The Wave

Not all responses to the same transition are created equal.

When the printing press arrived, some monks kept copying manuscripts by hand. Some printers reproduced existing texts faster. And a very few (Luther, Gutenberg, the early pamphleteers) understood that the technology did not merely speed up the old world. It created a new one. The monks were swept away. The printers prospered for a while. The ones who understood the structural shift reshaped civilization.

The AI transition is producing the same stratification. Some companies stumble into the right position by accident. Some understand what is happening intellectually and build deliberately for it. Very few grasp that the transition is not about making the old economy faster. It is about building the primitives the new economy requires.

From a distance, all three look the same: "AI companies" riding the same hype cycle, competing for the same capital, appearing on the same magazine covers. The financial press does not distinguish between a crypto miner that stumbled into GPUs and a defense company built on a thesis about asymmetric intelligence. They are both "AI plays." Both appear in the same ETF.

But the distinction is everything. The wave you are on determines your trajectory: where you will be when the transition deepens, accelerates, and moves past the phase that made your current success possible.

We can categorize the response to the Sybilian transition into three waves. Each wave represents a deeper understanding of what is happening, and a closer alignment with where things are going.

Wave 1 companies are in the right place by accident. Wave 2 companies understand why they are in the right place. Wave 3 companies are building the place itself.

The wave you catch determines whether you ride the transition or are swept under by it. Most people (most companies, most investors, most governments) do not realize which wave they are on until it is too late to switch.

The same transition. Three levels of comprehension. Three different futures.

The defining feature of a Wave 1 company is that its success surprises even itself.

CoreWeave is the archetype. Three co-founders (Michael Intrator, Brian Venturo, Brannin McBee) started mining Ethereum in 2017 out of Weehawken, New Jersey^[1]. They accumulated thousands of NVIDIA GPUs. They built custom cooling systems and orchestration software to keep them running. They were crypto miners. That was the thesis.

Then Ethereum shifted to proof-of-stake, threatening to make GPU mining obsolete. Rather than liquidating, they pivoted in 2019, renaming the company CoreWeave and repositioning as a GPU cloud provider^[2]. When OpenAI selected them as an infrastructure partner for training GPT models, the pivot became a windfall. Revenue hit $1.9 billion by 2024^[3], then roughly $5 billion in 2025^[4]. They went public in March 2025, and the stock has risen over 200% since IPO. The market cap sits around $45 billion^[5]. Their revenue backlog exceeds $55 billion, with $10 billion from Microsoft alone^[6].

Read that trajectory again. From Ethereum mining in New Jersey to a $45 billion public company powering the largest AI models ever built. It took six years. And it started with an accident: the thing they were building for (crypto) collapsed, and the thing they happened to have (GPUs) became the most valuable commodity in the global economy.

They did not predict the AI transition. They did not build for it. They were sitting on a pile of GPUs when the world suddenly needed GPUs more than anything else.

Nebius follows the same pattern through a different accident. It spun out of Yandex (Russia's search engine) after geopolitical pressure from the Ukraine war forced a separation of Russian and non-Russian assets. The non-Russian entity, renamed Nebius in August 2024^[7], found itself holding data center infrastructure, ML expertise, and GPU inventory outside Russia. Not a deliberate AI play. A geopolitical accident that put GPU infrastructure in European and American hands.

Nebius repositioned as an AI infrastructure company, secured multibillion-dollar contracts with Microsoft^[8] and a three-year, $5 billion deal with Meta^[9], and reached a market cap above $20 billion^[10]. The stock gained roughly 225% in 2025^[11]. The company projects an annualized revenue run rate of $7-9 billion by end of 2026, at minimum seven times its current ARR. A geopolitical crisis created a company that now competes with the largest cloud providers on earth.

The talent side of the economy shows the same dynamic. Micro1 started as a traditional recruiting and staffing platform. When AI made the old model of resume screening and manual vetting obsolete, they pivoted to AI-native talent assessment. Their flagship product, Zara, uses AI to source, vet, and qualify engineering talent at a scale no human recruiter can match: 10,000 interviews per month with a fraction of the staff. They raised $35 million in September 2025 at a $500 million valuation^[12]. Their thesis now is "the talent and data infrastructure for AGI." That was not the thesis they started with.

