Blog

The AI Economy Is Becoming an Industrial Economy

AI investment is moving beyond models and chips into cooling, power and construction. The companies that can turn equipment into commissioned capacity are moving closer to the centre of value creation.

Listen to the summary

1:37 · Read by an AI voice

Six outlined server racks standing on a thick lime foundation bar, on navy.

One of the clearest signals about the next phase of AI did not come from a model release or semiconductor announcement.

It came from SLB.

In August 2026, the energy-technology company agreed to acquire thermal-management manufacturer Kelvion for $4.1 billion. Kelvion develops heat exchangers and cooling systems for data centres and other industrial applications.

The transaction may appear unusual for a company still widely associated with oilfield services. It becomes more logical when viewed through the physical requirements of AI.

SLB expects its existing data-centre business to exceed two gigawatts of cumulative deliveries by the end of 2026. Adding Kelvion’s thermal-management technology will more than double the amount of revenue SLB can pursue for every gigawatt of capacity it helps deliver. SLB: acquisition of Kelvion

The acquisition is not simply an attempt to diversify into a fashionable market.

It reflects a larger change in where AI investment is going.

The first phase of the boom rewarded companies that designed models, manufactured advanced chips and provided access to computing capacity. Those layers remain important, but every additional unit of computing now requires a larger system around it.

Electricity must be generated and distributed. Heat must be removed. Equipment must be housed, connected, monitored and maintained. Projects must obtain land, permits, components, skilled labour and grid capacity.

AI is still a software revolution.

But building it at scale is becoming an industrial undertaking.

The marginal AI dollar is moving outward

The visible centre of the AI market is the processor.

Its physical centre is the data centre.

The International Energy Agency estimates that global investment in data centres nearly doubled between 2022 and 2024, reaching approximately half a trillion dollars in a single year.

Electricity consumption from data centres reached around 415 terawatt-hours in 2024. The IEA expects it to more than double to approximately 945 terawatt-hours by 2030, with AI accounting for the largest share of the increase. IEA: Energy and AI

Only part of that investment reaches chipmakers.

A data centre also requires transformers, switchgear, uninterruptible power systems, backup generation, batteries, network equipment, chillers, coolant distribution, heat exchangers and control systems. It requires construction firms capable of assembling those components into a facility that can operate continuously.

Cooling alone can account for between 7% and more than 30% of a data centre’s electricity consumption, depending on the facility. The IEA expects cooling and other supporting infrastructure to represent around one-fifth of the net increase in data-centre electricity use through 2030. IEA: energy demand from AI

This does not mean value is simply moving away from chips.

It means the value chain is expanding around them.

A more powerful processor is commercially useful only if it can be powered, cooled and placed into operation. As computing density increases, the systems surrounding the processor become more technically demanding and economically significant.

The marginal dollar of AI investment therefore begins to travel outward—from the chip to the rack, from the rack to the cooling system, from the facility to the grid and from the completed project into years of maintenance and operation.

A data centre is becoming a factory

Traditional data centres could often be treated as standardized buildings filled with standardized computing equipment.

AI infrastructure is becoming more specialized.

Accelerated processors place substantially greater electrical and thermal loads into each rack. Air cooling becomes less effective as density rises, encouraging operators to introduce direct-to-chip liquid cooling, coolant distribution units and larger heat-rejection systems.

The facility begins to resemble an industrial plant.

Its input is electricity. Its output is computing capacity. Its productivity depends on equipment utilization, thermal efficiency, reliability and operating continuity.

This changes the unit through which suppliers compete.

Selling a chiller, pump or power-distribution component remains valuable. But customers ultimately need functioning capacity delivered on schedule.

Equipment must be designed around the processor architecture, validated against expected workloads, integrated with the rest of the facility and commissioned under real operating conditions. A technically superior component that arrives late or creates integration problems may be less valuable than a complete system that can be brought online predictably.

The unit of competition is shifting from the component to the commissioned megawatt.

This helps explain why SLB described the Kelvion acquisition in terms of expanding its revenue opportunity per gigawatt.

The company is not trying to sell one more component into a data centre. It is attempting to capture a larger share of the complete infrastructure system.

Industrial companies are assembling complete positions

SLB is not acting alone.

In 2025, power-management company Eaton agreed to acquire Boyd Thermal for $9.5 billion. Boyd supplies liquid-cooling technology used in data centres and was expected to generate approximately $1.7 billion of revenue in 2026.

