From hyperscaler CAPEX to deliverable gigawatts: why the next phase of the AI build-out may depend as much on power grids, transformers and execution capacity as on semiconductors.
When investors think about artificial intelligence, a handful of companies tend to dominate the conversation: Nvidia, Microsoft, Alphabet, Amazon and Meta.
That makes sense. The first phase of the AI investment cycle largely rewarded the companies developing the models, supplying the computing power or financing its rapid deployment.
But as investment accelerates, the question is changing.
AI is no longer just a software story. It is becoming one of the largest physical infrastructure investment cycles of the decade: semiconductors, memory, servers, data centers, cooling systems, electrical equipment, transformers, power grids, generation capacity and raw materials.
And another reality is becoming increasingly important.
Having the capital — and even ordering the equipment — is not enough.
The infrastructure still has to be engineered, permitted, procured, built, connected to a sufficiently robust power system, tested and commissioned.
For investors, the interesting question may therefore no longer be simply who will win the AI race.
It may increasingly become:
Which resources will every competitor need, which ones are hardest to scale, and what could prevent hundreds of billions of dollars of committed capital from becoming operational computing capacity?
That is the central thesis of this article:
As computing capacity expands, some of the most important AI bottlenecks may progressively move beyond semiconductors into electrical infrastructure, power grids and ultimately the human and industrial capacity required to deliver the projects.
This builds directly on my earlier analysis of whether we are entering a new commodity supercycle. When demand can expand much faster than physical supply, the scarce input can become as important to investors as the final product itself.
There is, however, an essential caveat:
identifying a bottleneck is only the first step. Determining how much of it is already priced into markets is the second.
In Brief: Capital Moves Faster Than the Physical World
The scale of the AI investment cycle is becoming extraordinary.
Combined capital expenditure by Microsoft, Alphabet, Amazon, Meta and Oracle is expected to approach three-quarters of a trillion dollars in 2026. The companies do not consistently disclose AI-specific spending, but much of the surge in data-center, server, networking and cloud infrastructure investment is being driven by AI demand.
At the same time, the International Energy Agency’s updated 2026 outlook projects global data-center electricity consumption at around 950 TWh by 2030, roughly double the 485 TWh consumed in 2025 and equivalent to around 3% of global electricity demand.
The problem is that the different parts of the infrastructure do not operate on the same timetable.
Capital can be committed rapidly.
Industrial capacity cannot.
Some high-voltage cables now require two to three years to procure. Large power transformers can require up to four years, while specialised transmission components can take even longer.
The entire conversion process can therefore be thought of as:
Available CAPEX → available equipment → available power → adequate grid capacity → available skills → delivered project → operational compute
The chain can only move as fast as its slowest link.
Major institutional investors are increasingly looking at AI through a similar physical-infrastructure lens. BlackRock identifies power, memory, chips and data centers among the scarce inputs shaping AI investment opportunities, while also highlighting rising AI-driven demand for power grids.
From Announced CAPEX to Operational Capacity
Headline investment numbers are impressive, but investors need to be careful when adding together hyperscaler spending plans and announced data-center projects.
An announced dollar is not necessarily a dollar spent.
An announced gigawatt is not an operational gigawatt.
A useful way of analysing the build-out is to separate the project lifecycle into distinct stages.
| Stage | What it actually tells us |
|---|---|
| Announced | There is an intention to develop capacity |
| Planned | Technical and commercial development has begun |
| Permitted | Key regulatory approvals have been secured |
| Financed / contracted | Economic commitments are becoming substantially firmer |
| Under construction | Physical execution has started |
| Grid-connected | Power can actually reach the site |
| Commissioned | Systems have been tested and accepted |
| Operational | The capacity can produce compute and generate revenue |
Between 100 GW announced and 100 GW operational lies an extraordinarily complex industrial process.
For investors, the conversion rate of announced CAPEX into operational capacity may eventually matter almost as much as the amount of CAPEX itself.
This brings us back to a broader investment principle discussed in Why Markets Reward Surprises… Not the Best Companies:
A powerful economic trend does not automatically produce attractive investment returns if expectations already reflect it.
United States: A Massive Build-Out Meets Local Constraints
The United States sits at the centre of the current AI infrastructure boom.
Yet the important question is not simply how much generating capacity exists nationally.
It is where that capacity exists, whether the grid can move it, and whether new loads can connect to it.
The U.S. interconnection pipeline illustrates the scale of the wider grid challenge.
