Continued from here.1sorry this took so long from part I, it’s just an on-going thing where I re-consolidate my views every few weeks as the AI race continues to change. I thought AI and Robotics would be in the same post but then it got too long and now I will do Robotics later.
When it comes to analyzing stocks, I’ve always been a skeptic by nature. Poking holes in bullshit comes easier to me than spotting a trend early and having the imagination to believe in what it might become.
But there’s an equal and opposite trap: the techno-hypebeast who falls in love with every shiny new gizmo2(Meta? VR? De-Fi? NFT?) and convinces himself it’s the future of mankind. One person dismisses all signs of progress of a better future; the other sees the future in everything. Both biases have their costs.
The skeptical mindset comes from my first real exposure to the stock market being during the 2008 financial crisis, when the heroes of the day were contrarian Big Short figures like Michael Burry, Steve Eisman, and John Paulson. I spent more than a decade shorting OTC stocks and Nasdaq micro-caps—the absolute bottom of the barrel, fundamentally—where companies often resorted to manipulative pump and dump tactics to suck in naive investors. Chronic short-sellers tend to be skeptical of everything.
All else equal, you present me with a 60 minute presentation on the next game changing technology of the next 20 years and my gut reaction will often be something like “yeah it sounds nice BUT!”. Some of this just comes from ignorance too. I don’t have a technical background such as an engineering degree. I don’t understand how things are built. I don’t know what I don’t know. It’s just safer to be on the sidelines. But I don’t mistake my track record of dismissing obvious garbage for any kind of real foresight, either. When a technology appears to show real legs, I refuse to be contrarian for contrarian’s sake.
Over the last 24 months, I’ve slowly but surely become 100% AI-pilled—which runs completely against the grain of my entire trading career. I’ve always had a hard time drinking the Kool-Aid, even when I was early. Take crypto for instance–involved in 2013, long before the public had heard of it, yet I could never quite make the leap from “this is interesting” to “this changes everything!” 3(if you want to know my brief thoughts on crypto at the moment: I think we’re roughly at year fifteen and there’s still just no real use case for 99% of all coins besides speculation and gambling. The only major “use-case”, if you can call it that, that I truly believe in is bitcoin being hoarded as a digital gold, which will sustain based on the Lindy effect–which is to say it’s been around long enough to be a thing). But with AI, it’s easy to buy in. I use it. I see other people using it. It’s everywhere. It makes sense.
Society now has this big brain that can think and work for us and it’ll never tire or sleep. It will continue to learn and evolve and it can enhance almost anything. This is why I believe we are at a critical inflection point where it makes sense not just to pour capital into AI, but also into the future technologies that AI will unlock.
What I see happening reminds me of the technology tree from one of my favorite computer games franchises called Civilization. In Civilization, you build yourself as a society from the stone age into the modern age, and that periodic evolution occurs via the time and resources a player invests into developing new technologies. The tech tree starts with bare bones concepts like language, the wheel, and agriculture and then many centuries later, you develop modern tech like electricity, nuclear fission, and computerization. The end of the game usually allows the player to research technology our real world has yet to develop such as giant death robots, planet colonizing space technology, and self-sustaining ocean homes.
What I see, potentially, is this:
For decades, promise has exceeded progress in these so-called industries of the future–the categories on the right hand side. There would be startups and cool demos or pitch decks and then… nothing. There’s a term I learned for this: Potemkin technology. It is basically a technological façade: an impressive demo that creates the illusion that the future has arrived, while hiding how far it still has to go. The phrase comes from “Potemkin village”—something constructed to give an impressive outward appearance while concealing a much less impressive reality.
We were talking about fusion and quantum 50 years ago yet there are still no commercially mature fusion or quantum companies. What happens then is that industry people get jaded–you see many concepts are perpetually five years away and the permanent conclusion is that it’s vaporware and always will be. Talent abandons ship. Funding dries up. Products never commercialize. It only exists in science fiction. It’s easier to be a skeptic because the data says progress is too slow. ‘Slow’ becomes interchangeable with ‘impossible’.
But what if we can break through this Potemkin-jadedness with the assistance of a new force multiplying technology?
