The AI Economy is Not the Internet Economy
February 26, 2026. Published as an article on X on May 30, 2026, where discussion continues.
The AI economy will not resemble the internet economy because its core primitives are centralized and industrial scale. Frontier capability is produced by a scaling complex of compute, chips, energy, capital, and concentrated research talent, which is scarce and cornerable. This shifts the center of gravity upstream toward a small set of frontier labs and the bottlenecks that feed them.
Because the core product is cognition, improvements generalize laterally across markets and collapse category boundaries: in venture-scale markets, “AI apps” that simply package intelligence tend to be temporary wrappers that are bundled, displaced, or commoditized once labs treat the category as first-class.
The resulting market structure is a handful of credible frontier providers and winner-take-most product surfaces within major workflows, with equilibrium outcomes determined by whether supply remains capped or proliferates. Across these scenarios, durable value concentrates in lab equity, lab-adjacent acquisition targets, beneficiaries of the scaling complex, and the scarce physical and infrastructural inputs that constrain scaling.
“many shall run to and fro, and knowledge shall increase.” — Daniel 12:4
There is no reason the future should look like the past, or the present. Unlike in Thucydides, the classics, or Dalio, history does not repeat. Every epoch is singular. The classics imagined interminable cycles — the rises and falls of civilizations — because they existed in a static world, where things did not change that much. But we live in a technologically progressing civilization, even if narrowed to mean exclusively information technology today, and as a result, there may very well be no going back. Technology permanently changes this to an extent that the Bible even calls them miracles.
In particular, there may be no going back to the internet economy. People make the mistake of assuming that things will look similar. That because a certain category of winners won in the past, and it will be the same shape of winners that win in the future. Unfortunately this won’t happen. This time around things may look very different. The internet economy rewarded founders, angels, contrarians, and dropouts. The AI economy rewards researchers, scale allocators, and those aligned with capital intensity[1].
The reason for this is that the primitives of AI are dramatically different from the primitives of the internet. The internet was built on cheap, abundant computation and open protocols. Anyone could build on HTTP. Anyone could rent servers from the cloud. Infrastructure faded into the background and value accrued at the edges. Distribution was decentralized. Capital requirements were modest. A small team with a clever insight could compete with incumbents because the dominant variable was product-market fit and distribution velocity, not industrial scale.
AI reverses these conditions. Frontier intelligence is not cheap and abundant. It is expensive and scarce. Training state-of-the-art models requires tens of billions of dollars in capital, access to cutting-edge chips, gigawatt-scale energy, massive data centers, and dense clusters of highly specialized research talent. Inference at scale requires continued capital expenditure and infrastructure optimization. Intelligence is no longer just code. It is compute, energy, and capital intensity bundled together. These are not decentralized primitives. They are industrial primitives.
The internet democratized publishing and commerce because the marginal cost of replication approached zero and the infrastructure was widely accessible. AI, at the frontier, recentralizes capability because the marginal improvement in intelligence appears tightly coupled to scale. If performance improves with compute and capital input, then the entities that can mobilize the most resources compound advantage.
The key shift is this: in the internet era, insight beat scale. In the AI era, scale may produce insight. When capability is monotonic with capital intensity, the archetype shifts accordingly. The contrarian founder operating at the edge gives way to the research lab, the growth allocator, and the operator aligned with infrastructure and energy. The primitives determine the winners. And the primitives have changed.
If Peter Thiel were the archetype of the internet era: contrarian[2], scrappy, incentivizing dropouts, and the key skill being “thinking differently”, then Sam Altman is the archetype of the AI era, whose self described skill is not to think about something that nobody else is thinking about, but rather to think about how to take an existing thing and make it 10x bigger[3]. And then once it's 10x bigger, add another zero to that, ad infinitum.
In the internet era, the center of gravity sat at the edges because value was domain-specific. Airbnb succeeding did not imply it could build Uber. Stripe did not automatically become Shopify. Each category required its own product, its own distribution, and its own operational stack. The base layers — cloud infrastructure, internet protocols, and mobile platforms — enabled companies, but they did not collapse the map of markets.
