What compute, power and past technological revolutions can teach us about where AI’s economic value may ultimately go.
In my previous article, I looked at 150 years of technological change to understand what happens when a transformative technology becomes cheaper and more widely available.
Railways, electrification and the internet suggested a recurring pattern: scarcity attracts capital, capacity expands, costs fall, and some of the value created eventually moves elsewhere. This builds on my earlier analysis of the physical bottlenecks behind AI infrastructure.
I wanted to test whether AI might follow a similar path. For AI investing, this raises a fundamental question: which types of businesses are likely to benefit most?
The more I researched it, the less useful that question became.
A business can use AI to become more productive, improve its products or grow faster, while part of that benefit is passed on to customers, competed away, or absorbed by the capital required to deliver it.
For shareholders, the distinction matters: value created by AI and value ultimately retained by the business are not necessarily the same thing.
A better question emerged:
What determines who can retain the economic value AI creates?
Creating value is not the same as retaining it
Consider a business that uses AI to perform the same work faster. It may retain the productivity gain through higher margins. But competition could also push prices lower, customers could capture part of the saving, or additional spending on AI and cloud infrastructure could absorb some of the benefit.
The technology creates value in each case. Who retains that value is a different question.
RELX illustrates one possible mechanism. Its businesses combine information, analytics and professional workflows. In the first half of 2026, it reported underlying revenue growth of 7%, adjusted operating profit growth of 9% and adjusted EPS growth of 11% at constant currencies. Management also describes AI as helping it develop higher-value products faster.
RELX — First Half 2026 Results
These figures do not establish how much of RELX’s performance was caused by AI. They illustrate a broader idea: AI may be particularly valuable when it increases the usefulness of assets and capabilities a company already possesses.
Microsoft illustrates a different model. Its cloud and software businesses provide multiple ways to monetise AI, but delivering that capability requires enormous infrastructure investment. In its fiscal 2026 fourth quarter, Microsoft reported $41 billion of capital expenditure, with roughly two-thirds allocated to short-lived assets, primarily CPUs and GPUs.
Microsoft — FY2026 Q4 Earnings
Both businesses can benefit from AI, but the capital required to capture that benefit is very different.
This leads to another question: what exactly is the economic asset being built?
In AI infrastructure, the economic asset is the configuration
I initially assumed that AI hardware would depreciate economically very quickly as newer GPUs became more powerful and efficient.
But a GPU alone is not the economic asset. Customers need usable compute.
The transition between generations makes this easier to see. NVIDIA’s GB200 NVL72 is a liquid-cooled rack-scale system, while NVIDIA compares it with H100 air-cooled infrastructure. Moving between generations can therefore involve much more than replacing one chip with another. Servers, networking, electrical distribution, cooling and other parts of the data-centre infrastructure may also need to change.
An older configuration that is already installed, powered, cooled and connected can therefore retain economic value even after technically superior hardware becomes available.
The relevant question is not simply how quickly a GPU becomes technologically obsolete. It is:
How long can the complete usable configuration continue to generate sufficient economic value?
There is an interesting parallel with Bitcoin mining. Older mining hardware does not automatically become economically worthless when a more efficient generation appears. Research from the Cambridge Centre for Alternative Finance shows that the economic lifespan of ASIC mining hardware depends partly on the rate of technological progress and whether older equipment can remain profitable.
Cambridge — Bitcoin mining hardware and economic life
AI compute is more versatile, but the underlying principle is similar:
Technological obsolescence and economic obsolescence are not the same thing.
An older GPU may no longer be optimal for the most demanding AI workloads, but it can potentially migrate towards less demanding uses if its price and operating costs remain competitive.
Amazon and Meta show how uncertain infrastructure lifespan can be
Amazon and Meta provide an interesting real-world illustration through their accounting estimates of useful life — the period over which they expect their infrastructure to remain in use.
Effective January 2024, Amazon increased the estimated useful life of its servers from five to six years. The change reduced 2024 depreciation and amortisation expense by approximately $3.2 billion and increased net income by approximately $2.5 billion.
A year later, Amazon moved partly in the opposite direction. It shortened the estimated useful life of a subset of servers and networking equipment from six years to five, explicitly citing the increased pace of technological development, particularly in AI and machine learning. Amazon had also decided to retire certain equipment early, resulting in approximately $920 million of accelerated depreciation and related charges in Q4 2024.
At almost the same time, Meta moved in the other direction. In January 2025, it increased the estimated useful lives of most servers and network assets to 5.5 years. Meta subsequently reported that this reduced 2025 depreciation expense by $2.92 billion and increased net income by $2.59 billion.