AI is that game changing technology. You have to stop thinking of it as a chatbot when it’s actually a utility. This utility is a power that will make all current industries more efficient. This utility unlocks advanced robotics or “Physical AI” 4(which will be the part III of this blog series). It unlocks alternative power sources like nuclear SMR and fusion. It unlocks quantum computing. It unlocks smart warfare/defense technology. It unlocks advanced healthcare and gene-editing. There’s also a geopolitical angle here: the race to “win AI” against global rivals such as China has incentivized the development of our nation’s infrastructure—from improving the power grid to revitalizing our manufacturing base. These are seen as necessities to develop AI itself. Whether AI actually helps advance technologies of the future through its own capabilities, or simply creates the urgency and capital needed to push them forward 5(it’s both), is almost beside the point.
If our society didn’t unlock AI, maybe a truly advanced robot would take 100 years6(or “turns” in Civilization lingo) to develop on its own. Imagine an absurd scenario–in 2126 AD, Boston Dynamics finally trains the perfect humanoid robot via a massively long manual command tree that needs 10,000 state-of-the-art processors to work. This would only happen in a fictional computer game like Civ because in real life, such slow progress would quickly be labeled a waste of capital and then face abandonment. But with artificial intelligence, where the robot has a brain and learns tasks on its own? It’s now right around the corner–in my opinion, within 5 years.7The venture capitalists need that reasonable timeframe to know they will see the returns in their lifetime. Don’t believe me? Believe this guy instead:
“ Physical AI, as a large category, is the technology industry’s first opportunity to address a $50 trillion industry that has largely been void of technology until now… Now the question is how much longer? In three to five years, we’re going to have robots all over the place.” — Jensen Huang, March 20th 2026.
AI creates a flywheel for civilization. Better AI accelerates robotics, energy, healthcare, defense, quantum computing, and scientific discovery. Those advances, in turn, create better infrastructure, more energy, more data, more capital, and better tools to push AI even further. Each breakthrough makes the next breakthrough a little easier. It advances us into a new age on the technology tree.
I think this is a rare moment in time where you can win big just by participating and having common sense. You don’t need uncanny foresight to only pick out the ultimate winner as there will initially be many winners of varying magnitude. You don’t even need to be in AI itself if you don’t like the valuations; you can be in the next tier of tech tree that AI enhances, like applications or robotics. Just start investing. Now. Things are moving way too quickly to hesitate.8Anthropic and OpenAI have already switched pole position multiple times since I started drafting this months ago.
Now let’s talk history for a bit.
The History of AI
Here is the short history of Artificial Intelligence, just to understand how we got here.
First, you had the ancient era, characterized by man’s initial imagination of what the “machine” could be—captured in books such as Frankenstein and I, Robot, and in Alan Turing’s early ideas about whether machines could actually think. This was followed by the first breakthroughs in computing technology after World War II. This era is where our civilization had to grind through the earliest levels of the modern era’s tech tree just to make AI even visible as a future technology.
Then came the first AI era, centered around symbolic AI and the earliest neural network research. Researchers believed intelligence could be built by hand, one rule and one logical step at a time.
Then came the second AI era in the 1980s, when expert systems took off. The idea: capture the knowledge of human specialists and turn it into software that could reason like they did, albeit in a limited, domain-specific capacity.
Now we’re in the third AI era: the era of scaling large language models (LLMs), or the generative AI era. The basic neural network idea didn’t just make a comeback—it became the foundation of modern AI. Instead of trying to manually code rules, we now train massive neural networks on enormous amounts of data and compute. Then you give the model a prompt and it predicts the next token.
Because we are in the third AI era where we are investing all of our money in and hoping for success 9(and maybe the rise of AGI?), this is the one we need to break down at a more granular level.
The History of the Generative AI Era.
2012 — Deep learning has its proof-of-concept moment. AlexNet, a deep neural network trained on Nvidia GPUs, wins the ImageNet competition in a landslide, dramatically outperforming every other computer vision system. It proves that neural networks, when paired with enough data and compute, can beat decades of traditional techniques like expert systems or symbolic logic. This gives the AI field, long a fragmented collection of academic experiments, a real roadmap forward for a working general product.