Frontier AI changes that because the core asset is not a product but a general capability substrate. Put differently, the core product is cognition. A single model improvement can increase performance across many verticals at once: code, writing, customer support, research, analysis, and more. Once you can afford to push the frontier, entering adjacent categories often means productizing an existing capability rather than building a new company from scratch. This compresses what would historically have been an application layer.
The key distinction is between companies that sell cognition directly and companies that use cognition to produce something else. An AI app is a venture-scale company whose core differentiation is wrapping frontier cognition into a product: summarization tools, copilots, assistants, research tools, drafting engines, and other software whose primary value is access to intelligence. These products are structurally fragile because improvements in frontier models generalize across markets. If a category becomes important enough, the frontier labs can treat it as a first-class workflow and bundle the capability into the model surface and distribute it through existing interfaces.
What looks like an application becomes a feature, a plugin, or a thin distribution skin[5]. This is why venture-scale AI apps in large markets tend to be transitional: they are wrappers around a moving frontier. The exception is when a company is not actually selling cognition. Some businesses use frontier models as a production technology rather than as the product itself. An AI-native law firm, for example, is not selling intelligence. It is selling legal services[6]. The models allow the firm to reduce headcount, lower cost structures, and increase throughput, but the customer is still buying representation, compliance, and risk transfer. In those cases the AI is infrastructure inside the company rather than the value proposition to the customer.
| Profession | Before AI | After AI (Augmented) |
|---|---|---|
| Law Firm | 100 lawyers | 10 lawyers + AI systems |
| Investment Research | 20 analysts | 2 analysts + AI systems |
| Consulting Team | 50 consultants | 5 consultants + AI systems |
This argument assumes a regime of highly capable AI that dramatically compresses knowledge work but does not fully automate institutions. In that world, professions shrink but do not disappear. A law firm might operate with a fraction of the associates it once had, because many tasks that previously required junior lawyers can be handled by models. But the institution itself still exists because customers are buying representation, liability, compliance, and trust rather than raw cognition. The models act as production infrastructure inside the organization rather than replacing the organization entirely.
These companies can exist and may even thrive, but they are not an “AI app layer” in the historical sense. They are service businesses or operational companies whose economics are transformed by cheaper cognition. This distinction matters because it explains why the classic venture mental model struggles in this regime. If cognition remains scarce, frontier labs eventually absorb large categories into their product surfaces. If cognition becomes abundant, the ability to charge for access to intelligence collapses and value migrates to the surrounding infrastructure, workflows, and scarce bottlenecks. In both cases, the standalone AI app whose value is access to intelligence struggles to sustain a durable moat. The stack compresses from above.
Everything above rests on a hidden question: what is the highest-order bit for improvement? Is frontier capability still primarily a function of pre-training scale, or does advantage shift to post-training, data quality, and workflow proximity?
If pre-training scaling remains dominant, the logic of this essay becomes hard to escape. Capability continues to track compute and capital input[7]. The frontier moves because a small number of institutions can afford to push it. Those institutions generalize across categories, compress the stack from above, and erase app-layer differentiation faster than it can compound. In that world, “apps” in large markets are transitional by construction. They are wrappers around a moving frontier.
If post-training becomes the dominant driver, the story changes, but it does not revert to the internet economy. Post-training is less about raw quantity and more about quality: domain-specific feedback, curated traces, evaluation, iteration loops, and deep integration into real workflows. That shifts some leverage toward whoever is closest to the end user, because proximity can generate the highest-quality data and the fastest product-discovery cycles.
But even in that world, the bar for a durable independent “app layer” remains high. If a market is large enough, the labs have the incentive to treat it as first-class and to invest in the same post-training loops. So the only stable advantages are the ones that are path-dependent and hard to replicate quickly: proprietary feedback channels, entrenched distribution, and data that is endogenous to a particular workflow. In other words, post-training dominance does not automatically create a flourishing app ecosystem. It creates a pretty intense contest over who controls the data flywheel, and that could still mean significant frontier AI lab dominance.
