The whole @levelsiophenomenon reminds me a lot of the early Amazon Seller market. It looks like random garbage being sold at first glance, but is actually a really really big deal.
Back when Amazon Selling kicked off - it felt like a giant scam. Cheap Chinese products (the e-commerce equivalent of AI slop). Dicey unit economics.
For the longest period wall St analysts thought the entire e-commerce business actually wasn’t worth anything (or even negative value) and the bull case for the company was to spin it off leaving AWS
But after the 4 Hour Workweek - there was this cultural shift – people really liked the concept that they could make a living without the blessing of corporate America. Interestingly - this didn’t appeal the most to digital Nomads, or the people you’d think would have read Tim Ferriss work.
But rather - it saw hyper growth due to religious people, such as Orthodox Jews - who had a hard time fitting into the corporate world. They didn’t want to work less - but they wanted to define their own hours, and keep standards in line with their perceived cultural optimum
But when I first learned about this - I hadn’t yet been to seller conferences filled with bearded men with tophats and permanent sideburns. Rather I encountered this market abstractly as a Wall Street analyst covering TMT stocks with my coworker @swyx (who interestingly was featured along with his company in the @levelsio interview). The Seller business kept growing rapidly - to 50% of Amazon’s sales. And it was really weird products selling out - like Chinese charcoal toothpaste, reaching the top of Amazon.
These products were often so successful that they would cause major brands to pivot (Crest launched its own Charcoal toothpaste). But it still was weird. I didn’t really understand it. Part of the reason I was hired at the fund I worked at was bc I had some background in alt data. And the alt data for 3rd party sellers related things was always up and to the right.
(We also had a large position in Shopify for this reason - but it didn’t work out until long after I’d left)
Things clicked after a conversation I had with my friend Paul. He had a Korean friend who was making quite a lot of money doing Amway advertising arbitrage. For those of you who don’t know what Amway is - it’s the original Herbalife Multi Level Marketing Skem - which interestingly is behind former education secretary/ sister of Erik Prince’s $5+B fortune.
Paul got his friend to show us how he was showing money. He strolls into my apartment and gives me - what I can only describe as the worst energy drink I’ve ever had in my life. Like watermelon favored trash. Then he drops the economics on me. He was buying google ads on these Amway energy drinks for $0.30 on average. And he was making, on average $.7 of profit per click. So printing $.4 every time someone clicked an ad.
The profits were consistent, and entirely dependent on the number of clicks. It was digital arbitrage. “How did you even hear about Amway dude?”
“Oh my mom’s friends do it. They’re bored housewives”. That was his theory of how people even knew to search for these products. They’d been shilled in real life by door to door salespeople. And you could buy them online instead.
My gears immediately kicked into action. I signed up for an Amazon affiliate account. And I began running ads through a janky redirect so I wouldn’t get caught (doing this was against TOS). What I found made me nearly fall out of my chair.
On average - an ad to get someone on Amazon cost $.4. And on average, once they were on Amazon - in a generic fashion - their basket was $40 and they checked out 10% of the time. So this yielded the math:
Basket Size: $40 Conversion Rate: 12% Affiliate Payout: 8% Value Per Click ($40*.12*.08)=$0.38 Cost Per Click: $.4 Return on Ad Spend: 0% – didn’t make money by default
But then I realized you didn’t just need to advertise Amazon, you could advertise the specific products. Matcha Powder. Harry Potter books. New video games. Bed Bug Spray. Suddenly - with highly engaging ads you could get up to 20% checkout rate, and the cost per click would drop to sometimes 2-3 cents before normalizing. And different products had different basket sizes. With some optimization (choosing what products to run, and customizing copy) you could get:
Basket Size: $70 Conversion Rate: 12% Affiliate Payout: 8% Value Per Click: ($70*.12*.08) =$0.67 Cost Per Click: $0.3 Return on Add Spend = ($.67/$.3) >100%
This was a non scalable affiliate violating the terms and conditions of Amazon. But the info was real. And what was even more jarring - was that Amazon took 15% of every transaction.
If you plug 15% in the examples above - Value Per Click goes to $0.72 and $1.26. So an unoptimized Google Ad campaign generated $.72 of pure profit for Amazon on the low end and $1.26 on the high end. And you could generate clicks at huge scale at $.8-.9 a click.