According to Lawrence Berkeley National Laboratory’s 2026 Queued Up report, around 8,200 projects were actively seeking U.S. grid interconnection at the end of 2025, representing approximately 1,312 GW of generation and 749 GW of storage.
These are not data-center projects. They are generation and storage projects seeking access to the grid.
But they illustrate a crucial point:
Grid access is already a scarce development resource before the full impact of additional AI electricity demand is considered.
And grid constraints are inherently local.
The United States can have sufficient generating capacity at national level while a particular part of Virginia, Texas, Ohio or Pennsylvania struggles to accommodate another large data-center campus.
AI infrastructure is therefore increasingly a question of location and deliverability, not simply national energy supply.
China: The Same AI Race, Different Constraints
China is simultaneously building its own large-scale computing and energy infrastructure.
But comparing the United States and China solely through announced CAPEX or gigawatts would miss something important.
The two countries operate within very different industrial, regulatory, power-system and project-delivery environments.
China has an enormous manufacturing base and extensive experience delivering large-scale electrical and energy infrastructure.
The United States has extraordinary capital-market depth, a leading technology ecosystem and important advantages in advanced computing.
The constraints therefore do not necessarily emerge at the same points in the chain.
This leads to a broader investment insight:
The same dollar of CAPEX — or the same announced GW — does not necessarily have the same probability of becoming operational on the same timetable in different project environments.
Execution capacity itself therefore becomes part of the investment analysis.
Five Capabilities Have to Converge to Build AI
Operational AI infrastructure ultimately depends on five layers.
| Capability | What it includes |
| Technology | GPUs, memory, computing networks |
| Industrial capacity | Data centers, cooling, electrical equipment, transformers, cables |
| Energy | Power generation, storage, fuels |
| Grid infrastructure | Transmission, distribution, substations, connections, redundancy |
| Execution capacity | Engineering, procurement, construction, testing, commissioning, project management |
All five have to converge.
The best GPUs are useless without somewhere to install them.
A completed data center cannot operate without power.
Enough national generation is insufficient if the grid cannot deliver the required power to the site.
And having all the equipment available is insufficient if the infrastructure cannot be integrated, tested and commissioned.
Operational capacity is therefore not determined by the sum of available resources.
It is constrained by the least available critical resource.
The Three Physical Chains Behind AI Infrastructure
From raw materials to compute, an AI data center sits at the convergence of several physical systems that must ultimately operate as one.
The important point is that possessing these resources is not enough.
Semiconductors, electrical equipment and generating capacity still have to be engineered into a functioning system. Equipment must be procured, infrastructure built, grid connections completed, interfaces managed, systems tested and the final asset commissioned.
That introduces a fourth constraint cutting across all three physical chains:
Execution capacity.
This is why the AI investment opportunity extends far beyond technology companies.
How Do We Identify a Genuine AI Bottleneck?
Rather than assigning arbitrary star ratings, I prefer five questions.
| Criterion | The question investors should ask |
| Demand | How much incremental demand is AI actually creating? |
| Supply | How quickly can industry increase capacity? |
| Execution | How quickly can that capacity actually be built and commissioned? |
| Theme maturity | How widely has the market already recognised the AI connection? |
| Diversification | Are there other demand drivers independent of AI? |
Notice what is deliberately absent from this table:
valuation.
That comes afterwards.
A shortage can be completely real while still being a poor investment if investors have already priced in years of exceptional growth.
GPUs: The First Bottleneck Has Already Been Discovered
The first major constraint in the AI cycle was compute.
Accelerators remain indispensable, but markets understand this exceptionally well.
The strategic importance of GPUs is no longer hidden information.
This does not mean semiconductor leaders cannot continue to grow.
It means the market is already deeply aware of the opportunity.
Memory: A Real Constraint, But Increasingly Recognised
A powerful accelerator is only useful if data can reach it fast enough.
High-bandwidth memory has therefore become a critical component of AI architectures.
But this constraint is increasingly recognised too.
BlackRock explicitly includes memory among the scarce inputs in its 2026 AI investment framework.
Networking, Fibre and Connectivity
Running a few GPUs is one problem.
Making tens of thousands of accelerators work efficiently together is another.
Large AI clusters require extremely fast networking, fibre, advanced optical systems, cables and connectors.
The structural demand is substantial, but this part of the AI supply chain is no longer obscure to specialist investors.