In information goods, the textbook intuition is that competition pushes prices toward marginal cost. In software, that intuition is constantly defeated by increasing returns. Zero marginal cost means the leader can reinvest more, ship faster, and often become “the default,” which is why many software categories become winner-take-most rather than “race to zero[8].” At the same time, not all software concentrates. Some categories commoditize, splinter by persona, or support multiple durable winners because switching costs, integration costs, and taste remain real. The right mental model is bifurcation: software markets either concentrate into a few winners or degrade into a commodity[9].
Frontier AI pushes the system toward the concentration branch for structural reasons. The modern LLM paradigm has been characterized by empirical scaling laws: performance improves predictably as a power-law function of training compute, model size, and data size over many orders of magnitude[10]. If the scaling regime continues, capability is systematically purchased with compute and capital. The relevant competition is not “who has the cleverest feature.” It is “who can marshal and deploy the most compute, fastest, and keep doing it.”
Compute-optimal training results intensify the capital burden. Work associated with Chinchilla argues that many large models were undertrained, and that compute-optimal frontier progress requires scaling model size and training tokens together[11]. In plain terms, staying at the frontier is not a one-off invention. It is a continuing industrial program with recurring bills.
That program runs into physical constraints. Electricity is the simplest proxy. A DOE/LBNL report estimates U.S. data centers consumed about 4.4% of U.S. electricity in 2023 and projects roughly 6.7% to 12% by 2028, with AI servers and associated infrastructure as major drivers[12]. Training specifically can be even more extreme: a report summarized by Axios suggests the largest training runs may require on the order of 1–2 GW of power by 2028 (with higher upper-bound scenarios later), which is the kind of load that makes “who has power and permits” part of the competitive moat[13].
This is what “scale moat” means in AI. The barrier to entry is not just code. It is access to chips, energy, data center build-out, capital markets, research talent, and the organizational competence to translate those inputs into frontier systems. When the limiting reagents are scarce and cornerable, the number of credible frontier competitors is capped. That cap is oligopoly by construction.
But physical constraints do not only protect the labs. They also force the labs to share economics with the industrial buildout unless they vertically integrate it themselves. As marginal costs and physical bottlenecks begin to matter, the owners of power, land, grid interconnect, cooling, data centers, fabs, and specialized physical assets become price makers rather than passive suppliers. The frontier lab may own the model, but the model cannot scale without the real-world substrate. In that sense, the AI economy is not only a contest between model providers. It is a bargaining game between labs, hyperscalers, infrastructure owners, energy providers, capital markets, and the physical systems that constrain deployment.
Oligopoly then changes the terminal economics. It does not guarantee infinite margins, but it does mean margins are not automatically competed to zero. Prices and margins become a function of capacity scarcity, substitution, and how quickly new supply can be brought online. If you have a handful of frontier providers, the equilibrium often looks like few-take-most with durable positive margins. If the number of credible providers expands beyond a handful, price pressure rises and the system moves toward commodity behavior.
This sets up the next question: what equilibrium do we actually land in? The answer depends on whether pre-training scale remains the highest-order bit, whether “good enough” capability commoditizes, and whether open alternatives create enough credible supply to break the cap. That is why the next section is not a prediction, but a set of equilibria.
A scaling-law world is not the same as a recursive-self-improvement world[14]. Scaling laws imply friction: each major capability jump requires another tier of compute, power, data-center buildout, chips, and capital. RSI implies escape velocity: intelligence improves itself fast enough that the physical bottlenecks become secondary. If the next tier of frontier capability requires tens of gigawatts of compute, then the bottleneck is not merely algorithmic insight. It is industrial mobilization. In that world, AI can still compound dramatically, but it compounds through power plants, fabs, data centers, interconnects, and balance sheets. The base case is therefore not magical takeoff, but industrial acceleration: intelligence improves, but at the tempo of the physical systems required to produce it.
If the scaling complex is the bottleneck, then the AI economy does not have many stable end states. It has a few basins of attraction:
In this equilibrium, the pre-training scale remains the highest-order bit. Frontier progress continues to be purchased with compute, energy, and capital. That keeps the number of credible frontier actors small, because the entry ticket is industrial grade scale. The market stabilizes as a handful of labs with durable advantage, plus a long tail of orbiters.
