Caught My Eye…

1) SpaceX’s $60B Cursor Option

SpaceX said on April 21 that it secured an option to acquire Cursor, the AI coding startup, for $60B later this year, or pay $10B for a new partnership if the acquisition does not happen.

Cursor has become one of the leading AI coding products by turning the developer environment itself into an AI interface. The app sits closer to the work itself: where engineers write, test, debug, and increasingly delegate software tasks. It maintains a daily workflow used by high-value technical workers and creates a distribution channel through which model providers can compete for developer usage.

Cursor still sells access to Claude and GPT models even as Anthropic and OpenAI build their own coding products, which makes Cursor both a customer and a competitor to the frontier labs. And last month, two of Cursor’s most senior engineering leaders transitioned to xAI, where both report directly to Musk.

Reuters framed the deal as SpaceX’s deeper push into AI developer tools, while TechCrunch reported that Cursor had been preparing a $2B private fundraise at a $50B valuation, up massively from its $2.5B valuation last January.

SpaceX said the partnership would combine Cursor’s product and distribution with the Colossus training supercomputer, which is described as having the equivalent compute power of one million Nvidia H100 chips.

Both companies can find a win-win scenario where Cursor gets access to the massive amounts of compute that rivals the frontier labs, while xAI has a chance to break into the highly sought-after enterprise coding business with a SOTA model.

2) Table-Tennis Robot Beats Elite Players

Nature published Sony AI’s Ace research on April 22, describing an autonomous table-tennis robot competitive with elite human players. The paper says Ace played under official International Table Tennis Federation rules against five elite players and two professional players. It won three of five matches against elite players and remained competitive in the others. Sony AI later reported additional post-submission matches in which Ace defeated elite and professional players.

The reason this matters is not that a robot can hit a ball. Robots have played simplified versions of table tennis for decades. The hard problem is real-time interaction against an adversarial human, with spin, speed, imperfect perception, and less than a second between shots.

Nature notes that table-tennis balls can exceed 20 meters per second in high-level games and reach spin rates of 1,000 radians per second. Ace addresses that problem through event-based vision sensors, model-free reinforcement learning, and high-speed robotic hardware. Its custom robot platform has eight degrees of freedom, designed to give the racket enough position, orientation, and velocity control for competitive play.

In software, an AI system can pause, sample, retry, and correct. In table tennis, the world keeps moving, and the system must perceive the ball, estimate spin, plan a return, move the racket, and adapt to a human opponent in real time.

3) A New Way Life Produces DNA

Science published a Stanford-led paper this month showing some bacteria appear able to make DNA in a way researchers did not think life used. Normally, DNA is made by copying a template. One molecule pulls the two strands apart, and another builds a matching strand along each side. The result is two copies of the same DNA. But new research on how bacteria fight viruses shows that this rule has at least one exception.

In this case, the researchers found a bacterial system that can build part of a DNA sequence without following the usual script. It suggests biology may have more ways of making genetic material than experts have long assumed.

The discovery does not overturn modern genetics, nor does it mean that all life works this way. It seems to be a specialized trick used by bacteria, likely as part of their defense against viruses. But narrow exceptions can reshape science by revealing unknown mechanisms that researchers can now leverage. It opens a new chapter in bacterial immunity, and it gives synthetic biology a new enzyme architecture to study, even if practical engineering applications remain speculative.

Learn With My Friends and Me…

All In: SpaceX-Cursor Deal, SaaS Debt Bomb, New Apple CEO, SPLC Indictment, Colon Cancer Spike

Other Reading…

• Data Centre Delays Threaten to Choke AI Expansion (Financial Times)

• Fabric Is Harder Than Steel (Amy)

• The New Trade Order (Foreign Affairs)

On X…

To the supporters and friends of Social Capital:

This is our eighth annual letter, in which we share our investments, our observations on the year, and our thinking on technology, markets, and the future.

2025 was, in some ways, a historic year for Social Capital. Numerically, we did well. At the end of the year, our portfolio took an important inflection upwards when NVIDIA licensed Groq for $20B.

However, 2025 felt neither particularly successful nor unsuccessful for me in investment terms. Investment returns are a deeply lagging indicator, as is revenue of work done. If you had made a good decision, at some point in the past, it may bear fruit today or in the future. For Groq, because I made the decision to invest ten years ago, the results, while valuable and valued, do not accurately represent any meaningful work I did in 2025.

So what did I actually do?

Last year, there were three things I needed to do on the investing side. First, I had to set up my portfolio to thrive through the changes that were coming with the Trump Doctrine and the AI geopolitical reorganization (more on this later). Second, I wanted to create an unexploitable portfolio to thrive in this age of AI uncertainty. Third, I doubled down on a handful of businesses that were working.

On the entrepreneurial side, I really focused on scaling 8090 by clarifying the product strategy, sharpening a crisper go-to-market, closing some important deals, and refining 8090’s flagship product - Software Factory - so that it was in a position to be launched into the wild. We finally shipped it in February 2026.

2025 was about a cascade of tactical decisions. It was about me being in the engine room, tinkering. Did every tactical call pay off? No. But I kept my hands on the controls through all of it, and put us in position to do something big in the future. If Groq’s outcome is a clue, perhaps I’ll know in ten years if I made the right decisions in 2025.

Now here’s the really important part: when it’s raining money on your head, you need to be extremely grateful but realize that it’s temporary. That it’s not you. It’s not an accurate representation of today because it’s a reflection of all the big and small decisions that you got right years ago. You can’t let it disrupt your ability to see clearly. This is true in investing and entrepreneurship.

By the way, the discipline is the same in both directions: do not let the windfalls inflate your judgment, but do not let the losses distort it either. In 2024, I took a $400 million markdown in another portfolio company following a recapitalization after it ran out of capital. It stung but had no lasting emotional impact because I could see it coming. I’d already gone back, assessed the mistake, understood what I could’ve done differently… and moved on. So what have I learned from this?

I’ve said repeatedly that investing is a fundamentally internal struggle. But in many ways, entrepreneurship is as well. The feedback loops for outsized success are so long in both realms that for years at a time, the only person who can judge whether you’re on the right track is you. Everyone else is reading the scoreboard. But the scoreboard tracks lagging indicators. The inputs won’t show up for years. And the scoreboard is everywhere. It’s the press, your X feed, your peers, your team, your friends. The temptation to measure yourself by it is constant. In hindsight, it is clear to me that every major mistake I have made came from taking that scoreboard too seriously and giving in to the crush of outside expectations: when I self-sabotaged my path by replacing my own conviction with someone else’s. While hard to live by, all the greats show that the only proof that matters is your own. Because the moment you outsource the judgment, you outsource the power. And you cannot afford that.

Always think for yourself.

Once you stop outsourcing the judgment, and once you stop relitigating the losses or inflating the wins, once you let go of the future and the past, you’re left with a surprisingly simple question:

What should I do today?

Success is found in the present. Learn from the past, then move forward. Take a curious guess about the future and try something, then come back, see what you learned, and then repeat. There’s not much more to that. Sometimes, people just get in their own way by putting too much weight into these decisions. I try not to. At my best, I have no nostalgia. I think about the future, but I do my best to live in the present.

What the hell is going on – 2025 edition

Superintelligence is coming but it’s not here yet.

In my 2018 Annual Letter, I encouraged our readers to learn about two new technologies: artificial intelligence and computational biology.

“Over the next few decades,” I argued, “...we should see artificial intelligence successfully applied to problems that were previously too difficult using traditional software methods, such as self-driving cars, robotics, and drug design and discovery. AI can truly transform how we work, how we live, and even how we think.”

Good call.

In line with that thesis, I invested in companies applying AI or building the infrastructure that powers them.

Then ChatGPT landed in 2022, and AI became mainstream. A few years of remarkable progress followed: the duration of the tasks an LLM can confidently execute has gone from 10 seconds in 2022 to many hours in 2025. By 2025, AI lab CEOs were forecasting “artificial general intelligence” or even “superintelligence” arriving within the decade. It’s worth pausing here.

Let’s call today’s AI what it really is: an ultra-advanced, mathematically grounded autocomplete and logic checker. It’s computer software that predicts thousands of words into the future instead of one. That gap, from one word to thousands, turns out to be where something unexplainable happens. It is solving incredible problems that stumped us for decades. It is able to perform certain tasks like image classification or code generation as well or better than an expert. How it happens is still, at some level, a mystery.

When we can’t fully understand something, the temptation is to reach for supernatural language. In testing the most effective messaging for fundraising, parts of the AI industry decided that doomerism was the most effective. I get it. Fundraising requires narrative, and “we are building God” is a better pitch than “we wrote some very clever linear algebra, pirated the internet, and threw a bunch of compute at it.” This is what an industry still figuring out its place in society looks like. But let us not confuse the pitch with reality.

The cost of that immaturity, however, is now measurable: a recent NBC News poll put AI’s net favorability at negative 20 among registered voters. Only the Democrats and Iran, a nation we’re at conflict with, are less popular.

Because this feeling of “magic” isn’t exclusive to AI, nor is it new as it seems. We just weren’t alive for most of the last few rounds. Imagine discovering that an invisible force could light entire cities, power machines, and travel across continents (electricity). Imagine hearing a human voice travel across a wire for the first time (telephone). We weren’t alive to remember those. But we might remember the first time we pulled up a webpage and realized the entire world’s information was at our fingertips. That felt like magic too.

AI is no different. It feels unprecedented, but the experience of it is not. Every major technology feels like magic at first. We just forget how strange it once seemed, once it becomes ordinary (when was the last time you stared at your lightbulb in awe?).

To be clear: AI does not need to be superintelligent to be transformational. But it has to be trusted because low public trust leads to lower adoption and worse regulations, which in turn slows diffusion and may put America at a disadvantage relative to China, where 87% of the public views AI positively - or at least that is what we are told.

It would be much better if the AI industry said soberly: there is a lot of experimentation, the revenue is promising, but we do not yet know what is real, we are going to figure it out methodically, and we are going to work within regulated industries rather than flout them. That is a mature message. It is also the truthful one.

The AI Stack

If you look back at the evolution of humanity, there have been moments that enabled exponential expansions of human productivity. Every one of them also came with areas, in that expansion, where economic value concentrated. The assets at those layers are what I call fulcrum assets, because everything else pivots around them.

Historically, for oil, the fulcrum assets were refining and transport. For railroads, it was control of the network. For the internet, Global Crossing and WorldCom bet it was infrastructure and spent billions on fiber in the 1990s (they drastically overbuilt, and both went bankrupt). The fulcrum turned out to be the platforms that owned the users.

The question is: where are the fulcrum assets in the AI stack?

To find the fulcrum, you first need to define the stack. Take, for example, the first generation of the internet. In 1978, the ISO drew a map of the internet economy in seven layers, from physical cabling to application software. It was called the OSI model, and it demarcated where one company ended and the next began. That clarity built some of the greatest technology companies in history, and the framework preceded the industry. We need this for AI.

The AI stack is the complete architecture required to generate intelligence at scale, where every layer depends on the one below it, and fails without it. It begins with raw inputs such as critical minerals and energy and ends at the application layer.

As you can see, in our framework, the AI stack builds on a shared infrastructure base. Above that, it diverges into physical AI and digital AI, because great robots and great digital AI require different boundary conditions.

Digital AI trains on text, code, and images, data of which there’s a massive supply on the internet. Physical AI trains on sensory data: touch, gravity, friction, spatial depth. That data barely exists. It has to be collected from the real world, slowly.

The data are also different in kind. Text is words in a sequence, whereas sensory data is continuous streams of spatial and physical information coming in all at once. Models need different architectures to process each, the same way your brain uses different regions for language versus balance and movement.

Digital AI runs in data centers with abundant power and memory. Physical AI has to run on the robot itself, on a small chip with limited power. The robot can call a cloud model for reasoning, but its real-world model has to be local, because when it’s about to fall, it can’t wait for a server response. It needs to react instantly, the same way your reflexes work faster than your thinking.

Between the model and the application sits an execution layer. On the physical side, that means actuation: sensor fusion, kinematics, batteries. On the digital side, it means agent systems: tool use, orchestration, memory, APIs. Most applications will use both. In these scenarios, you’ll see a dominant model running the application, then calling on specialized models as needed. A hospital robot uses a language model to talk to patients but a spatial model to navigate hallways. A warehouse robot uses a vision model to move boxes but a language model to understand spoken instructions.

On the digital side, the model layer is commoditizing fast. I view foundation models as equivalent to electricity, where the model is not the moat. Rather, what you build on top of it is.

So if the fulcrum is not in the digital layer, where is it?

First, let’s consider silicon. Silicon is a sixty-year-old industry. There are bottlenecks inside it, but the asset itself is not the bottleneck. You can build a GPU, an ASIC, an FPGA, or a CPU on the same fabrication line, you are really just debating the process node. Memory looked like a chokepoint until people found ways around HBM with DRAM and SRAM. Networking looked like a chokepoint until photonics offered an alternative to InfiniBand.

Energy storage and actuation are different. A Panasonic line making NMC cells cannot make LFP prismatic cells. All the machines are different. All the tooling is different. You are operating at the level of chemistry: specific slurries, specific coatings, baked at specific temperatures. And without enough stored energy and fine-grained precision in movement, physical AI does not work. The robot just stops.

Said more simply: you cannot repurpose a factory that makes one type of battery cell to make another. Every battery chemistry requires its own manufacturing line, from scratch. And without sufficient energy storage and precise actuation, the robot does not move. Everything upstream becomes irrelevant.