2017 — Google invents the Transformer. In the landmark paper Attention Is All You Need, Google researchers introduce the Transformer architecture. Unlike earlier models, Transformers process information in parallel and use an “attention” mechanism to understand relationships between words. The result is dramatically faster training, far better scaling, and the foundation for virtually every modern LLM.
2022 — ChatGPT becomes AI’s iPhone moment. OpenAI packages years of LLM development into a simple chat interface that anyone can use. For the first time, hundreds of millions of people experience an AI that can process their question, seemingly reason and then respond. This ignites the modern generative AI boom as LLMs become the fastest-adopted consumer technology in history
2024 — The AI CapEx Supercycle Begins. After ChatGPT proves AI demand is real, Microsoft, Amazon, Google, Meta, and OpenAI collectively commit hundreds of billions of dollars toward GPUs, data centers, networking, and power, transforming AI from a software revolution into a full-scale industrial revolution. Combined hyperscaler capex reached roughly $261 billion in 2024 and has continued climbing sharply since.
2025 — The Agent Revolution. The debut of OpenClaw shows that AI has graduated from chatbot to coworker. Agents can reason through tasks, use software, call APIs, write code, and interact with computers much like a human employee. This marks the beginning of limitless digital labor—fundamentally changing how software and businesses are built. A year later comes the Hugging Face incident, offering an early glimpse of the risks posed by increasingly autonomous AI agents.
2026 — The Frontier Lab Wars. The AI race turns into an all-out arms race between two frontier labs. Anthropic overtakes OpenAI in private valuation as Claude Fable emerges as a dominant force in enterprise and coding. Both companies approach trillion-dollar valuations, backstopped by tens of billions of dollars in annual recurring revenue, and prepare for massive IPOs. Open-source models like DeepSeek and Moonshot AI wage LLM guerrilla warfare, trying to deliver 90% of the capability for a fraction of the cost. By September, OpenAI strikes back with GPT-6 “Astra,” which appears to leapfrog Claude Fable on several benchmarks.
So what does it all mean?
I think it is important to know the history so we can understand the downside. There’s been two AI eras before this and both ended badly, leading to an AI winter where the promise of general intelligence faded, funding greatly declined and almost all companies went bankrupt. Is this time different? Are we headed in the same direction? I don’t think so. The current AI boom is already many orders of magnitude larger than the first two in terms of economic, scientific, and cultural impact. It’s like comparing American football in the early 1900s to the NFL today. The former was a rough, nascent pastime played by factory workers in front of small crowds–it could have easily failed like jai alai or roller derby. The latter version is a multibillion-dollar institution woven into the fabric of our nation’s culture and its popularity has proven robust against multiple scandals as well as the competition of other billion dollar sports.10The NFL took about 35 years since the AFL/NFL merger to get to that point. Keep in mind it’s only been 3 years since ChatGPT! We have passed the point of no return.
Pay close attention to the CapEx Supercycle, because the amount of money being poured into this infrastructure buildout is historically significant. The first two AI booms were funded on a completely different scale. Much of the spending came from niche government programs, research labs, and agencies like DARPA. The technology industry itself was also far smaller, and the largest tech companies of the time represented a much smaller share of S&P 500 earnings than they do today. This time, the companies funding the AI boom are the largest and most profitable businesses in human history. They have a massive amount of brilliant talent at their disposal. They are spending far more on the AI buildout, as a percentage of GDP, than any other large capital expenditure project in U.S. history.
These are the giants of tech declaring “We believe this is the right thing to do.” Who am I to stand in the way and argue? Just follow the money.
In the Scaling Era of AI, that’s all we need to do. The largest and most profitable technology companies in the world are collectively betting hundreds of billions of dollars that scaling will continue to work–that more data and more compute creates more progress. For them, AI has become one giant experiment to see how far the scaling laws can be pushed. So far, the results still show an upward trajectory. Models continue to improve, capabilities continue to expand, and AI-driven revenue continues to accelerate.