The acquisition followed Eaton’s purchases of Fibrebond, which manufactures modular power enclosures, and Resilient Power Systems, which develops solid-state transformer technology. Eaton described the strategy as connecting power and cooling from the chip to the grid. Reuters: Eaton’s acquisition of Boyd Thermal

Vertiv announced a different but related transaction on the same day. It agreed to acquire PurgeRite, a company specializing in flushing, filtration and maintenance for liquid-cooling systems, for approximately $1 billion.

The target was not primarily a manufacturer. It provided the operational services required to make complex cooling systems function reliably. Reuters: Vertiv’s acquisition of PurgeRite

Schneider Electric has similarly expanded its liquid-cooling position through Motivair, which now provides cooling equipment, reference designs, spare parts, training and technical support for AI and high-performance computing facilities. Motivair by Schneider Electric

These companies are assembling combinations of technology, manufacturing, integration and service.

That combination matters because hyperscalers and data-centre developers face significant schedule risk. Every interface between suppliers creates another opportunity for incompatible specifications, delayed equipment or disputed responsibility.

An integrated provider can reduce the number of interfaces and accept responsibility for a larger part of the completed system.

This is a familiar industrial model.

Customers are not merely purchasing equipment. They are purchasing confidence that a complex asset will reach operation and continue performing after it does.

Project markets reward different capabilities

Technology companies are often evaluated through product-market fit.

Industrial companies must also demonstrate project-market fit.

The distinction is important.

A software product can be adopted by thousands of customers shortly after release. Distribution is largely digital, additional capacity can often be added incrementally and improvements can be deployed across the customer base.

Industrial infrastructure behaves differently.

A supplier must be included in facility designs, qualify for approved-vendor lists, reserve manufacturing capacity and coordinate delivery with a construction schedule. Its equipment may require certification, site-specific engineering, installation and testing.

Revenue arrives through projects that can be delayed, resized or cancelled. Growth can require additional factories, inventory, employees and working capital before the customer begins paying.

The companies positioned to succeed therefore need more than technological differentiation.

They need:

  • Manufacturing capacity that can expand without weakening quality.
  • Supply chains capable of supporting long project timelines.
  • Engineers who can integrate multiple systems.
  • Commercial teams able to influence designs before procurement begins.
  • Balance sheets that can support warranties and customer commitments.
  • Field-service networks that can maintain equipment after installation.
  • Credibility with developers, utilities, contractors and hyperscalers.

These capabilities take time to build.

SLB’s relevance to data centres does not come from having predicted the AI boom earlier than everyone else. It comes from having spent decades developing engineering, project-management and service capabilities for complex energy infrastructure.

The end market is new.

Many of the capabilities required to serve it are not.

The installed base can become more valuable than the first sale

The initial AI buildout is generating demand for equipment.

The completed infrastructure will generate demand for operations.

Liquid-cooling systems require fluid management, filtration, monitoring and maintenance. Electrical systems require inspection and testing. Pumps, batteries and power equipment need replacement parts and ongoing service.

Failure is unusually expensive.

If a cooling problem forces high-value processors offline, the cost is not limited to replacing a defective component. The customer loses computing capacity from an asset built to operate continuously.

This gives the service layer strategic importance.

Vertiv’s acquisition of PurgeRite illustrates the point. The company did not only expand its access to cooling technology. It acquired a capability required during commissioning and throughout the operating life of the system.

A supplier that installs equipment, monitors its performance and maintains it over time gains several advantages.

It generates recurring revenue. It develops operating data. It learns where systems fail. It becomes more deeply involved in the customer’s future facility designs.

The equipment sale creates an installed base.

The service relationship can turn that installed base into a long-term commercial platform.

This is also where digital technology returns to the industrial story.

Sensors, monitoring software, predictive maintenance and energy optimization can increase the value of physical equipment. But the software is most defensible when it is connected to proprietary operating knowledge and the ability to intervene in the real system.

The strongest industrial AI businesses may therefore combine hardware, software and field service rather than treating them as separate categories.

The bottleneck determines where value accumulates

AI infrastructure does not have a single permanent bottleneck.

It has a sequence of them.

Advanced processors were the most visible initial constraint. As chip supply expanded, electricity access, transformers, cooling equipment, construction capacity and grid connections became increasingly important.