There is an important qualification: Meta did not say that AI caused this extension. We therefore cannot conclude that AI shortened infrastructure life at Amazon while extending it at Meta.
The narrower conclusion is more useful:
There is no single lifespan that can simply be assigned to AI infrastructure. How long an asset remains useful depends partly on how each operator expects to use it.
Accounting useful life is also not the same as economic life. An asset can remain operational while becoming less profitable, or remain economically useful after newer technology has appeared.
If the economic life of an AI compute configuration is uncertain, the speed at which its capital can be recovered becomes particularly important.
Why payback matters for AI infrastructure
Nebius provides a useful example.
For new deals signed during Q2 2026, the company reported an estimated payback period of one year and ten months. Importantly, this is a forward-looking management estimate based on forecast costs and contracted future capacity, including capacity not yet built. It is not a realised historical payback.
Nebius — Q2 2026 Shareholder Letter
Whether those economics persist as capacity expands remains uncertain. Compute prices, utilisation, technology and financing can all change.
However, the underlying principle is important:
Scarcity does not need to be permanent. It needs to persist long enough for the capital committed to earn an adequate return.
A configuration that recovers its capital quickly is less dependent on assumptions about what AI hardware will be worth several years later.
The same principle applies beyond compute.
Power constraints operate on a different clock
A data centre does not simply need electricity generation. It needs sufficient power at the required place and time, together with the grid connection and infrastructure necessary to deliver it.
This creates a useful parallel:
A GPU is only one part of usable compute, just as generation is only one part of deliverable power.
Constellation Energy’s agreement with Meta provides an example. It covers 1,121 MW from the Clinton Clean Energy Center for 20 years beginning in June 2027 and supports a 30 MW increase in the plant’s output.
Constellation Energy — Meta 20-year nuclear agreement
These agreements do not prove that power scarcity will last for decades. They show something narrower: some customers are willing to make very long commitments around power supply.
Other constraints can respond faster. Schneider Electric, for example, reported first-half 2026 revenue of €21.2 billion, up 14% organically, and adjusted EBITA of €4.1 billion, up 22% organically, amid strong demand across its markets led by data centres.
Schneider Electric — H1 2026 Results
But electrical equipment is manufactured capacity: investment can expand supply and competitors can respond.
The important question is therefore not simply whether something is scarce today, but:
How quickly can additional capital remove that scarcity?
Compute, electrical equipment, grid connections and power generation operate on different supply-response times. Understanding these different clocks may be as important as identifying the bottlenecks themselves.
A framework for analysing AI economics
This research left me with five questions.
1. What economic value does AI create?
Does it reduce costs, increase revenue, improve a product or remove an important constraint?
2. Who can retain that value?
Can the business keep the benefit, or will customers and competitors capture much of it?
3. How much capital is required?
Is AI increasing the usefulness of assets already owned, or does capturing the opportunity require continuous investment?
4. How durable are the economics?
How quickly can technology, competition or additional capacity weaken today’s advantage?
5. How quickly can the capital be recovered?
When economic life is uncertain, payback becomes particularly important.
Together, these questions provide a framework for testing whether AI-driven economic value can persist.
From AI value creation to shareholder value
A company can benefit enormously from AI without all of that benefit ultimately accruing to its shareholders.
AI may increase productivity, reduce costs or improve products, but part of the value created can be passed on to customers through lower prices, competed away, or absorbed by the capital and financing required to deliver it.
Conversely, a company does not need to develop the most advanced AI technology itself to benefit. AI can increase the value of information, workflows or physical assets it already controls, sometimes without requiring the same level of additional capital.
Infrastructure adds another dimension. Scarcity can create attractive economics, but capital will attempt to remove it. What matters is therefore not only how valuable a bottleneck is today, but how quickly supply can respond relative to how quickly the invested capital can be recovered.
This brings the analysis back to shareholders. The businesses that benefit most from AI may not be those that retain the largest share of the economic value it creates. What ultimately matters is how much of that value remains after competition, additional investment and financing costs.
So rather than simply asking which businesses benefit most from AI, I would ask:
What value does AI create, who can retain it, how much capital is required, how durable are the economics — and can the capital be recovered before those economics change?
Together, these questions help explain whether AI can produce a durable improvement in business economics — and whether enough of that improvement can ultimately accrue to shareholders.
Disclosure & disclaimer
This article is for general informational and educational purposes and reflects my personal research and opinions. Companies are discussed as case studies to illustrate different business economics, not as recommendations to buy, sell or hold any security. Company guidance and forward-looking estimates are inherently uncertain. Investing involves risk, including loss of capital.
Disclosure: At the time of publication, I hold shares in Nebius Group and Schneider Electric.
Laisser un commentaire