NEW YORK, United States — The first great AI investment rush was built around chips, models and the companies developing them. The next one is increasingly about everything those systems need to operate.
Data centres are being built at extraordinary speed. Electricity has become a constraint. Cooling systems are becoming strategic equipment. Networking and memory are essential to linking increasingly powerful computing systems. Governments are also beginning to build their own AI infrastructure.
The shift is changing where investors are looking.
J.P. Morgan Asset Management estimates that hyperscale technology companies will spend about $700 billion on AI infrastructure in 2026, covering advanced chips, high-bandwidth memory, servers, power systems and new data centres. The bank has identified data centres, electricity, cooling and equipment as areas where investment opportunities could broaden beyond the companies most closely associated with AI models.
The scale of that spending is creating a market in which companies that may never develop an AI model can still become essential to the technology’s expansion.
The AI economy needs a physical foundation
Generative AI initially appeared to be predominantly a software story. The rapid development of large language models made computing power the critical commodity, putting semiconductor companies at the centre of the investment boom.
But increasingly powerful models require increasingly large physical systems.
Those systems need buildings, electricity, servers, networking equipment, storage and sophisticated thermal-management technology. They also require access to land and, crucially, connections to electricity grids.
J.P. Morgan estimates that hyperscaler capital expenditure will reach about $697 billion this year. The bank has described the financing of AI data centres as one of the defining capital-deployment themes of the current market.
That creates a different investment proposition from simply betting on which AI model will become dominant.
The companies supplying the infrastructure can benefit from the expansion of computing demand regardless of which application ultimately wins.
Electricity is becoming part of the AI trade
The biggest physical limitation may not be computing chips. It may be the electricity required to run them.
Data-centre construction is increasingly being shaped by the availability of power. J.P. Morgan has warned that electricity access, supply-chain limitations and permitting can determine whether projects are completed on schedule.
That makes power generators, grid equipment manufacturers, transmission businesses and companies capable of supplying electricity directly to large computing facilities increasingly relevant to the AI investment story.
The connection is already visible in major projects.
Nvidia has agreed to invest $1.5 billion in SB Energy and provide a guarantee of as much as $105 billion in support of an OpenAI data-centre project in Ohio. The facility is ultimately expected to provide as much as eight gigawatts of power, with 800 megawatts planned to be operational by 2028. SB Energy and SoftBank are also investing $4.2 billion in regional power infrastructure with AEP Ohio.
The numbers illustrate how the AI build-out is reaching beyond technology companies and into the traditional energy economy.
Cooling is becoming a major business
More computing power means more heat.
As AI servers become denser, conventional approaches to managing heat are being supplemented by more sophisticated liquid- and thermal-management systems.
One of the clearest signals came on August 31, 2026, when oilfield-services giant SLB announced a $4.1 billion agreement to acquire Kelvion, a company that manufactures heat-exchange and cooling equipment. SLB said the transaction would strengthen its data-centre business as demand for power and cooling infrastructure grows.
SLB expects its combined data-centre solutions business to generate between $4.5 billion and $5 billion in revenue by 2028, with $700 million to $800 million in adjusted earnings before interest, taxes, depreciation and amortisation.
The deal is significant not because SLB is abandoning its traditional energy business, but because it demonstrates how established industrial companies are positioning themselves around the infrastructure requirements created by AI.
Cooling, once an engineering concern largely invisible to technology investors, is becoming a commercial market of its own.
The network connecting the machines matters
The computing power inside an AI data centre does not operate as a collection of isolated machines.
Large AI systems depend on high-speed connections between processors, storage systems and servers. As workloads become larger, the movement of data inside the computing facility becomes increasingly important.
That has put networking companies and suppliers of related equipment in a position to benefit from the expansion.
Cisco, for example, has been deepening its relationship with Nvidia around AI data-centre infrastructure. The companies have expanded their work with Supermicro to offer high-density AI servers using both liquid and air cooling.
The broader opportunity includes networking equipment, storage, memory and the components required to move information rapidly through large computing clusters.
Andreessen Horowitz has also signalled the importance of this physical layer. The venture-capital firm recently established a $1.1 billion fund focused on hardware infrastructure for AI, including processors, memory chips, networking, data storage and robotics.