Data Centers: The Building Is Only the Beginning
Modern AI data centers are increasingly becoming factories for compute, with individual campuses requiring hundreds of megawatts and some proposed developments reaching gigawatt scale.
But constructing the building is only part of the problem.
For a data center to become operational, several conditions must converge:
land, permits, cooling, electrical equipment, construction resources, grid capacity and a secure connection.
This is becoming a financing issue as well as an engineering one.
Reuters reported in August 2026 that U.S. lenders are increasingly scrutinising permitting and community opposition as part of data-center credit risk. In the first quarter of 2026 alone, at least 75 U.S. data-center projects worth around $130 billion encountered organised opposition.
The value of a data-center site therefore increasingly depends on what exists around it, not merely on the building itself.
Cooling: Electricity Ultimately Becomes Heat
As compute increases, rack power density rises.
Most of the electricity consumed by computing equipment ultimately has to be removed from the facility as heat.
Cooling therefore becomes a critical part of the infrastructure, particularly as higher rack densities accelerate the adoption of liquid-cooling architectures.
But this is no longer an undiscovered theme.
The launch of the VanEck Data Center Supply Chain ETF (RACK) in June 2026 is revealing. Its strategy targets companies supplying the semiconductors, power, cooling systems and grid equipment required to support AI infrastructure.
Wall Street is already moving deeper into the physical AI value chain.
Electrical Equipment: Between the Grid and the Server
Between a high-voltage transmission line and a server sits a remarkable amount of electrical equipment:
transformers, switchgear, UPS systems, busbars, circuit breakers, power-distribution equipment, converters, cables and connectors.
An AI data center is therefore as much an electrical installation as an IT installation.
This part of the value chain also has an important investment characteristic:
Its demand does not depend solely on AI.
The same equipment is required for industrial electrification, renewable generation, energy storage, grid reinforcement and wider infrastructure investment.
That is exactly the type of characteristic discussed in How to Analyze a Commodity Like an Investor: some of the strongest structural cases involve multiple independent sources of demand competing for constrained supply.
Transformers and Cables: Industrial Lead Times Cannot Follow CAPEX Overnight
This is where the story becomes particularly interesting.
The problem is not simply that AI requires more electrical equipment.
It is that industrial supply cannot necessarily expand at the same speed as hyperscaler budgets.
The IEA reported in 2026 that average waiting times for new cables and large power transformers had almost doubled since 2021.
Some cables now require two to three years to procure.
Large power transformers can require up to four years.
Certain specialised high-voltage components can take even longer.
Transformer capacity itself requires factories, specialised machinery, copper, electrical steel and skilled labour.
The same applies to many types of high-voltage cable and switchgear.
This is where AI infrastructure begins to intersect directly with the broader thesis developed in Copper: The Strategic Commodity of the Decade?
Money can accelerate an order. It cannot instantly create industrial capacity.
Having Enough Power Generation Is Not Enough
This may be one of the most important distinctions in the entire article.
A country can have enough electricity generation nationally and still be unable to connect a new data center where the developer wants to build it.
Generation capacity is not connection capacity.
Between a power plant and a data center sit transmission lines, substations, transformers, distribution infrastructure and protection systems.
Every component has limits.
The relevant question is therefore not simply how many additional GW a country can generate.
It is:
How many MW or GW can actually be delivered to this specific site, at the required reliability level, and on what timetable?
The Grid Matters as Much as Generation
One megawatt somewhere in a power system is not necessarily equivalent to another megawatt.
A data center requires power at its connection point, with a level of resilience compatible with critical digital infrastructure.
Local network architecture therefore matters.
It would, however, be an oversimplification to reduce this to United States = radial / China = ring.
Grid topology varies by voltage level, operator, region and site.
The more useful question is:
What redundancy, transfer capacity and reconfiguration capability exists at the point where the data center wants to connect?
That local infrastructure — rather than theoretical national generating capacity — can determine the real quality of the connection.
An Engineer’s View: Not All MW Are Equal
In a radial architecture, losing an upstream element can interrupt downstream loads until the fault is isolated or supply is restored through another path.
More interconnected architectures can offer greater scope for reconfiguration after a fault.
For a data-center developer, the question is therefore not merely:
Can I get 500 MW?
It becomes:
Can I get 500 MW with the redundancy, transfer capacity and reconfiguration capability required after the loss of a critical grid element?
Two sites theoretically offering 500 MW can therefore have very different operational characteristics.
And that engineering difference becomes a financial difference.