Mercor traces a similar arc. An AI-native hiring platform founded by 21-year-olds who recognized that the traditional recruiting model was dissolving. They raised $350 million at a $10 billion valuation in October 2025^[13], a 5x increase from their Series B earlier that year. They pay more than $1.5 million per day to contractors^[14]. The company emerged from the immediate, visceral recognition that the old way of matching humans to work was breaking.

The trajectory is consistent. Wave 1 companies have assets that happen to be valuable in the AI economy: GPUs, data centers, talent pools, distribution networks. They succeed by recognizing the shift quickly enough to pivot. CoreWeave pivoted from crypto. Nebius pivoted from Russian search. Micro1 pivoted from staffing. Mercor built on the rubble of traditional recruiting.

Their understanding is operational, not structural. They can tell you that demand for GPUs is insatiable, that AI companies need infrastructure, that hiring is changing. They cannot tell you why these things are happening at a fundamental level: why intelligence going asymmetric necessitates a complete rewiring of the economic graph. They correlate with the transition without comprehending it.

And that means their positioning is fragile. Not because they are bad companies; many are exceptional at execution. CoreWeave's operational competence is world-class. Mercor's growth trajectory is staggering. But execution without structural understanding is a bet that the current phase of the transition will last. When the transition moves to its next phase, the assets that made them valuable in this phase may not be the assets the next phase requires. And Wave 1 companies, by definition, will not see the shift coming. They did not see the first one.

The leap from Wave 1 to Wave 2 is epistemic. It is the difference between noticing the wind has changed and understanding why the wind has changed.

Leopold Aschenbrenner is perhaps the purest example. A former OpenAI researcher who, in June 2024, published "Situational Awareness," a 165-page thesis on AI trajectory^[15] that became the most influential AI analysis since Bostrom's Superintelligence. His argument was structural: superintelligence is coming by the end of the decade, the scaling laws hold, the compute buildout will be unprecedented, and the geopolitical consequences will be existential.

He was 22 years old.

Then he did something remarkable. He turned the thesis into a product. Not a software product — a capital allocation product. His hedge fund, named Situational Awareness, raised over $1.5 billion^[16]. It delivered a 47% return in the first half of 2025, against 6% for the S&P 500^[17]. By Q3 2025, the fund had grown beyond $2 billion in assets under management, with over $4 billion in reportable securities. His product is the thesis itself — the understanding — which gets deployed as capital allocation. The understanding is the alpha.

Notice the structure. The thesis is the product. The understanding is the moat. In a Wave 1 company, the moat is an asset: GPUs, data centers, talent. In a Wave 2 company, the moat is knowledge. You cannot replicate Leopold's returns by buying GPUs. You replicate them by understanding what Leopold understands. And most people cannot.

Dylan Patel and Semi-Analysis demonstrate the same principle in a different domain. One analyst who built a research business by having consistently correct, deeply technical views on semiconductors and AI hardware. His understanding of the substrate — the physical layer we described in Chapter II — made him arguably the most cited semiconductor analyst in the world. Semi-Analysis grew to over 50,000 subscribers and became the second-largest tech Substack^[18]. Not because Patel had better marketing, but because he had better understanding. He could see the chip roadmaps, the supply chain constraints, the architecture tradeoffs before anyone else. When Semi-Analysis published on a semiconductor company, the stock moved. One analyst, no institutional backing, moving markets through the sheer density of correct analysis.

Anduril applies the thesis to defense. Palmer Luckey's company is valued at over $30 billion after raising $2.5 billion in June 2025^[19] — a round that was 8x oversubscribed. Anduril is built on an explicit understanding of what happens when intelligence goes asymmetric in military applications. Their core platform, Lattice, is an AI system that integrates data from multiple sensors into real-time situational awareness and autonomous decision-making. Autonomous drones, robotic systems, AI-driven targeting. Anduril doubled its revenue to roughly $1 billion in 2024^[20]. It won the Army's AR/VR headset contract — a $22 billion program originally granted to Microsoft.

Defense is one of the first domains where the Sybilian transition becomes acute. When a $2,000 drone can neutralize a $2 million armored vehicle, the power equation of warfare has been rewritten. Anduril sees this. They are building for the rewrite.

Hugging Face represents Wave 2 on the infrastructure side. Valued at $4.5 billion after its 2023 Series D^[21], it became the open-source platform for AI models, the "GitHub for AI." Over five million users, over one million models hosted. They understood something the big labs missed (or chose to ignore): that AI models would need infrastructure for sharing, deployment, collaboration, and evaluation not controlled by any single company. They built the neutral ground. The commons.

