Freda Duan: Rethinking AI Investment and Human Connection

Zhang Xiaojun welcomes Altimeter Capital partner Freda Duan back to the Business Interviews podcast to discuss AI and investing. Altimeter invests in both private and public markets. The conversation ranges from measuring tokens and competition among model companies to organizational change, anxiety, and human connection.

Zhang begins by asking how to understand tokens, a frequently cited unit in the AI era. Freda says people increasingly measure inputs and outputs in dollars per token, watt, or computation, but tokens themselves can be misleading.

She considers tokens per task more meaningful: models and applications may differ by tens or even hundreds of times in what they consume for the same job. One model might produce concise code while another produces much more; hidden reasoning and agent workflows also affect consumption.

Freda therefore rejects the idea that more consumption is always better. Users want results and efficiency. A finance chief who sees coding tool usage and employee productivity rise might increase the budget without seeing differences in the cost of completing the same task.

She breaks total AI spending into the number of users, tasks per user, tokens per task, and cost per token. She expects the first two to rise substantially as AI expands beyond programmers to more knowledge workers, while the latter two may improve.

Pricing cannot be understood from list prices alone. Citing Anthropic models, Freda says caching and the ratio of input to output affect actual cost; the specific figures she quotes were estimates at the time. She expects users to care increasingly about value for money.

In her view, waste in current token use and continued growth in total consumption can both be true. She compares boasting about consumption to boasting that a light bulb uses a lot of electricity: as the market matures, people may focus on achieving the same result with fewer resources.

On corporate AI spending, Freda cites reports and estimates concerning Meta and Uber and says spending by some leading companies is already noticeable. She stresses that comparing the scale of layoffs with model spending does not establish a direct causal relationship.

She expects the industry to move gradually from token-based pricing toward payment for outcomes, especially when results are measurable. An AI customer service provider that charges for resolved cases but not those transferred to a person, she says, can align supplier and customer around resolution and cost.

Not every outcome is easy to measure. Freda thinks sales conversion, collections, and insurance claims may adopt outcome-based pricing earlier, while creative work such as writing may retain usage pricing for longer.

On model competition, Freda has begun questioning the old assumption that the leading model changes hands every few months. Mature coding agents could create a feedback loop in which better AI helps develop the next generation; if a threshold is crossed, catching up might become harder.

She compares the change to horse-drawn carriages and cars: while early cars were unreliable, fast horses could keep up, but a reliable car would be difficult to match without an engine. She calls the analogy imperfect and says the threshold remains uncertain.

On Codex versus Claude Code, she suggests extensive debate may itself indicate that their real-world performance is not vastly different. She observes an aggressive pricing and access strategy from Codex and reflects that she spent too much time comparing the two companies instead of recognizing the market’s overall growth.

Freda distinguishes per-seat software subscriptions from usage-priced models. Subscriptions may encourage a customer to consolidate with one vendor; per-query or token pricing may encourage businesses to compare and switch models for different tasks. She therefore sees more than a traditional zero-sum contest.

She revisits an early underestimate of the coding AI market: investors calculated its size from the number of developers and a per-person subscription price, producing a relatively small figure. Subsequent revenue growth showed that framework was too narrow, she says; work that can be performed on a computer may represent potential demand.

When coding tool revenue rose rapidly, investors should have checked their original market-size assumptions. Freda says mistakes are inevitable in making judgments; the important thing is to update when the facts no longer fit the framework.

She discusses a “negative snowball” for model companies: if training costs multiply and gross profit from older models cannot keep pace, new investment continually depresses profit. Besides slower growth in training costs, she argues, much faster revenue growth could change that outcome.

Comparing OpenAI and Anthropic revenue requires attention to definitions, she cautions: annualizing a short period is not the same as stable recurring annual revenue, and gross and net revenue cannot be compared directly. Her figures and conversions are estimates from the time of the interview.

On computing capacity, Freda believes the market underestimated how much model revenue a unit of power could support. Drawing on outside estimates of capacity, she says the revenue ceiling remains hard to judge; her figures are observations and forecasts from the recording.

She recalls investors’ judgments as Anthropic revenue grew quickly: demand for coding might be larger and faster than expected, but margins and other uncertainties also prompted concern. She thinks model-layer products and business models may converge as companies pursue demand.

For xAI and Meta, Freda says outsiders have limited evidence for judging outcomes. She discusses data center rentals and possible uses for Meta’s models as investment observations at the time, rather than settled results.

Changes in software have led her to reconsider organizational size. Traditional enterprise software companies often have large sales and marketing teams; seeing model companies grow rapidly with smaller teams, she asks whether investors once overemphasized sales and how strong demand for intelligence may be.