Achieving the required resilience may require additional grid reinforcement, multiple incoming supplies, additional substations and transformers, storage or backup generation.
That means additional CAPEX.
It can also mean additional time.
The quality of a megawatt can matter as much as the quantity.
This is why simply adding up the “available GW” in an investor presentation can be misleading.
MW, GW and TWh: Power Is Not Energy
Data-center announcements often mix several units.
MW and GW measure power — the instantaneous rate at which a facility can consume or generate electricity.
TWh measure energy consumed over time.
The distinction matters.
The IEA’s updated 2026 outlook projects global data-center electricity consumption at around 950 TWh in 2030, compared with approximately 485 TWh in 2025.
But a data center does not merely require a certain amount of energy over a year.
It requires hundreds of megawatts to be available when needed, where needed and through a network capable of delivering them.
A country can therefore have sufficient annual electricity production and still be unable to accommodate a new 500 MW load in a specific location.
The Invisible Bottleneck: Skills
This may be one of the least visible dimensions of the AI infrastructure story in financial analysis.
Buying every required component does not build a project.
Between billions of dollars of CAPEX and an operational asset sit electrical engineers, technicians, construction teams, high-voltage specialists, control engineers, commissioning specialists, planners, contract managers and project leaders capable of coordinating thousands of interfaces.
That resource is constrained too.
The IEA’s World Energy Employment 2025 surveyed more than 700 energy companies, trade unions and education providers.
More than half reported critical hiring bottlenecks, particularly in applied technical roles such as electricians, pipefitters, power-line workers and engineers.
This constraint has something important in common with transformers:
It cannot be solved simply by allocating more capital.
A company can pay more to compete for existing workers.
It cannot create thousands of experienced electricians, grid engineers or commissioning specialists in six months.
And the people capable of leading complex multi-billion-dollar infrastructure programmes take even longer to develop.
Execution Capacity: Ordering the Equipment Is Not the Same as Delivering the Asset
This deserves to be treated as an investment risk in its own right.
A data-center or energy-infrastructure project is not simply a procurement list.
Between the investment decision and commercial operation lies an entire delivery chain:
Engineering → permitting → contracting → procurement → construction → interface management → grid connection → testing → energisation → commissioning
Any one of these can become the critical path.
The transformer may be available while the substation is not.
The building may be complete while the grid connection is delayed.
The grid may be available while commissioning resources are not.
The project ultimately becomes operational only when its last critical bottleneck has been resolved.
And the financial consequences are direct:
Delay → additional CAPEX → additional financing costs → delayed revenue → lower return on invested capital
For an infrastructure investor, execution capability should therefore be treated as an economic variable, not simply an operational detail.
Announced CAPEX → operational asset is not a financial conversion. It is an industrial process.
Electricity: The Fundamental Physical Constraint
At the end of the chain sits the same resource:
electricity.
The IEA expects global data-center electricity consumption to roughly double between 2025 and 2030, reaching around 950 TWh.
AI has therefore already become an energy issue.
The question is no longer whether AI will require significant additional electricity.
It is:
Where can sufficient generation be added, how quickly can it be built, and how can the power be delivered to the data centers that need it?
What Will Actually Power AI?
AI is sometimes presented as a straightforward nuclear-power story.
The reasoning is understandable: data centers need large quantities of reliable electricity.
But the actual supply mix is likely to be considerably more diverse.
Renewables, storage, natural gas, nuclear generation and grid reinforcement can all play different roles depending on geography, economics and project timelines.
This matters to investors.
AI electricity demand is not necessarily a bet on one generation technology.
It may be better understood as:
a structural increase in the value of available, deliverable and reliable power capacity.
Natural Gas: A Less Obvious Beneficiary
Natural gas has one major advantage in the United States:
much of the supporting infrastructure already exists.
Domestic production, pipelines, storage and established generation technologies can potentially provide dispatchable capacity on timelines that are difficult for some alternatives to match.
But the thesis needs to be stated carefully:
AI does not need natural gas. It needs reliable electricity that can be deployed quickly enough.
Gas is simply one possible solution to that requirement in certain markets.
Nuclear and Uranium: A Strong Story, But No Longer a Hidden One
Nuclear power matches several characteristics attractive to large data-center loads:
scale, high availability, firm generation and low operational carbon emissions.
It is therefore unsurprising that technology companies have shown increasing interest in nuclear supply.
But the investment story is now widely recognised.