As a result, I think the fulcrum assets of this era sit where the shared infrastructure meets the physical layer: critical minerals, chemical processing, energy storage, and actuation. But building the intelligence that sits on top of them is not free either. It requires deep talent, massive compute, close-to-infinite energy, and capital markets willing to finance all of it.

So who can actually build this?

Bipolar world

For the last 10 years or so, our geopolitical landscape has been formed more and more by the US-China power competition. Building frontier AI also happens to be so expensive and talent-intensive that only two countries can build it at the frontier.

The boundary conditions for frontier AI are: deep talent pools, massive compute infrastructure, close-to-infinite energy, and capital markets that can finance it. Only the US and China have the economic muscle to spend hundreds of billions of dollars on the computational and energy infrastructure that AI requires. Even if a third player wrote that check, they would still be years behind, because the US and China have been compounding since the beginning.

The result is a simple structure: the two superpowers at the center, each building its own intelligence stack. The US overwhelmingly closed-source, China overwhelmingly open-weight. Around them, a small number of countries that hold critical inputs: lithography in the Netherlands, critical minerals in Australia, energy and capital in the Gulf Countries, fabrication in Taiwan. Two planets and a handful of moons. The moons have leverage because they supply something that cannot be routed around.

Everyone else faces a simpler, less comfortable reality. If you do not control a critical input, you do not have a term sheet. You have a menu. And the menu has two items: rent the intelligence from China, which will come with a certain set of conditions, or rent the intelligence from the United States, which comes with a different set of conditions. Every country will have to go through a sorting function of choosing what they want.

Put simply: AI is the defining technological competition of the 21st century, the U.S. and China are the players, and a bipolar world is taking shape. But how did we get here?

Globalism: cheap goods, expensive consequences

First, some history. For most of the post-WWII era, U.S. trade policy was pragmatic and flexible. Presidents from Truman through Reagan used tariffs, sector-specific deals, and hard-nosed negotiations to solve discrete problems bilaterally. For example, when Reagan faced a flood of Japanese car imports in 1981, he leveraged tariff threats to extract voluntary export limits from Tokyo, eventually pushing Japanese automakers to build their cars on American soil. Trade policy was a tool for advancing American interests, not a rigid ideology.

That changed after the Cold War. The Soviet collapse, the 1991 Gulf War victory, and strong economic growth gave the United States its unipolar moment. With that moment came a rare domestic consensus that free markets, deregulation, and strong IP rights were simply the “right” economics.

As the world’s sole superpower, U.S. policymakers had both the ideological conviction and the hegemonic leverage to set rules for a truly global trade order. Under Clinton, Washington got the world to sign onto the WTO. The IMF and World Bank became vehicles for spreading neoliberal policy worldwide. In December 2001, China was admitted to the WTO despite not meeting many of its standards, on the promise that integration would follow. The zeitgeist was that economic openness would inevitably lead to political reform, as it had in South Korea and Taiwan.

While the U.S. was spending money on its protracted war in Iraq, China focused on scaling its economy and building wealth. It became known as the world’s factory. The short-term winners were American corporations and consumers, who saved an estimated $411,000 for every manufacturing job lost to Chinese competition, with middle and low-income households getting access to a mass supply of affordable goods.

As we know now, it came at a cost: America’s position eroded on two fronts.

Economically, we lost our resilience and domestic ability to make things. Since the 1980s, manufacturing has been claiming a smaller and smaller share of all American jobs. Chinese imports didn’t start that trend, but they made it worse. Between December 2000 and January 2026, manufacturing’s share of total U.S. employment dropped from 6.0% to 3.7%. On aggregate, U.S.-China trade was positive-sum. But China benefited disproportionately.

Geopolitically, the asymmetry was even starker. Between 2001 and 2024, China’s share of world GDP quadrupled from 4.0% to 16.8%, while America’s fell from 31.4% to 25.9%. And geopolitical power, while not zero-sum in absolute terms, is relentlessly relative. If your adversary grows faster than you, you lose comparative advantage, even if you yourself grow.

The return of bilateralism

Unlike previous technologies, AI creates a strategic imperative to gate access once you have it. And the moment you rent that intelligence, you cede sovereignty to whoever provides it.

So we have to see past the historical frameworks that have governed geopolitics and into a new way of organizing: not who protects you from a threat, but who gives you access to prosperity. Put another way, the organizing principle of geopolitics is no longer military alliances, but supply chain position.

So if you are a country that is not the United States or China, what do you write on your term sheet?

For most countries, the honest answer is: nothing. But there are five forms of leverage that still matter: critical minerals, endemic high-tech manufacturing, capital, energy, and data center infrastructure. Countries that hold any of these have something to negotiate with. The more of these layers they control, the better their deal. And countries that hold none will face an uncomfortable question: which version of intelligence do they rent, and what do they concede in return?

As a result, bilateralism is back. Why? Because the superpowers want flexibility: the ability to negotiate bilaterally with the countries that hold critical inputs, on their own terms, without being bound by multilateral frameworks that give equal voice to countries that bring nothing to the table.

Belt and Road was China’s strategic response to the old framework: lock in bilateral agreements, get its tentacles into critical infrastructure across Africa, Central Asia, and Southeast Asia, and become totally unexploitable. But they picked the wrong countries and the wrong infrastructure. They were playing to preserve optionality for long-term military and geo-economics in a world where everyone outsources to China and everyone wants to stay rich in the same way forever.

What they did not factor in is that America would simply tax everybody else and use that money to subsidize its way out of dependency. That is what tariffs are. You put everyone behind the wall first. Then you go country by country and say, “I will let you back in, but here is what I want.”

The natural objection is: won’t this push allies to diversify away from the U.S.? I think not. Because what would they be pivoting to? Leverage exists when options exist. You can’t pivot away from America giving you inter-related abundance,prosperity and productivity without knowing with certainty that China can fill the gap. And if AI is the next critical input to prosperity, there will only be one Chinese and one American option. Each country will then need to do a deal in their best interests.

The Trump Doctrine trumped the Monroe Doctrine

The Trump Doctrine is the American response to this new game. Many have called this a modern Monroe Doctrine, or even the Donroe Doctrine. It is not.

In 1823, President James Monroe warned European powers against further colonization or political interference in the Americas. It was a defensive posture: stay out of our hemisphere. For two centuries, it defined the baseline of American foreign policy in the region.

The Trump Doctrine is more ambitious. It expresses strategic optionality across four dimensions: economic resilience, meaning we no longer accept reliance on anyone but ourselves for what we need to grow and prosper. Geopolitical optionality, meaning engagement in global conflict is a choice, not a necessity. Fundamental security, meaning overwhelming military strength that acts as a deterrent. And internal population health, meaning a society that thrives.

Said differently, America wants the freedom to act on its own terms, in its own time, with no dependency that could be used against it.

Things happen in harmonics. After decades of China-based supply chains, the U.S. impulse is to swing the opposite direction and go fully nationalist and reshore production entirely. In the long-term, I see a more balanced approach with supply chains distributed across trusted allies such as Mexico, Canada, Australia, and more.

America is also training its capital markets to underwrite a new set of economics. Subsidies and tariffs have a direct effect: they make it cheaper to build at home and more expensive to build abroad. But they also set a precedent. Capital markets watch where the state puts its weight and recalibrates what they consider financeable. A decade ago, domestic manufacturing was a bad bet. Once policy signals otherwise, the risk models update. Independent of how politics evolve, American capital markets will carry The Trump Doctrine forward.

Why the race to intelligence is a race to peace

So, does this end in conflict?

Not necessarily.

Think about this geopolitical standoff as a heads-up match in poker. In poker, GTO, game theory optimal, is a strategy where you make your opponent indifferent to your actions. You are not trying to exploit them; you are acting to make yourself unexploitable. When both players adopt GTO, the game reaches a Nash equilibrium. Neither side can improve its outcome by changing strategy unilaterally. The result is stability through mutual indifference.

Right now, neither China nor America is at Nash. Both are exploitable. Each is racing to close its vulnerabilities, getting a little less exploitable with every move. You start highly exploitable, you change a thousand variables, and eventually you become unexploitable.

In the view of René Girard, we are more likely to compete over things when we are similar and want the same things, and less likely to fight when we are fundamentally different.

Look at America and China. At a deeply broad societal level, we are fundamentally differentially organized. The rewards in America are money. The rewards in China are recognition and power. One is individualist, the other Confucian. These differences are important and good because they increase the odds of peace.

But there is an overlap. Economically, both need energy and critical minerals. Militarily, China may no longer accept the extent of U.S. power projection in the Indo-Pacific, yet the U.S. may equally refuse to accept China’s growing military presence there. That’s a real risk. But reduce it to its simplest form: this is a negotiation. You close negotiations from strength. The U.S. may escalate to improve its position before any deal is reached.

So how does this end? Both sides reach a state where they cannot be leveraged by the other. We have raw intelligence. They have raw intelligence. Both deploy prosperity and abundance to their respective teams. Countries pick which side they want to be on. And because neither side is exploitable, neither side has a reason to fight. That is the Nash equilibrium.

We are not there yet. Both sides are still adjusting, closing one vulnerability at a time. But the trajectory is legible: the race to intelligence is the race to peace.

The collapse of terminal value

In our first annual letter in 2018, I remarked that “large internet companies are truly just hitting their stride,” and that “Big Tech will get bigger and will leave less room for obvious companies doing obvious things. The demands of innovation are going up.”

Big Tech absolutely did get bigger. The Magnificent 7, the world’s biggest tech companies, made up ~17% of the S&P 500’s value in 2018 when I wrote the above. In October 2025, their market concentration peaked at 36.2%.

The demands of innovation also increased by orders of magnitude.

The architecture of modern capital markets rests on a single, rarely examined assumption: that competitive advantages compound over time, and create persisting moats. That’s why, for most tech businesses, 60 to 80 percent of equity value lives in the terminal value, which is the estimated earnings beyond the period you can credibly forecast.

That number has always been a comfortable fiction. But increasingly, I believe AI is ending it.

The question every business has to answer now is:

“How will this not be unbundled by an LLM?”

If someone can replicate your core product in weeks, your moat is not a moat. And if your moat is not a moat, your terminal value does not exist. And if your terminal value does not exist, then 60 to 80 percent of your equity value simply ceases to exist. What remains is what you can prove you earn over the next handful of years and, consequently, a world where the future is worthless until it arrives.

We saw the first inning of this in 2025: a compression of multiples, the market quietly applying a larger discount factor to future cash flows. Most clearly, we saw this in public markets, which went long on silicon, short on SaaS.

Silicon companies, the ones building the physical infrastructure AI runs on, went up. Broadcom, SK Hynix, and the semiconductor supply chain. Software companies, the ones whose products can be replicated by a better prompt, went down. Duolingo fell 67.5 percent from its mid-2025 peak to the end of year. Adobe fell 24.6 percent from its 2025 peak.

For a decade, SaaS companies were the perfect private equity target: recurring revenue, contractual lock-in, predictable cash flows. From 2015 to 2025, private equity buyers bought over 1,900 software companies in transactions worth $440 billion. By year-end 2025, software and technology companies made up roughly 25 percent of the entire private credit market.

Private equity buys these companies with borrowed money. The SaaS cash flows pay back the loan. At least, they used to. If AI makes the product replaceable, the cash flows disappear, and the loan defaults. Put simply, the default rate on those loans tells you how far along the AI disruption actually is. Private credit is the canary in the coal mine.

But we are not yet in that new regime. Right now, we are compressing the price-to-earnings ratio. The real shift is when the market flips entirely from P/E to multiples of current free cash flow. We will know it’s there when companies only report GAAP, when there are no more adjustments to EBITDA, and when stock-based compensation gets treated as the real cost it always was.

If that transition happens, the implications cascade quickly. Growth investing stops working. The entire logic of growth equity is a bet on tomorrow: sacrifice free cash flow now, build a dominant position, harvest it for decades. That logic breaks when tomorrow is no longer financeable. Who funds a pre-revenue company at a billion-dollar valuation if there is no terminal value to grow into? Capital rotates to the physical world. The corporate world restructures around shorter horizons. M&A logic inverts. Short-termism stops being a pathology and becomes the correct response to the actual incentive structure.

I have written about these implications at length elsewhere, so I will not belabor them here. But the central paradox is worth stating plainly: the companies driving AI disruption are spending $300 to $500 billion per year on infrastructure that only makes sense over a seven-to-fifteen-year horizon. If markets reprice to short-duration multiples, that capex becomes very difficult to finance. Put another way, the disruption engine disrupts itself. The likely outcome is not a permanent new regime but an oscillating one: shorter cycles, higher volatility, terminal values compressing and recovering as AI investment ebbs and flows.

The partial move already underway is significant on its own. A full repricing would be the most consequential structural shift in capital markets since the postwar era.

Long-term investing in a short-term world

If AI shortens the window in which any business can call itself defensible, two questions follow. First, where does private capital go when it cannot trust ten-year projections? Second, who finances the things that still take twenty years to build?

Where private capital goes

First, to the fulcrum assets. Earlier in this letter, I argued that the lasting bottlenecks in the AI stack are physical: critical minerals, energy storage, and actuation. The same logic applies to capital allocation. Energy infrastructure. Farmland. Toll roads. Batteries. If your moat requires atoms, not bits, a better language model cannot unbundle it overnight. For now.

Second, to the small number of companies making history rather than reacting to it.