She extends the question to larger companies: if a small team can build a product, organizations should reconsider their structures and staffing. Her company size and revenue-per-employee figures are interview estimates, not proof that particular jobs are unnecessary.

Asked whether investors are abandoning software companies, Freda says valuation can fall even without a firm prediction that every company will deteriorate: greater uncertainty can raise the discount rate applied to future cash flows. She also sees gaps between some private valuations and public-market expectations.

A falling share price also affects stock compensation: maintaining the same value for employees may require issuing more shares. That creates tradeoffs among financing, retention, and dilution.

Which software is more vulnerable? Freda believes single-purpose tools whose value rests mainly in the interface, or applications handling highly structured data, may face earlier disruption. She contrasts a project management tool with Excel, whose cells have diverse meanings and uses and may be harder to replace.

She has revised part of her view on office software. AI can quickly generate a web-based presentation, but PowerPoint and similar tools may retain value as long as someone in a workflow needs to make direct, precise manual edits.

Data warehouses are not necessarily completely safe either. She describes comparing stock valuations with a model, tools, and data interfaces faster than a workflow that previously required a warehouse and queries. This is a specific experience, not proof that all data systems will be replaced.

One software opportunity, she argues, is capturing information that previously disappeared from business decisions. A CRM might record a customer discount without recording why it was chosen, who objected, or what the customer wanted; AI may help process that more original, unstructured material.

She advocates rethinking software for agents, drawing on her experience connecting chat tools and the needs for persistent, real-time communication. Interfaces built for people may not suit agents, though products still need to address interfaces and permissions.

On organizational change, Freda distinguishes technological diffusion from economic diffusion: model capabilities can improve quickly, while businesses may take years to absorb them and turn them into revenue and productivity. To her, studying organization also means asking when AI creates economic value.

She uses factory electrification as a historical analogy: at first, factories might simply put electric motors where steam engines had stood and retain the old layout; efficiency improved more visibly after production was redesigned. New technology does not automatically deliver immediate economic value, she argues.

She also cites the computer productivity paradox: computers entered offices before clear macroeconomic productivity gains, and businesses later redesigned operations around databases, networks, and supply chains. Many companies today, she believes, are still adding AI to old workflows.

Looking at organizations, she says hierarchy is not only about power: it also transmits, compresses, and interprets information up and down. Meetings, progress updates, and quarterly alignment partly address the cost of communicating between people.

In traditional product development, requirements pass from product manager to designer, developer, testing, and sales, with explanation and waiting at every step. If AI accelerates one stage, Freda says, the bottleneck may move to the next, so organizations need to rethink the entire process rather than optimize coding alone.

She imagines some teams changing from a relay race to a basketball game: a cross-functional squad of three to five could work in parallel, make most decisions itself, and escalate only major directional questions. Testing could be embedded in development, and product managers may need broader skills.

Zhang asks for concrete examples of successful transformation. Freda admits she is still looking and has not found a widely agreed answer. She speculates that businesses could capture more customer conversations and historical data for agents to help with decisions, but this remains a vision of future work.

Back in investing, Freda says practitioners spend much time finding information, cleaning data, comparing expectations, and understanding market positions. She asks whether an agent given sufficiently clean data and clear trading objectives could do better than a person. She is optimistic, while acknowledging that goals and preferences must first be defined.

One practical difficulty, she says, is fragmented data: even seemingly standardized financial information may require connecting a dozen or more providers. Another is understanding how different market participants make decisions.

She discusses the different roles of quantitative traders, retail investors, and institutions and believes retail investors deserve more attention. Her estimates of trading shares and generalizations about participant behavior are personal estimates, not universal rules or trading guidance.

She believes AI could narrow the information gap between retail and institutional investors and bring fundamental and quantitative research closer. News may be priced in faster, making an investor’s advantage depend more on understanding broad trends than simply seeing information first.

Will markets become more efficient or more volatile? Freda thinks both are possible: information could enter prices faster, while many agents reacting to the same signal could make trades more crowded. She presents these as hypotheses about changes in market structure.

In application startups, she surveys revenue in coding, healthcare, legal services, customer service, video generation, inference infrastructure, and chips. The figures are rough industry estimates from the interview and should not be compared as financial data on a uniform basis.

She has repeatedly revised her view of competition between model and application companies: she initially feared models would consume applications, then saw rapid application revenue growth. Recently, stronger tool use, connectors, and memory have renewed her concern that model companies will enter more application areas.

Some vertical application companies stress compliance, industry expertise, or their ability to handle long-tail needs. Freda thinks those can differentiate them, but relying only on compliance or last-mile work may still leave long-term profits and value capture vulnerable.

