Source note: this analysis is based on Abhay Venkatesh’s February 26, 2026 Google Doc, The AI Economy. The document is a scenario argument about frontier AI market structure, not a neutral market report, so the claims should be read as a framework rather than settled fact.

The technology industry has spent the last two decades operating under the economic laws of the internet. In that paradigm, compute was abundant, protocols were open, and distribution was essentially free. This environment heavily rewarded the scrappy founder. A small team in a garage could use zero marginal cost software to unseat entrenched incumbents, capturing value at the edges of the network. Because we just lived through this cycle, investors and operators naturally assume the artificial intelligence boom will follow the same pattern.

Abhay Venkatesh argues that this is a category error. His February 2026 report says the underlying primitives of the AI economy do not resemble the internet economy at all. Instead of a decentralized software ecosystem, AI may become a capital intensive, heavily centralized industrial complex. For investors and builders, applying internet era mental models to this new reality can lead to bad allocation decisions.

The Shift in Underlying Primitives

The core thesis of Venkatesh’s argument is that the building blocks of the future have changed. In the internet era, infrastructure faded into the background. Anyone could rent server space from Amazon Web Services and build an application on top of HTTP. Insight and product market fit were the dominant variables for success.

Frontier artificial intelligence reverses these conditions. Intelligence is no longer just code; it is a physical reality forged from tens of billions of dollars, cutting edge silicon, and gigawatt scale energy. These are not decentralized or democratized resources. They are scarce, physical, and highly cornerable.

When the primary driver of capability is industrial scale, the archetypal winner changes. The contrarian dropout who thinks differently is replaced by the institutional researcher, the growth capital allocator, and the infrastructure operator. In this regime, scale actually produces insight. The entities that can mobilize the most raw capital and physical infrastructure will inevitably compound their advantages.

The Illusion of the AI Application Layer

In the web and mobile eras, value accrued at the application layer because software solutions were domain specific. Building a successful short term rental marketplace like Airbnb did not give you the operational capability to build a payments processor like Stripe. Every category required its own unique product, distribution channel, and operating logic.

Frontier AI collapses these boundaries because the core product being sold is cognition itself. Venkatesh points out that improvements in a frontier model generalize laterally across multiple markets. A breakthrough in reasoning improves coding assistants, legal drafting tools, and customer support bots all at once.

This creates a structurally hostile environment for traditional venture scale software companies. If an “AI app” is merely a thin wrapper that packages frontier cognition into a specific user interface, its moat is fragile. The moment a market category proves lucrative, the frontier labs can treat that workflow as a first class feature. They can bundle the capability directly into their core product surfaces, weakening the independent application layer.

There is one major exception to this rule. Venkatesh distinguishes between companies that sell cognition and companies that use cognition to produce a traditional service. Consider an AI native law firm. This business is not selling “access to intelligence.” It is selling legal representation, liability transfer, and regulatory compliance. The customer does not care about the underlying model; they care about the legal outcome. For these companies, artificial intelligence acts as internal production infrastructure that reduces headcount and operating costs. While the traditional software wrapper gets crushed by the labs, the tech enabled service business can build a durable moat.

The Return of Physical Constraints

If the internet was defined by limitless digital expansion, the AI economy is defined by hard physical limits. The modern paradigm relies on empirical scaling laws, where performance improves predictably as developers throw more compute and data at larger models. Staying at the cutting edge requires a continuous, industrial scale program with recurring bills.

Venkatesh highlights the physical footprint required to maintain this trajectory. Training runs for next generation models are projected to demand between one and two gigawatts of power by the end of the decade. Data centers and AI infrastructure are expected to consume a growing double digit percentage of total United States electricity production.

This physical reality creates a hard moat. The barrier to entry is no longer just hiring smart engineers. It requires securing land, negotiating grid interconnects, sourcing specialized cooling equipment, and navigating local permitting environments. Because these limiting reagents are scarce, the number of credible actors capable of pushing the frontier is structurally capped.

However, this dynamic also forces AI labs to share their economic surplus with the physical world. The owners of power plants, data centers, and semiconductor fabrication facilities transform from passive suppliers into aggressive price makers.

Three Scenarios for Market Equilibrium

Given these constraints, how will the market ultimately settle? Venkatesh outlines three potential basins of attraction. These are speculative scenario frameworks, not guaranteed outcomes. The future will depend heavily on whether raw scale continues to yield capability breakthroughs and whether open source models can close the gap with proprietary systems.