The logical endpoint of Big Tech hitting its stride is the vertically integrated megacorp: a firm so embedded in physical supply chains that it becomes too big to disrupt. Elon seems to be assembling exactly this, with Tesla, SpaceX, and xAI converging into a single industrial entity.

Some of these companies are already financing themselves like sovereign entities. Microsoft, Amazon, and Apple have issued 40-year bonds. In 2026, Google even issued a 100-year bond that was oversubscribed tenfold.

At the same time, megacorps could face regulatory headwinds. In our 2019 annual letter, I argued Big Tech had become a modern railroad monopoly – hoarding talent, stifling innovation, and concentrating wealth and power in ways that demanded trust-busting.

That was seven years ago. Now look back at the Mag 7 chart earlier in this letter. We’re almost twice as concentrated now. But a new variable has entered the equation, which is that these same companies are spending hundreds of billions on the AI infrastructure that America needs to win the race against China. Much of that spending goes into data centers, silicon, and energy systems that will hold value regardless of which company built them. American regulators now face a genuine dilemma: the same concentration they once wanted to dismantle is now the engine driving national competitiveness. Trust-busting Big Tech and winning the AI race may be incompatible goals.

Who finances the long term

That leaves the second question: who finances the things that still require a twenty-year horizon if private capital will not?

Historically, governments have financed the things a country needed but that private markets refused to touch: the Interstate Highway System, the Apollo program, and the early internet. When the payoff horizon is too long or too uncertain for private capital, the state steps in.

That is what is happening now. The U.S. is already doing this, taking stakes in companies like Intel, negotiating tariff concessions for domestic investment, and subsidizing reshoring. China, through its model of capitalism with Chinese characteristics, always had deep state involvement in financing its strategic industries, and has committed to “never deindustrialize”. States holding large sovereign wealth funds – such as Norway, the Gulf States, and Singapore – are at an advantage. So are nations with low debt levels, should they choose to invest.

Put another way, industrial policy is making a comeback because AI is increasingly changing the world too fast for private capital alone to underwrite.

A closing observation

I opened this letter by mentioning that investment returns, as well as revenue, are deeply lagging indicators. What are the leading indicators? Curiosity, work ethic, being temporally present. And above all: thinking for yourself.

I had a young, very smart guy in my office recently. In our conversations, he’d always bring up others’ viewpoints: so-and-so tweeted this about rates. This person thinks that about China. He had a list of people he’d decided were smart, but what he was actually doing was was outsourcing his thinking to them. He was borrowing somebody else’s conviction.

Here’s what I told him:

Dismantle the idea that social proof matters. The only proof that matters is your own proof. Don’t trust anyone blindly. This is a key mental break you have to make. It doesn’t mean to look down on them, it doesn’t mean to judge them. But it doesn’t mean to blindly believe them either. It just means they are simply a human, over there saying something that may or may not be true.

We all went through this. I went through this evolution. First, I’d think: “This person is theoretically smarter than me, so I need to think what he thinks.” Then, I made a small change: “This person has a point of view that I should pay attention to. Now let me go and break this down from first principles and see where they’re right and where they’re wrong.”

When you do this over and over and over again, you notice patterns, you come to your own conclusions, and you make better decisions.

Put another way, don’t think you’re smarter than anybody, but also don’t think that anybody is smarter than you.

If you look at the great investors and the great entrepreneurs, that is the crucible test. They simply don’t care what anybody else says. Not because they think they’re better, but because they refuse to let somebody else carry what is fundamentally their burden.

I have many smart friends, but I can’t outsource the thinking to any of them. Because if I outsource the thinking, I outsource the decisions, and if I outsource the decisions, I inevitably outsource the power. And outsourcing power will make me miserable.

Which means you should do exactly what I told that young man to do: break this letter down from first principles. See where I’m right and where I’m wrong. If you just take my word for it, you’ve missed the entire point. Always think for yourself.

God is in the details.

Respectfully,

Chamath Palihapitiya

To receive access to all our recent and future Deep Dive reports, consider becoming a subscriber. The full AI Stack breakdown will come out next week.

We’ll also do Q&A on this letter in the members group chat:

Subscribe now

Appendix

2024 Social Capital Performance Summary

2023 Social Capital Performance Summary

2022 Social Capital Performance Summary

2021 Social Capital Performance Summary

2020 Social Capital Performance Summary

2019 Social Capital Performance Summary

2018 Social Capital Performance Summary

Disclaimer

Any opinions in this letter (this “Letter”) are Social Capital’s views and are provided for educational purposes only. None of the above is investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products. Offerings are made only pursuant to a private offering memorandum containing important information. Statements in this Letter are made as of the date of this Letter unless stated otherwise, and there is no implication that the information contained herein is correct as of any other time. Certain information contained in this Letter has been obtained from sources believed to be reliable and current, but accuracy cannot be guaranteed. References herein to specific sectors are not to be considered a recommendation or solicitation for any such sector. Projections or forward-looking statements contained in this Letter are only estimates of future results or events that are based upon assumptions made at the time such projections or statements were developed or made. There can be no assurance that the results set forth in the projections or the events predicted will be attained or occur, and actual results may be significantly different from the projections or statements. Also, general economic factors, which are not predictable, can have a material impact on the reliability of projections or forward-looking statements.

“Social Capital” refers to a collective strategy developed by Chamath Palihapitiya, the founder and CEO of The Social+Capital Partnership, L.L.C. (“Social+Capital”), and used in connection with the SC GP/LP Funds and Proprietary Capital. The “SC GP/LP Funds” refers to the historic venture capital funds managed by Social+Capital, which include (i) The Social+Capital Partnership, L.P. and The Social+Capital Partnership Principals Fund, L.P. (together, “SC Fund I”); (ii) The Social+Capital Partnership II, L.P. and The Social+Capital Partnership Principals Fund II, L.P. (together, “SC Fund II”); (iii) The Social+Capital Partnership III, L.P and The Social+Capital Partnership Principals Fund III, L.P. (together, “SC Fund III”); (iv) The Social+Capital Partnership Opportunities Fund, L.P. (“SC Opportunities I”) and (v) Social Capital Partnership Opportunities Fund II, L.P. (“SC Opportunities II”). The SC GP/LP Funds are not open to new investors.

“Proprietary Capital” refers to certain proprietary investments made by Mr. Palihapitiya starting in 2019 through entities owned or controlled by him. Proprietary Capital was not managed as an external fund and no third-party investor experienced Proprietary Capital’s performance. Proprietary Capital investments are made on an ongoing basis by Mr. Palihapitiya with cash on the balance sheet of Social+Capital rather than commitments of limited partners. The unaudited track record of Proprietary Capital does not represent performance records of Social+Capital, and only includes those investments made by Mr. Palihapitiya since 2019 that relate to the Social Capital strategy and does not include all of his personal investments. Proprietary Capital does not pay fees and expenses similar to what would be applicable to the SC GP/LP Funds or similar funds and, as a result, any metrics that are net of such fees and expenses are hypothesized and presented for illustrative purposes only. Proprietary Capital is not open to third party investors. For purposes of calculating Proprietary Capital’s track record, any realizations are treated as distributions and, to the extent additional investments are made with amounts so realized, such amounts are treated as additional contributions from limited partners.

Past performance is not indicative of future results and investors may lose investment capital. An individual limited partner’s metrics may vary from the reported metrics based upon the specific terms of the capital transactions relating to such limited partner.

“Gross IRR” represents the combined annualized internal rate of return from the inception of such SC GP/LP Fund and/or such Proprietary Capital investment, as applicable, through the applicable date on invested capital based on all contributions, and expenses payable, distributions (including tax distributions), and net unrealized value as of the applicable date, which excludes management fees and other fees net of each general partner’s carried interest (if any), in each case, of such SC GP/LP Fund or such Proprietary Capital investment, as applicable.

“Net IRR” represents the combined annualized internal rate of return from the inception of such SC GP/LP Fund and/or such Proprietary Capital investment, as applicable, through the applicable date on invested capital based on all contributions, and, for the SC GP/LP Funds deducting from such figure the actual management fees and other fees and expenses payable, distributions (including tax distributions), and net unrealized value as of the applicable date, which is net of each general partner’s carried interest, in each case, of such SC GP/LP Fund, as applicable. The “Total” Net IRR and “Total” Net TVPI reflects the composite performance of the SC GP/LP Funds, inclusive of actual fees, carry and expenses incurred, and Proprietary Capital, inclusive of the hypothetical fees, carry and expenses discussed above. Any performance information calculated on a model basis is for illustrative purposes only to show an estimate for the potential impact of fees, carried interest and expenses on gross returns, which impact will be substantial. A single limited partner’s actual net returns with respect to its interest in a fund will vary from the net performance reported herein for such fund based on the timing of actual capital contributions, distributions, taxes applicable to a particular limited partner, and any other amounts borne by or paid into the fund by such limited partner. As a result, the net figures reported for any SC GP/LP Fund and Proprietary Capital may not represent the actual investment experience of any individual limited partner.

“Net TVPI” represents (i) cumulative distributions plus residual value divided by (ii) paid-in capital. This excludes paid-in-capital from, and distribution to, the principal’s funds portion of the SC Fund I, Fund II, and Fund III, which is part of the overall fund size. The principal’s funds portion does not pay fees or carry.

“DPI” represents cumulative distributions divided by paid-in capital. This includes the paid-in capital from, and distributions to, the principal’s funds portion of the SC Fund I, Fund II, and Fund III. The principal’s funds portion does not pay fee or carry.

Proprietary Capital is composed of investments from permanent capital that are not subject to any fees or expenses. Proprietary Capital performance has been provided for illustrative purposes only, and is not necessarily, and does not purport to be, indicative, or a guarantee, of future results. The preparation of such information is based on underlying assumptions, and because it does not represent the actual performance of any single external investor or fund, it is subject to various risks and limitations that are not applicable to performance presentations of any single external investor or fund. Although we believe such performance calculations are based on reasonable assumptions, the use of different assumptions would produce different results. For the foregoing and other similar reasons, the comparability of this information to information of any single external investor or fund is limited, and prospective investors should not unduly rely on any such information in making an investment decision.

This Letter includes only a subset of Social Capital’s investments.

Caught My Eye…

1) Amazon’s $11B Globalstar Deal in the Race to Space

On April 14, Amazon agreed to buy Globalstar for $90 a share, $11.57B in total. Globalstar is a leading mobile satellite service (MSS) operator founded 35 years ago. As part of the deal, Amazon will acquire the satellite operations, infrastructure, and the spectrum licenses. In combination with Amazon’s satellite internet service, Leo (previously Project Kuiper), this will enable global coverage, even in more remote areas and between traditional cell networks.

The business is slated to launch more than 3,200 satellites, but is currently hovering around 200. In comparison, Starlink currently operates over 10,000 satellites.

The acquisition closes in 2027 pending regulatory approval and certain milestones. Amazon announced a parallel deal to carry Apple’s iPhone and Apple Watch features on Amazon’s Leo constellation. Apple also holds a 20% stake in Globalstar from its $1.5B investment in 2024. They invested to ensure Globalstar could expand its network to support certain iPhone features. This makes them simultaneously the announced retail customer on Amazon’s new constellation and seller.

2) IonQ’s Quantum Breakthrough With Light

On April 14, IonQ announced it had entangled qubits across two independent trapped-ion quantum computers using photons. This work was partly funded by the Air Force Research Laboratory. It also landed a U.S. defense research contract through DARPA’s Heterogeneous Architectures for Quantum (HARQ) program. These announcements led to a 20% jump in their stock price.

Frontier quantum efforts usually hit the same wall. More qubits can be packed onto a single chip, but errors and complexity grow faster.

By successfully linking two separate machines using light particles, IonQ demonstrated that you can potentially scale quantum much like we scale the modern internet. Connecting smaller, reliable units into a massive, coordinated network. The key breakthrough was finding a way using commercial machines on standard commercial fiber.

“Scaling quantum computation beyond the limits of a single chip is essential for realizing a future quantum internet.” - Niccolo de Masi, CEO of IonQ

Even though it was a demo rather than long-distance networking, modular quantum computing continues to advance.

Learn With My Friends and Me…

All In: OpenAI’s Identity Crisis, Datacenter Wars, Market Up on Iran News, Mamdani’s First Tax

Other Reading…

• Uber CTO Shows How Claude Code Can Blow Up AI Budgets (The Information)

• Against Cynicism (Will Manidis)

• Want to understand the current state of AI? (MIT Technology Review)

• Beyond the Sky (Colossus)

On X…

We are still completely underestimating how short we are in terms of the global demand and supply dynamics of a handful of critical minerals that we need.

What Are Critical Minerals

The U.S. defines critical minerals as materials that are essential to national security and economic stability, yet vulnerable to supply chain disruptions.

To build the 2025 critical list, the U.S. Geological Survey (USGS) modeled more than 1,200 supply-disruption scenarios, ranking minerals by how severely a cutoff could hit the U.S. economy.

This list grew from 35 minerals in 2018 to 50 in 2022 and 60 in 2025, as AI, defense, and industrial demand pushed more materials past the risk threshold.

The Input Layer Beneath Every Technology

Every EV battery, fighter jet, data center, cell tower, fertilizer system, and oil well depends on them.

The 60 U.S.-designated critical minerals underpin industries worth hundreds of billions across seven baskets: batteries, power grid, magnets, chips and displays, defense, agriculture, and industrial.

Whoever solves cheap energy wins the century, and critical minerals are the first input into that race. Right now, one country has a decades-long head start.

