AI Value Migration

The AI infrastructure layer captures 91% of value today. History — electricity, telecom, silicon — says this inverts to 20/80 within a decade. NVIDIA is in the exact position of Edison's General Electric, AT&T, and Intel at the peak of their respective dominance phases. The application layer winners are just beginning to emerge, and the enterprise that positions correctly for this migration will capture a disproportionate share of the most consequential value shift in a generation.

This is the chapter where four revolutions converge into a single thesis. In Chapter 18, we traced how Edison's vertically integrated electricity empire dissolved into commodity utilities while the appliance economy grew to $710 billion and the electricity-enabled economy reached $28.3 trillion. In Chapter 19, we documented how AT&T's network dominance eroded as over-the-top applications captured 5.7x the value of the carriers that built the pipes. In Chapter 20, we showed how the Wintel duopoly's equal partnership dissolved into a 21x gap as software captured all the growth while silicon commoditized. In every case, the pattern was identical: infrastructure dominates first, then commoditizes, then applications capture the value. The only variable was speed — and each revolution compressed the timeline further.

Now it is happening again. NVIDIA has grown from $144 billion to $4.3 trillion in seven years — a 30x increase driven by its 90% share of the AI chip market. Hyperscalers are spending $350 billion on AI infrastructure in 2025, with projections exceeding $600 billion by 2026. The infrastructure buildout is real and massive. But the first cracks in the 91/9 split are already visible. OpenAI's revenue has grown from $28 million to $20 billion ARR in four years. Anthropic has scaled from $10 million to $14 billion ARR in under four years. Cursor, an AI code editor that owns zero GPUs and zero models, went from $4 million to $1.2 billion ARR in twenty months. API pricing has collapsed 99.5% since GPT-4's launch. Open-source models have reached near-frontier quality. The signals are unmistakable: the infrastructure phase is peaking, and the application phase is beginning.

The forecast, grounded in the timelines of all three prior revolutions, points to a crossover around 2031 — the year when AI application spending matches AI infrastructure spending. After that, applications accelerate away, reaching an estimated 80% of total AI market value by 2035. This is not speculation. It is the pattern that has repeated in every technology revolution for 140 years, now compressed from decades to years because AI builds on all prior infrastructure simultaneously.

The Infrastructure Peak: We Have Been Here Before

The 91/9 split between AI infrastructure and AI applications in 2024 is not unprecedented. It is precisely the pattern. When Edison controlled generation, distribution, and the light bulb in the 1880s, infrastructure's share of the electricity economy exceeded 95%. When AT&T controlled the voice and data pipes in the late 1990s, infrastructure captured roughly 85% of telecom value. When Intel sat atop the PC value chain in the 1980s and 1990s, silicon held approximately 90% of the computing industry's market capitalization. In every case, the infrastructure share at peak dominance was between 85% and 95%. AI's 91% falls squarely within that historical band.

The more relevant question is what happened next. In every prior revolution, the infrastructure share fell to 20% or below at maturity. Utilities capture 1.8% of GDP today. Telecom carriers are worth one-fifth of the over-the-top applications they carry. Silicon companies (excluding NVIDIA's AI anomaly) are valued at one-quarter of software companies. The equilibrium is not 91/9. The equilibrium is 20/80. The current moment is the peak of infrastructure's relative value — the exact moment when the most value is still locked in the foundation layer, and the application layer is about to begin its exponential ascent.

The Model Layer: From Product to Platform

OpenAI and Anthropic occupy a unique position in the AI stack — they straddle the boundary between infrastructure and application. Their foundation models are infrastructure (commodity intelligence APIs that other companies build on), but their consumer and enterprise products (ChatGPT, Claude) are applications in their own right. This dual positioning is remarkably similar to Microsoft's in the early 2000s, when Windows was simultaneously the platform (infrastructure) and the product (application). The companies that own both the platform and the killer application on top of it tend to capture outsized value during transitions.

But the more profound signal is the speed. OpenAI reached $20 billion ARR in four years — the fastest enterprise software ramp in history. Anthropic is tracking even faster on a relative basis, growing 140x from $100 million to $14 billion ARR in just two years. These growth rates dwarf the scaling trajectories of Salesforce, Workday, or even Microsoft Azure. They suggest that the AI application layer is not merely growing — it is growing at rates that will compress the historical value migration timeline from decades to years.

The Application Layer Emerges

Cursor is the poster child for AI application-layer value capture, but it is far from alone. GitHub Copilot has reached 20 million users in three years, becoming the first AI coding tool to achieve mass adoption. Harvey AI has grown from $10 million to $195 million ARR in two years, demonstrating that vertical AI applications in domains like legal can scale rapidly. Perplexity AI is competing directly with Google Search. Glean is transforming enterprise knowledge retrieval. These companies share a common architecture: they own zero infrastructure, buy intelligence as a commodity input via model APIs, and build proprietary value through domain-specific data flywheels, fine-tuned workflows, and user experience design.

