AI Boom Favors Infrastructure Over Most Frontier Labs, Investor Argues
Reactions from ranked influencers
2 posts
Most AI companies are going to zero. You can know the thing, and still lose all your money. For example, you could have the insight that "cars will be a big deal" in the early 1900s, but making money from that insight was far from a sure thing. Over 1,900 US car companies have been started. Almost every single one has failed. To date, only two American car companies have avoided bankruptcy (Ford and Tesla). Right now, most people know AI is going to be a big deal. That is obvious. Making money from it, that's the hard part. Even a few months ago, the killer investment idea would have appeared to be to invest in the frontier labs. But as you can see with the new Kimi model, that can flip on its head in a matter of days. It's still unclear if there is value in the models themselves. Mark Twain put it best: "During a gold rush, it’s a good time to be in the pick and shovel business" Trying to pick winners is near impossible. What's easy, is selling picks and shovels: the GPUs, chips, electricity, and computing required to power the boom. Jeff Bezos likes to invert: instead of thinking about what will change, he instead focuses on what won't change. Here's a few things I believe will continue to be true in the the next five years: 1. People will use tools like Grok, Gemini, ChatGPT, Siri (lol) and Claude at increasing rates. The tools are undeniably useful, and history shows that making a technology cheaper and more efficient doesn’t reduce demand, it expands it. 2. Both training better models (training compute) and using them (inference compute) will require massive data centers. 3. We are majorly compute constrained and even before the AI boom, data center demand was growing rapidly. The choke point for all of this isn’t GPUs, it's power. Behind-the-meter generation, pre-existing high-voltage connections, and sites that can actually get energized in the next 24–48 months are scarce. This makes assets with secured power and land in the right places insanely valuable. And lucky for us, they all went on sale this week. My companies have invested in four stocks that we plan to hold for the long-term: $IREN - Previously misunderstood as a crypto miner. The stock re-rated after they signed a massive 10-year deal Microsoft. I talked about my investment on My First Million in January 2025. $CRWV - Similar to IREN, CoreWeave is one of the leading AI cloud providers. It's already locked in ~$100B in backlog from labs and enterprises that need capacity now, not in three years. Same story as the IREN-Microsoft deal: long-term committed revenue. $NUAI - Owns a huge portfolio of data center sites that are perfect for behind-the-meter development. See my post on this one from yesterday. $SPCX - I'm bullish on orbital compute over the next 5-10 years. And yes, before you yell into your computer any more, let's address some of your bearish points: "WHAT ABOUT LOCAL OPEN SOURCE MODELS?" I've thought about that, and while I agree that increasingly powerful workloads will be able to run locally, I think: 1. The cost of the hardware required to run frontier level models is still out of reach for the average person and will continue to be for some time (two year old Mac Studios with 512GB of ram are selling for $25-$35,000 on eBay). 2. Have you ever used your phone to edit photos in Lightroom or used Apple's AI image playground? Your phone quickly turns into lava. That is the GPU/CPU working overtime to run local models and it absolutely kills battery life. Given the huge portion of AI that occurs on mobile, the need for inference in the cloud (data centers) is going to continue to grow. 3. More and more tasks will be able to run locally with less and less compute/GPU. I won't argue that. But I believe our demand for intelligence is insatiable. If you're trying to win in business or create a life saving drug (or even just trying to file your taxes accurately), you want to use the best model. The question is, at what point does the return on intelligence end? Or do we just find increasingly wild tasks to give it? 4. Even "local" workflows are often hybrid. The heavy reasoning, tool use, retrieval, or multi-agent orchestration still routes to cloud. "WHAT ABOUT DATA CENTERS IN SPACE?" 1. I am a SpaceX shareholder and think orbital compute will be huge if they can pull it off. 2. This will take years, and Elon usually delivers a few years late. 3. Even if orbital works at scale, latency and the multi-year timeline mean it’s additive, not a substitute, for the next wave of demand. Companies will still sign terrestrial leases as insurance. Companies like IREN, Coreweave, and New Era Energy are finite resources that require upfront lock-in. Think of them as the last available industrial warehouses in a crowded city with no other availability where tenants are required to sign 10-year leases. If you want to play, you need to pay. If AI compute is the most in-demand service in the world, and not securing it could cost a company its moat, are people really going to bet that Elon will deliver data centers in space and wait 3-5 years to secure their compute or are they going to hedge their bets and sign deals with terrestrial data centers? My bet is the latter. "BUT AI ISN'T DELIVERING _____" I’m currently spending $30–40k a month on Anthropic and OpenAI credits. I treat it as payroll for an extra 30+ digital employees doing work that would cost $200–300k/month in human equivalents. There is no scenario where this genie goes back in the bottle. The demand is real, it’s compounding, and it’s only getting started. Mark Twain was right about gold rushes. Billions will be made by prospectors chasing the next big model, but most of them will go broke. The boring money is in the picks and shovels: the power, the land, the GPUs, and infrastructure that every serious AI effort will need more of for years to come. So, are you a gold miner or a pick axe salesman? PS: I'd love to hear any rebuttals if I've missed something. --- Important disclosure: I (and entities I control) beneficially own shares of IREN, CRWV, SPCX and NUAI. I wrote this after establishing these positions. I may buy more or sell at any time without updating this post. This is not investment advice or a solicitation to buy/sell securities. This is a high-risk, speculative situation and you can lose all your investment. Forward-looking statements and scenarios in this post are speculative and may not occur. Do your own research and read company filings.
