Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497]

Meta Platforms

Across the shows: 48 speakers in other episodes, 44 bullish and 23 bearish calls, past 30 days

• Cliff Sosin · CAS Investment PartnersBullish

  Meta's data and attention compound and are not destroyed by cheap intelligence, which should cut its costs and improve its ads.

  Meta has about $250B of revenue and spends about $250B a year, mostly on intelligence: data centers, buying inference, and engineers writing algorithms and code. Since the price of intelligence is falling sharply, the cost base should shrink. On the revenue side, creators get far better tools and advertisers can build better-targeted ads, which Meta can target more precisely. Meta's core inputs, data and attention, are complements to intelligence rather than substitutes, and they compound each other. Cheap intelligence will not reduce people's interest in watching short videos, and it may win Meta time from other activities. Sosin says he does not own Meta.

  “All of this is downstream of what they possess, which is not destroyed by ubiquitous intelligence, which is data and attention.”at 11:47

Grindr

• Cliff Sosin · CAS Investment PartnersBullish

  Grindr has lots of attention and data, so it could do far more with cheap, abundant intelligence.

  Grindr is a dating app that has been limited by its size. Its team and budget are finite, so it cannot build everything it could imagine. It has a large base of attention and data, which are the complements to intelligence. Using a mental model in which it could have 10,000 of Google's best engineers for five years, Sosin thinks Grindr could do much more. He hedges the view: the company is only a theoretical beneficiary, and he says he does not own it.

  “And it turns out they've got lots of attention, and they've got lots of data.”at 12:24

NVIDIA

Across the shows: 40 speakers in other episodes, 34 bullish and 13 bearish calls, past 30 days

• Cliff Sosin · CAS Investment PartnersBearish

  AI could commoditize Nvidia's chip moat, since its designs are essentially intelligence that others could replicate.

  Sosin argues that AI could commoditize even Nvidia. He says Nvidia quite literally makes files that get sent to Taiwan, and those files are very hard to make. But Nvidia's value is in a large number of people thinking hard about the problem, which is intelligence, and so in theory someone else could make them. He adds that as the frontier recedes, the cost of making something falls sharply. He treats this as a general risk rather than a firm call, and he offers it as a concern about Nvidia's moat.

  “Like NVIDIA quite literally makes flat files that they send to Taiwan. Their flat files are very hard to make. But what they have is a lot of people who think really hard about it. And so in theory, you could make them.”at 28:54

Taiwan Semiconductor Manufacturing Company

• Cliff Sosin · CAS Investment PartnersBullish

  TSMC's process moat is hard to replicate and less exposed to AI, though its success invites copying.

  TSMC did a very hard thing that depends on process know-how built over many years, which is very hard to replicate. Its advantages are physical and tacit, and they are much less amenable to AI because many of them are not legible to a model. Sosin says its attributes are better than most businesses on this point, though he does not claim to be a TSMC expert. He offers a mixed view: because demand is so large, competitors are trying to copy its moat, which he calls a poor fact for TSMC.

  “They did this super hard thing that requires all this process power built over many years, and it's super hard to replicate, and much less amenable to AI just because it is so physical.”at 29:26

Carvana Co.

• Cliff Sosin · CAS Investment PartnersBullish

  Carvana's differentiated infrastructure and brand make a very concentrated position comfortable; he likes it a lot.

  Sosin holds roughly 80% of his portfolio in Carvana and says he is comfortable with that because he sees its infrastructure, brand and overall setup as so differentiated that few scenarios work out badly. He argues that a portfolio-construction approach that spreads the position would have worked in general, but he is not looking for average results. He says he would not be upset if the stock rose a lot and he felt compelled to diversify, framing the choice as a trade-off between expected returns and knowingly taking an unusual risk. He compares it to finding Walmart or Home Depot early, where full-size positions would have been warranted.

  “Were it not for my sense that Carvana's infrastructure and brand and everything else is just so differentiated that there are few scenarios where it works out really badly, then I wouldn't feel so comfortable owning so much of it.”at 55:43

AI: overall outlook

• Cliff Sosin · CAS Investment Partners

  Intelligence cannot predict complex systems; progress past that needs experiments that take time.

  Intelligence is like a ratio, not an unbounded scalar. It helps interpolate between sparse facts about the universe, and it reaches a perfect answer only in domains like mathematics. Most important questions involve complex dynamic systems, like weather or how groups of people interact. Their small initial errors compound, as the butterfly-effect weather-model story shows, so simulation cannot give a closed-form answer, and intelligence has limited power there. Beyond the subset of predictions that a very smart system can make, progress needs more information, which requires experiments and touching reality. That bounds how far more intelligence can go, however much it grows.