This is precisely the pattern that played out in every prior revolution. The early application companies in the electricity era (Westinghouse appliances, General Electric consumer products) succeeded not because they generated their own power, but because they built the best products on top of commodity electricity. The early internet application companies (Google, Amazon, Netflix) succeeded not because they laid their own fiber, but because they built the best experiences on top of commodity bandwidth. The AI application companies succeeding today are doing the same thing with commodity intelligence. They are the Netflix and Meta of AI — built on infrastructure they do not own, capturing value the infrastructure providers cannot.

The Infrastructure Paradox: Spending More, Capturing Less

There is a profound irony in the current moment. Hyperscaler capital expenditure is accelerating — from $146 billion in 2022 to a projected $600 billion in 2026, a 4x increase in four years. Seventy-five percent of this spending is directly tied to AI infrastructure: GPUs, data centers, networking equipment, and cooling systems. The infrastructure buildout is real, massive, and accelerating. And yet, this very spending is creating the conditions for infrastructure's relative value to decline.

The mechanism is identical to what happened in telecom. When carriers spent $500 billion laying fiber optic cable in the late 1990s and early 2000s, they were building the foundation for applications that would capture all the value growth. The infrastructure investment was essential — without the bandwidth, Netflix and Spotify and Meta could not exist. But the investment commoditized the infrastructure itself. Bandwidth became cheap, abundant, and interchangeable. The value migrated to the applications that rode on top of it. The same dynamic is playing out with AI. Every dollar spent on GPU clusters is a dollar that makes compute more abundant, drives API prices lower, and expands the addressable market for AI applications. Model API pricing has already fallen 99.5% in under two years. The intelligence cost curve is declining at 10x per year — faster than Moore's Law, faster than bandwidth cost declines, faster than any prior technology cost curve in history.

The Crossover Forecast: ~2031

The forecast model projects that AI application spending will match AI infrastructure spending around 2031 — five to seven years from now. Infrastructure spending is still accelerating — hyperscaler capex nearly doubled from $380B (2025) to ~$700B (2026) — which may delay the crossover relative to prior technology analogies. This estimate is grounded in the timelines of all three prior revolutions, adjusted for compression. Electricity took approximately 40 years from infrastructure peak to application dominance. Telecom took 20 years. Silicon took 15 years. Each revolution compressed the timeline by roughly half, because each built on all prior infrastructure. AI builds on electricity, telecom, and silicon simultaneously — suggesting a compression factor of 3x to 8x relative to the silicon revolution, which translates to a 5-to-10-year inversion window.

The specific projections: infrastructure spending grows from $135.8 billion in 2024 to approximately $500 billion by 2035 — a healthy 3.7x increase that reflects continued demand for compute capacity. But application spending grows from $13.3 billion to an estimated $1.55 trillion — a 116x increase that dwarfs the infrastructure growth. The crossover happens around 2029, after which applications accelerate away. By 2035, applications account for roughly 75-80% of total AI market value, and infrastructure stabilizes at 20-25% — the same equilibrium ratio observed in mature electricity, telecom, and software markets.

The Pattern Repeats: NVIDIA = Edison = AT&T = Intel

Strip away the specific technologies and the pattern is identical across 140 years. A revolutionary infrastructure technology emerges. One company dominates the infrastructure layer, capturing 85-95% of the new market's value. That company becomes the most valuable or among the most valuable enterprises in the world. Then the infrastructure commoditizes, applications proliferate, and value migrates permanently to the layer above. The infrastructure company does not disappear — but its relative share of value declines from 90% to 20% as the application layer grows 5x to 100x larger than the infrastructure layer ever did.

Edison Electric controlled generation, transmission, and the light bulb. It became General Electric, one of the most valuable companies in the world. Then electricity commoditized and value migrated to the appliance economy ($710 billion), the industrial economy, and ultimately $28.3 trillion in electricity-enabled GDP. Utilities today capture 1.8% of GDP. AT&T controlled the voice and data network. It was worth $270 billion at its peak. Then bandwidth commoditized and value migrated to Netflix, Spotify, Meta, and the over-the-top application layer that is now worth 5.7x the carriers that built the pipes. Intel controlled the x86 architecture and was worth $509 billion. Then silicon commoditized through ARM and custom chips, and value migrated to Microsoft ($2.9 trillion), the cloud platforms, and SaaS ($300 billion market). Today software companies are worth 4.2x silicon companies, excluding NVIDIA.

NVIDIA controls the CUDA ecosystem and 90% of AI chip sales. It is worth $4.3 trillion. The question is not whether the pattern repeats. It is whether enterprise leaders recognize the pattern in time to position for the application layer before the crossover.