The IEA estimates that approximately 20% of planned data-centre projects could face delays if power-sector bottlenecks are not addressed.

New transmission lines can require four to eight years to build in advanced economies. Waiting times for critical components such as transformers and cables have doubled in three years. New gas-turbine deliveries can now involve lead times of several years. IEA: data-centre grid constraints

A data centre cannot open when 95% of its equipment is available.

The missing 5% can determine the value of everything already purchased.

That creates unusual economics for suppliers occupying a critical position in the project. A component representing a small share of total construction cost may possess substantial pricing power if its absence delays billions of dollars of computing equipment.

The strategic question is not simply which category is growing fastest.

It is which capability currently determines whether capacity can be placed into operation.

That position can migrate.

A company benefiting from cooling shortages today may face competition as manufacturing expands. Transformer capacity may remain constrained for longer. Construction bottlenecks may shift by geography, while changes in processor design can alter the technical requirements placed on the facility.

Industrial winners must therefore do more than occupy a temporary shortage.

They need to convert that position into deeper customer relationships, larger system responsibility and an installed base that remains valuable after supply catches up.

Industrial exposure is not automatically an advantage

The AI label can make almost any company adjacent to electricity, construction or cooling appear strategically important.

Adjacency is not a moat.

Industrial expansion carries its own risks.

Factories and service networks require investment before demand is fully known. A supplier can add capacity just as customer projects are delayed. Large technology companies can revise capital budgets, change equipment specifications or pressure suppliers on price.

Customer concentration is another concern.

A relatively small number of hyperscalers and data-centre developers account for a substantial share of AI infrastructure spending. These customers have purchasing power, sophisticated procurement teams and the ability to develop equipment jointly with competing suppliers.

Technical standards are also evolving.

A cooling architecture designed around one generation of processors may need to accommodate very different rack densities several years later. Equipment that cannot be adapted may lose relevance even while the overall market continues growing.

The valuation of acquired businesses creates further pressure. Paying a substantial premium for exposure to a growing category only works if the buyer can accelerate revenue, improve margins or sell a broader system.

A backlog is not the same as completed, profitable revenue.

The genuine winners will be companies that can translate AI demand into acceptable returns after accounting for manufacturing investment, project delays, customer concentration and continuing technical change.

The opportunity extends beyond industrial incumbents

Large industrial companies have an advantage, but they do not own the entire opportunity.

As AI infrastructure becomes more complex, new software and service categories emerge around it.

Developers need better systems for site selection, grid interconnection, equipment procurement, construction coordination and commissioning. Operators need tools that connect workload demand to energy use, cooling performance and maintenance schedules.

The most valuable products will not merely display infrastructure data.

They will improve an economically important result.

A system might reduce the time required to bring capacity online. It might identify a thermal problem before equipment is damaged, improve energy use per unit of computing or help operators shift workloads when power becomes constrained.

The commercial language is different from ordinary enterprise software.

The customer is not buying another dashboard. It is buying greater asset utilization, lower energy consumption, less downtime or a shorter construction schedule.

For founders, this means the physical AI buildout should not be viewed only as a market for equipment manufacturers.

It is also a market for products capable of coordinating the companies, data and decisions surrounding expensive infrastructure.

But the product must fit the industrial environment.

It must integrate with established systems, produce evidence that engineers and operators can trust and remain usable by people working under real operational pressure.

AI is changing the definition of a technology company

The boundaries between technology, energy and industry are becoming less useful.

An energy-services company can become a data-centre infrastructure provider. An electrical-equipment manufacturer can acquire liquid-cooling technology. A cooling company can develop monitoring software and processor-specific reference designs.

These combinations are not exceptions to the AI economy.

They are becoming part of its structure.

At BraidNest, we see this as a product-strategy shift as much as an infrastructure shift.

The value of a technical capability depends on the complete system required to deliver it. As AI moves into the physical economy, product decisions must account for manufacturing, installation, integration, maintenance and the operating environment surrounding the technology.

The next winners may look different from the companies that defined the first phase of AI.

They may have factories rather than consumer interfaces, engineering backlogs rather than user-growth charts and field technicians rather than purely digital distribution.

That does not make AI less of a technology story.

It means technology is becoming inseparable from the industrial system required to put it to work.

The AI economy is not leaving software behind. It is discovering how much industry is required to make software operate at scale.

  • Markets

All articles