Sequoia Capital's Alfred Lin: A Closed-Door Record from Stanford on the Technology Diffusion Faults, Ice and Fire Rules, and Valuation Reconstruction during the AI Supercycle

AI Deep Dive | Behind the Interview with Sequoia's Alfred Lin: The Biggest Opportunity in AI Lies Between Technological Advancement and Cash Returns

1. "Technology diffusion gaps" are primarily a cost of organizational transformation; those who can reduce it are the ones who can make money.

Just because a model can complete a task does not mean that a company can adopt it immediately. Companies also need to connect data, adjust permissions, validate results, reset job responsibilities, and determine who is responsible when errors occur. These costs do not automatically disappear with the refresh of model rankings.

Therefore, the value of AI application companies can be understood as digesting the complexity of technological changes for clients. Clients are purchasing a delivery system that can be integrated into daily work, including reliability, training, auditing, and ongoing maintenance.

OpenEvidence's enterprise solution emphasizes integration with clinical workflows, as well as the use of institutional identity verification and access control. These seemingly unremarkable features precisely reflect the conditions required for professional tools to transition from demonstration to actual deployment. OpenEvidence

However, slow diffusion has another side: if every client requires extensive customization for implementation, application companies may become labor-intensive service providers. To assess their commercial quality, one should observe whether the deployment time required for new clients is shortening and whether delivery costs can decrease with accumulated experience.

2. "Software will not die" does not mean that the current charging methods of software companies will not be rewritten.

Software still undertakes data organization, permission control, process execution, and result presentation. However, when intelligent agents can replace humans in operating multiple systems, the users of the software, purchasing logic, and charging units may all change.

Products that traditionally charge per employee seat may face a contradiction: as client businesses grow, the number of people using the software may decrease. If revenue depends on the number of seats, the productivity gains for clients may actually compress vendor income.

At the same time, charging by task is not a universal solution. Different task difficulties and computational consumption, as well as retries due to failures, can erode gross margins; charging by results requires solving attribution issues, such as whether sales growth comes from AI or from price, channel, and market conditions.

Thus, "opportunities exist at the application layer" must be further broken down into three questions: which workflow to control, why clients continue to pay, and whether the charging method can cover the complete delivery costs. Industry demand exists, but whether a certain business model can sustain profitability is a different judgment.

3. Slow diffusion also gives existing companies time to absorb new technologies.

Technological revolutions will not only help entrepreneurs. Mature companies that have clients, channels, data, and cash flow can also procure technology and embed it into their existing operating systems.

Walmart is a useful example, but the numbers need to be accurate: its net sales for fiscal year 1997 were approximately $104.9 billion, and for fiscal year 2026, approximately $706.4 billion, about 6.7 times the former, not 30 times. Amazon's annual total revenue surpassing Walmart's cannot be directly equated to the complete replacement of Walmart's retail business, as the two companies have different business structures. stock.walmart.com

This historical case illustrates that channels and fulfillment capabilities can continue to add value during technological changes. In the AI era, existing software companies may also leverage their installation base, procurement relationships, and business data to quickly catch up.

What entrepreneurs need to prove is not just "my model is more advanced," but rather: why existing suppliers find it difficult to serve this demand? Is it due to conflicting charging models, organizational structure limitations, or product architecture that cannot support it? These obstacles are more enduring than temporary functional advantages.

4. The faster the technological improvements, the more likely buyers are to wait; startups must make "using now" worth paying for.

When clients believe that the model will be stronger and cheaper in six months, they will raise the procurement threshold. Rapid progress creates demand but may also lead buyers to delay commitments, especially for projects with high deployment costs and long contract cycles.

Suppliers thus face a timing issue: if the product has not yet completed sales and deployment, the initially advertised technical advantages may have already become commonplace. Long-term procurement contracts cannot rely solely on the current model's performance.

Products that can withstand this pressure typically provide immediate and clear benefits and can continuously absorb upstream advancements. For example, clients are purchasing a service that continuously shortens contract processing cycles, rather than access to a fixed version of a model.

