Portfolio status as of 1/2/2024:
I’m up $1,856 since inception, for an IRR of 95%, beating the S&P 500 by a large margin mostly thanks to macro movements — higher leverage on lowering rate expectations — and a nice win with AMD. As a reminder, I’m not trying to hedge such macro movements, and there will be times when they move against the portfolio as well. For now, I’ll take the win.
At the start of Q4 2023 I made the following trades, as reported in the first Trade Report at the time:
leaving us with under $30 in cash.
The big outperform in the basket was AMD, which is up 35% in the period, after an optimistic outlook on their MI300X, as I anticipated.
AI chips continue to look like a no-brainer investment.
There are some reasons to doubt, though, so I’ll leave with a couple of notes on the durability of this current allocation.
What if AI doesn’t commercialize fast enough?
Companies are buying over $40 billion worth of AI chips this year, but actual generative AI revenue is…. a 2-3 billion?
There’s a big mismatch right now, one that cannot endure forever. How much are big tech willing to throw at AI before they need to see revenue and profit? Will they have an initial spurt of investment and then all pull back just as the GPU supply chain is ramping up? This type of catastrophic over-investment is common in technological revolutions, c.f. railroads, internet fiber. (Of course it is only catastrophic for the suppliers themselves, their pain is the general economy’s gain!)
There are two reasons to think this time is a little different. For one, GPUs are incredibly hard to make! You can’t just ramp up production, you can’t just enter the market with a big enough initial investment. Every step of the process requires extreme levels of expertise, extraordinarily expensive equipment, and years (or decades) of process optimization. Just look at Intel bumbling around! This doesn’t prevent boom-bust cycles, but it’s likely to draw out the boom over a longer time frame.
Second, the investors in this case are highly confident in the long term commercial opportunity of AI, and lucky for them, have enormously deep pockets that allow for patience. Google, Microsoft, Meta, Apple, Tesla, and Amazon command hundreds of billions of dollars of capex each year and all see AGI as an existential opportunity / threat. Google has been sinking billions into self-driving for 15 years without blinking. They’ll go to much greater lengths for AGI. (Not that my portfolio has skin in the TPU game currently. But the same applies to the other players.)
What if LLMs are a dead end?
What if LLMs are only good at generating cat poems in iambic pentameter, and don’t have any big disruptive commercial applications? What if people continue to prefer Google search over chatting with ChatGPT, as they have demonstrated so far that they mostly do?
This question will get a full post at some point, but is possibly relevant to my near-term GPU allocations, so I’ll address it here quickly.
The easy answer is that no one knows yet, all signs point to no they’re not a dead end, but either way it will take years to know for sure LLMs are actually a dead end, and in the mean time everyone will still be going crazy trying to test that hypothesis, given what is possibly at stake. So, there’s really no world where GPU investment falls off a cliff until we start to see GPT-6 plateau in 2025.
All that said, I am extremely skeptical that LLMs are a dead end. Look out for a future post on this topic.
Otherwise, I’ll be back with another Investment Review in 3 months to see how our latest allocations in Trade Report #2 that I released yesterday have performed.