  “To get past that, you then need to go out and get more information about the world, and that just takes time. You have to do experiments. Have to touch reality.”at 3:33

• Cliff Sosin · CAS Investment Partners

  Human judgment adds value to models today, but models' outcome distributions may soon be best.

  Sosin says humans currently add value because they understand some things models do not, from lived experience and from information not on the internet, and because humans can use models to make better judgments together. He also argues the future is intrinsically unpredictable, so models may not be able to predict it, but their distribution of outcomes may be the best that can be done. He says the gap between humans and models could diminish to the point where human judgment adds nothing, comparing it to adding one's judgment to a seven-day weather forecast.

  “I don't think models will able to predict the future... But the model's distribution of outcomes may be the best that you can do.”at 34:13

• Cliff Sosin · CAS Investment Partners

  AI agents will make prices move more often and sharpen markets by eroding price stickiness.

  Sosin argues that prices are sticky and many things are priced lazily, such as deposits. If everyone has agents constantly optimizing on their behalf, companies will rerun their financials continually and bid each customer for more than marginal cost, and agents will switch to the lowest-cost provider. He illustrates with garbage collection, where the price would vary by day or week as input costs change, instead of a fixed annual subscription with a small rise. He expects much more accurate and rapidly moving prices.

  “So I think if everyone... has agents who are working on their behalf, and constantly optimizing, ... I think the price of garbage collection will move around a lot more.”at 36:30

• Cliff Sosin · CAS Investment Partners

  AI will take over investing first at short timescales, since long horizons lack a verification loop.

  Sosin expects AI to encroach into investing starting at shorter timescales and then move to longer ones. The key constraint is that there is no verification loop. A friend's approach can be judged on short-term results, but not on twenty-year results. He suspects that investors working on timescales from seconds to months will increasingly delegate to models. He thinks many multi-strategy funds already combine quantitative strategies with qualitative human analysis, and that models add more dimensions of freedom to how investors can succeed.

  “I think it's gonna start at shorter timescales and work its way up to longer timescales... Because one of the challenges is that there's no verification loop.”at 53:58

AI infrastructure & data centers

• Cliff Sosin · CAS Investment Partners

  AI inference demand is too hard to forecast, and demand for AI is highly price-elastic.

  Sosin says he cannot forecast supply or demand within several orders of magnitude. How much more inference people use as intelligence gets better and cheaper is very hard to predict, and so is how much cheaper inference will get. He gives a personal example: he would make a video explainer for his daughter at small cost, but not at a real price, and he is less price-sensitive than most people, so AI demand will be very price-sensitive. He also argues that markets are U-shaped. Jevons paradox is a narrow example, and after enough price declines, the market shrinks. Fiber optics shows this: information-moving prices fell so far that the market stopped growing and then shrank, with Cogent a few-billion-dollar company that carries much of the internet's traffic.

  “I can't forecast supply or demand within several orders of magnitude. But this idea of how much more inference will people use as intelligence gets better and cheaper is a really, really hard thing.”at 24:12

• Cliff Sosin · CAS Investment Partners

  The AI infrastructure build-out will likely overshoot, and the last builder to stop will regret it.

  Sosin likens American society to a household saving for a house: resources are going into the AI build-out, electricians and plumbers are being diverted, and costs rise elsewhere in the economy. He expects the build-out to deliver a lot of value, but no one really knows where it leads, so the industry will keep pushing until it overdoes it. Because projects have long lead times and the situation is a winner's curse, the last person to stop building will regret it. He also argues that the technology is more dynamic than past infrastructure build-outs, which makes it hard to know where to invest and whether investors will earn money.

  “And undoubtedly, we're gonna get a lot of value for it. And we're gonna keep pushing it because no one really knows. We're gonna keep pushing it until eventually we overdo it.”at 28:09

• Cliff Sosin · CAS Investment Partners

  US capital markets redeployed about 3% of GDP into AI build-out remarkably fast, a notable feat.

  Sosin argues that US capital markets have redeployed roughly 3% of GDP into a highly speculative venture, the AI build-out, almost on a dime. He calls this remarkable and says no one else in the world has been able to move resources at this scale and speed. He ties it to real-world activity, such as people building data centers in Pennsylvania, and argues that the ability to move investment quickly matters because the inability to move capital can bind productivity growth even amid rapid technological progress. He adds that the scale could still be wrong, since it might be two or six percent.

  “I think we just redeployed... what is it, 3% of GDP right into this highly speculative venture basically on the turn of a dime. Just this remarkable thing. Of course, who knows? Maybe it should be two, maybe it should be six.”at 37:55

• Cliff Sosin · CAS Investment Partners

  The AI build-out could be overbuilt because cost per token will collapse, though demand may be enormous.

