Happy Sunday and welcome to Investing in AI. What a couple of weeks it has been in the AI space. First we got the Stargate announcement, a $500B AI infrastructure project that rumor has it will probably never actually happen at that scale. Then Deepseek was released, a language model out of China that matches or beats OpenAI’s o1 reasoning model but runs at a fraction of the cost. This is a pivotal moment for AI investors because it highlights the idea that GPU capacity isn’t everything.
The best analogy I see for this is the breaking of the “eyeballs” bubble in the 1999/2000 time frame. In the late 1990s, the internet was freaking people out. I was working on my MBA during that time and my professors were making crazy statements like “in 5 years you will never shop in a physical retail store again.” Since profits were elusive, companies were measured on how many eyeballs they had – meaning how many people were looking at their site every day.
AI has been through 3 mini-bubbles since AlexNet brought AI back into the limelight in 2012. The first was the AI PhD bubble. Investors had no idea where AI was going or what it could become, but were afraid of missing out so the proxy for investing was – did you have several PhDs with AI experience on your team? This didn’t end well as most of these investments had mediocre exits.
The second wave was the chatbot wave of 2017. It started with slackbots, and the idea was that language was about to have it’s AlexNet moment and language would see a similar breakthrough to what had happened in machine vision. I was a part of this with a startup called Talla and we all believed we were on the cusp of cracking the code. Turns out no, we were all too early and needed 1000x more compute and data. There were zero successful slackbot companies and the ones who made it pivoted away into more traditional application businesses.
The third mini-bubble was prompted by the release of ChatGPT in late 2022. While the core tech of GPT3, on which it was based, had been around for over a year, ChatGPT captured the imagination of the public and led to a mini-bubble around GPT wrapper companies. Similar to past bubbles, most haven’t done very well and the few who have pivoted away from being thin wrappers.
Since then, the AI community, and in particular the AI investor community, have been working from the stupid assumption that we just need more data and more compute to get to AGI. So the metric people have watched, particularly for public markets but to some extent in private markets – is GPUs. How many can you get access to? How large of a cluster are you building? That has been a proxy for how successful your AI will be in the coming years.
But like eyeballs, GPUs as a proxy is inaccurate. The release of DeepSeek proved this. (And also proved sanctions on AI chips don’t work, and can have counterintuitive second order effects). Deepseek was trained quicker and dramatically more cheaply than similar high performing models from the best AI labs.
We should have know this was coming. As I’ve written before, the human brain uses way less power and doesn’t use backpropagation. So we know there are other algorithms and paths to intelligence than our current Transformer + GPU approach.
We also know new hardware architectures are now in market that beat GPUs in performance in many ways, but the big AI labs and companies all believe being first to AGI is important. I’ve written before about why I don’t agree and why I don’t think AGI is such a hot opportunity. It boils down to – every physical system has tradeoffs. Generalized approaches work for markets that need flexibility but are always beaten by specialized approaches for markets that are big enough to support specialization. And there are a lot of those.
As an investor, it’s time to start thinking beyond GPUs and AGI. Bill Gurley was on CNBC last week and said we are moving from the “performance at all costs” phase of AI to the “optimization and ROI” phase of AI. I agree. And this requires an update of the mental models we use when thinking about where to invest.
Thanks for reading.