From a capital allocation perspective, it is important that the time required for clients to recoup their investment is shorter than the depreciation cycle of the technological advantage. If it takes a year to generate returns from deployment, while the product advantage is matched in three months, the sales commitment becomes very fragile. Diffusion opportunities belong to those who can continuously update value, not necessarily to those who package new models first.

5. Just because companies want to "own intelligence" does not mean they should all train models themselves.

Self-control encompasses multiple levels: controlling data, controlling business rules, controlling evaluation standards, being able to replace suppliers, and deploying models independently when necessary. Owning model weights is just one way to achieve this.

Open-source weights can increase choices but do not eliminate the costs of computing power, engineering, security, and maintenance. If the task volume is insufficient, building a system in-house may be more expensive due to low utilization than purchasing externally; if the internal team cannot continuously evaluate the model, the so-called control may remain nominal.

More valuable assets are often the company's own task samples, misclassifications, acceptance criteria, and feedback records. They can help companies assess whether different models are applicable and reduce future migration costs.

Therefore, "the business system does not allow a few model vendors to concentrate long-term" can only be a prediction. The market has a motive to break free from dependence, but that does not mean it will necessarily form a decentralized structure. Scale advantages in chips, cloud services, distribution entry points, and talent may still lead to high concentration in certain segments.

6. Circular investment cannot prove health solely through "currency circulation"; it must track the ultimate payers.

Chip suppliers invest in AI companies, which then purchase computing resources, helping to kickstart new markets. This arrangement does not inherently constitute a false transaction, nor does it inherently prove that terminal demand has matured.

The key distinction lies in whether the procurement funds come from ongoing operational revenue or from the next round of financing. If clients can only continue placing orders by obtaining new funds from suppliers or associated investors, then order growth simultaneously depends on product demand and capital market sentiment.

Comparing this relationship to a bank's money multiplier is also not rigorous enough. The Bank of England's explanation of modern money creation clearly states that bank loans create deposits, not simply transferring reserves at a fixed multiple; industrial equity investment cannot be directly equated to this mechanism. the UK's central bank

When examining circular investment, one should track cash paid by independent clients, contract cancellation conditions, sources of financing, and transaction pricing. The most telling stress test is: if no new financing occurs in the next twelve months, how much procurement can the supply chain maintain?

7. The risks of AI infrastructure stem from several mismatches in timeframes; debt merely exposes these issues earlier.

Data centers contain long-lived assets such as land, power facilities, and buildings, as well as computing devices that undergo faster technological iterations. If the entire facility is treated as an asset with a uniform lifespan, it is easy to overestimate its cash recovery capability.

Assuming a project invests $1 billion, with $600 million in servers and accelerators. If this portion of equipment needs to recover through operational cash flow within four years, it corresponds to an average of $150 million per year just for equipment investment, not including electricity, maintenance, financing costs, and other construction inputs. This is a recovery calculation, not equivalent to accounting depreciation.

The risk is that loan repayment arrangements may be rigid, while client contracts can be renegotiated; equipment may still be usable, but the rental price per unit of computing power may continue to decline. Thus, accounting profits and cash available for debt repayment may become separated.

Even if fully financed by equity, returns below capital costs will still destroy value. Debt increases the likelihood of forced sales, defaults, and risk transmission; one cannot assume that "no high leverage" is evidence of project economic viability.

8. Correct identification of super trends does not guarantee correct investment pricing.

Industry scale growth and leading companies becoming larger do not automatically support any financing valuation. Investors are purchasing future equity returns at a certain price, during which they must endure failures, dilution, and waiting.

Assuming an investor enters a company at a $1 billion valuation and exits ten years later at $5 billion. Ignoring subsequent dilution, a fivefold return corresponds to an annualized rate of about 17.5%; if the equity held is diluted by half during that period, the actual return is about 2.5 times, with an annualized rate of about 9.6%. This does not account for fund fees, taxes, and other failed projects.

Therefore, the larger the terminal market, the more it only indicates an increase in potential profit limits, not that every company has an increased probability of reaching the terminal point, nor does it imply that the purchase price is irrelevant.

For AI companies, market size, obtainable share, sustainable profit margins, and capital requirements should be estimated separately. The correct technological direction provides opportunities, while the appropriate purchase price and capital structure determine how much profit investors can retain.