Here, “AI apps” in large markets are transient almost by definition. If a market is big, the labs care. If the labs care, they bundle the workflow into the model surface, subsidize distribution with existing demand, and erase differentiation with release cadence. What looks like an application becomes a feature, a plugin, or a low-margin distribution skin. The app layer does not disappear because products are unimportant. It disappears because independent cognition products are not defensible.
The surplus concentrates in the core and its bottlenecks: frontier lab equity, chips, data centers, energy, and the supply chain of the scaling complex. The main risk to this equilibrium is provider proliferation: if the number of credible frontier labs expands beyond a small handful (roughly 3–5), pricing discipline breaks, margins compress, and the market shifts away from oligopoly toward commodity behavior.
The frontier labs look like a hybrid of cloud hyperscalers + operating systems + defense primes. There are a few providers, capability differentials matter, and distribution consolidates around the primary surfaces (chat, IDE/runtime, enterprise deployment). Pricing power is durable because the supply cap holds.
Apps in big markets mostly get cannibalized into first-class features, plugins, or distribution skins. The biggest “secondaries” and private markups concentrate in labs and in the scaling complex: chips, data centers, energy, grid interconnect, cooling, and land near power. The biggest venture outcomes are either (a) labs, (b) inputs to labs, or (c) orbiters that become acquisition targets because they accelerate distribution into a category the lab wants.
In this equilibrium, frontier progress continues, but the marginal value of additional frontier capability declines for most markets. “Good enough” intelligence saturates the bulk of real work. The frontier still exists, but most buyers stop paying a large premium for the last increments of capability.
The crucial point is that this does not automatically mean margins collapse. Supply can remain capped even if demand becomes indifferent. Why? Because running reliable, compliant, enterprise-grade intelligence at scale is still capital intensive, energy constrained, and operationally difficult. In this world, the model providers look less like mythical software monopolies and more like cloud infrastructure: large, durable, and profitable, but with rents bounded by substitution and procurement pressure.
Downstream, the “winner” is still not a durable cognition app in a huge cannibalizable market. What grows is everything adjacent to cognition: integration into messy systems, workflow ownership, distribution into enterprises, trust and procurement scaffolding, and operational embedding. The center of gravity remains upstream, but the economic action partially migrates into implementation and distribution layers because intelligence becomes abundant relative to the friction of deploying it.
The signature of this equilibrium is: buyers become model-indifferent, but the provider layer still has pricing power because supply remains structurally limited.
The labs look like cloud utilities: large, durable, and profitable, but with bounded rents because many buyers become indifferent to the last increments of frontier capability. Most categories stop paying for “best model” and start paying for reliability, compliance, uptime, and procurement trust. “Cognition” becomes cheap relative to the real bottleneck: deployment into messy systems.
The economic action migrates into integration, workflow ownership, and enterprise distribution, but not as a classic SaaS app layer. It looks more like: systems integrators with productized delivery, incumbents embedding models into existing software, and companies that own the operational choke points (identity, data permissions, audit, execution environments). In this world, labs do not capture everything, but they remain central providers because supply is still structurally limited.
This is the equilibrium in which some AI application companies can survive, but only by ceasing to be mere cognition wrappers. A Sierra-, Decagon-, or Harvey-style company can win if it becomes the trusted operational layer for a workflow: owning the customer relationship, integrations, data exhaust, compliance surface, procurement trust, and day-to-day deployment burden. Its moat is not that it has unique access to intelligence. Its moat is that it owns the messy institutional surface into which intelligence must be deployed. In that sense, the winner is less an “AI app” in the internet-era sense than a workflow institution rebuilt around cheap cognition.
In this equilibrium, credible supply expands beyond a small handful. Open alternatives become “good enough” for most workloads, sovereign providers proliferate, and additional labs enter[15]. The marginal buyer can credibly switch across many providers for most use cases. Once that happens, pricing discipline breaks and margins compress. The provider layer starts to look less like differentiated cognition companies and more like capacity markets selling compute-bound tokens.