The New Geopolitical Race

In 1992, China classified rare earth elements as protected strategic resources.

Deng Xiaoping famously declared, “The Middle East has its oil. China has rare earths.” 16 of the 17 rare earth elements are critical minerals, used in batteries, magnets, and defense technology.

China spent the next 15 years building the industry the West had let go of: cheap labor, lax environmental standards, and low costs pulled production to China.

By 2010, China had established dominant control over supply, and demand then exploded. Today, China controls about 85% of processing.

While the U.S. has minerals in the ground, it cannot easily process them due to a combination of complex bottlenecks that run deeper than they look.

This 103-page deep dive examines critical minerals landscape and economy from the ground up:

• What they are and why they matter

• The top-ranked minerals where a supply disruption would hit America hardest

• How the supply chain works from mine to finished product

• Where the biggest chokepoints and bottlenecks lie

• How the U.S. and its allies are trying to rebuild capacity

• Where private capital is beginning to find investable opportunities

This has been one of our most detailed and important deep dives this year. I hope you find it useful.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Terafab: Intel Joins Musk’s $25B Bet on Orbital AI

Elon Musk announced Terafab on March 21, a $25 billion chip fabrication joint venture between Tesla, SpaceX, and xAI in Austin. On April 7, Intel joined as a partner, contributing its 18A process node, a 1.8-nanometer-class technology, and the most advanced semiconductor manufacturing process within the United States.

80% of Terafab’s projected compute output is earmarked for a single chip called D3, a radiation-hardened processor built for orbital AI data centers. SpaceX has already filed an FCC application to launch one million data center satellites into low Earth orbit.

Musk reasons that AI workloads will be cheaper to run in orbit than on the ground within three years. Terafab’s structure reflects this split. One facility is dedicated to edge AI chips for Tesla Robotaxis and Optimus robots. The other focuses entirely on the D3 for orbital deployment.

2) Muse Spark: Meta Closes the Gap and Goes Proprietary

On April 8, Meta debuted Muse Spark, the first model from their Superintelligence Labs. It was built over nine months by a team led by Alexandr Wang, Meta’s Chief AI Officer. The new model is closed-source, signaling a shift in Meta’s frontier AI strategy. This is a notable break from the Llama playbook that helped make Meta the standard-bearer for open-weight AI.

Muse Spark performs competitively in multimodal perception, reasoning, health, and agentic tasks. For healthcare, Meta collaborated with 1,000 physicians to build Muse Spark’s clinical capabilities.

Meta says the model reaches Llama 4 Maverick-equivalent capability at 10x lower compute cost. The model now powers the Meta AI app and website, with rollouts to WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses planned in the coming weeks. Meta claims their models are scaling predictably and that Muse Spark is an early data point on the trajectory, with larger models in development.

3) OpenAI At $852B: Testing a New IPO Playbook

On March 31, OpenAI closed a $122 billion funding round at a post-money valuation of $852 billion, the largest private fundraising event in history. OpenAI was the fastest technology platform to reach 100 million users, and soon the fastest to 1 billion weekly active users (currently around 900M).

CFO Sarah Friar confirmed on April 8 that OpenAI will reserve a portion of IPO shares for retail investors.

“It has to be that everyone partakes, that it isn’t just that a very small group, and everyone else gets left behind.” - Sarah Friar CFO at OpenAI

In its pre-IPO private placement via JPMorgan, Morgan Stanley, and Goldman Sachs, OpenAI targeted $1 billion from individual investors and received $3 billion, described by those banks as the largest private retail placement they have ever executed.

OpenAI is heading toward a possible public filing in the second half of 2026 with a valuation nearing $1T. They project $280 billion in revenue by 2030, against $20+ billion annualized today.

Learn With My Friends and Me…

All In: Anthropic’s $30B Ramp and Mythos Doomsday, OpenClaw Ankled, Iran War Ceasefire, Israel’s Influence

My Groq Bet

Other Reading…

• Project Glasswing (Anthropic)

• Half of Planned US Data Center Builds Have Been Delayed or Canceled… (Tom’s Hardware)

On X…

Caught My Eye…

1) The Trade Map: AI Drove One-Third of Global Growth

McKinsey recently published “Geopolitics and the Geometry of Global Trade: 2026 Update”, covering economies that account for more than 90% of global commerce. The main finding was that AI-related trade, including semiconductors, servers, networking equipment, and data-center hardware, grew by close to 40% in 2025, compared with a 6.5% global average, and accounted for roughly 30% of total trade growth.

The April 2025 Liberation Day tariffs were expected to compress global trade, but the data showed that trade grew faster than the global economy. Instead, it changed how trades were routed more than volume. The US-China bilateral trade fell by roughly 30% in 2025, wiping out about $130 billion of Chinese exports to the US. But the U.S. replaced roughly two-thirds of that gap through imports from other countries. ASEAN absorbed much of the displacement: exports from Vietnam, Thailand, and Malaysia jumped nearly 14%, more than double the global average. India also added $15 billion in smartphone exports to the U.S. alone.

In response, China made some shifts in what it sells. McKinsey found consumer-goods exports declined while intermediate-goods exports rose. This reinforced a broader move from finished goods to components and materials inside global supply chains. This means supply-chain dependence is even harder to unwind than consumer-facing trade.

Europe also had major shifts in trade. Rising competition from China weighed heavily on Europe’s manufacturing base as imports from China increased by over $60 billion. Their auto industry was one of the hardest hit. Germany, the region’s largest carmaker for more than half a century, imported more cars from China than it exported for the first time in history. By the end of 2025, EVs assembled in China were about 15% of all EV sales in the EU.

2) The Massive Deal Wave in Pharma

Late March brought a sharp acceleration in pharma dealmaking. The two largest moves came on March 31: Biogen, a neurology-focused drugmaker, agreed to acquire Apellis, which sells treatments for rare diseases and eye disorders, for $5.6 billion. Eli Lilly, one of the largest diabetes and obesity drug companies, agreed to buy Centessa, a biotech developing a narcolepsy drug, for up to $7.8 billion.

Days earlier, Lilly also signed a collaboration with Insilico Medicine, an AI drug-discovery company, worth up to $2.75 billion. Around the same time, Novartis, a global Swiss pharma group, agreed to acquire Excellergy, a radiopharmaceutical company, for up to $2 billion.

We are seeing a broader pattern in pharma: companies facing future patent pressure are using M&A and licensing to add either near-term revenue assets or promising pipeline programs.

The Inflation Reduction Act, which lets Medicare negotiate prices for some drugs, adds more context. Small-molecule drugs can face Medicare price negotiation sooner than biologics, after 9 years from FDA approval, versus 13 years. That does not erase profits, but it can reduce the later-year upside for some drugs, making longer protected earning windows more valuable at the margin.

3) Whoop’s $575M Raise: The IPO Clock Starts

On March 31, 2026, Whoop, the wearable human performance device company, raised $575 million at a $10.1 billion valuation. This is nearly triple from its 2021 Series F valuation of $3.6 billion. The company said it now has more than 2.5 million members globally, that 2025 bookings grew 103% year over year, and that it exited the year at a $1.1 billion run rate.

“It’s our expectation this is the last round private round of financing that we’ll do,” - Will Ahmed, CEO and Founder of Whoop

The business model is quite distinct in consumer wearables. Whoop gives away a screenless wrist device and monetizes through a $239 annual subscription. This positions the product less as hardware and more as a recurring health platform. The investor list is just as interesting as the valuation. Abbott Laboratories participated in the round, alongside Mayo Clinic, Mubadala, and QIA. Abbott’s presence is notable because their FreeStyle Libre device is the most widely used CGM platform. Ahmed has publicly said noninvasive glucose monitoring is a priority and that current CGM technology has “problems at every layer.”

Whoop’s expansion plans show where it thinks growth will come from. The company says it is hiring for 600-plus roles in research and international expansion. They also see women becoming a much larger part of the member base, with 150% year-over-year growth in new female members. Within the niche, their main competitor is Oura, which also raised a massive $900 million round at an $11 billion valuation in late 2025.

Learn With My Friends and Me…

All In: SpaceX IPO, Iran War Fallout, Quantum Bitcoin Hack, The Space Opportunity

Other Reading…

• Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians (ARXIV)

• From Hierarchy to Intelligence (Jack Dorsey)

On X…

In 1970, a median home cost 2.7x the median household income. By 2024, that figure reached 5.1x. Today, 62% of middle-class wealth sits in an asset that 7 in 10 American households can no longer afford.

Home ownership in the U.S. is now the subject of executive orders, Senate votes, and protest movements. But most of the debate focuses on symptoms rather than the system that created them.

To understand why affordability looks the way it does today, we need to understand how housing became a financial instrument in the first place.

How Homes Became Wealth Storage

For most of human history, a home was just shelter.

The idea that a house should appreciate in value, generate wealth, and serve as the cornerstone of a family’s financial security is a relatively recent concept.

Policy didn’t accidentally inflate housing values. It systematically engineered demand, leverage, and incentives around ownership.

The FHA was created in 1934 to insure mortgages against default. The GI Bill of 1944 extended mortgage guarantees to returning veterans with reduced down payments.

Between 1944 and 1966, one-fifth of all single-family homes built in America were financed with GI Bill assistance, and homeownership rates jumped from 44% to 62%.

Real estate is the only asset class where Americans can routinely borrow 80% to 95% of the purchase price at favorable rates.

A 20% down payment on a $400,000 home gives you control of the full asset with $80,000. If it appreciates 5%, your return on equity is 25%, five times the market gain.

But leverage cuts both ways.

For a buyer who just put 20% down, a 20% price decline wipes out their entire equity position.

When 62% of household wealth depends on a single asset holding its value, every participant in the system, homeowners, local governments, banks, federal policy, is aligned around one outcome: prices must not fall.

That worked for decades. Then the system got tested.

Why The Market Is Freezing

For a decade, near-zero interest rates kept that system running, fueling demand and pushing prices higher. Then in 2022, with inflation running above 9%, the Fed raised rates at the fastest pace in 40 years.

In a normal market, that cools demand, brings prices down, and restores equilibrium. It did cool demand. Active buyers fell from 2.5 million to 1.4 million. But it also killed supply.

Homeowners locked in at 3% faced a $12,000-per-year penalty for selling and buying the same home at 7%. So they stayed put. Demand and supply collapsed at the same time, and prices never came down.

What Fixes Housing Affordability?

The evidence points to supply.

Austin permitted over 20 units per 1,000 residents at its peak, while San Francisco permitted fewer than 4. Austin’s median home costs about 5x median income. San Francisco exceeds 11x.

The pattern holds across different markets. Places that increase supply through permits and building have more stable prices.

But here’s the trade-off: affordability and wealth-building are in tension. Lower prices help buyers and renters. They hurt the 65% of Americans who already own. And homeowners vote at higher rates, donate more to local campaigns, and show up to zoning meetings.

Every policy choice picks a side. Zoning reform benefits future residents at the expense of current ones. Rent control helps today’s tenants while discouraging tomorrow’s supply.

And the questions only get more difficult from here.

• If institutional investors own just 0.5% of single-family homes, why have they become the political villain?

• If Houston can permit 6.5 homes per 1,000 residents while Los Angeles permits fewer than 1, is affordability a market failure or a policy choice?

• How are phantom mortgages pricing out a generation of buyers before they even apply to purchase a home?

• And if only 25% of millennials own a home by 30, compared to 43% of boomers at the same age, is America building a permanent renter class?

Housing affordability sits at the intersection of finance, construction economics, tax policy, demographics, and local politics. Few understand the underlying system that connects each piece.

Our research team at Social Capital spent some time studying housing affordability from first principles and developed a 110-page Deep Dive to bring you up to speed on the system that shapes the most important asset most families will ever own.

Let me know your thoughts in the group after reading.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Defense Startup Shield AI Hits $12.7B Valuation After US Air Force Deal

Shield AI announced a $2 billion financing package on March 26, including a $1.5 billion Series G, at a $12.7 billion post-money valuation. This is a 140% jump since their $5.3 billion valuation in March 2025.

Shield AI is building autonomy software that lets drones and aircraft keep operating when GPS or communications are jammed or unavailable. Its flagship software, Hivemind, has already been tested on platforms including the F-16 and was recently selected by the U.S. Air Force as a mission autonomy provider for its Collaborative Combat Aircraft program.

The company plans to use a part of the proceeds to acquire simulation software firm Aechelon Technology. Aechelon serves as the layer that works with the Department of War’s JSE (Joint Simulation Environment), where these systems can be trained and stress-tested in simulations before live deployment.

2) Google’s TurboQuant: The Quantization Breakthrough

Google Research introduced TurboQuant this week as a new way to compress one of the most expensive parts of large-model inference, the key-value (KV) cache. In Google’s tests, the method pushed compression as low as 3 bits per value with no accuracy loss, while cutting KV memory use by at least 6x and delivering up to 8x faster attention on Nvidia H100 GPUs.

What makes quantization so powerful is that a well-quantized model can lose far less quality than its reduction in bit depth would suggest.

As models handle longer prompts and more concurrent users, the bottleneck is often not raw compute, but rather memory traffic. The KV cache keeps growing with context length, which makes inference slower and more expensive.

TurboQuant matters because it attacks that bottleneck directly, online, without retraining or calibration. Benchmarks on open-source models like Gemma, Mistral, and Llama-3.1 variants showed dramatic efficiency gains while preserving full-precision performance. It is showing up first around open-source models because that is where outside researchers can benchmark and integrate new inference methods in public.

