I don't know where value capture is going to happen in AI

Do you?

Let me begin by stating upfront that these thoughts are deeply half-baked, and I don't claim any special expertise or differentiated insight into the latest and greatest in AI and LLMs. These are just some things I've been thinking about.

Outside of semiconductors, I'm having a hard time figuring out which part of the AI "stack" is going to gobble up excess profits over the next decade. Every bull thesis I come across feels somewhere in the realm of uncertain, lacking, or reliant on imperfect analogies. That doesn't mean they're wrong, but it leaves me feeling a bit… unsettled.

Let's go through a few areas of the market so you can see what I mean.

Model providers.

OpenAI is currently doing more than $3 billion of annual revenue. That's a lot, and certainly puts them amongst the fastest-growing technology growth stories of all time. But this eye-watering figure leaves a few things unsaid:

• What sort of gross margins are these offerings being sold at? How much capital expenditure was required to stand up these models?

• How are these revenues going to hold up in the face of incredibly rapid reductions in model pricing?

• What about competition from open-source or free, locally hosted models?

One thing that really worries me is the manner in which different foundation models are embedded (via API) into various web applications — often there is a simple drop-down where you can sub in whichever model you like, whether it's GPT, Bard, Claude, LLaMA2, whatever.

This worries me because on my right shoulder at this very moment is a devil who looks exactly like my freshman year Econ 101 teacher, whispering in my ear that substitute goods lead to price competition, not excess profits.

Source

So… I dunno. Will OpenAI eventually capture enough value to justify its $80 billion valuation? Or will any other model provider? I wouldn't bet against it, but I suspect that the business models that thrive in this new era will not be the ones that are currently in practice.

AI applications.

I suspect you don't need me to spell out the bear case: Most of these companies are thin wrappers on top of foundation models that provide no moats against competition from other solutions or the capabilities of the foundation models themselves. Also known as "whatever happened to Jasper?" If you buy this, then the question becomes: How many of these companies will exist in 3 years?

Source

I can certainly tell you that the demos for many of these products are incredibly impressive; watching them feels like seeing the future. But I also know that when the market values a certain type of company at 100x revenues, absolutely everything has to go right for that investment to deliver alpha. A future where foundation models can do 10x as many things as they can today, but are now free and accessible from any device regardless of whether it's connected to the web, seems like a tough one for many of these companies. Consider me cautious.

Picks and shovels.

Maybe the answer is AI/LLM infrastructure: Not the models themselves, and not the end-user applications built on top, but all of the scaffolding in between:

Source

Everything below the top row, basically (Source)

This seems like a plausibly good place to invest, perhaps even at 100x revenues. We are in a gold rush, after all.

On the other hand, many of these services are sold to the model providers. If building models ends up being a bad long term business model, then how exactly are these markets going to grow?

A possible bull case (I think?): Every enterprise in the world will run its own model that's been fine-tuned on its own business data, and the orchestration and upkeep of these models will fall on these enterprises themselves, causing enormous demand for these kinds of "ops" services.

Incumbents.

This is the "in a world with prolific AI, the ability to distribute (sell) software is more important than the ability to build it." Most of the value capture in AI will go to organizations that can push its capabilities into the hands of millions or billions of users, perhaps only monetizing them indirectly (e.g. via ads). Forget investing into AI startups, just buy stock in Microsoft, Google, Apple, Amazon, and Meta.

Source

I find this argument to be very aesthetically pleasing, for the following reason:

Over the past 10 years, the FAANG portfolio has returned 28% annually. This number happens to trounce the performance of, oh, maybe 75% of venture capital funds over the same period. Put another way, this means that a decade of VC practitioners sourcing, researching, thesis-building, diligencing, negotiating, attending meetings, recruiting, publishing blog posts, marketing the firm, tweeting, et cetera is all less valuable (in a returns sense) than simply guessing "big tech is going to get bigger" and then chilling at the beach for 9.99 years. Humbling!

The "incumbents win AI" argument simply rolls this forward by another decade.

The weird thing about this thesis is that it isn't really even about AI at all; it's really the idea that "monopolies gon' monopoly.” Which is sort of unsatisfying. What's the point of having a Substack if there's no value in my galaxy brained hot takes?

Nobody, which actually means everybody.

This is the "The End of Software" argument:

Software is expensive to create. You have to pay people to build it, maintain it, and distribute it. Because software is expensive to create, it has to make money. And we pay for it–software licenses, SaaS, per seat pricing, etc. Software margins have historically been an architectural envy–90+% margins and zero marginal cost of distribution.

Software is expensive because developers are expensive. They are skilled translators–they translate human language into computer language and vice-versa. LLMs have proven themselves to be remarkably efficient at this and will drive the cost of creating software to zero. What happens when software no longer has to make money? We will experience a Cambrian explosion of software, the same way we did with content.

Vogue wasn’t replaced by another fashion media company, it was replaced by 10,000 influencers. Salesforce will not be replaced by another monolithic CRM. It will be replaced by a constellation of things that dynamically serve the same intent and pain points. Software companies will be replaced the same way media companies were, giving rise to a new set of platforms that control distribution.

This argument basically says that since labor is the key input (today) in software production, and LLMs are going to drive the cost of labor to zero (highly debatable!), software is about to be so cheap and abundant that it will be tough to charge enough to even keep a software company in business at all.

Source

Let's set aside the whole "software is dead" piece of the argument for a second. How should we feel about a world where a significant portion of the means of production experience a radical decrease in cost? Is that dissimilar for a world with cheap oil, or cheap green electricity, or free/cheap Internet, or safe maritime trade routes in peacetime? Aren't these enormous deflationary forces that would lead to massive economic surplus?

In this world, it's not the AI companies that will capture economic value, it's that every company that utilizes these tools will capture more economic value. And same goes for every consumer, who gets to use them for free. (This sounds… utopian?)

If this is true, then it also suggests something subtle: AI is not a platform shift. It might be incredibly disruptive to certain business models and business practices, but it's not a re-platforming like the Internet, mobile devices, or the cloud. The fact that you can summon AI intelligence via an API call right into the same exact endpoint you were already using is the hint.

What I believe

1. AI is deflationary — it will lead to economic abundance

2. AI will be free or nearly free — that's the whole point

3. Because of #2, you can't "re-sell" AI output at a markup and expect to have a moat; you have to use it to solve problems your customers are hungry to fix even if they've never heard of "AI.” And they will pay you for solving the problem, not for the AI.

4. Investing in companies at 100x revenue valuations just to show you can is not a winning strategy