That creates a powerful feedback loop: more capital → better models → more valuable products → more revenue → even more capital to develop better models and better products and make gobs of money. No one knows where the limits are—or when that loop will eventually break. Until our tech overlords can identify that perilous edge of the waterfall, the incentive is still to keep scaling.
This was all I needed to understand to pledge my sword to the AI Bulls. I don’t need to be a scientist or an oracle of the future. I can just be a good little trader and follow the massive and historical amounts of money pouring in.
Yes, there are always question marks and worries along the way. It is hard to believe the current scaling paradigm can simply last forever until super-intelligence is achieved. It is understandable to be somewhat skeptical. But for now, we don’t know how long this lasts and what orders of incredible progress can be made until a real roadblock has been hit. The burden of proof is on the skeptics and bears to declare that progress has halted. As long as our tech overlords are true believers, so am I.
Let’s talk about investing approaches to AI.
Each category has a different dynamic at play. Infrastructure is the easiest to understand because you can physically see the money being spent from a top-down standpoint. Frontier labs probably have the greatest theoretical upside–the idea being the first to AGI will own the intelligence layer of humanity. They also have the most risk/volatility as they hurl money into a brutal arms race. Applications are where AI ultimately has to prove that it can generate a positive return on all this investment with actual utility to enterprises and consumers. If AI makes our world a better place, it will be this layer where that will be reflected.
The infrastructure companies build the factories. The frontier labs generate the intelligence ‘utility’ inside those factories. The application companies package that intelligence into something useful enough that people will pay for it. I believe there will be incredible returns investing in all three layers, but some risk/reward/moat profiles are easier to analyze than others.
AI Infrastructure — “the Picks and Shovels Trade”
How many times in the past 18 months have you heard that we are in a compute shortage? Probably a million times if you’re following anyone who has anything to say about AI. Demand for intelligence has been insatiable and it takes an incredible amount of power/chips/compute to bring it operational. So this CapEx or ‘buildout’ that everyone talks about is all about solving that shortage.
Let’s flashback to one year ago to January 2025. At the time I started seriously considering the AI infrastructure trade, NVDA had already gone up roughly 10x in three years. To the untrained eye11(read: me at this time), this meant the picks and shovels trade was more or less priced in.
Then DeepSeek released their heralded Model R1, which supposedly only cost $30 million12(this figure is very much disputed). This was absolutely not priced in and the AI infra trade crashed.13(NVDA drawdown from its early 2025 highs after DeepSeek impact was about 21%–about $600 billion of market cap)
Almost overnight, the conversation shifted from “there will never be enough compute” to “maybe none of this expensive compute is even necessary.” Investors started questioning the entire capital expenditure cycle and worried that the industry was building too many data centers for an overhyped technology with questionable returns. The infrastructure trade was declared dead. People were calling for OpenAI to be a zero for overbuilding14one source: https://www.youtube.com/watch?v=r8cpy1Gxe0o.
“Good thing I didn’t really invest in AI yet.” or so I thought… private valuations of the largest AI companies never budged during this entire time. I checked every single day.
Then the public markets slowly recovered. We went back to par. We went to new all time highs. Frontier labs and hyperscalers continued to follow their Capex-driven game plan. Neoclouds, memory names, semiconductor names kept climbing higher. The new narratives, like the HBM bottleneck and the need to build gigawatt-scale power, took over. The AI Infra trade was clearly NOT over. Demand continued to surge as models improved and agents/applications proliferated. Even today, as of September 2026, we are still massively compute constrained. Leading AI labs will still shut off your token usage if you’re draining too much from the well.
I think a lot about the DeepSeek moment. To some observers15(read: me at the time), it’s a perfect example of just how volatile and fragile the AI trade can be. To me, it encapsulates something else entirely: that Wall Street still doesn’t fully understand the AI trade. This is not a clean, black-and-white numbers story for the spreadsheet jockeys. It’s still fundamentally a technology story, and we’re still early in figuring out where that technology is going.