Better compression means more throughput from the same hardware, lower serving costs, and a better chance of running capable models on smaller clusters or edge devices. Semantic search is also faster and cheaper, which matters as search, recommendations, and AI assistants increasingly depend on understanding meaning, not just keywords. Despite the paper’s launch last year and their recent viral announcement, it is still early, and Google says the peer-reviewed work will be presented at ICLR 2026.

3) Meta’s TRIBE v2: Brain-Predictive AI

On March 26, 2026, Meta AI released TRIBE v2 (Trimodal Brain Encoder), a foundation model trained to predict human brain activity from video, audio, and language stimuli, built on over 1,000 hours of fMRI data from 720 subjects.

Instead of putting people in a scanner for every new experiment, researchers could first test ideas in software and narrow down which stimuli or hypotheses are worth studying in the lab.

That matters because brain research is slow, expensive, and noisy. fMRI does not directly read neurons; it measures changes in blood flow, and real scans are often degraded by motion and other physiological noise. A model that can generate cleaner predicted responses could help researchers design better experiments, compare how healthy and diseased brains differ, and eventually improve diagnosis or treatment planning for brain disorders. Meta’s own demo states that the goal is to provide a foundation for improving the diagnosis and treatment of brain disorders.

Meta released the model weights, codebase, paper, and an interactive demo under a CC BY-NC license. Meta emphasizes scientific applications, such as simulating experiments or aiding disease diagnosis. Without any retraining, TRIBE v2 can reliably predict the brain responses of individuals it has never seen before, achieving a nearly 2-3x improvement over previous methods:

Neuroscience has usually relied on narrow models built for one task, one modality, or one brain region at a time. TRIBE v2 gives a more general predictive layer across sight, sound, and language, then maps that shared representation onto individual brains.

Learn With My Friends and Me…

• All-In: Anthropic’s Generational Run, OpenAI Panics, AI Moats, Meta Loses Major Lawsuits

• My new video: Once I Understood This About Investing, My Life Changed.

Other Reading…

• Is Big Tech Facing a Big Tobacco Moment? (NYT)

• How I created Rosie’s mRNA Vaccine Protocol (Paul S. Conyngham)

• Innovation At The Speed Of Markets: How Regulators Keep Pace With Technology (FDIC)

• Artificial General Intelligence Forecasting and Scenario Analysis (RAND Corporation)

On X…

The breach cost over $2.45B and exposed medical information of 192.7M Americans, with 94% of hospitals reporting financial damage.

BlackCat (ALPHV) ransomware group accessed the system through a basic vulnerability:

The portal lacked multi-factor authentication.

While a criminal organization reached the records of 6 in 10 Americans through a single external access portal, the healthcare system itself still cannot move those records between two hospitals.

Up to 50% of patient records fail to match correctly when exchanged between systems. When a doctor treats a patient, information stored in another hospital’s system remains invisible at the point of care.

To understand how the system works and why it behaves this way, our Social Capital research team did a deep dive and examined the sector from first principles.

What emerged was a system held in place by three structural layers, each reinforcing the others, starting with a decision that has constrained progress in healthcare information for 60 years.

The 1966 Decision

In 1966, Massachusetts General Hospital needed a computer system that could serve 20 departments simultaneously without ever crashing during patient care. The available hardware was a PDP-7 with 4KB of RAM.

Three developers built a system called MUMPS that fused an operating system, a programming language, and a database into a single inseparable layer.

That fusion solved a critical problem: dozens of departments writing to the same patient record at the same time without conflicts or crashes.

It also created the constraint that still shapes most Electronic Health Record (EHR) systems today: because the layers were fused together, migrating away means replacing the entire system at once.

MUMPS was optimized to answer one question: what is happening with this patient right now?

Healthcare spending has grown from 6.9% of U.S. GDP in 1970 to 17.6% by 2024, and the questions EHR systems need to answer today are fundamentally different.

• Payers want to know which treatments work and which providers are expensive.

• Regulators want to measure quality metrics like infection and readmission rates.

• Researchers want to identify at-risk patients across entire populations.

These questions require data to move across patients, providers, and systems.

The existing architecture was not designed to move data cleanly across systems, so translation work shifted to humans and intermediary processes.

The $1 Trillion Administrative Layer

The U.S. spends roughly $1 trillion each year on healthcare administration, representing 25 to 30% of total healthcare spending and about 3.5x as much per capita as Canada.

That spending funds a massive translation workforce, including roughly 185,000 medical coders who convert clinical language into billing codes across more than 70,000 diagnosis codes and 10,000 procedure codes.

The administrative layer exists because each insurer maintains its own rules, contracts, and approval processes, requiring this translation to happen repeatedly across fragmented payers and providers.

In an economy where automation and AI are increasing efficiency across nearly every sector, why has healthcare resisted?

40 Years of Failed Interoperability

Getting different EHR systems to share patient data requires four things to happen:

1. Systems must be able to connect.

2. Data must be formatted consistently.

3. Systems must interpret what the data means.

4. Governance and consent must permit exchange.

Over the past four decades, there have been 6 major attempts to solve these issues.

Each one made progress, but none solved all four.

Consolidation Emerges

As a means to address both interoperability and complexity, market participants began to acquire more of the system.

• Epic consolidated horizontally, capturing 42% of the hospital EHR market overall and approaching 70% among hospitals with 500 or more beds.

• UnitedHealth consolidated vertically, capturing multiple layers of the healthcare system. It insures 52M lives through UnitedHealthcare, employs over 90,000 physicians through Optum Health, manages pharmacy benefits through OptumRx, and processes 40% of all U.S. claims through Change Healthcare.

This shifted coordination from across organizations to within them, but the core structural constraints remain.

Can LLMs Fix U.S. Healthcare?

The constraints described above emerge from three layers, each reinforcing how data moves, how it is used, and who captures value.

Across all three layers, a common limitation appears: the system depends on data that is fragmented, inconsistently structured, and difficult to interpret at scale.

Even today, most AI systems in healthcare focus on imaging rather than processing unstructured clinical language.

However, recent LLMs introduce a new capability: they can interpret clinical language at scale in ways previous systems could not, without requiring a rebuild of the underlying architecture.

Two parallel efforts are now emerging. AI is being integrated into healthcare institutions, while new tools allow patients to access and use their own data directly.

In January 2026, OpenAI and Anthropic launched healthcare products within days of each other at the J.P. Morgan Healthcare Conference, signaling how quickly both sides of this dynamic are developing.

Together, these forces apply pressure on the system from opposite directions.

Learn With Me

Previous reforms either required the underlying architecture to be rebuilt, vendors to cooperate, or institutions to share data on new terms. LLMs create a new possibility for change by enabling the interpretation of unstructured clinical language on top of the existing infrastructure. This is a major distinction, and it makes this sector worth studying now.

I had our research team at Social Capital explore the current constraints in Healthcare IT and innovation potential. In the Deep Dive, you’ll learn:

• How did a programming language built on 4KB of RAM become the foundation for systems managing 300M+ patient records, and why has no one been able to replace it?

• What does Epic’s proprietary data dictionary control, and why can a hospital export its own data but not read it?

• How did the Change Healthcare breach expose 6 in 10 Americans, and what does its structure reveal about the concentration risks in U.S. healthcare infrastructure?

• Why do 81.7% of prior authorization denials get overturned on appeal, and what does that reversal rate suggest about the system’s actual function?

• Which countries achieved full health data interoperability before the U.S. even digitized, and why can’t the U.S. replicate their approach?

• How are LLMs already being deployed in clinical trial matching, ambient documentation, and automated coding, and where are the limits?

This is the first part of a two-part series. Part Two will cover the race between platform companies and incumbents, the regulatory frameworks taking shape around clinical AI, the capital flowing into AI-native healthcare infrastructure, and whether the system absorbs this capability the way it absorbed every previous reform.

If you want to learn more with me, sign up below to read the full Deep Dive (and all our past releases).

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) NVIDIA and the $1T AI Buildout for the Agent Era

On March 16, 2026, at NVIDIA’s GTC keynote, Jensen Huang said, “I believe that computing demand has increased by 1 million times in the last 2 years… the inference inflection has arrived.” He now sees at least $1 trillion in revenue from Blackwell and Rubin between 2025 and 2027, up from the $500B forecast last year.

NVIDIA’s Vera Rubin platform integrates seven specialized chips into a unified rack, delivering 10x the inference efficiency and 4x the training performance of the Blackwell generation. This architecture features the purpose-built Vera CPU, which accelerates agentic AI and reinforcement learning workflows with twice the efficiency of traditional rack-scale processors.

NVIDIA also used GTC to show how its stack is shifting from model training toward low-latency inference and autonomous software systems. The new Groq 3 LPX rack, built around 256 LPUs, is designed to sit alongside Vera Rubin and handle the latency-sensitive parts of inference. The combined architecture can deliver up to 35x higher inference throughput per megawatt and up to 10x more revenue opportunity for trillion-parameter models.

On the software side, the company introduced NemoClaw, a new stack for the OpenClaw ecosystem. This will let users install Nemotron models and OpenShell in a single command. Privacy and security within an isolated sandbox will enable self-evolving, autonomous AI agents. Jensen described a future where NVIDIA would grow to 75,000 employees working alongside 7.5 million AI agents.

2) The Data Center Boom Hits a Power Bottleneck

At the end of 2025, the US disclosed data center pipeline reached 241 GW, an increase of 159% from the beginning of the year. But only a third of projects in the data center pipeline are under active development. Wood Mackenzie’s Q4 2025 data shows developers added 25 GW of new capacity last quarter, half the pace of Q3.

“Utilities just don’t necessarily have either the grid capacity or the generating capacity to be able to build it fast enough to accommodate these new large energy demand centers… It’s a bend in the trajectory that we’re now seeing companies realizing that they need to focus on projects at hand, rather than just endlessly adding new ones.” - Ben Hertz-Shargel, Global Head of Grid Edge for Wood Mackenzie

Overall, per-megawatt costs declined, but rack power densification has resulted in cost per square foot rising. This shows the industry’s shift toward more energy-dense facilities for efficiency and inference latency. Texas continues to maintain its lead in total pipeline capacity, while New Mexico, Indiana, and Wyoming saw the largest proportional growth due to their access to natural gas.

Reuters reports that hyperscalers are expected to spend roughly $650 billion in 2026, up over $200 billion since 2025. So in nominal terms, CapEx is hitting record highs, but it is becoming harder to expand at the same pace. The constraint has shifted from capital availability to the ability to secure energy, navigate interconnection rules, and deliver projects on time.

“Hyperscalers are increasingly partnering with developers and signing multi-gigawatt portfolios rather than single-asset PPAs” - Caroline Mead, SVP of Power Marketing at developer ⁠Engie North America

3) Crypto Gets Defined, and Equities Get Tokenized

On March 17, the U.S. Securities and Exchange Commission (SEC), issued an interpretative release clarifying how federal securities laws apply to crypto assets, with the help of the CFTC. The courts can deviate from it if they don’t like its rationale, but the SEC will follow up shortly with a proposed rule. The guidance introduces a formal taxonomy that separates crypto into five categories:

1. Digital commodities: Not Securities

2. Digital collectibles: Not Securities

3. Digital tools: Not Securities

4. Stablecoins: Not Securities

5. Digital securities: Securities

The release also outlines how assets can transition into or out of “investment contract” status and provides specific guidance on activities such as airdrops, staking, mining, and asset wrapping.

On March 18, the SEC approved Nasdaq’s proposal to support tokenized versions of traditional securities through a pilot program with the Depository Trust Company (DTC). Initial eligibility includes Russell 1000 stocks and major index ETFs. The structure preserves the existing market with tokenized shares being fungible with their traditional counterparts. They carry the same ticker and CUSIP and provide identical rights, including dividends and voting.

Learn With My Friends and Me…

• All-In: Jensen Huang: Nvidia’s Future, Physical AI, Agents, Inference, AI PR Crisis

Other Reading…

• National Policy Framework for Artificial Intelligence (White House)

• Beijing’s Big Problem: An Incredible Shrinking Economy (WSJ)

• Save us, Digital Cronkite! (Noah Smith)

On X…

Today, prediction markets have over $6 billion in weekly trading volume: more than 100x in just over 24 months.

The 2024 presidential election marked the move of prediction markets into the mainstream. The two largest prediction market platforms, Polymarket and Kalshi, saw over $4.2B in trading volume on the election event alone.

While traditional polling remained largely static within a margin of error, prediction markets reacted to new information in real-time, moving ahead of polls, calling swing states days early, and correctly pricing the eventual outcome.

This was more than just people placing bets. This was a demonstration of markets acting as truth machines.

How Markets Produce Signals

A prediction market begins with a simple, binary contract.

When an exchange lists a contract, it’s tied to a specific, verifiable event. Traders can buy either outcome, YES or NO, based on their belief in the likelihood of the event happening or not.

If the trader bought “YES” and the event occurs, the contract is worth $1. If the event does not occur, it’s worth $0.

When a contract trades at $0.65, the market is collectively pricing a 65% chance of the event occurring.

In other words, the price of a bet reflects the collective probability estimate of all participants, allowing the market to know more than any single individual. By aggregating dispersed information, these markets produce a measurable signal.

From a regulatory perspective, however, the legality of these markets was unclear until the Kalshi v. CFTC (Commodity Futures Trading Commission) case in 2024.

The eventual result of that case allowed market participation and capital to begin scaling.