Most AI companies are going to zero. You can know the thing, and still lose all your money. For example, you could have the insight that "cars will be a big deal" in the early 1900s, but making money from that insight was far from a sure thing. Over 1,900 US car companies were been started. Almost every single one failed. To date, only two American car companies have avoided bankruptcy (Ford and Tesla). Right now, most people know AI is going to be a big deal. That is obvious. Making money from it, that's the hard part. Even a few months ago, the killer investment idea would have appeared to be to invest in the frontier labs. But as you can see with the new Kimi model, that can flip on its head in a matter of days. It's still unclear if there is value in the models themselves. Mark Twain put it best: "During a gold rush, it’s a good time to be in the pick and shovel business" Trying to pick winners is near impossible. What's easy, is selling picks and shovels: the GPUs, chips, electricity, and computing required to power the boom. Jeff Bezos likes to invert: instead of thinking about what will change, he instead focuses on what won't change. Here's a few things I believe will continue to be true in the the next five years: 1. People will use tools like Grok, Gemini, ChatGPT, Siri (lol) and Claude at increasing rates. The tools are undeniably useful, and history shows that making a technology cheaper and more efficient doesn’t reduce demand, it expands it. 2. Both training better models (training compute) and using them (inference compute) will require massive data centers. 3. We are majorly compute constrained and even before the AI boom, data center demand was growing rapidly. The choke point for all of this isn’t GPUs, it's power. Behind-the-meter generation, pre-existing high-voltage connections, and sites that can actually get energized in the next 24–48 months are scarce. This makes assets with secured power and land in the right places insanely valuable. And lucky for us, they all went on sale this week. My companies have invested in four stocks that we plan to hold for the long-term: $IREN - Previously misunderstood as a crypto miner. The stock re-rated after they signed a massive 10-year deal Microsoft. I talked about my investment on My First Million in January 2025. $CRWV - Similar to IREN, CoreWeave is one of the leading AI cloud providers. It's already locked in ~$100B in backlog from labs and enterprises that need capacity now, not in three years. Same story as the IREN-Microsoft deal: long-term committed revenue. $NUAI - Owns a huge portfolio of data center sites that are perfect for behind-the-meter development. See my post on this one from yesterday. $SPCX - I'm bullish on orbital compute over the next 5-10 years. And yes, before you yell into your computer any more, let's address some of your bearish points: "WHAT ABOUT LOCAL OPEN SOURCE MODELS?" I've thought about that, and while I agree that increasingly powerful workloads will be able to run locally, I think: 1. The cost of the hardware required to run frontier level models is still out of reach for the average person and will continue to be for some time (two year old Mac Studios with 512GB of ram are selling for $25-$35,000 on eBay). 2. Have you ever used your phone to edit photos in Lightroom or used Apple's AI image playground? Your phone quickly turns into lava. That is the GPU/CPU working overtime to run local models and it absolutely kills battery life. Given the huge portion of AI that occurs on mobile, the need for inference in the cloud (data centers) is going to continue to grow. 3. More and more tasks will be able to run locally with less and less compute/GPU. I won't argue that. But I believe our demand for intelligence is insatiable. If you're trying to win in business or create a life saving drug (or even just trying to file your taxes accurately), you want to use the best model. The question is, at what point does the return on intelligence end? Or do we just find increasingly wild tasks to give it? 4. Even "local" workflows are often hybrid. The heavy reasoning, tool use, retrieval, or multi-agent orchestration still routes to cloud. "WHAT ABOUT DATA CENTERS IN SPACE?" 1. I am a SpaceX shareholder and think orbital compute will be huge if they can pull it off. 2. This will take years, and Elon usually delivers a few years late. 3. Even if orbital works at scale, latency and the multi-year timeline mean it’s additive, not a substitute, for the next wave of demand. Companies will still sign terrestrial leases as insurance. Companies like IREN, Coreweave, and New Era Energy are finite resources that require upfront lock-in. Think of them as the last available industrial warehouses in a crowded city with no other availability where tenants are required to sign 10-year leases. If you want to play, you need to pay. If AI compute is the most in-demand service in the world, and not securing it could cost a company its moat, are people really going to bet that Elon will deliver data centers in space and wait 3-5 years to secure their compute or are they going to hedge their bets and sign deals with terrestrial data centers? My bet is the latter. "BUT AI ISN'T DELIVERING _____" I’m currently spending $30–40k a month on Anthropic and OpenAI credits. I treat it as payroll for an extra 30+ digital employees doing work that would cost $200–300k/month in human equivalents. There is no scenario where this genie goes back in the bottle. The demand is real, it’s compounding, and it’s only getting started. Mark Twain was right about gold rushes. Billions will be made by prospectors chasing the next big model, but most of them will go broke. The boring money is in the picks and shovels: the power, the land, the GPUs, and infrastructure that every serious AI effort will need more of for years to come. So, are you a gold miner or a pick axe salesman? PS: I'd love to hear any rebuttals if I've missed something. --- Important disclosure: I (and entities I control) beneficially own shares of IREN, CRWV, SPCX and NUAI. I wrote this after establishing these positions. I may buy more or sell at any time without updating this post. This is not investment advice or a solicitation to buy/sell securities. This is a high-risk, speculative situation and you can lose all your investment. Forward-looking statements and scenarios in this post are speculative and may not occur. Do your own research and read company filings.