Every layer in the AI stack is necessary. That is the point of the framing.
But investors have often paid dearly for confusing a necessary industry with a good business.
Ports are necessary. Airlines are necessary. DRAM is necessary. Necessity alone does not guarantee pricing power or high returns on capital.
Jensen's framing raises a question it does not answer: which layers turn necessity into durable economics.
That is the investor's job.
Jensen's five layers are not equal in scale, complexity, or economic character. Each sits on the one below it. Each has a different relationship to capital, pricing power, and competition.
Energy is the base layer. Without power, nothing runs.
AI demand from data centres is real. Power availability, grid access, and cooling capacity are becoming strategic constraints. That makes energy infrastructure more important.
But importance is not the same as ownership quality.
Many energy businesses are regulated, capital-intensive, or dependent on long reinvestment cycles. They may benefit from higher AI demand. But the value created by that demand does not automatically accrue to shareholders.
The key question is not whether AI needs more power. It is who owns scarce power access, who earns regulated returns, and who bears the capital burden.
Chips are the most visible layer. They sit closest to the current bottleneck.
But in NVIDIA's case, the moat is not only the GPU. It is the system around the GPU.
CUDA, developer workflows, software libraries, and ecosystem familiarity create switching costs.3 Leaving NVIDIA is not just a hardware procurement decision. It means re-education, re-tooling, and re-integration across a highly specialised workflow.
That does not remove cycle risk. Semiconductor demand can overshoot. Custom silicon can improve. Margins can compress. But it explains why this layer has different economics from a standard hardware cycle.
The strongest chip businesses do not merely sell components. They become part of the operating system of an industry.
Infrastructure is where the capital bill becomes obvious.
Cloud platforms, hyperscalers, data centres, and networking systems all scale with AI demand. Revenue can grow quickly. But so can depreciation, maintenance, and reinvestment needs.
This is the structural risk of infrastructure businesses. They look attractive during build-out because demand is visible and capacity is scarce. As capacity expands, the economics can become more competitive.
The investor's question is whether infrastructure providers have durable differentiation. Do they control scarce sites? Do they have superior power access? Can they earn attractive returns after depreciation and reinvestment? Or are they financing the physical layer of someone else's economics?
Infrastructure may be essential. It may also be where capital goes to work without capturing a proportionate share of value.
Models are the intellectual property layer. This is where much of the public imagination sits.
Building a frontier model is expensive. Estimates put GPT-4's compute cost at roughly $78 million and Gemini Ultra at $191 million.4 Open-source pressure keeps pricing lower than the cost structure would prefer. Meta released Llama openly despite comparable training expenditure. DeepSeek V3 was reportedly trained for about $5.6 million in compute cost, though that figure should not be mistaken for the full all-in cost of building and operating a model company.5 Even so, it challenged the assumption that frontier capability must always require frontier spending.
The leading labs have a head start. Scale, data, enterprise trust, and product integration matter. Whether that head start compounds or narrows is still playing out.
The business question is not whether models are capable. It is whether capability converts into durable pricing power. That answer is not yet settled.
Applications are where AI eventually meets the customer.
Copilots, enterprise tools, healthcare workflows, coding assistants, industrial systems. The list is long and the productivity gains are real.
But this is the least useful layer to analyse thematically.
Some incumbents will embed AI into existing workflows and strengthen their positions. Some challengers will build better products from scratch. Some applications will delight users but fail to monetise. Others will save customers time but face low willingness to pay.
The application layer is where AI value becomes company-specific. The question is not whether AI applications will exist. It is who owns distribution, workflow, data, and customer habit. Those are the things that determine whether usage becomes revenue and whether revenue becomes durable profit.
Jensen's five layers explain what AI requires. They do not rank which layers are worth owning.
For investors, the stack is a starting point for harder questions.
Who has pricing power?
Who is a pass-through?
Where do the switching costs live?
Who bears the capital intensity?
Which businesses can keep what they earn?
Business importance is not investment quality. Thematic exposure is not owner economics. Growth in a sector does not mean value accrues equally to every participant.
The AI build-out is real. Jensen's framing makes that clearer. But the stack is not a thesis by itself. It is a map of dependencies, not a map of value capture.
For long-term owners, the work begins after the map is drawn.
The goal is to separate businesses that merely participate in AI from those that can retain economics because of it.
Not every necessary layer captures value. Not every beneficiary is worth owning.
Disclosure: Glavcot Insights and its contributors may hold positions in securities discussed in this article. All content is provided for informational and educational purposes only. This is not investment advice. Readers should perform independent research and consult qualified financial professionals before making investment decisions.
Glavcot Insights is now live. Join the free tier to get new research, updates, and future analysis directly in your inbox. (Check your spam folder if you don't see the confirmation email)