This scenario does not require models to be identical. It only requires that the buyer can substitute with acceptable performance and switching cost. When supply is abundant, cognition stops being the primary value-capture layer. The scaling complex still exists, but the rents migrate away from model providers and toward persistent bottlenecks: chips, power, facilities, distribution, integration, and the scarce assets that constrain buildout.
The signature here is: the model layer is no longer where profits concentrate, because intelligence becomes abundant enough and substitutable enough that it prices like infrastructure.
Frontier labs start to resemble airlines: huge revenue opportunity, heavy capex, operational complexity, price competition, and structurally thin margins. Not because the product is bad, but because the marginal buyer can credibly switch among many “good enough” providers and prices start clearing like capacity markets. In this world, intelligence becomes a commodity input and the surplus migrates away from “cognition providers” toward persistent bottlenecks.
The chip supply chain looks more like engine makers; cloud/data-center operators and energy providers look more like airports and fuel, capturing reliable rents; distribution surfaces (OS, enterprise channels, workflow incumbents) capture durable value because they control where demand routes. Labs still exist, but their economics look less like software monopolies and more like capital-intensive transport: high fixed costs, cyclical pricing, and limited ability to sustain high margins when capacity is abundant.
The most uncomfortable version of this equilibrium is that there is no durable secret at all[16]. Frontier models are technologically powerful but economically reproducible by any institution with sufficient capital, compute, and talent. In that case, the extraordinary valuations surrounding the labs are sustained partly by a cult-like system of incentives. Founders need the monopoly story to raise capital. Employees need it to justify sacrifice, identity, and equity value. Investors need subsequent rounds to validate earlier ones. Researchers gain status from treating their work as civilizationally singular. Governments gain leverage from treating the technology as strategically decisive. Nobody needs to be consciously dishonest; the system can sincerely reproduce a belief that no participant is individually incentivized to falsify.
The base case is that we land in Frontier Oligopoly, at least at the top of the stack. The best real-time indicator is where capital and infrastructure are concentrating. The largest financing events are not flowing to “apps.” They are flowing to the scaling complex itself. OpenAI’s newly announced $110B round[17], paired with exclusive third-party cloud capacity and multi-gigawatt compute commitments, is the market explicitly pricing in “only a few entities can afford to push the frontier.”
The second indicator is physical constraint. A credible frontier AI program is not just code, it is power, cooling, interconnect, and buildout timelines. DOE and Berkeley Lab estimate U.S. data centers consumed about 4.4% of U.S. electricity in 2023 and project roughly 6.7% to 12% by 2028, with AI servers a major driver[18]. Separate reporting focused specifically on frontier training suggests that the largest training runs may demand on the order of 1–2 GW of power by 2028[19]. When the binding constraints are industrial, the number of credible frontier actors tends to remain small.
A third indicator is the software dynamic: once a product surface becomes the default, it is not obvious it ever gets displaced. In most workflows, users don’t run “mixtures of models.” They pick a default and stay there. Even when raw capabilities are similar, frontier models often diverge over time in tone, reasoning style, and interaction patterns as a result of different training data, alignment choices, and product decisions. Users frequently develop subjective preferences for these differences — sometimes described informally as “vibes.” These stylistic divergences can reinforce switching friction, because a model that “feels right” to a user or organization may remain the default even when competing models reach similar benchmark capability. The default then compounds: it becomes the integration hub, the habit loop, the plugin ecosystem, the enterprise standard, the procurement approval, and eventually the distribution channel into adjacent categories.
This is why the base case is not merely “a handful of frontier labs.” It is “a handful of frontier labs attempting to own the default surfaces in major workflows.” If ChatGPT becomes the default interface for general chat, or Claude Code becomes the default interface for coding, the displacement threshold is not “slightly better.” It requires a discontinuity: a capability gap the incumbent cannot close, a distribution shock, or a trust event. Absent that, winner-take-most surfaces sit on top of the frontier oligopoly and make it even harder for an independent app layer to persist.
In fast-moving capability regimes it is normal for last-generation models to commoditize quickly, including through open-source alternatives. That does not necessarily imply the frontier itself commoditizes. The relevant economic object is not a particular model release but the institutions capable of repeatedly producing the next frontier system. As long as the capability gap between the frontier and the rest of the market remains meaningful for high-value workloads, those institutions can capture rents and consolidate distribution surfaces.
