Who Will Capture the Value of AI Infrastructure?

How Constraints, Capital and Risk Are Reshaping Infrastructure Economics

By Nune Igityan, Founder, Kensai · Published 14 September 2026 · Last reviewed 14 September 2026

In Brief

AI infrastructure is growing faster than the systems that support it. Chips, memory, packaging, power and grid capacity are all constrained. As investment arrives, those constraints move, and so do the conditions for capturing value.

• Usable capacity depends on a chain of inputs. Whatever is scarcest limits what the rest can deliver, and more than one input can be scarce at the same time.
• In this cycle, the bottleneck is moving through the chain. High-bandwidth memory and advanced packaging have become important constraints, while power and grid access are tightening further downstream.
• Whether profits move with the bottleneck is unproven. High margins where capacity is scarce point to bargaining power. They do not yet show a lasting economic rent.
• Contracts and financing shape who gains and who absorbs the losses. That is often not the firm that owns the assets.

1. The Scale of the Buildout

AI infrastructure investment is expanding at a pace with few historical parallels. PwC projects $31.6 trillion of cumulative data-centre capital expenditure through 2050, while the International Energy Agency (IEA) reports capital expenditure by the five largest technology companies exceeding $400 billion in 2025. The Bank of England describes the pace as historically unprecedented and notes the sector’s growing reliance on external financing. Recurring chip and equipment refresh cycles make this different from a one-off construction boom. (PwC, 2026; IEA, 2026; Bank of England, 2026)

The IEA identifies a mismatch between the speed at which AI is developing and the speed at which its physical and economic support systems can respond. Investment can relieve one constraint while capacity elsewhere remains difficult to add. Grid-connection waits can run five to ten years in many jurisdictions, making power capacity slower to relieve than some upstream constraints. (IEA, 2026) Where output is limited by what cannot be built quickly, competitive position comes to depend on securing constrained resources.

2. A System of Complementary Constraints

An AI system depends on a chain of inputs that are complementary rather than substitutable:

• advanced logic chips;
• high-bandwidth memory and advanced packaging;
• networking and interconnect;
• data-centre shells and cooling;
• electricity supply and grid connection.

Each is necessary and none is sufficient alone. A shortage at any point can limit utilisation across the system, and utilisation affects the economics. Several constraints can bind at the same time.

The IEA identifies limitations across electricity, grids, manufacturing capacity, chips and capital, while J.P. Morgan identifies power availability, supply chains and permitting as execution risks that can move project timing and cost materially. (IEA, 2026; J.P. Morgan, 2026)

Carliss Baldwin’s work on modularity explains how the architecture of a complex production system can create economically significant boundaries and transaction points. Where those boundaries arise, firms can occupy positions with greater influence over how the system is assembled and coordinated. (Baldwin, 2008)

A constraint at one point does not establish that the point governs the economics of the whole system. Several constraints can bind at once, so identifying a bottleneck describes a capacity condition rather than by itself determining who captures value.

3. The Bottleneck Keeps Moving

The constraints in this market are shifting. Gartner reports that the AI infrastructure bottleneck has shifted from accelerator silicon toward high-bandwidth memory and advanced packaging. The IEA reports tightening conditions across advanced chip manufacturing, memory, electricity infrastructure, grids and related equipment, with several constraints binding together. (Gartner, 2026; IEA, 2026)

As system performance becomes increasingly dependent on moving data between compute and memory, memory bandwidth and packaging can become practical limits on useful system capacity. In this cycle, investment that relieves one constraint has been accompanied by tighter constraints elsewhere. (Gartner, 2026)

Whether economic returns move along with the constraint is a different question. NVIDIA reported a 75% gross margin in the second quarter of fiscal 2027. In its earnings commentary, management said margins were expected to narrow as memory costs rose faster than expected, while the company said its supply commitments had more than doubled in the quarter, mainly for memory, and that price increases were expected to support margins later. (NVIDIA, 2026a; NVIDIA, 2026b) The evidence therefore does not support a mechanical transfer of rent to whichever layer is scarcest.

The three levels should remain separate. The shift in the bottleneck is well supported in this cycle. The economics around it are changing, as NVIDIA’s margin outlook makes visible. Whether returns systematically follow the moving constraint remains unproven.

