AI has become a physical-infrastructure cycle. The limiting resources are no longer only models and chips; they include power generation, grid connections, transformers, cooling, high-bandwidth memory, construction capacity and capital. The build-out is large enough to influence global investment patterns and local electricity systems. The most valuable opportunity map therefore sits across the full stack—from accelerators and data-centre design to electricity, equipment, financing and regional capacity—not only inside “AI software.”
Research question and scope
Core question: Where is economic value accumulating in the AI infrastructure build-out, what are the binding constraints, and which parts of the stack offer durable opportunities through 2030?
Scope: Global compute and data-centre capacity, semiconductors/memory, power demand, grids, onsite generation/storage, cooling/electrical equipment, capital and geographic location. Focuses on structural demand and bottlenecks rather than equity valuation.
Method: Uses IEA energy datasets, UNCTAD FDI, Stanford compute data, McKinsey capacity/value-chain studies and current capex/deal evidence. Forecasts are labelled as projections and compared across sources.
1. The build-out has reached macroeconomic scale
The IEA reports capital expenditure by five large technology companies exceeded $400 billion in 2025 and is expected to rise another 75% in 2026—more than global investment in oil and gas production. Its satellite tracking indicates AI-focused “factories” more than tripled capacity in eighteen months. S1 McKinsey estimates global data-centre spending could reach roughly $7 trillion cumulatively by 2030 under current build-out scenarios. S2
UNCTAD shows the investment footprint directly: announced FDI in data centres exceeded an estimated $270 billion in 2025 and represented more than one-fifth of global greenfield project value. S3 This is why AI infrastructure should be analysed as an industrial and investment market, not only a technology expense.
2. Compute capacity is expanding at extraordinary speed
Stanford’s 2026 AI Index estimates global AI compute capacity grew about 3.3× per year from 2022, reaching 17.1 million H100-equivalents by the end of 2025. Nvidia supplied more than 60% of total tracked capacity, with Google, Amazon, AMD and Huawei contributing additional accelerator capacity. S4 The same report estimates AI data-centre power capacity at about 29.6 GW by Q4 2025.
3. Data-centre demand is becoming the organising constraint
McKinsey’s June 2026 colocation analysis estimates global data-centre demand could rise from about 82 GW in 2025 to about 220 GW in 2030 under current adoption scenarios. S5 Its earlier scenario work places AI-ready demand as the majority of incremental capacity. S6 IEA’s electricity projection is consistent with the direction: global data-centre electricity use rises from 485 TWh in 2025 to about 950 TWh in 2030, around 3% of global electricity demand, with AI-focused data-centre consumption tripling. S1
| Metric | Latest evidence | Strategic implication |
|---|---|---|
| Hyperscaler capex | > $400bn in 2025; +75% expected in 2026 for five large tech companies. S1 | Capital intensity is becoming a competitive moat and financing challenge. |
| Greenfield FDI | > $270bn data-centre investment announced in 2025; >20% global greenfield value. S3 | AI infrastructure is reshaping international investment. |
| Compute | 17.1m H100-equivalents by end-2025. S4 | Accelerator capacity expanding faster than most physical industries. |
| Electricity | 485 TWh (2025) → ~950 TWh (2030) central IEA projection. S1 | Power is a location and growth constraint. |
| Capacity demand | ~82 GW (2025) → ~220 GW (2030) McKinsey scenario. S5 | Large construction/equipment pipeline required. |
| Grid queues | >2,500 GW of generation/storage/large loads waiting globally. S7 | Connection speed may be more important than nominal power price. |
Sources and notes
- IEA, Key Questions on Energy and AI (16 Apr 2026). [Source](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary)) — Current global energy/AI demand, capex, bottleneck and data-centre projections.
- McKinsey, The $7 trillion data center build-out (27 Mar 2026). [Source](https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-7-trillion-dollar-data-center-build-out-how-industrials-can-capture-their-share)) — Industrial value chain and cumulative investment scenario.
- UNCTAD, Data centres are reshaping the global investment landscape (22 Jan 2026). [Source](https://unctad.org/news/data-centres-are-reshaping-global-investment-landscape)) — 2025 greenfield FDI and data-centre investment.
