Li Lu on AI and Value Investing: Protect What Can Compound

AI can write code, conduct research, and coordinate tasks. An investor still has to answer an older question: how much value can a business ultimately create, and does the price make sense?

At Columbia Business School's Graham & Dodd event, Li Lu placed that question within a changing landscape. Capital is concentrating in a small group of technology companies. Chinese household wealth and industrial competitiveness have moved in different directions. Political relationships are shifting. Assessing businesses, understanding risk, and allowing value to develop remain necessary work.

Compounding connects these subjects throughout the discussion. Capital needs a foundation that survives. Knowledge and trust can accumulate in a similar way. AI can help execute work, while people still have to decide which work deserves years of sustained attention.

Six principles, ending with purchasing power

Li Lu organized his understanding of value investing around six principles.

• Shares represent ownership of a business. The subject of research is how the company operates and creates value. The share price is the available price for that ownership.
• Treat the market as a source of offers. Changing sentiment supplies opportunities, while a valuation needs evidence independent of price movements.
• Leave room for error and unexpected events. An uncertain future requires a margin of safety in both the price and the underlying judgment.
• Know the boundaries of competence. An investor needs to distinguish a well-understood problem from one that remains outside reliable understanding.
• Look for worthwhile opportunities with less competition. Crowding can reduce returns. Neglected businesses, industries, and markets may reward careful study.
• Measure wealth through its share of global purchasing power. Inflation and currency changes affect nominal balances. What those assets can ultimately command is the more durable concern.

In his account, the first three principles come from Benjamin Graham, the circle of competence from Warren Buffett, and the search for better hunting grounds from Charlie Munger. The final principle reflects his own experience investing internationally.

That helps explain his emphasis on enduring ownership of strong businesses. Continued value creation gives shareholders a chance to participate in growth and preserve or expand their purchasing power.

He also distinguished long-term investors, arbitrageurs correcting mispricing, speculators anticipating others' trades, and brokers or market makers. Individuals can combine these roles. The useful distinction is whether returns depend primarily on business value developing or on another participant changing the price.

More capable AI still leaves a role for informed judgment

Moderator Tano Santos raised a question for the students in the audience: what happens to equity analysts as AI becomes more capable?

Li Lu acknowledged the usefulness of generative AI. It can organize and apply accumulated human knowledge, making some previously scarce expertise more accessible. He described a progression from information tools to reasoning, coding, task execution, and collaboration among agents.

His concern was the way the narrative can skip intermediate steps. Increasing capability becomes a story about surpassing humanity, and that imagined destination becomes a justification for enormous investment.

Two parts of his argument deserve separate treatment. His account of how AI works with human data was explicitly provisional: an investor's interpretation. His insistence on human supervision was a position on governance. Neither establishes a scientific limit on future technical capabilities.

He opposed recursive self-improvement, or RSI, without human oversight. A system that generates objectives, decomposes them, and continually changes itself could become harder to control. He argued that development should account for employment, emotional dependence, security risks, economic costs, and society's capacity to adjust.

Santos then drew a professional implication: if AI outputs need evaluation, people who understand a domain, recognize errors, and communicate effectively could become scarcer resources. Li Lu broadly agreed while acknowledging additional nuances.

An abundance of answers increases the value of assessing their reliability. One implication of the exchange is that researchers need to accumulate more than report-writing speed. Understanding businesses, checking evidence, and rejecting a persuasive but unreliable answer remain substantive skills.

A powerful technology still has recurring costs

Li Lu emphasized that using AI brings continuing incremental costs. Deployment does not eliminate the resources needed to deliver subsequent services. Revenue, profit, capital requirements, and cost discipline therefore remain central to evaluating the businesses involved.

In his framework, the speed of development also depends on society's tolerance for risk. Technical progress, available capital, and public acceptance may move at different speeds, producing repeated expansions and contractions.

The discussion then turned to concentration in equity markets. Santos offered a rough comparison: the largest company's market capitalization had historically been around 2% to 4% of GDP, while his contemporaneous estimate for Nvidia was about 15%. This was the moderator's comparison during the discussion, rather than a live valuation measure for the article's publication date.

Li Lu responded by extending the time horizon. He estimated that the S&P had returned roughly 14% to 15% annually over about 16 years from 2010, following approximately zero or slightly negative returns from 2000 to 2010. Across roughly 26 years, he described the result as a high-single-digit return.

He used approximately 2% real economic growth, 2% to 3% inflation, and 4% to 5% nominal growth to illustrate why long-run business returns need support from economic activity. These were broad estimates in a conversation, without a complete index-return methodology or calculation. They do not establish an expected future return.

His larger point concerned where attention goes. When capital concentrates in a handful of popular companies, other good businesses may receive less attention. A technological wave expands what people consider possible and can widen differences in how assets are priced.

A valuable technology, a strong business, and an attractive purchase price require separate judgments. Recognizing the usefulness of AI does not settle the appropriate price for every company associated with it.

China's divergence requires looking at households and businesses

Li Lu described roughly 10 years of Chinese development as a migration toward higher-value manufacturing. He referenced Made in China 2025 and argued that technological self-sufficiency and industrial upgrading became more prominent after 2021.

His account included both gains and costs. Household balance sheets came under pressure while parts of manufacturing built new competitive strengths.

He offered rough figures for the household side: about 70% of a typical middle-class family's net worth associated with housing, and roughly 50% mortgage leverage on that housing. Against his estimate of a 40% decline in home prices, he described a potential 60% to 70% blow to family wealth. He used that account to explain weaker consumption, increased saving, and debt repayment.

These percentages came from his spoken explanation. The discussion did not provide a complete household sample or a balance sheet from which the calculation could be reproduced. They describe his framework and should not be combined into a formula applying to every household.

