Who Captures Value in the AI Value Chain?

The most visible product in the AI race may not be where the economic value accumulates. Value can move toward bottlenecks that are difficult to replace, slow to expand and reusable across customers. This guide follows the chain from chips and advanced packaging to lithography, electricity and the cloud.

The short answer

The strongest value-capture signal in AI is not concentrated in a single end product. It appears across layers where several difficult inputs meet: accelerators for advanced computing, the networks that connect them, advanced manufacturing and packaging, lithography tools, electricity connections and data-centre land and buildings.

Scarcity alone is not a durable advantage. Four conditions matter together:

The input must be difficult to obtain or slow to replicate.
The same capacity must be sellable to more than one customer or usable across several technology generations.
Demand must show up in revenue, usage, contracts or orders, not only in management commentary.
The capital, power, labour and financing required to expand capacity must not consume the surplus created by the bottleneck.

The compact conclusion is this: AI value is flowing most strongly toward the layer controlling the binding bottleneck, provided that layer can sell its control across customers and preserve a surplus after expansion costs. When the bottleneck is relieved, value moves to the next scarce input.

Why the question matters

AI is not a single product. A user request, a model's computation, the accelerator and memory carrying that computation, the factory producing the chip, the equipment used in production, the building processing the data, the electricity powering it and the contracts financing the capacity are all parts of the same chain.

Those layers have different economics. An application can grow quickly while its pricing power remains weak if competitors can copy the feature. A less visible manufacturing tool can set the pace of the chain if replacing it takes years. A data centre can have a long-term contract and still fail to operate if its power connection is delayed.

That is why “who is the AI market leader?” is incomplete. The more useful question is: Which input is necessary for the rest of the chain to operate, who controls it and how long can that control last?

How the chain works

The simplified mechanism is:

use case -> model training and inference -> compute demand -> accelerators, processors, memory and networking -> advanced packaging and manufacturing -> lithography and production equipment -> land, buildings, electricity and grid connection -> cloud capacity and software distribution -> customer revenue or productivity

Five questions help assess each layer.

1. Where is scarcity?

What is required to expand the product or service? A software feature may be copied within months. An advanced packaging plant, high-voltage connection or leading-edge lithography tool takes much longer to add.

2. Who controls it?

Owning a scarce input and being able to use it are not the same. A company may have strong demand while depending on another company, a single customer or one power connection for production.

3. Is the revenue signal real demand?

Revenue, orders, long-term contracts, usage and renewals are different forms of evidence. A reservation or contract indicates that capacity may be sold in the future. It does not by itself show that the customer is earning a return from that capacity.

4. How heavy is the capital burden?

A high gross margin does not prove that the same economic surplus remains after the investment and financing required to build capacity. Grid connections, leases, energy, depreciation, maintenance and interest must be read together.

5. Where does value move when the bottleneck is relieved?

New manufacturing capacity can reduce accelerator scarcity. Packaging, memory, electricity or cloud software may then become more binding. The value chain is better understood as a moving map of constraints than as a fixed ranking.

Compute and platforms: NVIDIA

NVIDIA's May 20, 2026 fiscal first-quarter results provide a strong company signal for the weight of AI computing demand in the supply chain. The company reported $81.6 billion of total revenue, $75.2 billion of Data Center revenue and a company-wide GAAP gross margin of 74.9%. Data Center revenue was about 92.1% of total revenue.

The calculation is:

$75.2 billion / $81.615 billion = 92.1%

NVIDIA also reported $60.4 billion of Data Center compute revenue and $14.8 billion of networking revenue. The mix shows that demand is not only for one chip. Accelerators, networking and software are sold as a broader system.

There is an important boundary. NVIDIA's Data Center revenue is not an AI-only revenue line. It covers products sold to different customers and for different workloads. The 92.1% figure therefore does not mean that 92.1% of NVIDIA revenue comes from AI. It shows how dominant the Data Center pool is within the company's reported revenue.