1. The Frontier Oligopoly (The Base Case) In this scenario, pre-training scale remains the dominant driver of intelligence. Only a handful of well capitalized labs can afford the entry ticket. They successfully capture the default product surfaces across major enterprise workflows. Because they maintain a capability lead, they retain strong pricing power.

In this world, the independent app layer is almost entirely cannibalized into native lab features. The economic surplus concentrates cleanly upstream. The biggest winners are the lab equity holders, the semiconductor supply chain, and the operators of the physical energy and data center bottlenecks.

2. The Capability Plateau Here, frontier progress continues, but the marginal value of that progress declines for the average business buyer. “Good enough” intelligence saturates the market. While running compliant, enterprise grade models remains operationally difficult and capital intensive, buyers stop paying steep premiums for the absolute cutting edge.

In this equilibrium, model providers begin to look like modern cloud utilities. They are large, durable, and highly profitable, but their profit margins are bounded by procurement pressure. The economic action migrates slightly downstream to systems integrators and workflow incumbents. The companies that win are those that own the messy, complicated deployment of models into legacy corporate infrastructure, rather than those pushing the raw boundaries of cognition.

3. Margin Compression This is the commodity outcome. Credible open source alternatives and sovereign backed labs proliferate, expanding the supply of intelligence beyond a tight oligopoly. When the marginal buyer can easily substitute one model for another without a severe drop in performance, pricing discipline collapses.

Under margin compression, frontier labs start to exhibit the punishing economics of commercial airlines. They face immense capital expenditure requirements, intense price competition, and structurally thin margins. Intelligence itself prices like a basic infrastructure commodity. Consequently, virtually all the durable profits migrate to the unyielding physical bottlenecks: the chip designers, the power providers, and the entities controlling enterprise distribution channels.

Second Order Consequences: The New Meta

If we assume the base case of a frontier oligopoly holds true, the secondary effects on business culture and investing strategy are profound. Venkatesh details several shifts that operators must understand.

First, alpha may compress in the venture capital asset class. In the internet era, outsized returns came from non consensus insights. An investor could turn a tiny seed check into a fund returning phenomenon. In a regime dominated by industrial scaling laws, the “obvious” trade is usually the correct one. The winning strategy is simply having enough capital to buy exposure to the scaling complex. Advantage shifts toward growth funds and public market crossovers that can write billion dollar checks, leaving early stage venture capital structurally disadvantaged.

Second, the market becomes hyper legible to insiders. In a decentralized web environment, an outsider in a dorm room could outflank an incumbent. Today, the map of opportunity is narrow and dependent on access to critical infrastructure. Those who sit close to elite talent networks, hyperscaler capacity, and large pools of capital have a multiplicative advantage. They can see where the frontier is moving and possess the resources to act immediately.

Finally, this economic structure drives extreme inequality and scarcity. As intelligence becomes industrialized infrastructure, wealth concentrates violently into the hands of those who own the capital and the bottlenecks. Meanwhile, broad categories of traditional knowledge work become highly substitutable. For the median worker, the AI economy presents itself primarily as relentless cost pressure rather than equity upside.

A Speculative Framework, Not Settled Fact

Operators and investors should treat this framework with the appropriate caveats. Venkatesh’s report is an argument about the physics of the current market, not a crystal ball.

The thesis relies heavily on the assumption that pre-training scale will remain the highest order bit for artificial intelligence progress. If the industry shifts toward a paradigm where post training, localized data flywheels, and workflow specific reinforcement learning matter more than raw compute, the center of gravity could shift back toward the application layer. Similarly, unforeseen algorithmic breakthroughs could suddenly decouple model capability from extreme energy usage, effectively democratizing the frontier once again.

Final Takeaways and Optimal Positioning

If the industrialization of cognition thesis holds true, the standard Silicon Valley playbook is broken. Building a thin software wrapper around a proprietary application programming interface is a dangerous game.

Instead, Venkatesh suggests that optimal positioning requires moving toward the extremes of the value chain. Investors should target equity in the frontier labs themselves, the physical inputs required to sustain them (silicon, land, and energy), or the underlying materials that feed those inputs. For builders, the most viable path is creating tech enabled service businesses that use abundant cognition to aggressively replace human labor, or building wedge products specifically designed to be acquired by the labs as distribution channels.

The AI economy is being built with concrete, copper, and specialized silicon. Recognizing that reality is the first step toward surviving it.

Source: Venkatesh, Abhay. “The AI Economy is Not the Internet Economy.” February 26, 2026. Google Docs.