Nuclear can remain a structural beneficiary of rising electricity demand without necessarily being an undiscovered opportunity.
Again:
A great economic story and a great investment at today’s price are not the same thing.
Copper: The Common Denominator
This brings the analysis back to the starting point of my commodity-supercycle work.
Almost every infrastructure layer discussed above uses copper:
data centers, transformers, cables, substations, transmission and distribution grids, power plants, motors, renewable generation and storage systems.
AI is therefore not the sole reason to be interested in copper.
It is another source of incremental demand arriving on top of several structural electrification trends.
I explored this relationship directly in Copper and Artificial Intelligence, while the broader supply-demand case is developed in Copper: The Strategic Commodity of the Decade?
This diversification of demand is important.
Even if some AI forecasts prove too optimistic, grids still need investment.
Even if some announced data centers are never built, electrification continues.
And even if one technology company loses the AI race, the copper already installed in the power system remains necessary.
Where Do the Most Interesting Constraints Sit Today?
The purpose of the following table is not to rank investments.
It is to identify where deeper analysis may be most useful.
| Value-chain segment | AI demand | Supply scalability | Execution difficulty | Other demand drivers | Theme maturity |
| GPUs | Very high | Medium | Medium | Medium | Very high |
| Memory | Very high | Low | Medium | Strong | High |
| Networking | High | Medium | Medium | Strong | High |
| Data centers | Very high | Medium | High locally | Strong | High |
| Cooling | Very high | Medium | Medium | Strong | High |
| Electrical equipment | High | Low–medium | Medium | Very strong | Medium–high |
| Transformers | High | Low | High | Very strong | Medium |
| Power grids | High | Very low | Very high | Very strong | Medium |
| Generation | Very high | Variable | High | Very strong | Variable |
| Natural gas | Indirect | Medium | Medium | Very strong | Low as an AI theme |
| Nuclear | Indirect | Very low | Very high | Strong | High |
| Copper | Indirect | Very low near term | High | Very strong | Medium |
| Technical skills | Cross-cutting | Low | Critical | Very strong | Low as an AI theme |
| Project delivery | Cross-cutting | Low | Critical | Very strong | Still relatively underrepresented in the AI narrative |
The key point is not that the bottom rows are automatically better investments.
It is that the physical and execution layers may deserve more analytical attention than they currently receive in mainstream AI narratives.
BlackRock and VanEck: Capital Is Already Moving Down the Stack
One of the most interesting findings from this research is that major asset managers are already moving beyond the simple “AI winners” framework.
BlackRock identifies power, memory, chips and data centers among the scarce inputs shaping AI investment opportunities, while also highlighting rising AI-driven demand for power grids.
That framing matters.
It shifts the question from:
Who develops the best model?
towards:
What scarce resources will every model ultimately require?
VanEck’s RACK ETF expresses a related idea through listed equities. The fund targets companies supplying the semiconductors, power, cooling systems and grid equipment required to support the data-center infrastructure build-out.
This does not mean BlackRock or VanEck will necessarily be right.
It does show that the institutional AI conversation is already moving down the infrastructure stack.
The centre of gravity is beginning to shift from the companies spending the CAPEX towards the companies and assets required to deploy it.
One layer, however, remains much harder to capture through a simple thematic equity basket:
execution capacity itself.
AI Is Also Becoming an Infrastructure-Finance Story
The change goes beyond public equities.
In August 2026, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create compute-financing platforms intended to mobilise more than $500 billion of third-party capital for AI infrastructure.
This is a remarkable development.
The AI boom increasingly resembles the financing of large infrastructure assets:
data centers, physical equipment, long-duration usage commitments, debt structures, asset-backed financing and long-lived cash-flow-producing infrastructure.
This changes the investment question again.
It is no longer simply:
Which stock should investors buy?
It also becomes:
Who finances the infrastructure, who owns the asset, who carries the execution risk, and who ultimately captures the cash flows?
For investors in infrastructure and private markets, that may prove to be one of the most important developments of the next phase of the AI cycle.
The Indicators I Would Watch From Here
If this thesis is correct, investors should gradually spend less time counting model launches and more time tracking how successfully capital is being converted into operational infrastructure.
I would focus on seven indicators:
- Hyperscaler CAPEX — not just the absolute number, but its growth rate and the return it produces.
- GW that have actually progressed — announced, permitted, financed, under construction, grid-connected, commissioned and operational.
- Locally available MW and GW — not national capacity, but actual capacity at the connection point.