One popular bear argument around AI infrastructure is “circular financing”: Nvidia and the hyperscalers invest in companies that turn around and spend the money on their chips and compute. Hurp-derp, they’re manufacturing demand and inflating the books! But these companies are trying to build billions of dollars of infrastructure before it can produce the equivalent revenue. The young companies actually building it don’t always have the balance sheets to move fast enough, while Nvidia and the hyperscalers have tons of capital and desperately need more capacity to come online. So they help finance the buildout. This isn’t particularly exotic—aircraft, auto and equipment manufacturers have financed customers forever. If customers are using the compute and capacity remains scarce, then it’s not fake demand. This is simply a capital-intensive industry bootstrapping itself during a massive buildout.
Two of my earliest names—Lambda and Crusoe—fall into the neocloud category, cloud providers that are purpose-built for AI. Every time you hear about datacenter buildouts or leasing–there’s a good chance these two might be involved in some capacity. With AI demand exploding, these specialized cloud providers seemed like a no-brainer to me. They’re basically AI factories for the next century of AI usage16(remember: think of it as a utility). I also thought they would IPO the soonest in the AI stack17(this thesis was violated when Anthropic/OpenAI filed S-1 much sooner than I anticipated). These companies need to raise money for their own capex (get the GPUs, build the datacenters) and what better to do that than to tap public markets for credit? CoreWeave was the first major neocloud to make the jump, going public in early 2025. Lambda and Crusoe are expected to follow within the next 6–12 months, with others like Nscale and Fluidstack potentially on a similar path.18(I also want to add that I absolutely should have invested in Nebius as it was public and sitting right there, big miss by me)
I also like inference services like Baseten, Fireworks, and Together AI because the next phase of AI infrastructure increasingly shifts from building models to running them. Training is episodic; inference happens every time someone actually inputs a prompt. As AI gets embedded into more software—and agents begin generating vastly more calls on their own—the amount of inference compute required will explode.
As the inference market gets bigger, another part of the AI infrastructure stack is heating up: chip startups purpose-built for inference acceleration. There are seemingly a dozen of them now, all attacking the same basic problem—how do you run increasingly powerful models faster and cheaper than you can on general-purpose GPUs? I personally like Positron. Etched is also getting a lot of hype and has 4x’d in less than a year.
The picks-and-shovels trade is usually the easiest one to understand in any technological gold rush, and for better or worse, it’s where I’ve placed most of my bets.
AI Infra names I’m in: Crusoe, Databricks, Lambda, Positron, Radiant, Sygaldry
Names I’m considering: Baseten, Fireworks, Etched, Nscale, Fluidstack, VAST Data, TogetherAI, Modal, Fractile, MatX
The AI Labs aka “the Holy Grail trade”
The second investment path are the AI labs. At the elite frontier, you have just two: OpenAI and Anthropic. Then behind them is xAI, Google DeepMind, and Meta who have fallen slightly behind, at least for now. Then you have the Neolabs, which are the smaller upstarts trying to do something slightly different or maybe open source: DeepSeek, Moonshot, Mistral, AMI Labs, Safe SuperIntelligence, ThinkingMachines, Hark, among many many others.
This is probably where the greatest amount of value could theoretically be created.
Here is the Holy Grail of all bull cases: if one company builds a genuinely superior general-purpose intelligence, it would become the most important company in human history. It would not simply sell software. It would sell labor, research, judgment, creativity and decision-making. I don’t know what price you could even put on this company.
The more realistic base case, in my view, is that models become semi-commoditized. No single lab runs away or even attains AGI, but AI still becomes enormously valuable. Instead, we get a heterogeneous market. Cheap and open-source models handle routine tasks, while expensive closed frontier models are deployed where the marginal increase in intelligence actually matters. Companies might use one model for coding, another for research, another for customer service, and a cheap open-source model for millions of mundane internal tasks. The frontier labs still build enormous businesses, but competition prevents any one of them from being the Holy Grail.
The bear case? The scaling era of LLMs hit a real wall well before anything resembling AGI, technological progress stalls, and eventually all models are completely commoditized. Open-source models remain close enough to the frontier that nobody can sustain pricing power. Customers freely move between models based on price and token prices end up in a race to the bottom. Frontier models can’t sustain their margins and valuations collapse.
There is a lot of uncertainty now. It is too early to tell what these companies will be in “steady state” (when progress isn’t so rapid) but revenue and adoption does continue to accelerate.