  Sosin describes two opposing views with equal confidence. On one side, he uses a huge number of tokens himself, sees how he could use many more, and thinks everyone will use as much as he does, so capacity will be very short and the build-out will make fortunes. On the other side, he notes that the amount he earns per token is going up, that he is an unusual user, that people will adapt more slowly than he expects, and that the cost per token will collapse. That second story suggests the build-out may be totally overbuilt. He says he can tell both stories with equal certainty.

  “...the cost per token is gonna collapse. And so maybe, actually, we're totally overbuilt.”at 51:54

Software & SaaS

• Cliff Sosin · CAS Investment Partners

  AI could be bad for certain software companies, since their products are crystallized intelligence.

  Some companies produce crystallized intelligence, meaning fixed artifacts that embody a set of decisions. A general-purpose AI model is a general intelligence tool that can rebuild those artifacts. Many software tools are hardened projections of intelligence, so they face a risk from cheap general intelligence. Sosin says he is not the first to make this point, and he presents it as a logical implication of treating intelligence as cheap and ubiquitous.

  “Then there's also companies where what they produce fundamentally is some sort of artifact that is just some sort of crystallized intelligence. I'm by far not the first person to realize that AI could be bad for certain software companies.”at 7:17

• Cliff Sosin · CAS Investment Partners

  Cheap intelligence will squeeze software companies that extract rents, lowering costs for their customers.

  Sosin argues that many economic rents are held by intermediaries, using a software company that charges bowling alleys as an example. As intelligence becomes cheap and ubiquitous, those rent-taking companies get smoked, which is bad for them but lowers the cost of bowling for consumers. He says this leads to a flattening effect on income, because lower-income people spend a larger share of their income on such goods. He presents this as the benefit side of the AI cost-deflation story.

  “One is there's all these rent takers in the economy. Think you're the software company that charges bowling alleys, and those guys all get smoked, which sucks for them, but it does lower the cost of bowling.”at 50:02

AI pricing & unit economics

• Cliff Sosin · CAS Investment Partners

  The price of intelligence should fall about 100x over the next five years.

  Sosin says the price of intelligence is falling by a factor of about 100 over the next five years. He treats that as a first-principles input: if intelligence gets that cheap, the cost base of companies that buy it, such as Meta with its roughly $250B of spending, largely goes away. He cites the broader trend of rapidly falling model costs as evidence. He describes the 100x figure as a rough, approximate estimate.

  “the price of intelligence is falling by a factor of a 100 or something over the next five years, first principles,”at 10:37

Frontier AI labs & model race

• Cliff Sosin · CAS Investment Partners

  Frontier labs' pricing power erodes as models commoditize; only a narrow set of hard tasks needs the frontier.

  Each new frontier model unlocks a larger addressable market, but once evaluations saturate, the tasks commoditize. Customer-service chatbots, for example, would never use a frontier model. A wave of hard tasks needs the frontier, and behind it, most work does not. Cheap Chinese models have shown rapid commoditization at every step, and token costs have collapsed. Even if a lab is the cost leader, the market for tokens behind the frontier may be small, and the margin for a cost leader is set in a shrinking commodity market. Demand may also become rarefied: few use cases need very elite intelligence, and many need good intelligence, which is competitive and commoditized. Sosin also says his friends who are frontier bulls may have an information advantage, and he finds the claim that labs can keep producing cheaper tokens hard to judge.

  “Certainly, the trend has been rapid commoditization of these things if you were to look at the rise of Chinese models in every step.”at 14:33

Consumer AI products

• Cliff Sosin · CAS Investment Partners

  AI agents will make switching easy, shifting volume to low-cost providers at lower margins.

  If everyone has an AI agent, many decisions will be offloaded to it, because it finds the pennies consumers leave on the table. Sosin contrasts choices such as a bank deposit, which an agent can fully decide, with picking a sweater, which needs personal judgment. For services like garbage collection, an agent can negotiate a small saving and switch the household provider, and the first-order effect is a move to the low-cost provider. The second-order effect is that the low-cost provider then gets more customers on a street and can roll up the market, creating a better business that wins more volume at much lower margins.

  “So the first order effect is that you move to the low cost provider. ... That actually creates in some sense a better business, where the low cost guy wins a bunch more volume at much lower margins.”at 31:35

US Treasuries & yields

• Cliff Sosin · CAS Investment Partners

  Long-term rates are higher partly because the AI build-out requires paying people to save for the future.

  Sosin says he is not a macro expert but offers one piece of the puzzle. Building a society's future capacity means consuming less today, and the AI build-out is entirely about future consumption: a plumber building a data center is not producing present consumption. He argues that one reason rates are higher is that society has to induce people to save for that future, and the way to do that is to pay them through higher rates. He calls this one piece of a very complex puzzle.