4. Scarcity, Margin and Economic Rent

Current profits are concentrated at several constrained positions, consistent with scarcity creating bargaining power but not sufficient to establish a durable rent. TSMC reported a 60.3% operating margin in the second quarter of 2026, while SK hynix reported a 76% operating margin. SK hynix attributed its result to broad increases in memory prices, including conventional memory, alongside AI demand. These are comparable operating-margin measures, but they arise from different businesses and should not be read as evidence that one layer necessarily earns a higher economic rent than another. (TSMC, 2026; SK hynix, 2026)

Why constrained positions can convert into bargaining power is well covered in the economics literature. Teece’s work on complementary assets holds that returns from an innovation may accrue to whoever controls the assets required to commercialise it, depending on the appropriability regime and the strategic positions of the firms involved. Jacobides and Tae find that conditions within industry segments are associated with changes in those segments’ share of total sector value. Whether value moves along a chain therefore concerns how the total is divided, which a single firm’s margin cannot show. (Teece, 1986; Jacobides and Tae, 2015) Buyer concentration can also affect how that surplus is shared: large buyers with alternatives, including the option to develop their own chips, can exert more bargaining leverage.

Margin and economic rent are different measurements. Accounting margins describe operating outcomes. Economic rent is the return to a resource above what is required to keep it in its current use. Economic profit is the return remaining above the cost of capital and the risks borne by the firm. Abraham et al. (2024), using a thirty-year panel of Belgian firms, find that most of the average price-cost margin covered fixed costs, with only a small share representing excess profit. This economy-wide evidence suggests that a high margin in a capital-intensive position can reflect fixed-investment recovery rather than rent.

High margins may reflect scarcity, but they may also reflect technology leadership, product mix, accumulated research advantage or ecosystem dependence. The AI-specific record is too short to separate a structural rent from the profitability of an intense investment cycle. Scarcity can therefore create bargaining power without guaranteeing lasting economic rent.

5. Capital Responds to Scarcity

High returns attract capital, and capital is now expanding fabs, memory and packaging capacity, data centres and power infrastructure. TSMC has increased capacity investment for leading-edge and advanced packaging, while SK hynix is expanding memory and advanced-packaging capacity in response to supply constraints. These commitments are intended to add capacity over time; they are not yet delivered capacity. (TSMC, 2026; SK hynix, 2026)

AI infrastructure capacity responds to established mechanisms, although their outcome in this market remains unproven. Bertolotti, Lanteri and Villa model investment-goods producers with market power and show how markups can decline as capital accumulates. Their calibrated semiconductor-demand model attributes the observed equipment-price increase mainly to rising marginal costs and to a smaller extent to rising markups. Price rises in a constrained market can therefore reflect the cost of expanding capacity, even when markups also rise. OECD research finds infrastructure effects to be nonlinear, with possible over-investment and inefficient use in some sectors. Röller and Waverman find a positive causal link between telecommunications infrastructure and growth, strongest once networks approach critical mass near universal service. Both growth studies concern returns to the wider economy, not supplier margins. (Bertolotti, Lanteri and Villa, 2024; Égert, Koźluk and Sutherland, 2009; Sutherland et al., 2009; Röller and Waverman, 2001)

Infrastructure history also shows how restructuring and entry can alter market power. Joskow’s assessment of electricity liberalisation draws out lessons from changes in competition and regulation. Demirer, Fradkin and Tadelis document rapid supply expansion and falling prices in the emerging market for model intelligence. Cheaper model access can also raise demand for compute, so falling prices downstream need not weaken scarcity upstream. (Joskow, 2008; Demirer, Fradkin and Tadelis, 2026)

The IEA shows extraordinary investment occurring alongside constraints that continue to bind. Capital can weaken scarcity, depending on the relative speeds of capacity growth and demand and on whether architectural change creates new constraints as investment closes old ones. The historical evidence therefore describes mechanisms that can change market power rather than a settled path for AI infrastructure.

6. Infrastructure Risk

The economics of this market also depend on who absorbs the loss if demand, technology or timing disappoint. That exposure is less visible than capital expenditure and it is changing quickly.