Still, this category is difficult for me. The valuations are enormous, the competitive landscape changes every few months and the capital requirements keep rising. It’s too early to know whether these companies are developing durable moats through brand, distribution, or network effects—or whether users will simply migrate to whichever model is best and cheapest at the time. Right now, I suspect the value resides in alternative neolabs who can capture the value that OpenAI/Anthropic won’t but you need a strong understanding of AI architecture to analyze their quality.
One example is the emerging group of companies building world models, which represent something of an alternative—or perhaps a complement—to the current LLM consensus. Their basic thesis is that predicting the next token is not enough to build truly intelligent systems. An AI that wants to operate in the physical world needs to understand how that world actually works: how objects move, how actions produce consequences, how environments change over time, and what is likely to happen next. Instead of simply asking, “What word comes next?”, a world model is trying to answer something closer to “What happens next?” I’m probably not the guy to make this my investing alpha but I nonetheless threw a flier into the AMI Labs pre-seed round.
AI Labs I’m in: Anthropic, AMI Labs
Names I’m considering: OpenAI19(I completely overthought this one but I probably won’t invest at where it is now, which is over 1T), ThinkingMachines, Safe Superintelligence, WorldLabs, Prometheus, DeepSeek, Moonshot, ReflectionAI
AI Applications aka “the Wrapper Trade”
Applications are where AI eventually has to show an ROI to a business or a consumer.
There was a survey done at MIT called The GenAI Divide: State of AI in Business 2025. The researchers reviewed 300+ publicly disclosed AI initiatives, conducted 52 structured interviews, and surveyed 153 senior leaders. The headline finding was: 95% of organizations were getting zero measurable return from their GenAI initiatives, while roughly 5% of integrated pilots were producing substantial measurable value.
There are still plenty of skeptics that will cite studies like this as evidence that AI is an impractical novelty. My counter? I love reading this study. It means we’re still early. It means a huge portion of the enterprise world still hasn’t figured out how to properly deploy these tools and create surplus value. Some 72 year old CEO’s sole attempt at “deploying AI” might just be prompting “help my business please”. Thus there exists a massive distribution, education, and implementation opportunity.
There’s an entirely new category of AI-native application companies, one layer below the models, that were basically born yesterday. These are not generic chatbots looking for a problem, but software designed around the actual workflows of lawyers, doctors, programmers, and knowledge workers. Real engineers then embed these applications into existing workflows for non-technical users.
Look at Harvey in legal, where they aim to greatly reduce the enormous amount of legal grunt work such as research, contract review, due diligence, document analysis and drafting. Harvey says more than 200,000 lawyers across 2,400+ organizations use its platform, including more than 75 of the Am Law 100.
In medicine, OpenEvidence is building something similar for doctors: an AI-native medical information and clinical decision-support platform that lets physicians instantly find medical literature instead of manually digging through papers and guidelines. It now incorporates material from the New England Journal of Medicine, JAMA, Nature, Cochrane and NCCN, among others, and says it is used by medical professionals across 10,000+ U.S. care centers.
Software engineering is where AI has indisputably made the most visible progress so far. Coding was an almost perfect early use case: the work is digital, the output is easily testable, and enormous amounts of code already exist as training data. Cognition’s Devin is an attempt to turn the coding copilot into an actual software-engineering agent: give it a task and let it navigate a codebase, write code, test it and iterate with substantially less human hand-holding.
The application layer has a very obvious bull and bear case happening at once.
It is bullish because the cost of building software is collapsing. A small team can now create products that would have required an entire engineering department a few years ago. Software can be personalized for narrow industries and obscure workflows that were previously too small to serve. In the past 10 months, we have seen previously heralded SaaS companies from the pre-AI era re-rated significantly lower due to this existential threat.
It can be bearish for exactly the same reason. If your company can build an AI application in six weeks, so can 500 other companies. So can Anthropic and OpenAI and they have better distribution and more compute. Existing software companies can bundle similar capabilities into products that enterprises already buy. A thin wrapper around somebody else’s model might not be a durable moat.