This meant that all the Wall Street analysts were wrong, and that the Amazon e-commerce business was virtually guaranteed to be a money printer. If you could break even on fulfillment - just from the seller fee you’d be printing money on the first customer you acquired on Google via a paid ad. Insane LTV/CAC
The math gets even more insane when you realize what happens when a Seller goes out of business. Amazon takes their inventory, liquidates it and sells it. Even more absurd - because Sellers are small businesses - they often can entirely dodge major compliance audits/ product safety checks. So Amazon essentially wins no matter what, the seller wins sometimes, and the consumer and society always loses but thinks they’re getting things cheaply and conveniently so are cool with it. (aka a “Good Business” kek)
This is why I quit my job at a fund and started an Amazon Advertising business. It was - essentially, a sure thing.
Tobi Lutke indicated this was also how he got the confidence on Shopify as a business model, and was key to his early pitches to investors. It was literally a mathematical certainty Shopify would have positive unit economics - bc the cost per click was very consistent, the conversion rates were consistent and the profit output was also very consistent
But this gives you a sort of high level picture of how the system works: 1] You need a ton of people, like Orthodox Jews, who want to opt out of the system. You need these people to make ads, source weird products, design copy, and even shill their neighbors and networks on products so the ads are recognizable and get clicked 2] You have a platform that profits regardless of whether or not they succeed, and thanks to limited liability - the platform always wins (even if the individuals fail) and can therefore invest in the Capex to grow the platform 3] So long as the system is able to generate a profit per click on top level customer acquisition it can quite literally print money, and there is no stopping its growth. It is a perpetual motion machine
So during COVID, what happened is you had a massive influx of people who opted out of the system. At the same time as everyone was trapped at home buying things online. And free money from the govt. So you had a triple cylinder rally which wasn’t sustainable - but even when it normalized, the perpetual motion machine remained in place and will continue to remain in place
The price action was so extreme because everyone is operating on quarter to quarter numbers and doesn’t understand the statistical stability of the underlying drivers
But – e-commerce is the old story. It’s useful because it helps us understand the frame for the new story.
LevelsIO is the Amway moment. He’s the 4 minute mile for AI slop. The 4 hour workweek for a new generation of disillusioned post capitalists. He’s basically creating products with tokens as an input cost in a rapid, iterative fashion. But he’s doing it so that it generates a predictable profit per click. Once he figures out paid growth, he can spin up n order perpetual motion machines.
And unlike e-commerce where the cost structures are increasing due to supply chain issues and geopolitical tension with china, the margin profile is tethered to Moore’s law. The cost per token for GPT4o, Llama etc are constantly going down - and the quality of the output is constantly going up. And this will continue for the next several years
This creates a new model. You have four variables:
Value Per Token - this is the amount of money you can make for a unit of compute. This can come from selling photos, porn, trading capital markets, doing meeting summaries - whatever it is. But we have confidence this will probably increase over time as models get smarter
Cost Per Token - this is how much it costs to generate the value. This should also always be dropping. More importantly it will be heavily subsidized by the aggressive equity valuations of the chip makers and AI companies
Customer Acquisition Cost - this is where it gets really interesting. In e-commerce, there’s no reflexive link between the product and the ad. You have to hire copy people. But AI is excellent at generating content, and increasingly generating ads and retargeting engagement like chats to lower churn.
Number of “Societal Opt Outs” - you need the LevelsIO characters, living the digital nomad life. The Orthodox Jews. The structurally unemployed Otaku. But once again - with e-commerce there’s no extremely reflexive link here. There’s a very real chance that AI will cause mass white collar structural unemployment by automating a lot of jobs. This not only increases the number of people iteratively applying AI to make money in random ways, but will also increase the likelihood of government style MMT programs.
The reason why LevelsIO is so important is that he proves that a 1 man operation can generate positive unit economics across the entire stack fully accounting for developer cost. People are looking skeptically at the products as AI slop - when the reality is that e-commerce started the exact same way, and was similarly under appreciated in its scope.
Nvidia and other chip companies should pay attention to him, fund him, and others like him. And I believe they will - derisking the bet even further for early adopters
The implication of this is somewhat exciting.
So long as indie producers can generate a high Profit Per Token, and use AI to drop acquisition costs - the entire AI space will enter a decade long bull similar to e-commerce between 2014-2024. The profitable categories - just like ecommerce - will be ever changing, but the underlying logic is identical.
My personal conviction is that the highest value per token is decisively in the crypto / speculation space but there will be big winners across major categories - companionship, pornography, video games, generative Art, virtual worlds and more.
You’ll obviously need to see more levelsIO crop up to build conviction in the thesis but - I think with the number of sites he’s shipped, by himself, it’s pretty clear the underlying dynamics are identical to early e-commerce all the way down to the weird sociocultural digital nomad stuff.
It’s exciting because it’s a cash flow category in a rising tide of speculative froth, with now proven unit economics. What I call - a sure thing.