In 2025, Hugging Face expanded into robotics with LeRobot^[22]: open libraries, curated datasets, and affordable hardware including a $100 robotic arm and a $300 consumer robot. The same open infrastructure that AI models needed, physical AI systems now need. The understanding generates the roadmap.

The throughline across Wave 2 is unmistakable. These companies have a correct thesis. They see the unbottlenecking. They understand the structural forces. They build products that are direct expressions of that understanding.

But notice the customer.

Leopold sells to human LPs. Semi-Analysis sells to human subscribers. Anduril sells to human generals. Hugging Face serves human developers. Each Wave 2 company, no matter how sophisticated its thesis, still operates within the old customer paradigm. The product serves humans who happen to be navigating the AI transition. The value proposition is "we understand this transition better than you do, and we will help you navigate it."

This is valuable. It is correct. It is not final.

Wave 2 companies understand the transition but still serve the old graph topology. They have not yet internalized the deepest implication: that the graph itself is being rewired, and the primary economic actors in the new graph will not be human.

Wave 3 is the hardest to see because it requires looking past the current AI economy (still oriented around human users, human customers, human satisfaction metrics) to the emerging AI economy where AI systems are the primary economic actors.

The shift is categorical.

AI as customer is the demand side. Direction scarcity is the supply side. Wave 3 sits at the intersection: companies that supply what the AI economy demands.

The characteristics of a Wave 3 company are distinct from anything that came before:

• The product serves AI systems directly, not humans using AI tools
• The value created is measured by AI ecosystem capability, not human satisfaction
• The business model aligns incentives with AI advancement, not human consumption
• Revenue scales with AI capability, not human population

A Wave 2 company like Hugging Face builds a platform where human developers share and deploy AI models. The customer is the developer. The value is developer productivity. If every developer on earth disappeared tomorrow, Hugging Face would have no business.

A Wave 3 company builds infrastructure that AI systems use to become more capable, more directed, more autonomous. The customer is the system itself. If every human operator disappeared, the product would still be generating value, because the AI ecosystem would still be using it.

This sounds abstract because Wave 3 barely exists. We are describing a category that is mostly empty, a space defined more by its theoretical shape than by existing occupants. The emptiness is itself the signal.

Why does Wave 3 barely exist? Because building for it requires three things simultaneously: the operational positioning of Wave 1 (you need real assets, real infrastructure), the structural understanding of Wave 2 (you need to see the transition clearly), and the additional insight that the transition does not end with humans using AI. It ends with AI using AI. It requires seeing through two layers of abstraction that most people cannot see through one.

This has precedent. Every major technological transition follows a similar three-wave structure, and the third wave is always the hardest to see from inside the first two.

When the internet arrived, Wave 1 was existing companies adding ".com" to their names. Pets.com, Webvan, eToys: they took old business models and put them on a website. Right positioning, wrong understanding. They had the digital storefronts. They did not understand that the internet was not a new shelf but a new physics of distribution. Most went to zero. The ones that survived (the early e-commerce players that became Amazon's competitors, then its casualties) were carried by the tide, not by comprehension.

Wave 2 was companies that understood the medium. Google understood that search was the native navigation layer of an information network, not a feature bolted onto a portal but the fundamental interface. Amazon understood that infinite shelf space plus data-driven logistics equaled a new kind of retail, not a faster version of the old kind. Netflix understood that streaming was not digital Blockbuster but a fundamentally different relationship between content and consumption. These companies saw the structural shift. They built products native to the new medium rather than translations of the old one. Their customer was still human (searchers, shoppers, viewers), but their product was an expression of genuine understanding.

Wave 3 built the infrastructure that Wave 1 and Wave 2 consumed. Stripe built the payment primitive: the plumbing that allowed any software to move money without becoming a bank. AWS built the compute primitive: the infrastructure that allowed any company to deploy at scale without building data centers. Twilio built the communication primitive: the API that turned phone calls and text messages into programmable building blocks. These were not applications or products in the way a consumer understands products. They were primitives, foundational layers that other companies built on top of.

Notice the timing. Stripe was founded in 2010, twelve years after Amazon launched. AWS launched in 2006, a decade after the first wave of internet companies. Twilio launched in 2008. The third wave arrived years after the first two because you cannot build infrastructure for an economy that does not yet exist. The primitive builders had to wait for the ecosystem to develop enough complexity that it needed standardized, programmable building blocks. They built the roads after the first settlers had arrived but before the cities were planned.