Even so, she says she still invests in application companies. She observes that some firms have set up growth funds that provide liquidity to early-stage funds and that expectations of listings and acquisitions have left some investors with capital; financing conditions can change.

On the popularity of Neo Labs in Silicon Valley, she says many researchers have left established AI labs to start new teams. Investors, she believes, seek returns outside the high valuations of model companies and consider possible exits through acquisitions or talent hiring.

Whether a fund can invest in this category also depends on its size and portfolio construction: each lab may raise a large amount, making it hard to cover every team while undesirable to bet on only one. Her example of spreading investment across several labs illustrates capital constraints, not a general allocation recommendation.

She is also watching companies that reduce inference costs and build infrastructure for agents. Email, phones, browsers, identity, payments, and compliance were largely designed around human usage; agents’ different request rates and interaction patterns could create new product needs.

In agentic commerce, she is more optimistic about complex business-to-business transactions. Consumers may enjoy browsing, while cross-border procurement involves quantities, pricing, certification, import and export compliance, and payment terms. She thinks agents may help with such communication-heavy tasks.

On autonomous driving and robotics, Freda admits progress has been slower than she expected. She is interested in whether open models can make autonomous driving capabilities available to more automakers and notes efforts to help systems understand complex road situations.

She is more cautious about robotics: improvements from adding data are different from a clear scaling law that can be extrapolated and generalizes across tasks and environments. Hardware manufacturing also presents challenges, she says, making investment judgments highly uncertain.

Revisiting her previous assessment of US stocks, Freda believes she was right that model-company revenue would become a market focus but wrong to expect capital to shift gradually from semiconductors to AI applications. She openly treats the latter as a mistake to revisit.

She worries that major technology companies, especially cloud providers, could face pressure on free cash flow after sustained capital spending. Although cloud revenue is growing, she thinks increased competition and a redistribution of value across the computing stack could change the economics.

On chips and advertising, she discusses cloud providers commercializing their own chips and new entrants potentially reaching spending outside conventional ad budgets. Retail promotion fees illustrate why advertising need not be a purely fixed-sum contest among existing platforms.

For potential large IPOs, Freda estimates market capacity using hypothetical sizes and sees capital rotation as the more important question: new listings could attract investment and put pressure on existing large technology stocks. This is scenario analysis, not a prediction that the listings will occur or produce a particular outcome.

She estimates that layoffs may affect earnings per share differently across companies: software firms with high labor costs could be more sensitive than some large technology companies. Social effects, AI spending efficiency, and returns on capital investment could also influence market sentiment.

On employment and deflation, Freda resists a definitive conclusion. Past technological revolutions repeatedly prompted employment fears, but their speed and breadth differed; governments might also respond with fiscal or other policies if disruption is severe.

She reviews historical cases including mechanized weaving, declining agricultural employment, and the automation of telephone operators, arguing that humanity has experienced many major transitions. AI may spread faster, but its employment effects are hard to predict precisely; she reminds herself to stay open and prepare for different scenarios.

Freda is optimistic about long-term productivity and sees opportunities to start companies: tools lower the barrier to building products, while needs and competitive structures are changing. She cautions founders against mechanically applying her framework for mature companies to early-stage ventures.

The discussion turns personal. Freda admits to considerable anxiety as new models, tools, and workflows keep appearing. As a technology investor she feels she should try them herself, yet often worries about falling behind. She also recognizes that comparison on media and social networks can amplify anxiety.

She recounts crying after struggling to install a tool late at night: one technical problem triggered fears that the era might leave her behind. But change also brings excitement and opportunity, she says, and sharing such experiences might help others feel less alone.

She connects anxiety to caring deeply and fearing an uncertain future. Layoffs, economic effects, and competition are hard to predict completely, so she wants to form opinions while leaving room to adjust them as facts change.

Zhang asks what can be relied on in an uncertain era. Freda believes greater intelligence and efficiency are beneficial over the long term and may make tools available to more people, while acknowledging that the redistribution of resources may be painful.

Freda says she once enjoyed meeting industry experts, but now many conversations aimed at exchanging information can begin with a question to AI. She does not consider human conversation unimportant; she is distinguishing knowledge gathering from other kinds of connection.

She recalls an unplanned conversation with a friend about the courage to act, regrets, and what really matters. Although it provided little new factual information, she felt it was one of her most meaningful conversations in a long time.

She therefore tends to entrust some information and analysis to AI and make meetings simpler: observing one another, sharing personal journeys, and discussing sincere or even uncomfortable topics. Zhang says the program also hopes to build that kind of connection with listeners.

Finally, Freda reflects on the name “Language Is World.” Language both conveys information and creates human connections; AI may take on more of the former, she believes, while the latter becomes more precious.