NVIDIA's August 17, 2026 PORTS-Pike announcement shows how the bottleneck can move beyond the chip. The company says an AI factory requires not only chips, memory and networking but also advanced packaging, land, power and a completed shell. Buying the fastest processor is not enough if the power and building to run it are unavailable.

The support mechanism in that announcement also needs a narrow reading. NVIDIA describes its support as covering defined portions of lease and power payments plus a residual-value commitment. That does not mean it is taking on the full cost of the site or every tenant obligation. PORTS-Pike is therefore best used as a company case showing how capital and infrastructure constraints can be managed, not as proof of NVIDIA's future revenue opportunity.

NVIDIA's value may come from two different sources: selling scarce compute capacity and offering that capacity as part of a wider system. The first advantage can weaken as competitors add supply. The second may last longer if customers face high switching costs across chips, networking and software. Customer concentration, alternative architectures, new capacity and product cycles will determine which advantage matters more.

Manufacturing and advanced packaging: TSMC

TSMC shows how computing demand becomes scarcity inside a fab and an advanced-packaging plant. The company reported $40.2 billion of revenue, a 67.7% gross margin and a 60.3% operating margin in the second quarter of 2026.

TSMC said that technologies at 7 nanometres and below accounted for 77% of wafer revenue. High-performance computing represented 66% of net revenue. Both figures matter, but neither is AI-specific. High-performance computing covers workloads beyond AI, so calling the 66% figure AI revenue would be wrong.

In its second-quarter discussion, TSMC management said advanced-packaging capacity was so tight that it was limiting customers' growth. The company also said it was adding capacity. Read together, the two statements show a current revenue opportunity that can be reduced over time by capacity investment.

The value-capture test for TSMC is therefore not only its gross margin. The harder question is whether customers' need for leading-edge manufacturing and packaging is growing faster than TSMC can invest, yield and expand. If so, scarcity can support pricing power. If new plants and alternative suppliers add capacity quickly, revenue can grow while some of the surplus returns to investment costs.

Production equipment: ASML

ASML represents a layer far from the end user but difficult to replace. The company reported €9.326 billion of net sales, a 54.0% gross margin and €2.918 billion of net income in the second quarter of 2026. Installed-base management sales were €2.762 billion.

Installed-base management sales were about 29.6% of net sales:

€2.762 billion / €9.326 billion = 29.6%

The result shows that the equipment layer includes an ongoing service business alongside new machine sales. Recurring service revenue can increase the economic value of the installed base and the depth of the customer relationship. It is not AI-only revenue, however, and it remains tied to customer capital spending.

ASML says extreme ultraviolet lithography systems are unique to its product portfolio. That is a strong company signal about technical difficulty, not an independent competition study. The company also says AI investment is supporting demand for advanced logic and memory chips.

The counterevidence is just as important. ASML said it planned to add about 30% to low-NA EUV capacity for 2027 and about 30% to immersion DUV capacity. Those are company plans, not completed capacity. They show the mechanism clearly: scarce equipment can create value, but high value also encourages new capacity. For scarcity to last, technical advantage must be reinforced by production, service and customer switching costs.

Cloud and distribution: Microsoft

Tracking physical bottlenecks does not mean that software and cloud layers cannot capture value. In many cases, the cloud is where customers place AI inside daily workflows and pay for the capacity.

Microsoft reported $59.3 billion of Microsoft Cloud revenue in its fiscal fourth quarter of 2026, up 27%. Azure and other cloud services grew 43%, while commercial remaining performance obligations reached $678 billion. Microsoft also reported more than 30 million paid Microsoft 365 Copilot seats.

These indicators show that AI and cloud demand is not only a hardware-supplier story. Contracted work, cloud growth and customer distribution can create another value pool above the physical layers. None of these figures, however, shows AI-only profit. Cloud revenue includes many services, remaining performance obligations do not reveal the timing or margin of future revenue and paid seats do not prove usage intensity or customer productivity.