- Stanford HAI, 2026 AI Index — Research & Development. [Source](https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development)) — Global compute capacity, supplier concentration and AI data-centre power capacity.
- McKinsey, Colocation data centers: The infrastructure race behind AI (30 Jun 2026). [Source](https://www.mckinsey.com/mgi/our-research/colocation-data-centers-the-infrastructure-race-behind-ai)) — 82 GW to ~220 GW global demand scenario and location economics.
- McKinsey, AI power: expanding data center capacity to meet demand. [Source](https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/ai-power-expanding-data-center-capacity-to-meet-growing-demand)) — AI-ready capacity share and demand scenarios.
- IEA, Electricity 2026 — Grids / Executive Summary. [Source](https://www.iea.org/reports/electricity-2026/executive-summary)) — Grid queues and required investment.
8i TECHNOLOGIES
The AI Infrastructure Economy
Compute, data centres, semiconductors, power, grids, cooling and capital: where the next AI value pools are forming
Reference date: 15 September 2026
See what matters next.
| Executive answer AI has become a physical-infrastructure cycle. The limiting resources are no longer only models and chips; they include power generation, grid connections, transformers, cooling, high-bandwidth memory, construction capacity and capital. The build-out is large enough to influence global investment patterns and local electricity systems. The most valuable opportunity map therefore sits across the full stack—from accelerators and data-centre design to electricity, equipment, financing and regional capacity—not only inside “AI software.” |
|---|
Research question and scope
Core question: Where is economic value accumulating in the AI infrastructure build-out, what are the binding constraints, and which parts of the stack offer durable opportunities through 2030?
Scope: Global compute and data-centre capacity, semiconductors/memory, power demand, grids, onsite generation/storage, cooling/electrical equipment, capital and geographic location. Focuses on structural demand and bottlenecks rather than equity valuation.
Method: Uses IEA energy datasets, UNCTAD FDI, Stanford compute data, McKinsey capacity/value-chain studies and current capex/deal evidence. Forecasts are labelled as projections and compared across sources.
1. The build-out has reached macroeconomic scale
The IEA reports capital expenditure by five large technology companies exceeded $400 billion in 2025 and is expected to rise another 75% in 2026—more than global investment in oil and gas production. Its satellite tracking indicates AI-focused “factories” more than tripled capacity in eighteen months. [S1] McKinsey estimates global data-centre spending could reach roughly $7 trillion cumulatively by 2030 under current build-out scenarios. [S2]
UNCTAD shows the investment footprint directly: announced FDI in data centres exceeded an estimated $270 billion in 2025 and represented more than one-fifth of global greenfield project value. [S3] This is why AI infrastructure should be analysed as an industrial and investment market, not only a technology expense.
2. Compute capacity is expanding at extraordinary speed
Stanford’s 2026 AI Index estimates global AI compute capacity grew about 3.3× per year from 2022, reaching 17.1 million H100-equivalents by the end of 2025. Nvidia supplied more than 60% of total tracked capacity, with Google, Amazon, AMD and Huawei contributing additional accelerator capacity. [S4] The same report estimates AI data-centre power capacity at about 29.6 GW by Q4 2025.
3. Data-centre demand is becoming the organising constraint
McKinsey’s June 2026 colocation analysis estimates global data-centre demand could rise from about 82 GW in 2025 to about 220 GW in 2030 under current adoption scenarios. [S5] Its earlier scenario work places AI-ready demand as the majority of incremental capacity. [S6] IEA’s electricity projection is consistent with the direction: global data-centre electricity use rises from 485 TWh in 2025 to about 950 TWh in 2030, around 3% of global electricity demand, with AI-focused data-centre consumption tripling. [S1]
| Metric | Latest evidence | Strategic implication |
|---|---|---|
| Hyperscaler capex | > $400bn in 2025; +75% expected in 2026 for five large tech companies. [S1] | Capital intensity is becoming a competitive moat and financing challenge. |
| Greenfield FDI | > $270bn data-centre investment announced in 2025; >20% global greenfield value. [S3] | AI infrastructure is reshaping international investment. |
| Compute | 17.1m H100-equivalents by end-2025. [S4] | Accelerator capacity expanding faster than most physical industries. |
| Electricity | 485 TWh (2025) → ~950 TWh (2030) central IEA projection. [S1] | Power is a location and growth constraint. |
| Capacity demand | ~82 GW (2025) → ~220 GW (2030) McKinsey scenario. [S5] | Large construction/equipment pipeline required. |
| Grid queues | >2,500 GW of generation/storage/large loads waiting globally. [S7] | Connection speed may be more important than nominal power price. |
4. The bottleneck stack
The build-out is constrained by multiple interdependent supply chains. IEA reports a recent shortage of high-bandwidth memory expected to persist at least through end-2027, a 70% surge in gas-turbine orders in 2025, and tightening constraints in power equipment and advanced chip manufacturing. [S1] The implication is that a data centre is only as fast as its slowest critical component.