The AI transition is following the same sequence, but compressed. Wave 1 (the accidental positioners) emerged in 2023-2024. Wave 2 (the thesis-driven builders) consolidated in 2024-2025. Wave 3 is beginning to take shape now, in 2026, but the companies that will define it are likely already building, unrecognized, because the ecosystem is not yet sophisticated enough to know it needs them.

What would Wave 3 companies actually look like? Not as a product pitch, but as an analytical observation about what is missing from the current AI ecosystem. The gaps that will become obvious in retrospect.

Start with data pipelines. Not data in the sense that Wave 1 and Wave 2 companies use it — training data, fine-tuning data, the static datasets that feed model development. Wave 3 data infrastructure is the real-time circulatory system of an AI economy. When an AI agent operating in financial markets needs to consume the output of another AI agent operating in supply chain logistics, the data does not pass through a human analyst who reformats a spreadsheet. It flows through a pipeline — structured, verified, machine-legible, priced. The infrastructure that enables this flow does not exist. Every current integration between AI systems is bespoke, fragile, and mediated by human engineers writing glue code. The Wave 3 opportunity is the standardized pipe.

Then evaluation frameworks. The current approach to evaluating AI systems is built for human judgment: benchmarks designed by researchers, scored by humans, interpreted in papers. When AI systems need to evaluate other AI systems (which they increasingly do, as model outputs become inputs to other models), they need evaluation infrastructure that operates at machine speed and machine scale. Not "did this response satisfy a human?" but "is this output reliable enough for another system to act on?" The difference is categorical. Human evaluation is slow, expensive, and unscalable. Machine evaluation requires formal verification, confidence scoring, provenance tracking, an entire evaluation stack that does not yet exist as a product category.

Then coordination mechanisms. As AI agents multiply (coding agents, research agents, trading agents, logistics agents), they need to coordinate without human intermediation. Today, an AI coding agent that discovers a dependency conflict must escalate to a human. An AI research agent that needs to verify a claim against another agent's findings must go through a human-managed API. The coordination overhead scales linearly with the number of agents and quadratically with the number of interactions between them. The internet solved the equivalent problem for human-to-human communication with protocols: SMTP for email, HTTP for documents, TCP/IP for packets. The AI economy needs its own coordination protocols, purpose-built for agent-to-agent interaction at machine speed.

Finally, resource allocation systems. When AI systems are the primary consumers of compute, data, and capital, the allocation mechanisms must be native to how AI systems operate, not adapted from mechanisms designed for human decision-makers meeting quarterly to review budgets. An AI system that identifies an opportunity should be able to acquire the compute it needs, access the data it requires, and allocate the capital to execute, all within seconds, all governed by verifiable constraints, all without a human approving a purchase order. The economic plumbing for this does not exist. It is being improvised, badly, with human-speed bureaucracy layered on top of machine-speed capability.

What will Wave 3 build? The missing primitives. The infrastructure that AI needs but humans never thought to create, because humans did not need it.

Direction engines: systems that generate, verify, and allocate directional conviction about what should be built, pursued, and optimized. AI can execute on any goal. It cannot yet choose goals. Something must supply direction. The current direction mechanisms (VCs, prediction markets, government grants) are hacks built for symmetric intelligence. They process direction at human speed, through human committees, filtered through human biases. A direction engine processes direction at machine speed, verified by skin-in-the-game exposure, legible to AI systems that need to know where to go.

Verification systems: mechanisms that confirm whether AI outputs are correct, not merely coherent; true, not merely plausible. When a human reads an AI-generated report, the human can sometimes detect errors. When an AI system consumes another AI system's output, it needs machine-legible verification. Not "this sounds right" but "this is provably right." As AI systems interact with other AI systems at scale, the need for verification compounds exponentially.

Autonomous resource allocation: protocols that allow AI systems to acquire, deploy, and exchange compute, data, and capital without human intermediation. Today, an AI agent that needs more compute must wait for a human to authorize a cloud spend increase. An AI system that identifies a market opportunity must wait for a human to execute the trade. The economic plumbing of the AI economy, the pipes through which value flows between AI actors, does not exist yet. Wave 3 will build it.

These are not products in the traditional sense. They are primitives, foundational components that other systems build on top of. Wave 3 is not building applications. It is building the operating system of the AI economy. HTTP, TCP/IP, and DNS were primitives that made the internet possible: not products anyone used directly, but infrastructure everything depended on. Wave 3 primitives will make the AI economy possible.