- The quality of those MW — redundancy, resilience, reconfiguration capability and reinforcement requirements.
- Equipment lead times — transformers, cables, switchgear, turbines and other critical equipment.
- Interconnection queues — generation, storage and large-load connection pipelines, interpreted carefully according to what each dataset actually measures.
- Skills and project delivery — specialist hiring, wage inflation, contractor availability, commissioning resources and schedule slippage.
This gives us a much more useful investment funnel:
CAPEX → announced GW → financed GW → constructed GW → connected GW → resilient GW → commissioned GW → revenue → return on capital
Every arrow represents a potential delay, cost overrun or cancellation.
What Could Prove This Thesis Wrong?
A useful investment thesis should explain how it could fail.
The first warning sign would be a sustained reduction in hyperscaler CAPEX, particularly if AI-related revenue fails to justify current investment rates.
The second would be a much faster-than-expected improvement in the energy efficiency of AI models and computing hardware.
The third would be a rapid expansion of transformer, cable and electrical-equipment manufacturing capacity sufficient to eliminate current supply constraints.
The fourth would be a meaningful acceleration in grid connections and transmission development.
The fifth would be the emergence of persistent excess data-center capacity.
Finally, investors need to watch the economics of the CAPEX itself.
The transition of major technology companies towards much more infrastructure-intensive AI spending is already placing greater emphasis on free cash flow and return on invested capital.
The thesis does not require today’s bottlenecks to last forever.
It only requires them to persist long enough to drive several years of additional investment and shift part of the economic value towards the companies and assets capable of resolving them.
A Bottleneck Does Not Mean a Stock Is Cheap
This may be the most important investment distinction in the article.
A shortage of transformers does not automatically mean transformer manufacturers are undervalued.
Hundreds of billions of dollars of required grid investment do not automatically make every electrical-equipment company attractive.
And structural copper demand does not mean every copper miner is a good investment.
What this analysis identifies is where to look.
The next questions are company-specific:
How much exposure does the company actually have to the constraint?
How much additional capacity can it build?
What margins can it earn?
How durable are those margins?
What competition can emerge?
How capital-intensive is the expansion?
And how much of the expected growth is already embedded in the share price?
Ultimately:
What price are investors being asked to pay for that growth?
Only then can a structural theme become an investment thesis.
Conclusion: The Next AI Race May Be a Race for Execution Capacity
The first phase of the artificial-intelligence boom was dominated by models and semiconductors.
The next may be much more physical.
Hundreds of billions of dollars of CAPEX have to become data centers.
Those data centers become hundreds of MW and eventually GW of incremental electricity demand.
That power has to be generated and delivered.
The grid needs sufficient capacity and resilience.
Transformers, cables and electrical equipment have to exist.
Raw materials have to be produced.
And all of it has to be engineered, procured, constructed, integrated, tested and commissioned by people with the necessary skills.
This creates a fundamental asymmetry.
Capital can be mobilised in months.
Industrial capacity can take years.
Grid infrastructure can take longer.
And the experience required to deliver hundreds of complex infrastructure projects simultaneously cannot be manufactured overnight.
That asymmetry is the heart of the thesis:
Capital can be mobilised quickly. The equipment, grids and skills required to turn it into operational capacity cannot.
For investors, the question may therefore no longer be simply who wins the artificial-intelligence race.
It may increasingly be:
Where will the hundreds of billions of dollars of AI CAPEX actually flow? Which bottlenecks will prevent that capital from becoming operational assets? And which companies possess the industrial, technical and execution capabilities required to resolve them?
Artificial intelligence is a computing revolution.
But to transform the economy, it first has to be built, powered, connected and delivered in the physical world.
The next phase of the AI boom may therefore be less a race for capital than a race for execution capacity.
Primary Sources
- International Energy Agency — Key Questions on Energy and AI, 2026
- International Energy Agency — Building the Future Transmission Grid, 2026
- International Energy Agency — World Energy Employment 2025
- Lawrence Berkeley National Laboratory — Queued Up: 2026 Edition
- BlackRock Investment Institute — 2026 Midyear Global Investment Outlook
- VanEck — Data Center Supply Chain ETF (RACK)
- Reuters — reporting on hyperscaler capital expenditure, AI infrastructure financing and U.S. data-center development
Data and sources checked as of 17 August 2026. This article is intended for general educational and analytical purposes and does not constitute personalised investment advice.
Laisser un commentaire