Cliff Sosin: What Remains Valuable When AI Intelligence Gets Cheap

Cliff Sosin says he can tell two plausible stories about AI infrastructure: demand might grow so quickly that capacity stays scarce, or falling token costs and slower adoption could leave the industry overbuilt. His point is the difficulty of forecasting either outcome with confidence.

Patrick O’Shaughnessy asks Sosin to explain his phrase “just intelligence.” Sosin challenges the assumption that intelligence increases without meaningful limit until it can solve everything. He sees intelligence as a tool for making predictions from available information.

Sosin compares better intelligence to filling in gaps between scattered facts. It can be powerful where a problem is amenable to inference, but he argues that many questions involve complex, changing systems rather than a neat closed-form answer.

A simulation can start with imperfect information or accumulate errors as it runs. Sosin argues that, in systems such as weather or social interactions, this limits the detail and horizon of reliable predictions regardless of how capable the reasoning system is.

To learn more, he says, a system must gather new facts and run experiments: it has to touch reality. At the same time, he expects intelligence itself to become dramatically cheaper, making capable tools available for many ordinary tasks.

O’Shaughnessy asks why complex systems constrain even highly capable AI. Sosin recounts the weather-model story in which rounded initial conditions produced a sharply different forecast after several simulated days, illustrating sensitivity to initial conditions.

The butterfly-effect example does not make simulations useless. Sosin says approximations can be extremely valuable, but argues that precision and predictability remain constrained by incomplete measurements and the behavior of complex systems.

In Sosin’s scenario, people and companies could have teams of highly capable agents working on problems continuously. He expects this to affect businesses differently, because some help customers overcome a lack of expertise while others benefit from customers not optimizing.

For example, a consumer may accept a decent bank rate rather than spend time searching for a slightly better one. An agent that compares options continuously could make the small savings worth pursuing, changing the economics of businesses that rely on inertia.

Sosin describes some software as a fixed piece of intelligence embedded in a product. If general-purpose models make such functions much cheaper to reproduce, the value of the software business could be challenged, even though the specific outcome varies by product.

His investment question is therefore: if everyone has capable intelligence, what remains scarce? He names data, attention, emotional attachment, physical proximity and scale. He also includes regulatory protection as a possible advantage, without presenting it as a social good.

Meta is his example of a company with complements to intelligence. He describes its core product as personalized television: a large pool of content presented to an audience whose preferences can be learned and matched.

Sosin offers a deliberately simplified thought experiment about Meta’s spending. Data centers, inference, engineering and algorithms all contribute to producing or applying intelligence. He explicitly cautions that this reduction should not be interpreted literally.

If intelligence gets cheaper, Sosin suggests creators could produce more content, recommendations could improve and advertisers could target more effectively. Meta might reinvest the savings rather than eliminating costs. These are hypothetical consequences, not projections established by the interview.

The key assets in his argument are attention and data: AI can help a platform use them more effectively, while cheap intelligence alone does not automatically make an audience disappear. The two can reinforce each other as recommendations improve.

Sosin also mentions Grindr as a hypothetical example of a smaller company with attention and data but limited engineering resources. He says he does not own the stock; the thought experiment is what the business might build if technical capacity were much cheaper.

O’Shaughnessy challenges the commoditization thesis: frontier labs might retain models that are materially better and use distillation to serve both expensive and cheap applications. Sosin agrees that protected superiority and lower costs could make such a lab an excellent business.

His uncertainty is about how much demand requires the very best model. As capability improves, tasks that once needed frontier models may be served by cheaper alternatives. New difficult tasks may open up, but it is hard to know how large they will be.

Sosin proposes a possible diamond-shaped demand frontier: early improvements expand the set of useful applications, while the demand for ever more exceptional intelligence might eventually narrow. He offers this as a possibility rather than a prediction he can verify.

O’Shaughnessy raises Jevons paradox: lower prices might induce far more use. He also asks whether models could generate their own new questions and thus create demand for intelligence, rather than relying only on people to identify uses.

Sosin agrees that falling prices can increase consumption. But he argues that total market revenue can eventually fall if price declines outpace the rise in quantity. He uses communications bandwidth as an analogy, while acknowledging AI need not follow the same path.

On scientific discovery, Sosin expects models to connect existing research and suggest useful experiments. He nevertheless expects physical experiments to remain necessary for many questions: better hypotheses do not automatically provide the missing observations.

Sosin describes using several model subscriptions behind a tool that builds things for him. He no longer always knows which model is doing each task; to him this is an example of a user experience that can obscure the identity of the underlying model.

O’Shaughnessy asks about automated laboratories and robotics that connect AI to real-world experiments. Sosin welcomes the idea, but says even automated experiments encounter physical speed and capacity limits. Breeding and observing animals, for example, still takes time.

Automation can improve the selection of experiments and accelerate learning without making success certain. Sosin regards those improvements as significant, while rejecting the inference that intelligence alone instantly removes the need to test claims against the world.

The conversation returns to infrastructure investing. Sosin says forecasting both inference demand and improvements in supply is extraordinarily hard. Even a correct overall forecast would not identify which layer of the value chain captures the returns.

He gives personal examples of generating educational videos and analyzing exercise form. Those uses are appealing at very low cost, he says, but he might not pay much for each one. Price sensitivity complicates any estimate of total consumer spending.

O’Shaughnessy asks why so many talented entrepreneurs are building supply if the economics are uncertain. Sosin says a large, important infrastructure project necessarily attracts both talent and capital; their presence demonstrates its importance, not that every investment will pay off.

Rather than choosing which infrastructure provider wins, Sosin tries to think about the businesses that may emerge once the infrastructure exists. He compares the question to looking beyond railroad builders toward the industries enabled by railroads.