| Constraint | Why it binds | Opportunity created |
|---|---|---|
| Advanced accelerators/HBM | Frontier AI requires scarce chips and memory bandwidth | Chips, memory, packaging, testing, alternatives |
| Power connection | Data centres scale faster than transmission/permitting | Grid tech, flexible connections, onsite power, storage |
| Transformers/switchgear | Long lead times and rising power density | Electrical equipment manufacturing and services |
| Cooling | Rack density exceeds conventional air-cooling economics | Liquid cooling, heat rejection, water/thermal optimisation |
| Construction/skilled labour | Campus timelines collide with regional labour constraints | Modular/prefab systems, specialist contractors |
| Capital | Projects too large for balance sheets alone | Private capital, structured finance, project/debt markets |
| Social licence | Power prices, water, land and local opposition | Community benefit models, siting intelligence, transparent tariffs |
5. Power density is forcing a redesign of the data centre
IEA projects that by 2027 an advanced rack could have peak power demand comparable to 65 households; AI server power density increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027. [S1] McKinsey reports a shift toward 800-volt DC architectures as operators seek to reduce conversion losses and manage much higher power density. [S8]
This creates value pools in power electronics, busways, switchgear, batteries, liquid cooling, thermal management and higher-voltage components—industrial categories that historically sat far from the centre of software strategy.
6. Grids are becoming a decisive market-access layer
IEA’s Electricity 2026 report says more than 2,500 GW of generation, storage and large-load projects are stalled in grid queues worldwide. Annual grid investment would need to rise roughly 50% from today’s $400 billion by 2030 to meet forecast electricity demand. [S7] The location question for AI infrastructure is therefore increasingly “where can power be connected on time?” rather than simply “where is electricity cheap?”
- Grid queue reform and capacity auctions can become competitive advantages for regions.
- Non-firm connections and demand response can accelerate time-to-power.
- Battery storage can turn data centres from pure loads into flexible grid assets.
- Onsite gas/nuclear/renewables may reduce dependence on queues but introduce new equipment, permitting and financing constraints.
7. Energy efficiency improves—and demand still rises
The IEA finds energy use per simple AI task has been falling by at least an order of magnitude annually in recent years, yet more intensive reasoning, video and agentic workloads can use hundreds or thousands of times more energy per query than simple text generation. [S1] This is a classic rebound dynamic: efficiency lowers unit cost, which enables more ambitious and more frequent use.
8. Capital and returns are becoming constraints too
Reuters’ July 2026 analysis of major US technology companies found AI infrastructure spending increasingly pressuring free cash flow, with investors demanding evidence that large capex programmes translate into AI/cloud revenue. [S9] IEA similarly notes data-centre investment has grown too large to be financed solely from company balance sheets and will depend increasingly on capital markets and expectations of return. [S1]
| 8i interpretation The infrastructure cycle is not demand-risk-free. Physical bottlenecks can delay supply in the near term, while weak AI monetisation could tighten capital before planned capacity is completed. Opportunity assessment must therefore combine demand, bottleneck value and financing durability. |
|---|
9. Geography: the next map of AI capacity
UNCTAD reports data-centre FDI concentrated in a small number of locations led in 2025 by France, the United States and South Korea, with large emerging-market projects in Brazil, India, Thailand and Malaysia. [S3] Stanford reports the United States hosts far more data centres than any other country, while leading-edge chip fabrication remains highly dependent on TSMC. [S4]
A current illustration is Google’s September 2026 announcement of roughly €13 billion in new AI infrastructure in Finland, coupled with a long-term nuclear power agreement and grid/energy investment. [S10] The strategic package—cold climate, power, grid investment, energy contracting and political support—shows how infrastructure location has become a multi-factor competitive proposition.