The Bank of England identifies growing debt financing of AI infrastructure, declining free cash flow among hyperscalers, and a potential maturity mismatch between long-lived debt and the shorter or more uncertain economic lives of AI hardware. It notes that frontier data centres can become outdated as chip requirements change, while debt may remain outstanding against assets with a shorter economic life. (Bank of England, 2026)

Two arrangements illustrate that the party owning an asset and the party exposed to its value need not be the same. CoreWeave’s 2025 filing discloses asset-level debt supported by take-or-pay contracts and redeployment risk where infrastructure depends on concentrated customer demand; while the contract runs, take-or-pay passes the cost of unused capacity to the customer. Meta’s El Paso venture is owned 80% by BlackRock-managed funds while Meta retains 20% and leases the campus; Meta also provides residual value guarantees whose aggregate threshold approaches development cost. (CoreWeave, 2026; Meta Platforms, 2026) Ownership has moved largely to institutional capital, while much of the risk that the campus loses value remains with Meta. The structures separate ownership, financing and use.

Ownership, control and exposure are therefore three separate positions. Newbery and Pollitt’s analysis of British electricity restructuring found that producers gained more than the entire cost reduction, while consumers and government lost, illustrating how market organisation can alter who captures the resulting surplus. Contracts and financing structures affect how surplus and downside are allocated. (Newbery and Pollitt, 1997)

The ability to absorb infrastructure risk may confer advantage, since balance-sheet strength can allow a firm to commit capital earlier and hold capacity through periods of weak utilisation. That remains an inference. An integrated firm may also absorb more risk because it has more capital locked into assets whose value depends on demand it cannot control.

7. Where Could Economic Value Accumulate?

The major control points are economic positions rather than a ranking, and each carries an unresolved question.

Accelerators and their software ecosystems are a clear current case of strong value capture, arising from scarce compute combined with ecosystem dependence, software compatibility and switching costs. The position is also absorbing rising memory costs. How far custom silicon and alternative architectures erode that combination is open. Leading-edge manufacturing rests on capital intensity, process knowledge, supplier ecosystems and qualification requirements. Memory and advanced packaging have gained importance as data movement becomes more critical and now appear to be taking a larger share of the surplus in the AI hardware chain; whether that position remains scarce as capacity arrives is unresolved.

Hyperscale integration combines capital, customer relationships, utilisation, distribution, proprietary silicon and downstream services, raising the possibility of capture at the level of the system. The structural response is observed; its competitive effect is inferred, since no evidence establishes that spanning multiple layers produces superior risk-adjusted returns. Power-connected capacity is an emerging downstream constraint, where the scarce asset may be grid connection, permitted capacity, suitable sites, transformers or long-term power arrangements. The allocation of value at that layer will depend on contracts, regulation and local infrastructure conditions.

What separates these positions is a set of conditions that can point in different directions for the same firm:

• how constrained the resource is, and how easily customers can substitute away from it;
• how long new supply takes to create;
• who holds bargaining power over price and terms;
• whether the firm can capture value elsewhere in the system;
• whether it can finance expansion and withstand weak utilisation;
• who bears the downside when conditions change.

Three propositions provide a way to test the argument. First, watch whether NVIDIA’s margins recover as its price increases take effect, or continue to fall as memory costs rise. Recovery would indicate bargaining power to pass higher memory costs to customers; continued decline would be consistent with value shifting toward memory. This can be observed within several reporting cycles. Second, watch whether memory prices and margins hold when new capacity comes online. If they hold, demand is absorbing added supply; if they fall, investment is weakening scarcity. This can also be observed within several reporting cycles. Third, over several years, watch what happens as long-term capacity contracts are renewed and the hardware financed in this cycle ages. If firms that guaranteed asset values absorb losses without losing position, the case for risk capacity as an advantage strengthens. If integrated firms fare worse when demand or asset values weaken, concentrated exposure may be the stronger explanation.

8. The Market Is Still Being Built

Demand creates constraints. Constraints attract capital. Capital changes the constraints, and technology changes them independently. The economics of the infrastructure layer also shape which firms can afford to operate and compete in AI-mediated markets. As constraints move, bargaining power, economic value and risk can move with them, though not automatically and not in fixed proportion.

