Prospecting in the AI Gold Rush
There’s received wisdom — which ChatGPT corroborates — that it was more profitable during the California Gold Rush to support those prospecting gold than it was to be a prospector yourself. Supporting goods included everything from mining equipment and clothes to housing, food, and banking, and these entrepreneurs were significantly more profitable because they sold bare necessities that every prospector needed, whether the prospector was successful or not. Prospectors on the other hand rarely struck it rich because gold was genuinely difficult to find.
One obvious lesson you might take away is that in the current AI hype cycle (or AI gold rush, if you will), the smart opportunity is to spend less time using AI to build new applications (prospecting) and to spend more time building tools for people building with AI (mining supply, in gold rush terminology).
Looking at the number of LLM infrastructure and LLMOps companies, this seems to be a widely shared opinion. A few months back, a clever Twitter hot take (which I can’t find anymore) quipped that there were more companies building LLM fine-tuning services than there were companies with fine tuning use cases. We ourselves were originally in the LLMOps space; the first iteration of RunLLM was a developer platform for LLM-powered apps.
There will undoubtedly be very large companies built in the LLM infrastructure space, many of which have likely already been started; however, we’ve come to believe that it makes more sense to be a prospector in the AI gold rush than to be a supply company. Let’s dive into why.
Innovations in mining technology. What prompted the California Gold Rush was not discovery of a new precious metal that warranted new mining techniques. Mining technology looked the same in 1848 as it did in 1849, but it turned out there was a new place to go mine gold. Once you decided to mine, you needed the same picks, shovels, axes, and pans that you’d always needed. This meant that mining supply companies had no need to innovate — they were selling tried & tested tooling that their customers knew they wanted in a greenfield market.
The same is quite obviously not true today. What kicked off this hype cycle is fundamentally new technology. Companies building with AI aren’t sure yet what their hardest problems are. There are clearly some good guesses — prompt experimentation, model cost arbitrage, etc. — but as of today, there’s only a few companies mature enough with generative AI to understand these problems well. As the space matures, we’ll see new problems emerge and the standard splits between larger enterprises and SMBs. In the meantime, everyone is just figuring out the basics, and selling a solution to a prospect who doesn’t understand their problem is a losing proposition.
Gold is a limited commodity. In the Gold Rush, striking it big was all about finding gold. Gold you found was gold that I could not then find, and the more people that started prospecting, the worse everyone’s odds got. Once anyone found some gold, value was generated by selling the gold itself, which you can obviously only do once.
There’s a level at which this might be true in AI — if all 8 billion of us on Earth started generative AI companies, we’d be in a bad place. (But maybe we could all sell to each other, like a really large YC batch?) In reality, today, there’s a huge number of unexplored applications of AI. Unlike with gold prospecting, finding the nugget of an idea (pun intended) isn’t the end goal — it’s just the beginning. It’s the value you generate from productizing that idea that’s the end goal, and you can sell that nugget many many times. Of course, as your expertise around AI + your idea increases, you’ll undoubtedly find new applications.
It’s always been easy to sell gold. Putting the previous two points together, the path forward becomes obvious. Once you found some gold in the gold rush, selling it was easy. Gold has been in high demand for nearly all of recorded human history, so there’s no doubt that finding gold would be a profitable endeavor. It was just that your chances of finding gold became smaller and smaller as time passed.
If you’re building a new product, especially a B2B product, you’ll need to think carefully about your target customer. Most enterprises today know they should be using AI, but they don’t quite know what to do. Many of them don’t have the expertise on staff to feel confident shipping AI applications, so now matter how easy you make it to build a copilot or fine-tune a model, they won’t quite know where to start. That’s why selling picks & shovels is so difficult.
On the other hand, if you promise to 2x the productivity of a marketing team or reduce engineering headcount needs by 20%, it’s not a question of whether your solution will be in demand — it’s simply a question of whether you can deliver.
Those types of eye-popping productivity figures are the promise of AI, and enterprises believe there is a ton of low-hanging fruit. If you can find the AI gold that enterprises want, you will very likely be able to sell it. Given the rate of innovation in the space, your chances of finding gold are higher than ever.
This isn’t to say there won’t be successful LLM infrastructure companies. There undoubtedly will be. In broad strokes, however, we believe that most picks & shovels (GenAI infrastructure) companies today are too early and too undifferentiated. Meanwhile, the ability to innovate on applications of AI feels close to unbounded. As models and techniques improve, the list of possible applications is only growing.
Whether or not you agree with us, if you’re thinking about diving into a new problem in the space, say hi — we’d love to talk about it!
Yes, there is a great rush in the AI application space, and everyone wants to add AI to their portfolio, and get the box ticked. There is Gold, Silver and Bronze to be found in every layer of the AI stack today - Hardware, Platform , Models, Apps. But the skills required vs financial resources required in each of these layers are very different. Whether you approach from top-down or bottom-up. The requirements of the enterprises also differ vastly and so we will see "No one size fits all" slogan and enterprises need to keep their eyes wide open and choose the right ingredients for their business outcomes.