AI is like Free Trade

How should we "play AI"?

The biggest question in investing right now is "how should we play the AI trade?"

One school of thought says that you should do everything possible to put capital into the 4 or 5 foundation model companies that matter, regardless of entry price. If OpenAI develops the singularity, it will be worth something approximating the GDP of the entire planet. Why spend any time worrying about valuation multiples when the prize is essentially world domination?¹

Another school says: Invest into any company built on top of LLMs. The world is changing quickly, and these businesses are growing faster than the winners of any prior generation of technology (supposedly, at least). Buy anything with the word "AI" attached to it. Ignore the haters who diss your "wrapper" companies.

A third school says to just buy Nvidia. And ARM. And Broadcom. And Meta/Google/Amazon/Apple/Tesla/Microsoft. You want to be in the silicon game, baby. Chips are the ultimate picks-and-shovels play for AI.

There are other schools too: Short everything AI because this is just a bubble. Short everything in general because AI will accelerate societal breakdown and World War 3. Buy SaaS. Sell SaaS.

I have a slightly different mental model when it comes to the longer-term AI trends.² Here's how I think about it: AI more than anything is a broad, massively deflationary force that will create winners and losers, but overall will create more consumer and producer surplus than direct value capture by the companies that constitute the AI value chain.

Put another way: AI is less of a discrete, patentable technology that can be exploited directly by the businesses that build it, and more of a social technology that will lead to diffuse improvements in the standard of living for society as a whole, even as it creates losers in specific pockets of the economy. In this way it is more analogous to a worldwide reconfiguration like Free Trade than a monopolizable asset like railroads.

Why Free Trade? Well, the effect of something like NAFTA is to rotate production to lower cost regions in order to reduce input costs (and eventually finished product costs). This leads to an overall surplus — Producers can procure lower-cost raw materials and intermediate goods to either increase profits for themselves and/or drive more competitive pricing in order to win share, which leads to lower prices for the consumer. Despite the overall surplus it does cause harsh losses in specific parts of the economy, in this case the domestic workforce that is outcompeted by lower-cost foreign manufacturing (in the U.S. this is the hollowed-out manufacturing base).

What's happening with AI? In this instance, white collar cognitive labor is being substituted for a machine that has been trained by crunching down and synthesizing the brainpower of billions of other people throughout history. The distilled outputs of this machine can be procured for fractions of a penny.

Say what you want about slop and hallucinations and "glorified auto-complete" — obviously this technology is wildly imperfect, no disagreement from me there. But let me ask you a question: Do your "Made in China" jeans ever have a button or a pocket sewed on incorrectly? Does that lead you to instead spend $250+ on American-made jeans, or do you just shrug and deal with it? At a certain point the cost savings are worth the inconsistency in craftsmanship, as evidenced by the $20 Wranglers that millions of people buy at Walmart even as American denim plants continue to shut down.

Note that I didn't say this is good, just that it is happening. It is the nature of capital to seek its highest rate of return (or as Ranjan Roy says, "money is always swimming towards yield"). Now that the AI genie is out of the bottle, will it be implemented up and down the economy to drive down costs? It is inevitable.

$50,000 worth of labor for $1.68

I was struck by a post from Simon Willison two weeks ago that described how cheap it was for him to accomplish a pretty major task with Gemini that would likely take a human a few months to do manually. Apologies for the long blockquote but you really need to read it all to let it sink in:

Here’s a fun napkin calculation: how much would it cost to generate short descriptions of every one of the 68,000 photos in my personal photo library using Google’s Gemini 1.5 Flash 8B (released in October), their cheapest model?

Each photo would need 260 input tokens and around 100 output tokens.

• 260 * 68,000 = 17,680,000 input tokens

• 17,680,000 * $0.0375/million = $0.66

• 100 * 68,000 = 6,800,000 output tokens

• 6,800,000 * $0.15/million = $1.02

That’s a total cost of $1.68 to process 68,000 images. That’s so absurdly cheap I had to run the numbers three times to confirm I got it right.

How good are those descriptions? Here’s what I got from this command:

• llm -m gemini-1.5-flash-8b-latest describe -a IMG_1825.jpeg

Against this photo of butterflies at the California Academy of Sciences:

A shallow dish, likely a hummingbird or butterfly feeder, is red. Pieces of orange slices of fruit are visible inside the dish.

Two butterflies are positioned in the feeder, one is a dark brown/black butterfly with white/cream-colored markings. The other is a large, brown butterfly with patterns of lighter brown, beige, and black markings, including prominent eye spots. The larger brown butterfly appears to be feeding on the fruit.

260 input tokens, 92 output tokens. Cost approximately 0.0024 cents (that’s less than a 400th of a cent).

This increase in efficiency and reduction in price is my single favourite trend from 2024. I want the utility of LLMs at a fraction of the energy cost and it looks like that’s what we’re getting.

This is what I meant by a massively deflationary force.³ How much would you personally demand to be paid to write 4-6 sentences each that describe 68,000 photographs? You probably wouldn't do it at any price because it sounds so dreadful, but if you were forced to price it out it would probably be something like $50,000 (one description every 3 minutes means that this task would take 3400 hours; at $16 California minimum wage that's $54,400. By the way, that would take you about a year and a half to accomplish in 40 hour workweeks.) This crunched-up-humanity machine just did it for $1.68 and probably a few minutes.

Again, you can talk about how this is all just uninspired regurgitated slop, but this is just one example of how the technology is undoubtedly a lower cost way (a dramatically lower cost way, actually) to accomplish cognitive labor that could in some other instance employ a human. Even if the quality of the output is terrible⁴, the cost savings overwhelm the quality control problem by many orders of magnitude.

Does this example make you want to be an investor in OpenAI, the company that had to spend billions of dollars on data centers to crunch all of the text on the Internet into a distillate goo that they could sell to people like Simon Willison for $1.68? Or would you rather be Simon Willison, the guy who just bought $50k of labor for less than two bucks? It seems to me that 99.99% of the economic surplus in this transaction went to the consumer (Simon). That’s not me saying OpenAI is a bad business; I’m just saying that it’s getting really good to be Simon. Here’s the best news: We’re all about to be Simons.

We're about to enter a world where a pair of jeans costs 5 cents — maybe you have to throw out 2 or 3 pairs before you find one that is sewn correctly, but who cares? (Temu is a thing, after all). This is going to change everything.

Rather than asking ourselves how to "play the trend"⁵, what we should be doing is pondering how this is going to remake the world — what new opportunities are going to be possible⁶, and what cherished things we're going to lose.

1. Though once the singularity arrives, will money even be valuable anymore? Unclear how IRRs will work once you cross the event horizon.

2. Which is not to disagree that some of these are going to be incredibly lucrative trades.

3. Elsewhere in Simon's piece he discusses how some of the newer, smaller models (such as Llama 3.3 70B) can be run on a laptop. This means you don't even have to pay OpenAI or Google their few cents for the inference costs.

4. And what happens when it's actually good?

5. If you could go back to 1992 when NAFTA was signed, would you want to "play the trend" by buying stock in a bunch of shipping container companies? Or would you decide instead to found Amazon?

6. If Free Trade led to a populist backlash that metastasized in the form of Donald Trump, what will happen in response to the mass adoption of AI? Maybe you should stop reading this Substack and go figure out how to run for President.