10. Where the value pools are forming
| Value pool | Attractiveness logic | Key watch items |
|---|---|---|
| Compute/accelerators | Persistent AI demand and performance race | Alternative accelerators, export controls, utilisation |
| HBM/packaging | High bandwidth and advanced packaging bottlenecks | Capacity additions, yield, vendor concentration |
| Data-centre development | Large capacity requirement and time-to-market premium | Power access, permitting, capital cost |
| Power equipment | Transformers/switchgear/turbines bottleneck build speed | Lead times, new manufacturing capacity |
| Cooling/power architecture | Rising rack density forces redesign | Liquid cooling adoption, 800VDC standards |
| Grid/flexibility | Connection queues constrain deployment | Grid-enhancing tech, batteries, demand response |
| Energy supply | Firm power increasingly strategic | Gas, nuclear, geothermal, renewables + storage |
| Finance | Projects exceed balance-sheet comfort | Debt capacity, project structures, utilisation/ROI evidence |
11. 8i assessment
| 8i view AI infrastructure is one of the largest investable ecosystems created by AI and one of the clearest bridges between Technology Intelligence and Market Intelligence. The most defensible opportunities are likely to be in scarce enabling capabilities with multi-year bottlenecks—not simply in the highest-profile AI asset. |
|---|
12. What businesses and investors should do now
- Map exposure to the full bottleneck stack, including power equipment and grid access.
- Use time-to-power and time-to-connect as core site-selection variables.
- Separate training and inference requirements; inference increasingly changes geography and network design.
- Stress-test projects for lower AI monetisation and higher financing costs.
- Evaluate local social/policy constraints alongside energy price.
- Look for enabling suppliers with backlog and difficult-to-replicate manufacturing/service capability.
- Treat energy and AI strategies as one integrated infrastructure decision.
Source register
Sources are prioritised toward primary institutions, transparent methodology, observable behavioural data, large-sample surveys and current production evidence. Vendor evidence is identified as such and used cautiously where it provides proprietary usage data or live implementation detail.
[S1] IEA, Key Questions on Energy and AI (16 Apr 2026). Source — Current global energy/AI demand, capex, bottleneck and data-centre projections.
[S2] McKinsey, The $7 trillion data center build-out (27 Mar 2026). Source — Industrial value chain and cumulative investment scenario.
[S3] UNCTAD, Data centres are reshaping the global investment landscape (22 Jan 2026). Source — 2025 greenfield FDI and data-centre investment.
[S4] Stanford HAI, 2026 AI Index — Research & Development. Source — Global compute capacity, supplier concentration and AI data-centre power capacity.
[S5] McKinsey, Colocation data centers: The infrastructure race behind AI (30 Jun 2026). Source — 82 GW to ~220 GW global demand scenario and location economics.
[S6] McKinsey, AI power: expanding data center capacity to meet demand. Source — AI-ready capacity share and demand scenarios.
[S7] IEA, Electricity 2026 — Grids / Executive Summary. Source — Grid queues and required investment.
[S8] McKinsey, The shift to 800-volt DC at data centers (30 Jul 2026). Source — Higher-voltage architecture and power-density implications.
[S9] Reuters, AI investment boom puts Big Tech free cash flow under pressure (22 Jul 2026). Source — Current financing/returns pressure.
[S10] Reuters, Google €13bn Finland AI infrastructure/nuclear deal (9 Sep 2026). Source — Current European location/energy case study.
[S11] McKinsey, Frontiers of compute: reducing AI inference costs (25 Jun 2026). Source — Hyperscaler capex and semiconductor/inference architecture.
[S12] UNCTAD, World Investment Report 2026. Source — Strategic-sector investment and geographic concentration.
[S13] IEA, Electricity 2026 — Grids. Source — Detailed grid bottleneck and queue-management analysis.
[S14] McKinsey, Beyond compute: infrastructure that powers and cools AI data centers. Source — Cooling/power equipment and supply-chain opportunity.