So if you are forced to pick one equilibrium as the base case, it is this: a small number of frontier labs, roughly a handful, competing at the edge of capability while co-evolving with the cloud oligopoly and the energy buildout. Provider proliferation remains the primary risk. If credible frontier supply expands beyond a small handful (roughly 3–5), pricing discipline breaks and the economics drift toward commodity behavior.
If the base case is frontier oligopoly plus winner-take-most product surfaces, then the first-order conclusion is obvious: most of the surplus concentrates upstream. The second-order effects are less obvious, but more important, because they change what strategies work, what institutions win, and what the culture believes is “the game.”
Three consequences fall out of the structure almost regardless of which equilibrium we land in: alpha compresses, systems become more legible to the actors closest to the frontier, and the gains concentrate into scarce bottlenecks.
In the internet economy, large outcomes often came from non-consensus insight and distribution hacks. The archetype was the contrarian founder and the early angel. In a scaling regime, the dominant variable is not insight. It is access to scale. When capability improves as a function of industrial inputs, the set of winning strategies narrows. There is less room for cleverness at the margin, because the frontier moves on a schedule determined by capital, chips, power, and research talent.
This compresses “alpha.” The obvious thing is more often correct than people want it to be. The best trade is frequently just exposure to the scaling complex, not a complicated thesis about the app layer.
One way to see this shift is in venture fund math. In the internet economy, a $500M fund could generate extraordinary carry by finding a handful of early 20× outcomes. In the AI economy, the largest companies may absorb enormous amounts of capital to build the scaling complex itself. If a firm can write $500M–$1B checks into frontier companies and those companies compound 2–3×, the carry math can look comparable to the classic venture model. In that regime, the advantage shifts toward funds large enough to access frontier companies rather than those optimized to discover early-stage outliers.
The shift is already visible in how frontier AI companies are capitalized. Several of the most ambitious new labs — such as Thinking Machines, Reflection AI, Flapping Airplanes, Applied Compute, and others — have raised unusually large seed or early rounds by historical venture standards. These rounds are often led by a small group of growth-oriented investors capable of writing very large checks. For much of the venture ecosystem, the companies that may ultimately define the frontier are simply inaccessible at the earliest stages.
Early-stage venture and angel investing still exists, but it is less structurally privileged than it was in the internet era, because most app-layer moats get vaporized by bundling and release cadence. Growth capital, public market exposure, and infrastructure ownership become relatively more important.
This is the cultural meta shift:
Old meta:
New meta:
The point is not the names. The point is what they represent: the archetype shifts from “think differently at the edge” to “align with scale and compound it.”
By hyper-legibility I don’t mean that everything becomes transparent in the abstract. I mean that the world becomes legible to insiders in a way that compounds their advantage. When the regime is dominated by scale, coordination, and access to the scaling complex, the “obvious” path becomes the winning path, and it is obvious primarily to the people already inside.
In the internet era, a contrarian outsider could plausibly outflank incumbents. Distribution was fragmented. Infrastructure was cheap. The map of opportunity was wide. In the AI era, the map narrows and the pipes matter. The people who sit closest to frontier labs, elite talent networks, top capital allocators, and hyperscaler infrastructure can see what is real, what is next, and what will be funded. They do not need to be contrarian because they have privileged alignment with the compounding function.
That is why “obvious things” compound more. Being an insider is not just helpful, it is multiplicative. Access to the right compute, the right data center operators, the right distribution channels, the right enterprise buyers, the right research talent, and the right capital partners turns a good idea into an inevitable one. The outsider can still build something, but they are running uphill against a regime that is increasingly path-dependent.
In this world, status and network are not superficial. They are infrastructure. A Stanford dropout can have a 1000× advantage over an outsider because they inherit proximity to talent, capital, and credibility. A growth investor born into an elite capital network can have a 1000× advantage because the regime rewards the ability to write large checks into the scaling complex and to place those checks where the compounding loops already exist. This is not “unfairness” as a moral claim. It is a description of how compounding works in concentrated regimes.