He also cited deposits of approximately RMB 165 trillion, equivalent to about 120% of GDP. That is a stock of deposits, rather than the share of annual income newly saved.

On the industrial side, he argued that the objectives of Made in China 2025 had been achieved, pointing to new energy, electric vehicles, batteries, storage, and heavy equipment as examples of accumulated competitiveness. This was his assessment of the policy's results.

For semiconductors, he argued that external restrictions created orders for domestic suppliers, supporting a cycle of revenue, profit, research, and product improvement. He was positive about self-sufficiency in mature nodes while acknowledging remaining gaps in advanced nodes. His optimism should not be converted into a claim that every technical bottleneck has been resolved.

He also suggested that AI coding could help close software ecosystem gaps, including efforts to build alternatives around systems such as CUDA. That was an argument about potential progress, without evidence that replacement had been completed.

In financial terms, he estimated that annual new property sales had fallen about 70% from their peak, a reduction on the order of RMB 10 trillion. He compared that with approximately RMB 10 trillion in annual manufacturing upgrades and technology investment. This was a comparison of magnitudes, not evidence that every unit of lost property spending had moved directly into manufacturing.

The most practical distinction in this section concerned a company's position: is it being supported as a policy objective, or is it bearing the cost of achieving another objective? Industries can move between those positions. Past support does not guarantee future benefit.

Political change does not remove economic interdependence

Li Lu distinguished political arrangements from economic relationships. Political structures can shift, while the division of labor, markets, and supply chains built by businesses have their own sources of persistence.

He argued that leaving a common market carries substantial costs. Trade barriers may redirect transactions, with some relationships continuing through third countries. Cross-border dependencies in semiconductors served as one example of that interdependence.

This is a perspective on economic resilience, rather than a reason to disregard political risk. During the questions, he emphasized examining a country's objectives, history, and actual record. He applied that requirement to China, the United States, and other markets.

He also compared the emphasis of Chinese and American AI companies. In his characterization, Silicon Valley focuses more on extending the technological frontier, while Chinese companies put more weight on what the technology can do for consumers and business customers. That generalization does not describe every company's choices.

Applied to business research, these observations suggest concrete questions: where demand originates, which suppliers are essential, how policy changes affect customers and costs, and whether the company can continue creating value within those constraints.

Growth and value can face the same questions

An audience member asked whether value investing could apply to growth and technology companies.

Li Lu grounded his answer in the relationship between price and value. A stagnant business returning cash, a rapidly expanding company, and an opportunity with a small chance of a large payoff all require an assessment of what could be created relative to what is paid.

He connected those judgments to future cash flows and discounting, with a crucial condition: the investor must be able to understand and estimate those cash flows. Rapid growth does not remove uncertainty or the boundaries of competence.

Elsewhere in the questions, he noted constraints on AI from chips, electricity, and labor. That makes the earlier cost discussion more concrete. Expanding digital capabilities still requires physical resources.

His position on crypto was strongly negative. He argued that it lacks the enforcement foundation supplied by sovereign states and predicted that it would eventually disappear. That is his stated position and forecast, rather than an established outcome.

Long horizons need compatible evaluation and capital

A participant early in an investment career asked a more immediate question: how can someone committed to long-term value survive a period of underperformance?

Li Lu did not offer a shortcut to closing the performance gap. He argued that trying to pursue long-term value while delivering superior results every quarter or month can gradually shift behavior toward speculation about the market.

A long-term approach needs compatible capital and relationships. Beyond having independent reasons for a judgment, an investor needs partners who understand the method and can tolerate periods of disagreement with market prices. Maximizing assets under management can conflict with that requirement.

He acknowledged that this approach does not suit everyone. For those uncomfortable with the pressure of independent judgment and temporary divergence, he presented indexing as an alternative.

An investment horizon has to exist in funding arrangements and evaluation practices as well as in a strategy document. His answer makes patience an issue of relationships and organization, alongside individual temperament.

Compounding depends on keeping the accumulated foundation

Near the end, Li Lu returned to how he learns.

With limited knowledge early in his career, he started with comparatively understandable balance sheets and gradually moved toward more complex businesses. A circle of competence needs boundaries, but sustained learning can expand them.

He linked the accumulation of knowledge with the accumulation of capital. Understanding more can improve the ability to assess and protect assets. That is an account of his experience, not a guarantee that more learning produces investment returns.

Trust mattered just as much. He recalled Munger's emphasis on dependable relationships and acknowledged the fortunate breaks in his own investing history. Outcomes cannot all be credited to individual ability.

Compounding therefore extended beyond money to knowledge, reliable relationships, and the health that allows meaningful work to continue. Each can make today's effort part of tomorrow's foundation.

The discussion leaves a useful standard for choosing work over a long horizon: whether accumulated knowledge remains useful, trust continues to deepen, and the businesses owned keep creating value. AI can accelerate many steps. Those foundations still require time and people willing to protect them.

Original source: 36th Annual Graham & Dodd Event with Li Lu ’96, Founder and Chairman of Himalaya Capital, published by Columbia Business School.

The speaker is Li Lu, founder and chairman of Himalaya Capital; the moderator is Columbia Business School professor Tano Santos. The event took place on 2026-10-05, as recorded on the official event page. The video was uploaded on 2026-10-15 and runs 1:28:17.

This article is an original synthesis of the video's supplied English subtitles. Dates, names, and terminology were checked against the official Columbia Business School transcript. The subtitle track is marked as non-automatic but contains transcription errors. Market and macroeconomic figures retain their status as the speakers' estimates; forecasts and assessments remain attributed to the speakers.