovo shipped 286,000 x86 servers globally, up 45.6% year-over-year, ranking first worldwide. By revenue, it grew 96% to $8.26 billion, just 0.6 percentage points behind Dell. In the latest fiscal quarter, the Infrastructure Solutions Group (ISG) posted revenue of RMB 57.9 billion, up 98% year-over-year, with an operating profit margin of 9.1%—an all-time high. AI server backlog orders more than doubled in a single quarter to over RMB 360 billion. Management added a critical note: most of these orders are deliverable within one year, contingent on supply chain readiness.

A detail within the order structure: one new cloud client started with a budget of just a few hundred million dollars, growing to tens of billions within nine to ten months. Such ramp-up speed is unprecedented in the traditional IT era.

CFO Ken Zheng offered an honest accounting on a panel, not shying away from the thin margins of AI servers: new cloud clients yield 25% to 30% gross margins, with some overseas scenarios higher; enterprise clients deliver over 30% gross margins and operating margins of 18%; the overseas business deeply tied to NVIDIA faces margin pressure, relying on scale. The path to net margins of 5% or even 8%+ lies not in any single blockbuster product but in revenue structure rebalancing: storage business margins of 60% to 70%, with a long-term target of $5-6 billion in revenue and $1 billion in profit; services include equipment leasing and high-margin pure software increments, with "hardware plus services" synergy that management says the market hasn't fully priced; further out is enterprise inference, which Zheng calls "AI's opportunity for the next decade, just beginning."

The timeline: revenue of $100 billion in one to two years (potentially achievable sooner) with net margins above 3%; $130 billion in three to five years with net margins above 5%; $150 billion in five-plus years with net margins above 8%.

03 The Market Cap Equation

Beyond the scorecard, I want to deconstruct Lenovo's market cap logic from the mechanism perspective of a "global AI infrastructure leader." Market cap is simply three numbers multiplied: Revenue × Net Profit Margin × Valuation Multiple. Each variable has its own driving mechanism.

Revenue is driven by market share and delivery scheduling. Global shipment leadership isn't just a title—it creates scale advantages in procurement, manufacturing, and delivery, supporting further market share gains. Wang Qing's framework: the global server market in 2026 is approximately $700 billion, with the U.S. at $500 billion and China at $100 billion. Lenovo is already second globally by revenue, just 0.6 points behind Dell. Every percentage point of market share movement represents tens of billions in revenue. With RMB 360 billion in backlog orders mostly convertible within a year, "growth expectations" become "growth schedules"—a harder basis for valuation than any long-term story. ISG's single-quarter revenue of RMB 57.9 billion annualizes to approximately $32 billion, not yet including the long-term $5-6 billion storage opportunity and services growth.

Net profit margin is driven by structural quality. ISG's OPM has reached 9.1%, but this blends three business types: thin-margin new cloud deals, high-margin enterprise clients (OPM 18%), and even harder-margin storage (60-70% gross margin) and services. As the mix shifts toward the latter two, net margins climb from the current 2-3% range toward 3%, 5%, and 8%. Two amplifiers: first, shipment leadership itself is a cost advantage, spreading procurement and manufacturing thinner; second, supernodes exclude "pure server vendors," concentrating competition among full-stack players and improving pricing power. Structure, scale, and barriers—three forces aligned.

Valuation multiple is driven by framework switching. The market assigns one multiple to a hardware manufacturer with 2% net margins and another to an "AI systems plus services" company—Dell enjoys the latter. Lenovo currently trades at roughly one-seventh of Dell's multiple, "but our revenue and net profit aren't seven times different," as Zheng noted. This gap breaks down into three discounts: the PC company historical label, the low-net-margin manufacturing identity, and market skepticism about order quality (new cloud client creditworthiness). These three factors are being disproven quarter by quarter through order conversion and margin improvement.

Multiplying the three components yields a rough calculation (not a forecast): $100 billion revenue × 5% net margin = $5 billion net profit. At that point, if the market still applies a "low-margin hardware" multiple, market cap growth relies mainly on profit expansion. If the multiple shifts toward systems and services, both profit and multiple move in the same direction. This is why value-revealing stocks often move in steps rather than steady climbs—each time net margins cross a threshold (3%, 5%), the market reassigns a valuation framework. Even if the current 1/7 gap narrows by half, the upside is more than threefold. Zheng's confidence that "we are completely undervalued by the market" rests here: three variables each have clear mechanisms driving them, not a bet on a trend.

A caveat: multiple migration depends on timely order conversion and margin improvement. Any link breaking would cause all three to contract together. This is why the market now only pays for what can be recovered.