Microsoft supports a more balanced conclusion: Distribution and customer relationships can create a second value-capture layer above physical infrastructure, but durability must be tested through usage, renewal and pricing.

Electricity, land, grids and financing

The least visible layer of the AI chain may be physical power infrastructure. The International Energy Agency reported that global data-centre electricity demand grew 17% in 2025, while AI-focused data-centre demand grew 50%. It expects total data-centre electricity consumption to rise from 485 terawatt-hours to 950 terawatt-hours by 2030. The implied compound annual growth rate between those endpoints is about 14.4%.

That is not a linear law of demand. The same IEA report says energy use per simple AI task has fallen by at least an order of magnitude annually in recent years. Video generation, reasoning and agentic tasks can use hundreds or thousands of times more energy than simple text generation. Efficiency and usage intensity can therefore offset or exceed one another.

The IEA identifies bottlenecks across electricity supply chains, grid connections, advanced chip manufacturing, high-bandwidth memory and capital. It also says that an advanced AI server rack could have peak power demand equivalent to 65 households by 2027 and that AI-server power density rose 11 times from 2020 to 2025. The economic question is not only how much electricity is consumed but how much power can be connected to the right place at the right time.

Financing adds a third constraint. The Bank for International Settlements says hyperscaler bond issuance topped $100 billion in 2025 and describes leases, capacity offtake agreements, guarantees and off-balance-sheet vehicles as economically debt-like burdens. These structures can speed construction. If demand is slower than expected, they can also distribute risk to banks, private-credit funds, insurers, suppliers or customers.

Data-centre capacity should therefore not be read only as “how many gigawatts are being built?” The more important questions are:

Is the power connection secured or still at the application stage?
Which party carries the land, building and lease obligations?
What happens to rent, guarantees and capacity commitments if demand falls?
Can the capacity be sold to many customers or only one?

The strongest counterargument: applications capture the value

The strongest counterargument is that hardware and electricity scarcity may be temporary. Over the long run, applications and distribution channels that embed AI into workflows may capture the economic surplus.

There are three reasons to take that argument seriously. First, falling model and compute costs can make the same capability available to more software companies. Second, customers pay for faster sales, lower costs or better decisions, not for a chip itself. Third, an application company that controls the customer relationship and a specialised workflow may retain the surplus even if infrastructure providers lose pricing power.

The application layer still has to pass the scarcity test. If a feature is easy to copy, switching costs are low and usage does not convert into pricing, a large user base may not become a large surplus. Microsoft's paid Copilot seats are therefore a meaningful adoption signal. They are not enough to establish usage per seat, renewal or margin.

A July 2026 BIS model study describes a scenario in which the investment race could run about 50% above the socially efficient level. This is not an observed sector statistic or a definitive forecast. It is a useful counterweight: not every bottleneck investment creates economic value. Companies may overpay for capacity to secure future market share, while specialised assets can be difficult to resell if demand slows.

Applications are not automatic winners, and physical infrastructure is not an automatic loser. Value accumulates where customers cannot easily do without the function and competitors cannot easily copy it. Today that function may be an accelerator. Tomorrow it may be a power connection or a software workflow.

How to read the layers

Layer	Available signal	What it supports	What it does not prove
Compute and networking	NVIDIA Data Center revenue is about 92.1% of reported total revenue	Strong demand and broad system sales in compute infrastructure	That all revenue is AI revenue or that the advantage is durable
Manufacturing and packaging	TSMC reports 77% of wafer revenue from 7nm-and-below technologies and 66% from high-performance computing; management reports tight advanced packaging	Leading-edge capacity and packaging can become binding constraints	That high-performance computing equals AI or that the shortage will persist
Lithography and services	ASML installed-base management sales are about 29.6% of net sales	Technical difficulty and recurring service revenue can coexist	AI-only profit or an independently proven economic monopoly
Cloud and distribution	Microsoft Cloud revenue of $59.3 billion, 43% Azure growth and $678 billion of commercial remaining performance obligations	Demand, contract visibility and distribution can create value	AI-only margin, usage intensity or customer return on investment
Electricity and financing	IEA's 485 TWh to 950 TWh projection and BIS's $100 billion-plus bond and debt-like-structure evidence	The bottleneck can move into physical and financial layers	Which company will capture the most profit there

This is not a company ranking. Each row has a different measurement boundary. A value-capture claim becomes useful only when scarcity, control, demand evidence, capital burden and substitution risk are considered together.