That investment reflects a broader recognition that AI’s next bottlenecks may increasingly occur outside the software layer.
Enterprise AI is moving from experiments to operations
Infrastructure is only half the story.
The other major shift is happening inside companies as businesses attempt to incorporate AI into ordinary work.
Research published in August using data from ChatGPT Enterprise found that AI use among organisations has been growing through both the arrival of new corporate users and increased usage by existing customers. The research examined more than 1,500 organisations and more than 17 million messages, finding applications across writing, technical work, communication and information synthesis.
That creates opportunities for businesses that help companies deploy AI rather than merely sell access to a model.
The next generation of enterprise AI is also increasingly focused on agents — software systems capable of carrying out tasks across applications rather than simply responding to individual prompts.
That change could create demand for software that manages permissions, monitors agents, protects corporate information and establishes controls around autonomous systems.
Security could become an AI growth market
The more authority businesses give AI systems, the more valuable security becomes.
AI agents can be granted access to customer information, source code and other business systems. That creates a new security problem: companies need to know what their AI systems are doing, what information they can access and what happens when they behave unexpectedly.
The market is already attracting capital.
Obsidian Security raised $85 million in a Series D funding round in August, giving the company a valuation of $1.1 billion. The company focuses on security for AI agents operating across enterprise applications.
Obsidian said nearly 70 per cent of its customers already allow AI agents to interact with business data.
That suggests another layer of the AI economy is emerging: not simply software that uses artificial intelligence, but software designed to control and secure the artificial intelligence being used by other businesses.
Governments are building their own AI capacity
Another potential source of demand is national investment.
Countries are increasingly treating computing capacity as strategic infrastructure rather than simply as another commercial technology service.
The European Union, for example, awarded a €387.8 million contract on August 31 to France’s Bull to develop LUMI-AI, a new AI supercomputer in Finland. The system is expected to become operational in the second half of 2027. It will use AMD AI chips, IBM storage and Nokia networking equipment.
The project is part of the European Union’s AI Factories programme. EuroHPC, the European organisation responsible for the programme, has allocated €8.2 billion through 2027 to develop 19 AI Factories across 12 existing supercomputers.
That spending creates another investment theme: AI sovereignty.
Governments that want domestic computing capacity need chips, data centres, power, networking, storage and cloud infrastructure. The result is a market that can extend well beyond Silicon Valley.
The opportunity comes with risks
The expansion of AI infrastructure does not guarantee that every company participating in it will make money.
Building a data centre requires enormous amounts of capital, and projects can face delays involving electricity connections, construction, equipment availability and regulatory approvals.
J.P. Morgan has highlighted those execution risks, noting that power availability, supply chains and permitting can materially affect project schedules and financing structures.
There is also the question of whether spending will continue at its current pace.
Companies have already committed vast sums to AI infrastructure, while investors are increasingly scrutinising whether the resulting computing capacity will generate sufficient returns. Recent analysis has raised concerns about the size of the capital expenditure cycle and the debt being used to finance some data-centre projects.
For investors, that means identifying demand is only the beginning. The financial strength of a company, the quality of its contracts, its competitive position and the economics of the infrastructure it supplies remain important.
The next AI winners may look less like AI companies
The most important change in the AI investment landscape may be that the definition of an AI company is becoming much broader.
A power company supplying electricity to a data centre may be participating in the AI economy. So may a manufacturer of transformers, a cooling-equipment supplier, a networking company, a memory-chip producer, a data-centre developer or a security business monitoring autonomous software agents.
The transition is already visible in the behaviour of investors and corporations. J.P. Morgan has described the opportunity as broadening beyond semiconductors into data centres, power, cooling and equipment, while venture capital is moving into physical AI infrastructure and industrial companies are acquiring businesses positioned to supply the data-centre boom.
That does not mean the original AI winners have disappeared. Chips and computing power remain fundamental to the industry.
But the economics of AI are becoming increasingly physical.
The first phase was about building the intelligence.
The next phase may be about building everything required to keep that intelligence running.
For investors, that could mean that some of the most consequential opportunities in AI will be found not in the next chatbot, but in the power lines, cooling systems, servers, networks, security platforms and industrial infrastructure behind it.