04 From Chip Stacking to System-Level AI Sovereignty

Returning to industrial logic, supernodes represent the hardest component and carry the most significance for China's industrial policy.

The reasoning is straightforward. DeepSeek's 671 billion parameters can run on an 8-card cabinet; Kimi 3.5, with 2.8 trillion parameters, already exceeds 8-card capacity; next-generation models with 10 trillion parameters require 32-card supernodes or clusters.

Chen Zhenkuan's view: supernodes "are not a choice but a necessary AI infrastructure." His timeline: about five AI chip companies will support supernodes this year, with all leading players capable next year—"Next year will definitely be the big year for supernodes. This year is about feasibility; next year is large-scale commercialization." Lenovo's product line is accelerating: 40-card inference products are already launched, with 128-card and 256-card models under development, sample machines by end-2026 supporting 10,000-card training clusters. The company has already won a bid from a major domestic internet firm, with delivery in H2.

Supernodes fundamentally change the nature of competition. A single supernode cabinet must be split into computing, power, storage, and connection cabinets, with optics, liquid cooling, AI storage, and both vertical and horizontal scaling all coordinated within one system. Chen put it bluntly: "A pure server vendor today cannot make a supernode. You need to integrate server, storage, networking, operating system, and software teams to define a single product. Comprehensive capability, all-around capability—this is a certainty."

This statement must be viewed within the context of U.S.-China tech competition. The advanced process gap is a near-term reality. But computing competition has never been a single path. When single-point process technology can't catch up, densifying inter-card connectivity, expanding cache, and improving system-level efficiency—using engineering and integration to offset single-chip gaps—is both necessary and viable for China under current constraints. The country's software engineer dividend, world-leading manufacturing and delivery capabilities, and relatively abundant electricity all align with this path.

Chen offered a comparison: the U.S. faces electricity shortages; China faces chip shortages. One Ulanqab data center's power supply equals the average of six U.S. states. With a $700 billion global server market—$500 billion in the U.S. and $100 billion in China—China's share shouldn't be just one-fifth of America's. This gap itself represents growth potential for coming years.

Policy implications are already emerging: AI infrastructure competition is no longer about chips alone but about comprehensive competition in electricity, grids, land, industrial standards, and systems engineering. When supernode production lines require five tons per square meter load capacity, 30 megawatts per base, and dedicated 110 kV substations, "computing sovereignty" moves from concept to planning documents—power planning, industrial land use, and manufacturing upgrades all become fronts for AI infrastructure policy.

Chen also noted a variable: each chip company defines its own interconnect protocol—"12 companies have 13 protocols." Lack of standardization raises system costs across the industry. This is both a commercial issue and a question of standard-setting influence, requiring industrial coordination to resolve.

05 The Decade of Inference

Another consensus at the summit concerned timing.

Song Yuchun of Lenovo Capital's assessment: AI represents the early stage of the Fourth Industrial Revolution. Generative AI has truly exploded for less than four years, and the market has mainly seen the first wave of infrastructure buildout. "Actual penetration into thousands of industries hasn't really begun." His fund has invested in over 330 companies and 26 IPOs over a decade, with over 150 in AI spanning chips, infrastructure, large models, agents, autonomous driving, and embodied intelligence. He called AI the best-performing sector in terms of deal count, investment amount, and returns over this decade—and the best may still be ahead.

On the enterprise side, Zheng's observation was more specific: large multinational companies are still in early AI adoption, needing first to address internal data, organizational structure, and business process integration. Companies that have undergone mergers face system integration measured in years. But once these foundations are laid, enterprise-side inference demand will surge. Combined with data security driving preferences for local private deployment, enterprise computing budgets could see concentrated release around 2027.

He shared one more detail. When overseas CFOs discuss AI optimization, their first reaction is often "cut 25% of headcount." His perspective was different: if revenue reaches $150 billion with net margins of 5-8 points, "I wouldn't just avoid layoffs—I'd hire more people. Just AI talent." In the public debate about AI and employment, this is a rare, numerically specific counter-argument from an operator's mouth.

06 Conclusion

Risks are on the table: financing costs, crowded trades, upstream component price increases—investors at the summit didn't avoid them. Wang Qing's reminder was sobering: fundamental evolution and capital market pricing don't always move in sync.

But return to that five-ton floor. Every industrial revolution in history didn't arrive through keynote speeches—it arrived through foundations, power grids, steel beams, and load calculations.

Today's AI has reached that point.

Disclaimer: The above content (including images and videos) is uploaded and published by a user of NetEase Hao, a social media platform that provides information storage services only.