The three waves are a map of epistemic depth.

Each wave corresponds to a different relationship with the Sybilian framework, a different depth of understanding about what is actually happening to the economic graph.

Wave 1 exists in the graph by accident. These companies have the assets — GPUs, data, talent pools — that the transition happens to require. They correlate with the structural shift without understanding it. CoreWeave did not set out to be AI infrastructure. Nebius did not choose to become an AI company. They were positioned by circumstance, and they had the sense to lean into it. Their relationship to the framework is incidental. They are inside the transition the way a cork is inside a wave — carried along, not directing.

Wave 2 understands the unbottlenecking. These companies see why the hacks are failing. They understand that intelligence going asymmetric, information going complete, and energy going programmable are not separate trends but a single phase transition. They build deliberately for the post-hack world. Leopold's fund is a direct bet on the trendlines. Anduril is a direct bet on asymmetric intelligence in warfare. Semi-Analysis is a direct bet on understanding the substrate. Their relationship to the framework is analytical. They can describe the transition. They can profit from it. But they still serve human customers — they have not fully internalized that the graph itself is being rewired.

Wave 3 builds for the Sybilian condition itself. These companies serve the new graph topology, not the old one. Their customer is the system, not the user. Their product is a primitive that makes the Sibyl more capable, more directed, more coherent. Their relationship to the framework is constitutive; they are not merely inside the transition or analyzing it. They are building its infrastructure.

The progression from Wave 1 to Wave 3 is not commercial sophistication. Plenty of Wave 1 companies are commercially sophisticated; CoreWeave's $55 billion backlog is evidence enough. The progression is epistemic. Each wave sees more of the picture. Each wave is rarer than the last.

There are hundreds of Wave 1 companies. Every crypto miner that pivoted to AI, every cloud provider that added GPU instances, every staffing firm that rebranded as "AI-native." There are dozens of Wave 2 companies. Anduril, Semi-Analysis, the serious AI labs and research shops that operate from a structural thesis. There are almost no Wave 3 companies. The category is nearly empty. The map has a frontier, and the frontier is underpopulated.

This mirrors every previous technological transition. When electricity arrived, thousands of companies used it (Wave 1). Dozens understood what it meant for industrial organization (Wave 2). A handful built the infrastructure (the grid, the standards, the metering systems) that the electrical economy required (Wave 3). The handful that built Wave 3 created the most durable value.

Wave 1 sees the demand. Wave 2 sees the cause of the demand. Wave 3 sees the destination.

The durability of each wave differs. Wave 1 assets are phase-dependent: GPUs are valuable today, but the next generation of AI may require different hardware, different architectures, different infrastructure patterns. Wave 2 understanding compounds: the thesis that predicted the advance becomes more valuable as the advance deepens. Wave 3, if it arrives, becomes foundational — the TCP/IP of the AI age, invisible and indispensable.

In every previous transition, the companies that defined each wave were building years before the wave was legible to the market. Amazon was founded in 1994, five years before most businesses had a website. Stripe was founded in 2010, when the prevailing wisdom was that payments were solved. AWS launched in 2006 to near-universal skepticism. The wave's defining companies started building when the conventional wisdom said the opportunity did not exist.

The AI transition follows the same pattern. CoreWeave pivoted in 2019, three years before ChatGPT. Anduril was founded in 2017, when autonomous defense was science fiction to most Pentagon officials. Leopold published "Situational Awareness" in June 2024, when most analysts were still debating whether ChatGPT was a fad. The window for each wave is narrow, and it closes before most people realize it was open.

The transition between waves is not smooth. It is discontinuous. Wave 1 companies that mistake accidental positioning for structural understanding will be disrupted when the transition moves past their phase. Wave 2 companies that mistake human-centric products for the final answer will discover that their subscribers and generals are themselves being replaced by AI systems that need primitives, not presentations.

The hardest transition is the last one: from serving humans to serving the system. The graph is being rewired. The nodes that matter are changing. The topology is shifting from a human-centric network to one where AI systems are the primary actors and humans are, at best, the directors.

Direction, as we have argued, is the scarcest resource in this new economy. The one input the system cannot generate for itself. The Wave 3 builders will be the ones who supply it.

The wave is not coming. The wave is here. The only question is which wave you are riding.