The Liquidity Network Effect

Liquidity is essential for a market to function. The market needs an adequate number of willing buyers and sellers for a given contract at various prices.

Without sufficient liquidity, even small trades can cause large price movements, which result in poor price discovery and wide bid/ask spreads. In other words, there can be a large gap between the price people are willing to pay and the price others are willing to sell at.

Conversely, when more traders enter the market, liquidity improves, price signals sharpen, and more traders follow. This is the liquidity network effect in action. As trading volume increases, the market becomes more efficient.

While the CFTC ruling removed constraints on institutional liquidity from entering centralized exchanges like Kalshi, decentralized markets were also growing in parallel.

Polymarket, which is decentralized, grew its liquidity through structural advantages: near-instant settlement, global access without traditional banking relationships, and the ability to spin up new event contracts quickly without per-contract regulatory approval.

Now, both Kalshi and Polymarket are formal participants in the U.S. market, following Polymarket’s acquisition of QCEX, a CFTC-licensed exchange, in July 2025.

Despite different architectures and settlement structures, both exchanges have continued to grow, driven by regulatory clarity that legitimized the category.

This also helped to unlock broader distribution.

Unlocking Consumer Distribution

Prediction markets are now accessible through platforms like Robinhood, Coinbase, DraftKings, and FanDuel.

This unlocks 100M+ retail accounts that are already linked to bank accounts, giving distribution of prediction markets through platforms people are already using without needing to acquire users from scratch.

With new technology enabling easier access, regulatory clarity, and increased trading volume that improves price signals, prediction markets have exploded and are expected to continue growing.

Research featured by CNBC projects prediction markets could approach $1 trillion in annual trading volume by 2030, or roughly $19.2 billion per week.

But with that growth comes a new set of questions.

In a recent All-In interview with SEC Chairman Paul S. Atkins and CFTC Chairman Michael S. Selig, we explored key issues: exposure to manipulation, the path to regulatory clarity, the industry’s long-term outlook, and the role AI will play in shaping these markets.

To go further, our research team at Social Capital developed a 60+-page deep dive into prediction markets.

When you read it, you’ll learn:

• The architecture of prediction markets, their key differences, advantages, and disadvantages.

• The pivotal moment for prediction markets that unlocked a wave of institutional investment.

• Key markets in the space and how the competitive landscape is shifting.

• How different players in the ecosystem are capturing value.

• How prediction markets can be used in the private sector to surface latent intelligence.

• The endgame for prediction markets in the U.S.

• Applicable insights on how to interpret and use this new information signal.

If you want to learn with me, sign up below to read the full Deep Dive (and all our past releases).

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Yann LeCun’s New AI Lab, AMI, Raises $1.03B for World Models

Advanced Machine Intelligence (AMI) launched March 10, 2026, with a record $1.03 billion seed round at a $3.5 billion pre-money valuation, making it Europe’s largest seed round ever. Founded by Yann LeCun (post-Meta) as chairman, the team also includes Alexandre LeBrun (ex-Nabla CEO) as CEO and Saining Xie (ex-DeepMind and Meta) as Chief Science Officer. The lab drew co-leads from Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions.

AMI’s bet is that the next important AI systems will not come from scaling next-token prediction alone. Instead, it pursues world models built on LeCun’s Joint Embedding Predictive Architecture (JEPA). This enables systems to learn from sensory/physical data, maintain persistent memory, reason causally, plan under constraints, and operate safely in the real world.

AMI targets robotics, manufacturing, transportation, and healthcare. Nabla (LeBrun’s prior company and LeCun’s investment) is the first partner to gain early access to agentic clinical AI.

This vision aligns with LeCun’s February 2026 arXiv paper “AI Must Embrace Specialization via Superhuman Adaptable Intelligence“, co-authored with Goldfeder, Wyder, and Shwartz-Ziv. The paper argues AGI concepts are misguided; human intelligence is specialized, not general.

It proposes Superhuman Adaptable Intelligence (SAI): rapid adaptation to surpass humans on any key task, plus tackling human-infeasible domains. SAI favors SSL and world models for grounded reasoning/planning, positioning AMI as a post-LLM path where breakthroughs come from architectural innovation rather than just more compute/data.

2) Netflix Makes Major Bet on AI Filmmaking Tools

Netflix announced on March 5 that it acquired InterPositive, the AI filmmaking technology company founded by Ben Affleck in 2022. The entire 16-person team (engineers, researchers, creatives) is joining Netflix, and Affleck is staying on as a senior advisor.

Financial terms were not disclosed at the announcement, but Bloomberg later reported the deal could be worth up to $600 million, making it one of Netflix’s largest acquisitions. The company had been operating in stealth, building tools specifically “by filmmakers, for filmmakers.” Netflix already piloted the technology on its own production “El Eternauta,” reporting 10x efficiency gains over traditional VFX pipelines.

InterPositive’s core innovation is per-film custom AI models: it ingests raw production dailies, trains a bespoke model on that specific project’s footage and style, then hands filmmakers an autonomous post-production engine.

The tool can relight entire shots, remove stunt wires frame-perfectly, reframe missed angles, enhance or generate backgrounds, insert VFX elements, and fix continuity errors, which can save studios months and millions of dollars. Because the model is trained on the actual film’s data, outputs stay stylistically consistent and director-controlled rather than generic.

The acquisition gives Netflix an in-house post-production AI capability built around project-specific models and filmmaker-controlled workflows. In its announcement, Netflix framed the deal as “innovation for filmmaking, by filmmakers”, keeping human creatives at the center.

At the same time, Affleck described the tools as purpose-built for production environments rather than general-purpose generative media. That makes the transaction less about consumer-facing AI and more about integrating specialized production software into the studio pipeline

3) Mastercard Brings Circle and 85+ Firms Into Crypto Payments Push

Mastercard launched its Crypto Partner Program on March 11, bringing together more than 85 crypto-native firms, payments providers, and financial institutions in a formal collaboration group.

Participants include Circle, Binance, PayPal, Ripple, Gemini, Paxos, SoFi, Worldpay, Polygon, Solana, Fireblocks, and BitGo. Mastercard said the program is designed as a forum for product input, standards, and collaboration. The program helps shift digital assets from experimental side rails into production-grade integration with Mastercard’s 150+ million global merchant network and core card infrastructure.

The company’s stated focus is on practical payment use cases, especially cross-border remittances, B2B transfers, payouts, settlement, and other forms of money movement, where digital assets can add speed and programmability.

Mastercard’s explicit goal is to fuse the “speed and programmability of digital assets” with established card rails so payments feel seamless to end users. Circle’s inclusion is front and center with their Chief Commercial Officer, Kash Razzaghi, emphasizing in the accompanying Mastercard Newsroom interview that stablecoins (starting with USDC) are moving from trading/investing into payments and store-of-value layers, especially in high-inflation markets.

Circle’s role helps explain the stablecoin angle. The partnership explicitly aims to make stablecoin usage feel like typing an HTTP request, where users just send dollars, unaware of the blockchain layer beneath.

This marks an institutional tipping point at which crypto and traditional finance converge to upgrade the underlying financial infrastructure. After years of parallel tracks, regulation has cleared the runway, and giants like Mastercard are now betting that collaborative infrastructure will drive the next leg of growth through cheaper, 24/7, borderless transfers that approach zero marginal cost.

Learn With My Friends and Me…

• All-In: Iran War, Oil Shock, Off Ramps, AI’s Revenue Explosion and PR Nightmare

Other Reading…

• Fed to Loosen Capital Requirements for Big US Banks (Financial Times)

• US Supreme Court Declines to Hear Dispute Over Copyrights for AI-Generated Material (Reuters)

• When Your Life’s Work Becomes Free and Abundant (Aditya Agarwal)

• Challenges and Research Directions for Large Language Model Inference Hardware (ARXIV)

• Autoresearch (Andrej Karpathy)

On X…

At the time, U.S. Census data showed deaths exceeding births in more than 73% of U.S. counties. Even after stripping out pandemic-related deaths, population growth had still neared all-time lows.

Since then, the U.S. total fertility rate has fallen another 4% to a record low of 1.6, extending a long-term decline of 57% since 1960.

And the U.S. is not an outlier.

110 out of 204 countries now sit below the total fertility replacement level of 2.1.

By the year 2100, it’s expected that over 90% of countries will fall below this level.

When Elon posted about this last year, it got 44 million views:

Historically, population decline has been driven by external and involuntary shocks like famine, disease, or war. Notable examples include:

• The Black Death in the 14th century

• The collapse of the Roman Empire

• China’s Great Famine

Unlike past population declines, today’s is voluntary and structural.

To understand why, my research team at Social Capital examined the issue from first principles.

What emerged in our research was five structural forces shaping fertility decisions across modern societies, starting with the most fundamental constraint:

1. Biology: The Limiting Factor

Human fertility is age-dependent.

The average woman in developed countries now has her first child around age 29. But by this point, fertility has already begun to decline. At that age, if 100 healthy couples try to conceive in a given month, only about 20–25 will succeed.

Meanwhile, population-level testosterone in men is declining approximately 1% every year. Low testosterone levels are associated with impaired sperm generation. And this decline persists even after controlling for age, health, and lifestyle.

That means a 30-year-old man today has roughly 20% lower testosterone than a 30-year-old from his father’s generation.

Both sides of the reproductive equation are under pressure.

While biology defines the limits of human reproduction, the next force changes how people interact with those limits:

2. Technology: Reshaping Human Reproduction

Modern reproductive technologies have fundamentally decoupled sex, relationships, and childbirth.

• Technologies of decoupling, such as prevalent contraceptives and legal abortion, separated the act of sex from its potential reproductive consequences.

• Technologies of extension, such as IVF, egg freezing, and fertility monitoring, enabled people to extend reproduction beyond the natural window of peak fertility.

• Technologies of substitution, such as dating apps, digital content, and algorithmic social platforms, have quietly reshaped how people form (or fail to form) relationships at all.

The net effect is that modern technology is exerting a downward pressure on fertility outcomes, largely because avoiding or delaying pregnancy is simpler and more economically accessible than enabling it after fertility has declined.

Which brings us to the next force:

3. Economics: Childbearing Costs More Today

In pre-industrial agrarian economies, children worked and contributed economically to the household. In modern economies, children are primarily a financial cost rather than an economic asset.

In the United States, it’s estimated that it costs roughly $310,000 for a middle-income household to raise a child to age 17, excluding college.

At the same time, key family expenses such as housing, childcare, and education have grown faster than household income, increasing the financial trade-offs associated with parenthood.

If financial constraints were the dominant driver, one would assume wealthier countries would have the highest birth rates. In reality, the opposite pattern appears.

Many of the world’s richest countries have fertility rates below 1.5, while several of the poorest maintain rates above four children per woman.

To understand that dimension, we need to examine the role of cultural norms and social incentives that shape reproductive decisions.

4. Culture: A Mass Shift In Desire

French philosopher René Girard argued that humans desire what others around them desire.

According to this theory, if people prioritize career achievement and personal independence over raising a family, those desires propagate socially.

The inverse is also true: if more people around you have children, the social pull toward parenthood increases.

There’s data to support this. According to research published by the American Sociological Review, a woman’s probability of having a child increases significantly in the 12 to 24 months after a close friend gives birth, peaking at around 2 years.

In short, fertility decisions, in either direction, are socially contagious.

But culture alone does not determine fertility outcomes; there’s another factor we should consider:

5. Policy: How Government Intervenes

When governments respond to population decline or growth, policy typically falls into two categories: soft policies and hard policies.

• Soft policies act indirectly, shaping fertility through economic security, cultural norms, and incentives around family formation.

• Hard policies rely on legal mandates and enforcement to directly alter reproductive behavior.

Whether through incentives or coercion, institutional interventions often generate backlash and produce questionable long-term influence on fertility.

The 5 Forces Converge

With these forces applying pressure across multiple dimensions of human reproduction, the demographic time bomb I referred to in 2021 has continued ticking.

The data is striking, but the cause and effect behind each of these forces runs much deeper. That is why my research team put together a 100-page Deep Dive to investigate further. Here are the questions we set out to answer:

• What does male fertility data reveal about our environment that almost no one is talking about?

• How did three distinct technologies each reshape the relationship between sex, relationships, and reproduction, and how did their effects compound?

• Which policies have meaningfully moved the needle on fertility historically, and what does China’s failed reversal tell us about their hard limits?

• How does the decision to have children spread through social networks, and can a negative feedback loop ever be reversed?

• How do some communities sustain high birth rates inside low-fertility economies?

• What does a 1968 mouse experiment with unlimited food, water, and space reveal about why wealthy societies stop reproducing?

If you want to learn more with me, sign up below to read the full Deep Dive (and all our past releases).

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Global equity markets exceed $150 trillion in value, yet trading hours remain limited, settlement still depends on layers of intermediaries, and access to many of the highest-growth companies remains restricted to a small group of investors.

These structural limits define how capital moves, who can participate, and how quickly ownership can change hands.

Market infrastructure providers are already exploring how tokenization could modernize this system. Institutions, including NYSE, Nasdaq, and DTCC, have begun developing tokenized equity and settlement infrastructure.

With the adoption of equity tokenization, these barriers are beginning to lift.

Since the start of 2025, the market capitalization of equity tokens has risen nearly 3.5x, reflecting a broader shift toward tokenized real-world assets.

This expansion has occurred alongside the rise of stablecoins. These currency-pegged tokens have grown more than tenfold in under five years and now form the primary settlement layer for on-chain financial activity:

While stablecoins serve a different function than equity tokens, their rapid adoption demonstrates that tokenized financial instruments can achieve meaningful scale when they deliver clear infrastructure advantages.

