Why AI Economics Still Favor Chips and Infrastructure

Usage is exploding. The harder question is where durable economic value will settle.

Generative AI looks like software on the surface: a user types a request, an application responds, and the product can reach hundreds of millions of people almost instantly. Economically, however, the industry still behaves much more like a capital-intensive infrastructure buildout than the classic software boom.

That is the central tension in a Stanford MS&E 435 lecture by Apoorv Agrawal, an investor at Altimeter Capital and the course instructor. His framing is deliberately provocative: the AI stack is an “inverted triangle.” In the mature cloud-software model, infrastructure supports a much larger application economy. In today’s generative-AI market, Agrawal argues, the strongest economics still sit lower in the stack—in chips and the infrastructure built around them.

The important point is not that applications do not matter. Their usage is growing rapidly. The point is that adoption and value capture are not the same thing.

Why the stack still looks upside down

Traditional software economics were powerful because the marginal cost of serving one more customer could approach zero. Once the code existed, distributing another copy was cheap. Generative AI changes that mechanism. Every additional query, generated image, coding session or agentic task consumes inference compute. Better models can become more efficient, and hardware costs can fall, but the cost of serving the next unit of usage does not disappear.

Agrawal uses that difference to explain why the current AI stack looks unlike the cloud-software stack. In his lecture, he estimates that roughly $350 billion of incremental revenue added across the generative-AI ecosystem over the prior two years was heavily concentrated in semiconductors, with about three quarters flowing to that layer. He also estimates that application-layer gross margins can be dramatically lower than classic SaaS margins.

Those figures should be treated as the speaker’s analytical model, not as audited industry statistics. The broader direction, however, is supported by public company results.

NVIDIA reported $96.2 billion in quarterly revenue in its latest reported quarter, including $89.0 billion from Data Center, while posting a 75.0% gross margin. That does not prove that “all AI profits” belong to chipmakers, but it does show how unusually strong the economics of the semiconductor layer remain while the rest of the industry is still building around it.

At the same time, the companies buying and operating that infrastructure continue to spend at extraordinary scale. Microsoft reported $41 billion in quarterly capital expenditures, with roughly two thirds tied to short-lived assets such as CPUs and GPUs. Its cloud gross margin fell to 65%, with the company explicitly citing continued AI-infrastructure investment and growing product usage as pressures, partly offset by efficiency gains. Alphabet has guided to $175–185 billion of capital expenditures for 2026, with the vast majority aimed at technical infrastructure.

This is the most important fact pattern behind the “inverted triangle” idea: demand is real, but serving that demand is still expensive.

Mass adoption does not solve the business model

The temptation is to assume that massive user growth will automatically repair the economics. It may not.

Google says the Gemini app has surpassed one billion monthly users. That is an extraordinary distribution milestone. But scale creates two effects at once: more opportunities to monetize and more inference load to serve. A product can become ubiquitous before its unit economics resemble those of mature software.

This is why Agrawal’s question—“Where’s the money in AI?”—is more useful than a simple ranking of which model is best. The durable winners may be determined by several variables that are only partly visible in headline user counts: inference efficiency, pricing power, distribution, enterprise bundling, advertising, vertical integration and the ability to move workloads onto custom silicon.

For application companies, the crucial test is not simply whether usage grows. It is whether revenue per unit of useful work can grow faster than the fully loaded cost of providing that work.

The strongest counterargument: today’s economics may be transitional

There is a major reason not to mistake the current structure for destiny.

Compute costs can fall. Model architectures can improve. Custom accelerators can reduce dependence on the most expensive general-purpose GPUs. Companies with their own chips, cloud platforms and consumer distribution can internalize costs that look punishing when viewed at a single layer. New business models can turn enormous free-user bases into profitable demand.

There is already evidence of movement in this direction. Alphabet has said engineering and hardware optimizations pushed the cost of AI Mode responses to their lowest level since launch. Microsoft likewise reports that efficiency gains are offsetting part of the margin pressure from AI infrastructure and product usage.

That does not invalidate the inverted-triangle thesis. It refines it.

The current AI economy is infrastructure-heavy because the industry is still paying to build the substrate on which applications run. The real uncertainty is how quickly efficiency, custom silicon and monetization can move economic power upward toward the application layer.

The next phase of the AI cycle

The first phase of generative AI was about capability: can these systems do useful work?

The second phase has been about scale: can companies make that capability available to hundreds of millions of people and enterprises?

The next phase is economic. It will ask whether the application layer can turn abundant usage into durable margins without simply passing ever-larger compute bills down the stack.

That is why the most consequential AI metric may not be benchmark performance or even user growth. It may be the relationship between the value created by one more unit of AI work and the cost of computing it.

When that relationship improves decisively, the triangle may finally start to flip.

Primary Sources

Editorial method: Apoorv Agrawal’s estimates of industry value capture are attributed speaker claims, not audited sector totals.

한국어로 읽기