Note that this is not investment advice but a thought experiment about portfolio construction.

Assuming that compute, in particular, AI compute, will be very valuable in the future, I would include the following categories in an “evergreen” portfolio (I am not limiting myself to public markets, but including also large or late stage private companies where acquiring a stake might be possible on the secondary market). For our evergreen portfolio, we will allocate equally to each area of the stack, from components in the chip supply chain, all the way to AI model developers. For the sake of this thought experiment, I’m going to assume that obtaining secondary equity in private names is possible (or that they may list in the coming years, like ARM).

The five tranches, and their suggested weightings within the portfolio are as follows:

• Extended chip supply chain (10%)
• Chip manufacturers (20%)
• Datacenters (10%)
• AI clouds (35%)
• Foundation models (25%)

I’m focusing mainly on AI clouds, as I believe they are the ultimate picks and shovels play for AI, and there are a lot of pure play AI names in the sector. Foundation models are important, but we haven’t encountered clear winners in the space yet. Certain chip manufacturers also have meaningful AI exposure.

Extended chip supply chain:

The standout here is TSMC. They are exposed to geopolitical tensions, but they are the standout in high quality chip manufacturing, with 57% market share in semis, followed by Samsung, UMC, Global Foundries, and SMIC.

TSMC’s critical supplier is ASML, the Dutch company that provide the photolithographs that TSMC uses in their foundries. You would also look at Applied Materials (manufacturing for chip fabs), Lam Research (wafer), Tokyo Electron Limited (assorted semis equipment), KLA, BASF, DuPont, and Merck KGaA (chemical supplies)

Based on AI as a share of their business, I would allocate within this tranche of to TSMC (50%), UMD (20%), Global Foundries (20%), ASML (10%).

Chip manufacturers:

The standout of course is NVIDIA. They dominate GPUs for modern clouds and for contemporary models. These chips, like the H100, or the forthcoming H200, tend to have a relatively long depreciation period because if you train a model on a specific chip, you will want to do inference on that same chip. The other competitors here are ARM, Intel, AMD, Qualcomm, and the hyperscalers like Apple, Samsung, Google, Microsoft, and Huawei which are all building their own chips.

Because the hyperscalers are trading richly already, and chips are a relatively small part of their business, we will focus less on them here. Within this tranche, I would allocate to NVIDIA (50%), AMD (20%), Intel (20%), and ARM (10%).

Datacenters:

We are interested in brick and mortar datacenter operator that are able to build tier 4 datacenters. This is a vital part of the AI supply chain and likely the bottleneck in the coming 24 months. The most important ones are Equinix, IBM, NTT Comms, Switch, KDDI corp, Sungard, Global Switch, and T Systems (dutch telekom). Here I allocate 50% to Equinix, 30% to Switch, and 20% to NTT comms, based on their respective AI focus and geographic exposure.

AI clouds:

This is one of the most interesting tranches. We have the megaclouds run by Microsoft, Oracle, Amazon, and Google, followed by boutique names like Core Weave, Lambda, Paperspace, Cirrascale, and Vultr. Additionally you have crypto names like Akash and Render.

Microsoft gets 20% of the tranche since they are so AI focused and have OpenAI exposure. Core Weave (disclosure: I am an investor from seed) is the standout in the boutique name, so they get 40%. 5% apiece to Lambda, Paperspace, Cirrascale, and Vultr (acquiring stakes in these names might be harder), and 10% across the major crypto render tokens as a hedge, in case decentralized rendering takes off. Overall, I believe this category will create the most value in a long term AI portfolio.

Foundation models:

This one is harder to predict. Major players like Google have yet to meaningfully enter the game. Meta is open sourcing their models. There’s still a huge amount of churn in foundation models and it’s very unclear who the winners will be. But nevertheless we need exposure. The big names here are OpenAI, Google (Deepmind), Anthropic, Huggingface, Midjourney, Meta, Stability AI, Microsoft, Inflection, and Cohere. Since we already have Microsoft exposure, and Google and Meta aren’t really AI companies, I would focus on OpenAI (30%), Anthropic (20%), Huggingface (10%), Stability (10%), Inflection (10%), Stability AI (10%) and Cohere (10%).