Equity tokens represent the next test: whether tokenization can extend beyond payments into ownership of financial assets.

What Is an Equity Token?

Equity tokens are not simply traditional stocks placed on a blockchain.

A traditional stock represents ownership in a company.

An equity token is a blockchain-based asset that represents either shares of a company or a structured claim linked to those shares, with ownership tracked and transferred using distributed ledger technology (DLT).

Tokenized Equities Can Solve 3 Market Gaps

1. 24/7 trading: Markets are moving from a 24/5 (or less) schedule toward continuous, 24/7 availability.

Even today, about 11% of U.S. equity trading occurs outside regular market hours.

A 24/7 market structure more rapidly prices in new information during after-market hours, and better reflects a global shareholder base, with foreign investors now owning roughly 15% of U.S. equities.

2. Ownership: In traditional finance, equity ownership is recorded across multiple intermediaries, including brokers, clearing firms, and central securities depositories.

Tokenization reduces reliance on these layers and allows ownership to be tracked directly on a shared ledger.

This changes equity from a static record into a programmable financial asset.

An owner can pledge the asset as collateral for an on-chain loan. They can use it to secure credit. They can place it into automated liquidity pools to generate yield.

In traditional markets, similar actions typically require multiple intermediaries and additional settlement steps. Each of these intermediary interactions comes with brokerage fees and commissions, ultimately costing the equity asset holder. Even modest reductions in post-trade friction could save the equities industry an estimated $5 to $10 billion annually.

3. Access: While the first two advantages apply primarily to public equities, tokenization also addresses the access constraint in private markets.

Under current securities rules, many private offerings are restricted to accredited investors. This generally requires either $1M in net worth excluding a primary residence, or $200,000 in annual income, or $300,000 jointly with a spouse.

Private companies must also manage shareholder counts to remain private. U.S. regulations require SEC reporting once a firm exceeds 2,000 shareholders of record or 500 non-accredited investors.

In addition, institutional venture capital funds often require commitments of millions from limited partners.

As a result, most investors have had little access to the highest-growth private companies before they reach public markets.

Equity tokenization has the potential to address this access gap.

Equity tokens can be issued through several structural models, though today the most common approach relies on special purpose vehicles (SPVs).

In this structure, an SPV holds the underlying shares while tokens represent a financial claim on that vehicle. This allows issuers to offer investors exposure to private companies that were historically limited to venture capital firms and institutional investors.

As one example, Robinhood recently announced a promotional distribution of OpenAI and SpaceX tokens to eligible users in the European Union.

These tokens provide exposure to two of the world’s most sought-after private companies. However, they do not represent direct ownership of OpenAI or SpaceX shares. Instead, the tokens represent a financial claim linked to an intermediary structure.

This highlights a central challenge in equity tokenization: what a token represents is not always standardized.

Different issuers can design tokens with materially different economic rights. In Robinhood’s case, it is not publicly clear whether the SpaceX token provides exposure to preferred equity or whether it could convert into common shares if SpaceX eventually goes public.

Preferred and common equity differ in liquidation priority, voting rights, and return profile. Without clarity on these terms, investors cannot easily price or compare tokens tied to the same company.

As a result, many tokenized private equity products provide economic exposure rather than direct ownership. Because the token sits on a separate legal layer from the underlying shares, investors must understand the structure before assuming what they actually own.

Despite these structural ambiguities, investor demand for private market access continues to grow. This comes within a broader landscape in which companies are staying private much longer.

Surveys show roughly 90% of Americans would be willing to allocate a portion of their retirement savings to private assets, with particularly strong interest among Gen Z and Millennial investors.

Equity tokenization promises greater access to private markets, continuous liquidity, and new ways to structure financial ownership.

But with this new level of access comes a new kind of complexity.

When two tokens can reference the same company in fundamentally different ways, and economic exposure does not always mean ownership, how do investors separate real opportunity from structural risk? And what about regulation?

Our research team at Social Capital examined these questions from first principles in our Equity Tokenization Deep Dive. Inside, we cover:

• The core benefits equity tokenization introduces to capital markets

• The tokenization trilemma and the three structural designs attempting to solve it

• The market map: major issuers, what they offer, and how their structures differ

• What self-custody means in practice and the trade-offs of moving tokens off a broker platform

• How private company tokens are structured and what investors are actually buying

• The regulatory landscape across the U.S., EU, and other major jurisdictions

• Whether a single token design can satisfy regulators, protect ownership rights, and integrate with the broader financial system

I hope you enjoy reading this Deep Dive we put together below.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

While AI is compounding its ability at a historic rate, an aging population and declining fertility rates are already causing labor shortages.

These trends, combined with declining costs of robotics hardware, underpin a compelling case for humanoid robots and physical AI.

According to Morgan Stanley, the humanoid robot market is set to exceed $5 trillion by 2050.

Even in 2025, the larger robotics space saw $21 billion of VC capital invested.

And with a steady increase in patent activity mentioning “humanoid” over the past few years, these machines are already walking onto factory floors.

In November 2025, the Figure 02 humanoid contributed to the production of 30,000+ BMW X3 vehicles.

Viral videos of humanoids like Tesla’s Optimus running across factory floors and Boston Dynamics’ Atlas doing roundoff back handsprings have sparked interest in the sector.

Despite early excitement, the question remains:

When is it reasonable to expect humanoid robots to experience a ‘ChatGPT moment’ of mass adoption?

To answer, let’s explore how humanoid robots work and their potential use cases.

How humanoid robots work

A humanoid robot is an artificial helper designed to resemble and replicate human behavior.

They typically feature bipedal locomotion and embedded AI. They can interact socially and perform tasks without major safety risks.

Four interconnected systems give humanoid robots human-like perception, intelligence, and movement:

• Sensors (The Sensory System): Cameras, lidar, microphones, and tactile sensors provide the robot with continuous readings of its environment, just as your sensory system does.

• Control Unit & Software (The Brain): The intelligence layer that processes sensor data to execute complex functions.

• Power System (The Metabolic System): The battery that supplies and manages energy to every component, continuously.

• Actuators (The Muscular System): Mechanical devices that convert energy into rotary or linear physical motion (the joints, the grip, the movement)

This is a flow of how they usually work:

Why humanoids instead of single-purpose machines

For most of human history, productive output was a function of human muscle. Agriculture, manufacturing, logistics, and construction were all built around the physical limits of the human body.

Because humans did the work, the built world standardized around human form: doorways, staircases, countertops, and tools are all designed for two arms, two legs, and hands that grip.

Redesigning every factory, warehouse, and home around task-specific machines would be unfeasible.

A humanoid robot that fits into existing infrastructure doesn’t need the world to change around it.

Near-term use cases focus on structured, predictable settings, enabling a robot to learn quickly, make mistakes cheaply, and improve rapidly.

Ranked by timeline and complexity, here’s how the use cases will likely break down:

Humanoid Robots will have the highest impact in these 7 areas:

1. Domestic Assistance: Supporting mobility needs, handling household chores, and providing medication reminders.

2. Manufacturing: Assisting assembly tasks, moving tools and parts, inspecting finished products.

3. Security & Monitoring: Patrolling facilities, investigating alerts, and assisting in emergencies.

4. Customer Service & Reception: Greeting and directing visitors, answering questions, and managing check-ins or bookings.

5. Facility Maintenance: Conducting routine inspections, performing minor repairs, cleaning, and sanitizing spaces.

6. Healthcare: Assisting nurses, delivering supplies or meals, monitoring patients.

7. Warehouse and Logistics: Picking and packing items, loading and unloading goods, and moving inventory in warehouses.

By 2050, Morgan Stanley estimates that more than 1 billion humanoid robots could be working globally, with a market size of over $5 trillion. This is one of the biggest opportunities in the AI era.

However, the path from exciting viral hype demos in controlled environments to real-world deployment at scale is neither straightforward nor solved yet. Operating beyond predefined conditions requires breakthroughs in hardware design, systems engineering, and scaled manufacturing.

That gap is why our research team at Social Capital spent months studying humanoid robots from first principles. Not just what looks impressive today, but how these systems are built, where the constraints lie, and what a credible deployment roadmap could look like. Here are a few things we covered in our 119-page Deep Dive:

• The history of robotics and how we got to the humanoid form factor.

• The three biggest risks standing in the way of the mass deployment of humanoid robots.

• The full timeline: what adoption looks like in 5, 10, and 20+ years, industry by industry.

• The geopolitical race to build humanoid robots, and which countries are already pulling ahead.

• How global adoption is unfolding differently across the U.S., China, Europe, and emerging markets.

• A deeper breakdown of how these systems actually work: the sensors, the motors, and the AI making real-time decisions.

Physical AI will be a major component in our AI Stack Deep Dive and 2025 Annual Letter.

Let me know your thoughts in the group chat after reading.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive below ↓

Read more

Caught My Eye…

1) Meta’s Multi-Billion, Multi-Year NVIDIA Partnership

On February 17, 2026, Meta and NVIDIA announced a long-term infrastructure partnership to expand Meta’s AI-optimized data centers. The agreement covers millions of NVIDIA Blackwell GPUs and next-generation Rubin GPUs. Meta will also integrate NVIDIA Spectrum-X Ethernet into its networking infrastructure. This will speed up their deployment at scale, unlocking gains in AI training efficiency and shortening time to insights.

Meta says it will deploy NVIDIA Grace CPU-only servers at large scale, with NVIDIA’s next-gen Vera CPUs planned later. This represents the first large-scale NVIDIA Grace-only deployment, when hyperscalers typically source general-purpose CPUs from Intel and AMD.

Meta and NVIDIA described this initiative as deeper co-design, positioning it as a systems-level effort to improve performance per watt rather than a single component purchase.

This partnership lands inside Meta’s newly raised 2026 capex outlook of $115–$135B, up roughly 75% from 2025, and is almost entirely focused on its Superintelligence Labs efforts. Zooming out, the infrastructure race is still going strong, with multiple hyperscalers raising their 2026 capex guidance.

2) The Fed Pricing in AI and Prediction Markets

On Feb. 18, 2026, the Fed released minutes from its Jan. 27-28 meeting. Officials kept rates at 3.50%–3.75% (10–2, with two dissenting for a 25 bp cut). This shows a patient approach amid steady growth, a balanced labor market, and inflation above target.

The Committee viewed AI as a key factor in their decisions. The minutes mention a “strong pace of AI-related investment,” linking productivity gains to economic support.

However, they also noted that firms are cautious about hiring due to uncertainty over how AI and automation will affect labor demand. They flagged AI as a source of forecast uncertainty and a potential financial risk, citing high valuations, concentration, and AI projects funded through opaque private markets that require oversight.

Federal Reserve Governor Michael Barr echoed this view the day before. He emphasized the uncertainty about monetary policy’s impact on structural factors such as AI. In the short term, he expects painful labor market disruptions, but in the long term, he believes the labor market will eventually adjust, creating new jobs and enhancing others, which will boost real wages.

That same week, a new Fed working paper suggested they should monitor prediction markets. The paper found that Kalshi can provide real-time probability distributions for inflation, unemployment, GDP, and specific Fed funds outcomes that traditional tools often miss.

Since 2022, Kalshi’s median/mode have accurately matched the next meeting’s decision the day before FOMC meetings, outperforming fed funds futures. Additionally, Kalshi’s CPI forecasts, especially the median/mode, have surpassed Bloomberg consensus. Notably, the distributions showed a bias: upside CPI surprises influenced expected policy paths more than downside surprises.

3) J&J’s New 2 Million sq ft Cell Therapy Facility

On February 18, 2026, Johnson & Johnson announced a $1B+ investment to build a next-generation cell therapy manufacturing facility in Montgomery County, Pennsylvania. The site is designed to expand J&J’s U.S. production capacity for advanced therapies that support its medicines portfolio and pipeline, including treatments in cancer, immune-mediated, and neurological diseases.

The company positioned the project as both a technology and workforce play. This investment aims to advance cutting-edge manufacturing processes and train a workforce with skills in advanced technologies shaping the future of medicine.

J&J estimates it will create more than 500 skilled biomanufacturing jobs once fully operational, and more than 4,000 construction jobs during development. J&J also estimates a ~$10B annual economic boom to the state, with 10 facilities totaling more than 2 million square feet across manufacturing, research, distribution, and office operations.

J&J tied the investment to its previously announced plan to invest $55B in the U.S. through early 2029 across manufacturing, R&D, and technology. The goal is to support manufacturing the vast majority of its advanced medicines in the U.S. for U.S. patients.

Learn With My Friends and Me…

• All-In: David Sack’s Epstein Debate Panel

Other Reading…

• Kalshi and the Rise of Macro Markets (Federal Reserve)

• Multi-Agent Cooperation Through In-Context Co-Player Inference (ARXIV)

• Measuring AI Agent Autonomy in Practice (Anthropic)

On X…

I hope you’re enjoying and learning from all the Deep Dives and What I Read This Weeks my research team and I have been producing this year.

It’s been a lot of work, but also a lot of fun, and we’re just getting started. I can’t wait to share all the fascinating topics we have lined up.

And I have something new I want to share with you…

I’m launching my YouTube channel!

Click on the image to watch or click the link here: https://youtu.be/0-LAT4HjWPo

I received a lot of incredibly positive feedback on my health video on All In, so my team and I decided to start a channel for me to share lessons and insights from my 30+ years in business, investing, and life, and to turn a few of our Deep Dives here into videos to make them more accessible and digestible. Aiming for 1x video a week every Friday.