Hyper-legibility, then, is the death of romantic contrarianism. The edge shifts from “seeing what no one else sees” to “being positioned where the obvious becomes actionable.” The winners look less like lone geniuses and more like institutions.
If intelligence becomes industrial infrastructure, returns flow toward capital and bottlenecks. That creates a predictable distributional pattern: upside concentrates in the institutions that own the scaling complex and the scarce inputs that feed it.
In the oligopoly base case, wealth concentrates first in private markets. Equity is captured upstream and early. The labor market bifurcates: a small number of people become massively leveraged (researchers, operators, allocators), while broad categories of knowledge work become more substitutable. The median worker experiences AI as cost pressure, not as equity upside.
Even in the margin-compression equilibrium, the story doesn’t become egalitarian. If cognition becomes cheap, the surplus migrates to whatever remains scarce. Chips, power, facilities, land, grid access, distribution surfaces, and enterprise control points become the durable rent collectors. Scarcity does not disappear. It changes form.
This is why the AI economy tends to feel politically unstable: it industrializes cognition, concentrates early gains, and pushes wealth into bottlenecks that are hard to democratize quickly.
For the short to medium term, the optimal positioning continues to remain[23], however:
This is why the highest-fidelity model of the AI economy is not a stack diagram, but a dependency graph: who needs whom, which inputs are scarce, who can set prices, and where substitution is actually possible. The important question is not merely “which company has the best model?” It is “which actors control the bottlenecks that the best model must pass through?” Labs, hyperscalers, power owners, data-center developers, fabs, equipment suppliers, enterprise distribution owners, and physical-world operators all sit at different points in the graph. The surplus flows to the nodes that are both necessary and hard to substitute.
The app layer is structurally trapped. If frontier cognition remains scarce, the labs eventually treat large categories as first-class and compress “apps” into features or distribution skins. If cognition becomes abundant, it stops being the profit center and the ability to charge for “AI” collapses. In both cases, the standalone AI app whose value is “access to intelligence” struggles to sustain a durable moat. The only exception is when the app is secretly a path to (2): a wedge that becomes acquisition, distribution, or a choke point the labs decide they need.
The most interesting opportunities may therefore sit at the seams between the digital and the physical. Models and software cannot directly access the physical world without people, workers, scientists, laboratories, hospitals, oil wells, hardware, mining facilities, manufacturing shops, machines, permits, and process expertise. These assets operate on older scaling laws: capital, land, labor, regulatory permission, operational know-how, and the actual machines. As cognition becomes cheaper, these physical and institutional bottlenecks may become more valuable.
The game theory suggests that you should be long models and infrastructure:
| Scenario | Career | Capital if Long AI |
|---|---|---|
| AI takeoff | Labor collapses | Capital takes off |
| AI plateau | Labor stable | Capital neutral to slightly down |
It seems like AGI is real. For the longest time, we thought that the crazy people were those who were crazy about AGI, but now, in 2026, the people who did not believe in AGI were the crazy ones. There is not an AI bubble, but however, there is an “everything but AI” bubble.
The long term post-AGI future will look dramatically different than the present. We can speculate what that might look like, but they will remain speculation. We will not comment here because this essay is hard enough to write, and projecting the near term is challenging enough as it is.
Anything is possible, abundant intelligence, the full commoditization of software, a return to the world of atoms, political instability, new economic primitives to deal with the massive inequality generated by AGI, and more. We’ll live in a dramatically different world than the one we live in today. We've lived through some pretty crazy times: DEI, COVID, Trump 2016, Biden ZIRP Lockdown, Crypto, Trump 2024, and AGI has every indication of being even crazier[26].
Thank you to Moin Nadeem, Colin Flaherty, John Suh, Jon Goldsmith, Aashay Sanghvi, Kshitij Kulkarni, Ricky Moezinia, Ishani Thakur, Daniel Levine, Will Quist, Sami Senapathy, Paul Kwan, Erik Torenberg, Alex Danco, Erik Schluntz, Sharon Zhou, Andi Peng, Luke Qin, Malika Aubakirova, James da Costa, Caitlin Kalinowski, Rahul Mitra, and others for their feedback in creating this document.