What would change the conclusion?

The conclusion would change if:

New accelerators, custom chips or more efficient models quickly reduce compute scarcity.
Advanced packaging, memory, manufacturing and lithography capacity grow faster than demand.
Delays in permits, transformers or grid connections last longer than expected.
Application companies begin to disclose AI-specific revenue, usage, renewal and margin separately.
Cloud contracts and capacity reservations fail to convert into actual usage and cash generation.
Debt, leases, guarantees and capacity commitments create a larger economic burden than the visible balance sheet suggests.
Export controls, energy regulation or new safety rules change the substitution map across the supply chain.

What to watch next

For the next earnings release or infrastructure announcement, track these signals together:

How Data Center, compute, networking, cloud and application revenue are defined.
Whether orders, capacity reservations and commercial contracts become actual usage and renewals.
Whether lead times for advanced packaging, high-bandwidth memory, lithography and power connections are lengthening.
When new capacity is expected to come online and which party carries the investment cost.
Whether gross margin, service revenue or contract visibility survives energy, depreciation, lease and financing costs.
Whether dependence on one customer or a small group of customers is increasing.
Whether AI-specific revenue and profit disclosures replace company-wide or platform-wide proxies.

Conclusion

Value in the AI chain does not automatically go to the company making the largest investment or sitting closest to the end user. Current evidence shows different forms of scarcity in computing, advanced manufacturing, packaging, lithography, electricity and cloud distribution.

NVIDIA offers a signal of strong demand for compute and networking. TSMC shows the importance of leading-edge production and packaging. ASML combines technically difficult equipment with recurring service revenue. Microsoft shows how cloud growth and contracts can create a value layer closer to the customer. The IEA and BIS make clear that the chain cannot operate without power, grids and financing.

None of these signals identifies a permanent winner by itself. The best reading is this: Scarcity creates revenue. Durable value requires the controlling layer to preserve its irreplaceability, customer breadth and post-capital surplus while it expands capacity. When the bottleneck is relieved, value moves to the next bottleneck.

Methodology and limitations

The research uses a data cutoff of August 20, 2026. The company cases are NVIDIA's fiscal 2027 first-quarter results, TSMC and ASML's second-quarter 2026 disclosures and Microsoft's fiscal 2026 fourth-quarter results. The system-level evidence comes from the IEA's April 2026 report and BIS studies published in March and July 2026.

Calculations were reproduced from extracted-value records stored in the source snapshots. NVIDIA Data Center revenue is 92.1% of total revenue, ASML installed-base management sales are 29.6% of net sales and the endpoint-implied annual growth rate for the IEA's 2025-2030 total data-centre electricity projection is 14.4%. These are descriptive calculations, not valuation, share-return or economic-rent rankings.

Company-reported Data Center, high-performance computing, cloud, Copilot and installed-base services figures are not presented as AI-only revenue or profit. TSMC's high-performance-computing category is broader than AI. ASML's lithography uniqueness is a company product description, not an independent competition review. NVIDIA's PORTS-Pike announcement is one company case and includes forward-looking framing. IEA projections are estimates, and BIS's 50% result is a model outcome.

Raw HTML, PDF or filing bytes were not added to this package. Instead, the package records direct official links, source dates, retrieval time, extracted values and canonical hashes of the extracted payloads. This makes the evidence path and its limitations visible without claiming that the raw-archive gap has been removed.

Sources and further reading

Not investment advice; for research and educational purposes.