  “this idea that rates are higher because we're inducing people to save for the future. And the way you do that is you pay them.”at 42:27

Venture capital & startups

• Cliff Sosin · CAS Investment Partners

  It is the best time to start a company, with cheap capital and AI tools letting startups go far fast.

  Asked whether it is the best time to start companies, Sosin agrees. Capital markets will fund highly experimental projects at fairly low cost of capital, and AI tools let a company started without legacy baggage go very far very fast. He also notes that AI makes running a company easier, since you can do much more with far fewer people. He adds a caveat: if something is easy for you, it is easy for everyone else, so a startup needs a wedge, such as lock-in or proprietary data, to create enduring value. He says the best outcomes often look like they make no sense at first.

  “it's the best time to start a company until tomorrow.”at 45:09

Earnings & corporate profits

• Cliff Sosin · CAS Investment Partners

  Aggregate corporate margins likely fall with AI, though full-stack vertical businesses could raise margins.

  Sosin asks whether aggregate corporate margins go up or down and says the themes discussed would argue for down, because many businesses earn margin through information advantages that consumer surplus erodes. He cautions that the base rate from 1990s predictions may be poor and that he could be wrong. He then argues the picture is mixed: cheap ubiquitous intelligence lets businesses build complete vertical stacks, such as a full-stack pet care company spanning food, insurance, hospitals and vets, that were historically split across layers because of coordination and principal-agent problems. Such businesses could be higher margin and more efficient, with more consumer surplus. He cites Carvana as an early example of this kind of build.

  “And I think everything we're talking about would argue down.”at 47:53

Big Tech platforms & digital ads

• Cliff Sosin · CAS Investment Partners

  Aggregation theory breaks down with agents, since ranking first stops mattering when agents scan all results.

  Sosin argues that aggregation theory relies on consumers looking only at the first result on Google, which lets the aggregator capture value. If agents review everything in search, being ranked first no longer matters. Instead, quality, price and service become more salient. He suggests that this is one reason the 1990s internet research reports that predicted margin shifts turned out wrong, since aggregation economics were the main reason they missed. He concludes that this weakens the aggregator advantage in an agent world, though he presents it as a view that could be wrong.

  “a lot of aggregation theory breaks down because that basically relies on the idea that consumers only look at the first thing on Google. But it turns out if your agent looks at everything on search, then being ranked first doesn't matter.”at 48:24

AI & jobs

• Cliff Sosin · CAS Investment Partners

  AI will commoditize professional intelligence work, pushing high earners down to their next-best alternative.

  Sosin describes a professional class whose role is essentially intelligence: taking large bodies of information and applying it to circumstances in a fairly straightforward way. If those tasks are commoditized, everyone gets the best available intelligence, which is good for consumers and especially for people who cannot afford top advice today. But the professionals who earned extra rents because they were smarter then have to fall to their next-best alternative. This mechanism is what he uses to argue that AI could flatten the income curve.

  “they take big corpuses of information and apply them to certain circumstances in a fairly straightforward way. And if those all get commoditized... They then have to drop to their next best alternative.”at 50:40

Wealth, inequality & billionaires

• Cliff Sosin · CAS Investment Partners

  Cheap intelligence could flatten income by eroding rents, or steepen it by rewarding initiative; he sees both.

  Sosin presents two stories. In the first, AI concentrates income because a few labs and data holders earn increasing returns. In the counter-argument, cheap ubiquitous intelligence erodes rent-takers. Software companies that charge bowling alleys get squeezed, lowering costs for consumers, and lower-income households spend a larger share of income on such goods, which flattens the distribution. Professional intelligence roles also get commoditized, pushing those earners to their next-best alternative. The opposite story is that AI acts as a bicycle for the mind for smart, motivated people, which could raise returns to that group and steepen the curve. He says he can tell both stories with equal certainty.

  “Maybe that actually causes increasing returns to that subset of people, and it steepens the curve.”at 51:31

Education

• Cliff Sosin · CAS Investment Partners

  Kids should learn foundational skills without models first, so they can direct models well later.

  Sosin argues that more intelligence and more crystallized knowledge make it easier to communicate with and direct a model, since you cannot ask a model to assume a stochastic process unless you know what one is. So he thinks children should spend substantial time separated from models, learning foundational skills such as multiplication that build on each other, so they can understand what the model is doing. He criticizes teachers who block AI for essays, arguing that students should be given harder, more ambitious work that uses models fully, then be made to justify and own the output.

  “And so increasingly, I'm of the view that we should spend a bunch of time with kids basically separated from these models,”at 58:40