The same cycle can contain persistent scarcity and subsequent margin compression. Scarcity can persist while demand grows faster than capacity and weaken when supply catches up. The market has not completed a full cycle, so which condition dominates remains unknown.

For capital allocators, the tractable questions are narrower than who is winning:

• which constraints are structural and which are transitional;
• how long replication takes at each point in the chain;
• who has contracted to bear the risk if demand or technology changes.

Some constraints may prove lasting while others prove temporary. The useful distinction is between constraints whose underlying sources remain difficult to replicate and those that investment, substitution or architectural change can relieve.

The distinction will become clearer as new capacity arrives, contracts are renewed and system architecture evolves. The current evidence does not settle which constraints are structurally difficult to replicate, which can be relieved by investment or substitution, or how long the evidence will take to distinguish them.

Key Definitions

Bottleneck — A point in a system of complementary inputs where capacity is especially constrained and can limit what the wider system can deliver. Several constraints can bind at once.

Economic rent — A return to a resource above what is required to keep that resource in its current use. It is distinct from economic profit, which is the return remaining above the cost of capital and the risks borne by the firm, and from accounting margin.

Value creation and value capture — Value creation is the surplus generated by an activity. Value capture is the share of that surplus a firm retains. The two are frequently divided differently from how they are produced.

Complementary assets — The resources required to commercialise an innovation. Where they are scarce and tightly held, returns can accrue to their owners rather than to the innovator. (Teece, 1986)

Risk allocation — The contractual and financing arrangements affecting who bears the downside if demand, technology or timing disappoint. Distinct from ownership.

Questions This Article Answers

Who is capturing the value of AI infrastructure today?

Several constrained positions are currently earning unusually high margins, including accelerators and their software ecosystems, leading-edge manufacturing, and memory and advanced packaging. Their current profitability is well evidenced. The accelerator position is also absorbing rising memory costs. None of this establishes that the positions are durable, because the AI-specific record is too short to separate structural rent from the profitability of an intense investment cycle.

Does economic value move when the bottleneck moves?

Not mechanically. The bottleneck has shifted toward memory and advanced packaging in this cycle, and current evidence suggests some movement toward memory as memory costs affect adjacent margins. Systematic migration of economic rent remains unproven.

Will AI infrastructure commoditise as capacity expands?

The mechanism is real but conditional. Theory supports the proposition that investment can weaken scarcity and compress markups. Infrastructure history shows that restructuring and entry can change market power and who gains from it. Whether investment weakens scarcity depends on whether supply expands faster than demand and whether architectural change creates new constraints. Current evidence shows extraordinary investment alongside constraints that continue to bind.

Why does infrastructure risk matter alongside infrastructure ownership?

Because ownership, use and exposure can be separated by contract and financing structure. Long-term guarantees, take-or-pay agreements and asset-level debt can leave the downside with a firm or investor that does not ultimately own the asset. The extent to which the ability to absorb that risk creates competitive advantage remains an inference.

What determines who captures value in AI infrastructure?

How constrained a resource is, how easily it can be substituted, how long new supply takes to create, who holds bargaining power over price and terms, whether a firm can capture value elsewhere in the system, whether it can finance expansion and absorb weak utilisation, and who bears the downside when conditions change.

References

Abraham, F., Bormans, Y., Konings, J. and Roeger, W. (2024) 'Price-Cost Margins, Fixed Costs and Excess Profits', The Economic Journal, 134(663), pp. 2655–2684. Available at: https://doi.org/10.1093/ej/ueae037

Baldwin, C.Y. (2008) ‘Where do transactions come from? Modularity, transactions, and the boundaries of firms’, Industrial and Corporate Change, 17(1), pp. 155–195. Available at: https://doi.org/10.1093/icc/dtm036

Bank of England (2026) Financial Stability Report, July 2026. London: Bank of England. Available at: https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026

Bertolotti, F., Lanteri, A. and Villa, A. (2024, revised 2026) Investment-Goods Market Power and Capital Accumulation. Federal Reserve Bank of Chicago Working Paper 2024-13. Available at: https://www.chicagofed.org/publications/working-papers/2024/2024-13