Here are a few videos we have planned:

- Breakdown of my Groq's investment from the early founding days to $20B deal with NVIDIA
- My best and worst investments
- Electric Grid and Solar explainers
- Humanoid Robotics explainers

To the extent that you’re interested, check out my first video and leave a comment to share what you think and what you would like to see from me.

If the community enjoys them and I get more views than Jason, I’ll keep doing it 😉!

Only joking, Jason, love you.

Watch the video, share it with a friend, and let me know in the comments what you think!

See you on YouTube!

Chamath

For most of history, investing was about relationships, judgment, and timing. It was about who you knew, what you believed, and whether you had the conviction to act. That edge worked when information was scarce, and markets were slower.

Today, public and private information are instantly priced in. As such, strategies that depend on reacting to news, reading filings, or “having an opinion” compete in a game where the half-life of new information is measured in milliseconds.

No human, no matter how smart, experienced, or well-networked, can manually process that firehose of data in real time and execute on trades accordingly, with perfect discipline.

Unsurprisingly, the best long-term returns in modern investing come from a fund mainly managed by code, not by humans.

The chart shows a top-performing autonomous system, outpacing the S&P 500 over many years.

That is the Renaissance Medallion Fund, which has earned 39% annual returns since the late 1980s.

How did a fund run by code double the returns of the market?

Medallion’s edge came from consistently repeating small statistical advantages across thousands of instruments simultaneously. Tiny edges, captured over and over through rapid trades in highly liquid assets.

Human investors can closely track only a limited number of instruments. They can’t monitor thousands of markets in parallel, spot micro-opportunities in milliseconds, and execute that process across thousands of trades consistently.

The shift to code-driven systems started when researchers began building mathematical frameworks to work around them.

Brief history of autonomous investing

Between 1952 and 1973, three breakthroughs made automation inevitable.

1. In 1952, Harry Markowitz proved portfolio construction could be mathematical.

2. In 1964, William Sharpe introduced the Capital Asset Pricing Model (CAPM), giving investors a way to measure risk, compare returns to a benchmark, and quantify performance with what became known as the Sharpe ratio.

3. In 1973, Fischer Black and Myron Scholes published The Pricing of Options and Corporate Liabilities (the Black-Scholes equation for pricing options), laying the foundation for automated investing.

To automate investing, you need:

1. Clear input data.

2. Predictable output results.

3. Rules that don’t rely on human guesses.

4. The ability to adjust using formulas alone.

Black-Scholes’ method provided all four. This meant options could now be priced and traded entirely by formulas and machines, replacing the old method of human estimation.

Between the 1960s and 1980s, Edward Thorp turned these ideas into practice at Princeton/Newport Partners. He used math-based investing to generate positive returns.

Jim Simons took Thorp’s probabilistic framework and scaled it at Renaissance’s Medallion Fund, often considered the greatest fund of all time by annual returns.

Medallion reportedly earned over 65% gross and about 39% net annually from the late 1980s, with tightly managed capital. It closed the fund to outside investors in 1993 because additional capital would dilute returns by pushing trades into less liquid, lower-edge opportunities.

By comparison, Berkshire’s audited record shows roughly 19.9% annualized returns since 1965. Over the same period, the S&P 500 delivered roughly ~10% annualized returns

The Five Levels of Autonomous Investing

Renaissance’s Medallion Fund returns came from code that automatically sized, hedged, and executed trades across thousands of instruments. The specific algorithms Renaissance built remain secret, so we aren’t able to look at the specific methodology or formulas.

However, the research suggests that the edge was not attributable to any single algorithm. It was how much of the investing process could be handed to machines, and how those connected pieces worked together. To understand where that stands today, it is helpful to examine autonomous investing at five levels.

Level 1: Manual Investing

The edge came from information scarcity. Who you knew, what you believed, and whether you had the conviction to act. The edge was largest when markets moved slowly enough for humans to process them.

Level 2: Algorithmic Investing

Computer programs execute pre-defined rules automatically. A signal comes in, a trade goes out. No learning, no adaptation. The edge here is speed and discipline.

Level 3: Automated Investing

Systems manage complex tasks and automate entire workflow stacks. Fintech software and APIs integrate data feeds, portfolio models, and execution into a single continuous process, reducing friction from manual handoffs but without adding intelligence.

Level 4: Autonomous Investing

Machine learning models that operate within objectives/constraints. Able to update based on new data without explicit reprogramming.

Level 5: Agentic AI Investing.

A fully autonomous system that can plan, choose actions, use tools, monitor outcomes, and self-correct across multi-step workflows. In October 2025, seven frontier AI models each managed $10,000 in real capital with zero human intervention at Alpha Arena.

At Learn With Me, we decided to research and write this Deep Dive to understand what autonomous investing is, how it integrates research, risk, compliance, and execution into a single AI-driven stack. This is now the primary source of competitive advantage.

In the full Deep Dive, we cover:

• The new agentic infrastructure.

• Regulation and global macro perspectives.

• Three main risks for autonomous investing.

• Analysis of Renaissance’s Medallion Fund.

• How autonomous traders detect anomalies with the precision of human intuition.

• Where major investment firms such as JPMorgan, BlackRock, and Bloomberg fall across the three layers of autonomous investing.

Hope you enjoy reading,

Chamath

P.S: Please let me know what you think in our Substack group chat, and if we should share more specific investment research.

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive below ↓

Read more

Caught My Eye…

1) Gemini 3 Deep Think: A New Era of Reasoning

On February 12, 2026, Google DeepMind announced a big upgrade to Gemini 3 Deep Think, a specialized reasoning mode designed to solve the world’s most complex scientific, research, and engineering challenges.

This model enables researchers to interpret complex data and engineers to model physical systems through code. In a featured example, the Wang Lab at Duke University optimized fabrication methods for complex crystal growth to discover new semiconductor materials. Deep Think successfully designed a recipe for growing thin films larger than 100 μm.

The performance metrics released with this launch are record-breaking across every major technical domain. Gemini 3 Deep Think set a new high-water mark of 84.6% on ARC-AGI-2, a benchmark designed to test a model’s ability to adapt to entirely new tasks. This is almost double their previous score of 45.1% 3 months ago. Currently, only seven humans have a higher elo than Deep Think’s score of 3455 on Codeforces.

Additionally, Deep Think can now interpret messy experimental data, assist in autonomous scientific research, and turn a hand-drawn sketch into a functional 3D-printable CAD file. This shows that scaling laws continue through inference-time computation. One application that showcases this is their math research agent, Aletheia, which autonomously follows a specific process to get the highest-quality answer.

2) Agentic Wallets: Giving AI Agents Purchasing Power

On February 10, 2026, Coinbase announced the official launch of Agentic Wallets, a first-of-its-kind infrastructure designed specifically for autonomous AI agents. While previous AI models could advise or summarize data, they often hit a wall when a task required a financial transaction. These new wallets remove the human bottleneck, enabling AI agents to independently hold funds, execute trades, and pay for resources such as API keys and compute power without needing manual approval at each step.

“Now agents can spend, earn, and trade autonomously and securely.” - Brian Armstrong, CEO of Coinbase

The technology is built on the x402 protocol, an open protocol for internet-native machine-to-machine transactions, that has already processed over 50 million transfers. To ensure safety, Coinbase has integrated Smart Security Guardrails that allow developers to set session spending caps and transaction limits.

This launch signals the arrival of the Agentic Web, where autonomous bots can manage DeFi portfolios, rebalance liquidity at 3 AM, or monetize their own generations in the creator economy. Coinbase is lowering the barrier for developers by offering gasless trading on its Base network and believes we are now in the age of AI agents that can act, not just advise.

3) Isomorphic Labs Unveils The Successor to AlphaFold

On February 10, 2026, Alphabet’s biotech subsidiary Isomorphic Labs announced the launch of the Isomorphic Labs Drug Design Engine (IsoDDE). Its predecessor, AlphaFold 3, won a Nobel Prize for predicting static 3D structures of molecules. IsoDDE shifts focus from observation to engineering. This new platform is “full-cycle,” designed to predict how new drugs bind to their targets and whether they have a therapeutic effect.

IsoDDE more than doubles AlphaFold 3’s accuracy on out-of-distribution benchmarks.

This means it can predict the structure of new protein-ligand systems that the model didn’t see during training. The engine can find cryptic pockets using just an amino acid sequence. Most proteins are dynamic, which means they have binding sites that appear briefly as they move. Earlier AI models only detected a smooth surface, leading researchers to label many aggressive cancers and neurodegenerative diseases as undruggable. The IsoDDE uses advanced reasoning to mimic this molecular motion and uncover these cryptic areas. This allows scientists to design precise drugs targeting KRAS, a cancer-driving protein, and to use Cereblon, a substrate receptor.

In August 2025, Deep Origin, another biotech firm, presented results at the Aging Research and Drug Discovery (ARDD) conference. Their physics-informed docking engine, DODock, outperformed AlphaFold 3 on the same benchmark as IsoDDE.

“We’re glad to see Isomorphic Labs reaching performance levels comparable to what we presented months ago. More accurate docking is good for drug discovery, full stop.” - Michael Antonov, CEO of Deep Origin (prev: Co-founder of Oculus and Scaleform)

Learn With My Friends and Me…

• All-In: Debt Spiral or NEW Golden Age? Super Bowl Insider Trading, Booming Token Budgets

Other Reading…

• Your AI Conversations Aren’t Privileged (Moish Peltz)

• Sabotage Risk Report: Claude Opus 4.6 (Anthropic)

• xAI Welcome to All Hands (xAI)

On X…

But it can act on and express only ~10 bits per second. This includes tasks like moving, thinking, and speaking. This is like being connected to fiber-optic internet, but only being able to respond through a 1990s dial-up modem.

An upgrade is needed, especially in a world racing towards increasingly capable AI.

In 2016, Elon Musk founded Neuralink to develop a scalable Brain-Computer Interface (BCI): a system that establishes a direct communication pathway between the brain’s electrical activity and an external device, bypassing traditional pathways (i.e., nerves and muscles).

Neuralink progressed from wired implants in rodents to a wireless implant in the first human in only 5 years.

Brain chips, however, were unpopular with the general public. In 2021, a Pew Research Center study found that 78% would not want one for improved information processing, 56% said widespread use would be bad for society, and only 13% considered it a good idea.

However, therapeutic use was widely accepted. 77% favor brain chips for paralysis treatment, and 64% favor them for age-related mental decline.

And that’s where Neuralink began.

In January 2024, Neuralink conducted the first human implantation of the N1 chip in Noland Arbaugh.

Since then, Noland Arbaugh has used the N1 daily to regain digital independence, navigating the internet, learning online, communicating publicly, and controlling computers for work and play, including games such as chess and Civilization VI.

So how does Neuralink actually work?

We go into further details in the Deep Dive below, but at a high level, the process of turning a thought into a digital action follows a five-step path:

1. Listening for Neural Signals

The process begins with 1,024 electrodes distributed across 64 flexible threads. These threads are implanted into the motor cortex, the brain region responsible for movement. Each electrode acts as a microscopic microphone, detecting electrical ripples from nearby neurons.

2. Amplifying and Digitizing the Signal

Raw brain signals are exceptionally faint, at roughly 15,000 times weaker than a standard AA battery. On-chip amplifiers within the N1 implant boost these signals to measurable levels while filters remove background noise. The chip then transforms the signal waves into a long stream of data for the computer to read.

3. Identifying Spikes

The N1 needs to identify meaningful signals in the data stream, known as spikes. It analyzes the unique shape of each waveform to attribute the activity to its specific source neuron.

4. Packaging and Transmitting Data

Transmitting raw neural data would be too slow and power-intensive. Instead, the N1 chip organizes the detected spikes into small digital packets. This compressed data is then sent wirelessly via Bluetooth to an external device, such as a smartphone or computer.

5. Decoding Intent into Action

Before a user can control a computer, the system undergoes a ~20-minute calibration. During this time, the user imagines specific movements (like moving a hand or squeezing) while the software learns which neural firing patterns correspond to those intents. Once trained, machine learning algorithms decode these patterns in real-time, moving a cursor on a screen dozens of times per second.

As of late January 2026, Neuralink has enrolled 21 participants in its worldwide clinical trials. These individuals use the implant for an average of 50 hours per week, controlling computers, playing games, browsing, and handling daily tasks entirely through thought.

21 participants represent a meaningful advance over the early animal trials. Yet Neuralink’s stated target is far larger: 10,000 implants per year by 2030, at an estimated total cost of $40,000 to $50,000 per patient.

Reaching that volume will require consistent long-term thread performance, further automation of the R1 surgical robot, streamlined regulatory pathways, and acceptance within medical systems. The challenges ahead are still substantial.

Neuralink also intends to expand beyond motor control. Future capabilities include restoring tactile sensation, enabling natural speech for individuals who have lost it, and restoring vision through a planned product, Blindsight, that directly stimulates the visual cortex with camera input.

What Neuralink has already proven makes this one of the most fascinating areas in technology right now. That is why we decided to research it in depth and create this Deep Dive.

Neuralink is pursuing one of the most difficult frontiers in technology.

The implications include reversing neurological disabilities, modulating circuits involved in depression or anxiety, enhancing memory and cognitive speed, and ensuring humans remain competitive as AI capabilities advance rapidly.

I invite you to learn with me. Let me know what you think in our Substack group chat.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive below ↓

Read more