The best AI application businesses will probably own something that the big frontier models do not: proprietary data, customer distribution, regulatory approval, a trusted brand, pricing structure that aligns with the customer, and forward deployed engineers who embed value at the company. They might also target specialized verticals where the frontier labs have little incentive to compete directly. If models commoditize, the surplus value might leak to applications.
Right now, as of writing this, I’m also investigating the emerging consumer AI assistant category. The hype company that all the tech geeks are talking about is called Instinct. Instinct is basically trying to build the AI personal assistant that handles the part of life we don’t want to. Instead of prompting a chatbot every time you need something, you give Instinct access to your digital life—email, messages, calendar, screen, location and other apps—and then text or call it like you would a human assistant. These products may soon take device-form (glasses, watches, earbuds) with some kind of microphone attached to take instruction and collect context.
There will be unbelievable winners in the application layer. There will also be many thin wrapper companies that briefly go viral and then disappear. I’m still researching and have yet to really fire my bullets in this category.
AI Applications I’m in: Hark20(they’re basically doing what Instinct is showcasing)
Names I’m considering: Glean, Harvey, OpenEvidence, Clay, Cognition, Factory, Higgsfield, Suno, Sierra, ElevenLabs, Instinct, Perplexity
The Physical AI trade
We are now about 3 years into the AI trade. Let’s say you don’t like it for whatever reason… instead, we can try to skate to where the puck is going.
That, my friends, is where the money is going next.
I honestly just need to hit the publish button so let’s just save the rest of my uncollected thoughts for Part III: Physical AI and Robotics and Other Frontier Technologies.
It won’t be a perfect process.
I originally thought I could craft this complete manifesto on how AI will take over the world. But all I really have are my own learning-on-the-fly experiences, along with these loosely collected principles (mostly borrowed from other tech investors) that I’m now passing along to you. I did the best I could and I hope it was semi-coherent. I want to call back to my first post.
I’m probably just a know-nothing bum with no real skills to create or source long-term value–please keep that in mind as you peruse through this post, which may come off as a day trader LARP-ing as a venture capitalist. My investments are not recommendations. Don’t read this as a guide to success—it might just be a roadmap to wrecking your car.
I don’t know everything. I am not qualified to write about the technology in any real depth. Not even close. I’m just humbly reading about AI developments every single day and constantly find myself in awe of the founding teams and engineers building this world. I don’t even know if I believe AGI is possible solely through scaling LLMs.21I don’t think you need to get all the way there to make incredibly profitable investments anyway. I don’t understand every part of the tech. I almost certainly have blind spots. I might be directionally correct and still pick some wrong investments.
But as someone who has never had much of an investing temperament, and definitely never had much of a techno-hypebeast personality, I gotta say this: I have never been more convinced that this is the defining technological trend of our lifetime. I think the societal impact of AI will exceed cloud computing, social media, and dotcom all combined.
Agnostically speaking, just looking at it through the eyes of a trader of 15+ years experience, I don’t see anything resembling the insanity that usually marks the top of a bubble. We haven’t come remotely close to a 2021-equivalent year of frothiness. I still see plenty of skepticism, much of it based on views that already appear factually invalid. I think there are still more people who hate AI than people actively trying to get rich off it.
I’ve already changed my views on AI so many times, and I’m sure I’ll change them many more times from here. When my views were outdated, I missed some incredible opportunities—like being able to invest in OpenAI around the $150 billion range or, an even worse miss here–Anthropic around $60 billion. I missed great public market investments like the HBM stocks and Nebius22(I do have some AI names but they were mostly lucky holdings from pre-2020 like GOOGL, TSM and NVDA). The skeptic inside of me kept saying “nah, play it safe. you don’t know what you don’t know”.
I decided that willful ignorance cannot be the right call anymore. It’s time to put in the work, even if I’m starting significantly behind in terms of technical knowledge. The more I’ve learned, the more I’ve continued to buy in. And incidentally, the more my own AI usage has increased as well.
Maybe I’m just slowly drinking more and more of the Kool-Aid and becoming the sucker on the bid that the old, skeptical me would have been shorting into. We will see what happens. Hopefully, a decade from now, I’m not writing the post-mortem on this post and lamenting the worst investing decision of my life.