CoreWeave (2026) 2025 Form 10-K. U.S. Securities and Exchange Commission. Available at: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm

Demirer, M., Fradkin, A. and Tadelis, N. (2026) ‘The Emerging Market for Intelligence: How Firms Buy and Sell AI’, Journal of Economic Perspectives. Available at: https://www.aeaweb.org/articles?id=10.1257/jep.20261506

Égert, B., Koźluk, T. and Sutherland, D. (2009) Infrastructure and Growth: Empirical Evidence. OECD Economics Department Working Papers. Available at: https://doi.org/10.1787/225682848268

Gartner (2026) The Next AI Bottleneck Has Moved: HBM and Advanced Packaging, Not GPU, Will Set AI Silicon Infrastructure Economics in 2026-2027. Available at: https://www.gartner.com/en/documents/7896177

International Energy Agency (IEA) (2026) Key Questions on Energy and AI. Paris: IEA. Available at: https://www.iea.org/reports/key-questions-on-energy-and-ai

Jacobides, M.G. and Tae, C.J. (2015) ‘Kingpins, Bottlenecks, and Value Dynamics Along a Sector’, Organization Science, 26(3), pp. 889–907. Available at: https://doi.org/10.1287/orsc.2014.0958

J.P. Morgan (2026) Financing the Infrastructure Behind U.S. Data Center Growth. Available at: https://www.jpmorgan.com/insights/banking/capital-markets/financing-ai-infrastructure-data-centers

Joskow, P.L. (2008) ‘Lessons Learned From Electricity Market Liberalization’, The Energy Journal, 29(Special Issue 2). Available at: https://doi.org/10.5547/ISSN0195-6574-EJ-Vol29-NoSI2-3

Meta Platforms (2026) Meta Announces New Strategic Venture with BlackRock to Develop Data Center in El Paso. Available at: https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-New-Strategic-Venture-with-BlackRock-to-Develop-Data-Center-in-El-Paso/default.aspx

Newbery, D.M. and Pollitt, M.G. (1997) ‘The Restructuring and Privatisation of Britain’s CEGB — Was It Worth It?’, The Journal of Industrial Economics, 45(3), pp. 269–303. Available at: https://doi.org/10.1111/1467-6451.00049

NVIDIA (2026a) NVIDIA Announces Financial Results for Second Quarter Fiscal 2027. Available at: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027

NVIDIA (2026b) CFO Commentary on Second Quarter Fiscal 2027 Results, 26 August 2026. Available at: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27cfocommentary.htm

PwC (2026) Global Data Centre Outlook 2026–50. Available at: https://www.pwc.com/gx/en/1/services/consulting/technology/data-centre-outlook.html

Röller, L.-H. and Waverman, L. (2001) ‘Telecommunications Infrastructure and Economic Development: A Simultaneous Approach’, American Economic Review, 91(4), pp. 909–923. Available at: https://doi.org/10.1257/aer.91.4.909

SK hynix (2026) Q2 2026 Business Results. Available at: https://news.skhynix.com/en/q2-2026-business-results/

Sutherland, D., Araújo, S., Égert, B. and Koźluk, T. (2009) ‘Infrastructure Investment: Links to Growth and the Role of Public Policies’, OECD Economics Department Working Papers No. 686. Available at: https://doi.org/10.1787/225678178357

Teece, D.J. (1986) ‘Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy’, Research Policy, 15(6), pp. 285–305. Available at: https://doi.org/10.1016/0048-7333(86)90027-2

TSMC (2026) Second Quarter 2026 Results. Available at: https://investor.tsmc.com/english/quarterly-results/2026/q2

About the Author

Nune Igityan is an AI Market Strategist and founder of Kensai. Her work examines how AI is reshaping markets and competitive advantage — how buying decisions form, where economic value accumulates, and how organisations compete in AI-mediated markets. She previously held senior product marketing and growth roles at Google and Meta.

How to Cite This Article

Igityan, N. (2026) Who Will Capture the Value of AI Infrastructure? How Constraints, Capital and Risk Are Reshaping Infrastructure Economics. Kensai. Available at: https://www.kensai.uk/research/who-will-capture-the-value-of-ai-infrastructure