9. High financing does not overturn lean principles; it merely requires a more accurate distinction between "must invest" and "not yet validated."

Basic models, drug development, and hardware projects may indeed require higher minimum capital investments. One cannot require all startups to initiate with the same low-cost approach, but neither can one deduce that all tracks should rush with heavy investment just because a few projects are expensive.

The effective part of lean is to eliminate critical errors at the lowest possible cost. For application companies, it may first validate whether clients are willing to pay; for research companies, it may first prove that key experiments or technical paths are valid.

The value of large financing is to allow companies to cross irremovable technical thresholds or to rapidly expand advantages after validation. Its risk is that it obscures real demand: team size, computing power consumption, and brand presence may be growing, yet clients still do not form stable usage.

The amount of financing should correspond to the uncertainties that need to be eliminated in the next phase. Each new expenditure should answer: what evidence does it add, and what risk does it reduce? "Competitors raised more" is not sufficient justification for investment.

10. The most useful way to realize "ice and fire" is to translate vision into phased bets.

Vision is responsible for proposing a future worth pursuing, while operational discipline is responsible for determining the boundaries of each investment step. The balance between the two relies not only on two complementary partners but also on decision-making mechanisms.

For example, a project can receive a six-month budget, but the next phase of investment must meet pre-agreed technical, customer, or cost metrics. This avoids short-term profits stifling long-term exploration and prevents "long-termism" from becoming a reason for unlimited additional funding.

Amazon's history should not be oversimplified as "first profiting from books, then using profits to expand all categories." Its historical reports show that it had already entered the music and video business in 1998, with a net loss of approximately $124.5 million that year, and the net loss further expanded in 1999. s2.q4cdn.com

This indicates that operational discipline does not mean the company must first achieve comprehensive profitability. It is more about a continuous understanding of customer behavior, unit costs, capital needs, and financing capabilities: one can afford losses but must know what the losses are purchasing and what evidence will prompt them to stop investing.

11. Fields where giants have not yet entered are merely a time window, not a moat.

"Giants are busy with infrastructure, so entrepreneurs can enter vertical markets" has practical implications but cannot serve as a long-term competitive strategy. Once demand is large enough, giants can reallocate resources, acquire teams, or integrate similar functions for free into existing products.

More solid opportunities arise from structural mismatches between giants and the market: services require extensive industry knowledge, client procurement methods are unique, data acquisition is difficult, or charging models conflict with existing businesses.

Entrepreneurs need to leverage the window to accumulate capabilities that are difficult to replicate quickly, such as long-term customer relationships, complex workflows, validated quality systems, and feedback data that can be legally and sustainably used. Simply accessing a model or adding a few prompt templates is usually insufficient to achieve this.

At the same time, one should also be wary of another trap: giants may be unwilling to enter because the market is too small or delivery is too expensive, rather than because they have not yet discovered the opportunity. It should be verified whether vertical depth enhances customer value or merely increases service costs.

12. Venture capitalists look for extreme successes, but entrepreneurs must manage all their life risks.

Top investors value obsession, resilience, and non-consensus judgment, which has its portfolio logic: a few massive successes can cover a multitude of failures. However, founders typically have only one company, and what they invest includes time, reputation, and career opportunities.

Therefore, the risks that a fund can bear cannot be directly translated into the risks that an individual should bear. The premise of daring to go against consensus should be having verifiable unique information, rather than a habitual denial of mainstream views.

The so-called endurance of pain should also distinguish between continuous learning and repeated losses. If the reasons for customer rejection are already clear, and the product has not improved, continuing to invest will not become reasonable simply because of persistence.

A more useful entrepreneurial discipline is to write down three things in advance: what you believe in, what evidence supports it, and what facts should prompt a change in direction. The truly rare ability is to remain sensitive to new evidence while maintaining long-term goals.

Core judgment: The most important time gap in the AI era occurs between when technology is already available, organizations have not yet completed their transformation, and capital is already demanding returns. Companies that can shorten the distance between the first two have opportunities; those that can align financing timelines, delivery capabilities, and customer cash flows are more likely to turn opportunities into long-term value.