He describes a capital cycle in which society directs workers and resources toward an investment in future capacity. Construction can create value but also raise costs elsewhere. Because the right amount is unknown and projects have lead times, the buildout may overshoot.

Rapid technical change adds another uncertainty: designs and cost advantages may shift while infrastructure is being built. Sosin speculates about AI’s effects on chip design, but does not claim that current leaders have already lost their advantages.

Asked about TSMC, he says its physical manufacturing capabilities and accumulated process knowledge may be harder to replicate with intelligence alone. He also notes that intense demand encourages others to try to build alternatives, which creates its own competitive uncertainty.

O’Shaughnessy asks whether physical proximity and local density become more valuable when intelligence is abundant. Sosin answers with a stylized waste-collection example, using it to show how agents could change switching and costs.

Today a household may not bother changing waste collectors to save a small amount. An agent could compare offers and switch on its behalf. If more customers on one street choose the same provider, route density could reduce that provider’s cost further.

The result could be more volume for an efficient operator but lower margins from tougher price competition. Sosin stresses that these effects can point in different directions within the same business; there is no single universal outcome.

Asked how agents might change his own investing, Sosin expects humans and models to complement one another for a while, like teams of humans and computers in freestyle chess. Human experience and information absent from a model can still help judgment.

Over a longer horizon, he can imagine the gap narrowing until a model’s forecast is better without human adjustment. He does not claim that this would make the future perfectly predictable; rather, the best available distribution of outcomes might come from a model.

The hosts discuss whether AI could replace prices as an organizing mechanism. Sosin argues that price combines a signal about scarcity and value with an incentive to act, making it a powerful way to coordinate economic activity even when intelligence improves.

O’Shaughnessy distinguishes prompting, scheduled agents and continuously proactive agents. Sosin thinks widespread agents could make prices more responsive, with buyers switching more readily and sellers updating offers as their costs change.

The discussion turns to American capital markets. Sosin praises their ability to direct large amounts of money and labor toward speculative AI infrastructure quickly. He treats the precise amount that should be invested as unknown.

Sosin says capital allocation is ultimately tied to people and physical resources, not just financial numbers. Workers building a data center are doing something different from what they would otherwise have done, and that shift can enable future production.

He expects some bad behavior to surface in any large investment cycle, but argues that governance and access to diversified public markets are underappreciated. Individual investors can hold shares in many companies without running each one themselves.

Asked why the United States has been innovative, Sosin offers a personal theory combining culture, institutions and a speculative genetic component. The genetic claim is his conjecture, not evidence established in this conversation. He also points to institutions that can make risk-taking and failure easier.

On interest rates, Sosin says he is not a macroeconomic expert. He suggests that a large investment in future productive capacity can require more saving today, but presents this as one part of a complicated picture, not a complete explanation of rates.

O’Shaughnessy asks how Sosin would advise companies in his portfolio to use AI. Sosin starts with hands-on experimentation: executives and employees, not only developers, should try building things so they understand what the tools can do.

His next suggestion is to work toward read-only access to company information for an authorized agent. That requires system-by-system integration and careful permissions. Broader access, write capabilities and changes to systems would require additional controls and processes.

Sosin acknowledges that these apparently simple steps can take years. Established companies have legacy systems to integrate, while new companies can start with fewer constraints. Even so, a startup needs an enduring advantage if competitors can use the same tools.

The hosts ask what durable advantages may be underrated. Sosin is unsure that any one category is systematically mispriced. He considers whether cheaper intelligence could reduce margins by increasing consumer choice and competition.

He notes that similar predictions accompanied the early internet, yet aggregation created powerful businesses. If an agent compares many options instead of relying on the first search result, however, he thinks the mechanics of some aggregation advantages could change.

Conversely, cheap coordination and software might make it easier to build integrated businesses across an entire value chain. Sosin uses pet care as a hypothetical example and points to Carvana as a company that built a more integrated model before current AI tools.

The effect on inequality is also ambiguous in his account. Lower costs and access to expertise could benefit consumers broadly, while people whose work is commoditized could lose income. At the same time, AI might amplify the returns to initiative and ownership.

Sosin returns to the two AI infrastructure stories. His own heavy use makes a shortage seem plausible; better efficiency, falling prices and slower adoption by others make overcapacity plausible too. He says he can make a convincing case for either.

O’Shaughnessy describes a friend who selects investments himself but lets an AI system determine position sizes. The friend reports stronger results so far. Sosin cautions that a short experiment cannot establish a long-term investing record.

Sosin expects automation to advance from shorter investment horizons toward longer ones, and sees room for quantitative strategies combined with human qualitative analysis. Models add possibilities, but their effectiveness must be judged over an appropriate period.

Asked about his highly concentrated portfolio, Sosin explains his confidence in Carvana’s infrastructure and brand. He recognizes the tradeoff: concentration involves unusual risk, and histories of great concentrated winners leave out those who failed.

On education, Sosin argues that knowledge still matters when models are available. A person who understands a field can ask better questions, direct a model more effectively and evaluate its output. Without that foundation, communication itself becomes harder.

He therefore favors teaching children foundational skills they can use independently, even if tools can perform the calculations. Knowing basic concepts allows a student to understand and direct more advanced work with a model later.

At the same time, Sosin wants students to undertake more ambitious projects with AI rather than merely submitting generated essays. Teachers should ask students to explain and defend the work, making sure they understand and take ownership of it.

O’Shaughnessy closes by thanking Sosin for offering a perspective that differs from forecasts of unlimited AI demand and perpetual infrastructure shortages. The discussion leaves the central questions open rather than declaring a single outcome certain.