Not a Medium member? You can read it for free using the official Friend Link.

For years, the defining story of AI infrastructure was the race to secure GPUs. Companies competed for Nvidia accelerators, high-bandwidth memory, advanced networking equipment, and the capital required to assemble enormous computing clusters.

Now an AI company is reportedly spending billions of dollars on something that seems far removed from language models: gas-fired power generators.

On November 8, 2026, the Financial Times reported that Anthropic had agreed to a roughly $3.4 billion deal involving commitments for 1.7 gigawatts of Caterpillar generator capacity, originally ordered by Texas-based OnePWR Solutions. The arrangement reportedly includes engineering and construction capabilities associated with installing the equipment.

Reuters covered the report, while noting that it could not independently verify the transaction and that the companies had not responded to its requests for comment.

The reported deal is remarkable not simply because of its size, but because of what an AI company would be securing. Anthropic isn’t acquiring another model architecture or faster accelerator. It is reportedly moving deeper into the physical infrastructure required to operate the compute it expects to need.

That makes more sense when you consider a question increasingly shaping AI infrastructure planning.

How many accelerators can a company actually power?

A GPU Is Only One Layer of Compute Capacity

From a user’s perspective, an AI product is almost entirely software. You submit a prompt, a model processes it, and a response appears. The physical infrastructure is hidden behind an API, much as cloud computing hides servers and networking equipment from the applications running on top of them.

But the hardware requirements underneath modern AI systems are becoming difficult to ignore.

Training and serving large models requires accelerators connected through high-speed networks, installed in racks, and supported by electrical distribution and cooling systems. The facility must have sufficient grid capacity or another dependable source of electricity. That electricity must also be delivered at the voltage, reliability, and power quality required by the equipment.

The dependency chain looks something like this:

Every layer depends on the capacity and readiness of the infrastructure beneath it.

A company might have purchased thousands of GPUs, secured the necessary networking equipment, and completed a data-center building. Yet if the facility cannot obtain enough electrical power, the available hardware cannot become fully usable computing capacity.

There is also an important difference between the capacity of a power-generation project and the compute capacity it can support. A generator rated at one gigawatt does not translate directly into one gigawatt of usable IT load. Cooling, electrical conversion, facility operations, redundancy, and other factors affect the relationship.

That distinction matters because the next phase of AI infrastructure isn’t simply about buying more hardware. It is about assembling complete systems that can operate that hardware reliably.

Anthropic acknowledged the scale of the challenge in February 2026, stating that training a single frontier model would soon require gigawatts of power. The company also announced commitments intended to cover grid-upgrade costs associated with connecting its data centers and address electricity-price impacts on consumers.

The scarce resource is increasingly not just an accelerator.

It is deployable compute with dependable power attached.

Electricity Doesn’t Scale at Software Speed

The difficulty becomes clearer when we compare the speed of technology deployment with the timelines of the electricity industry.

A company can provision additional cloud instances relatively quickly when capacity is available. New accelerators may take much longer to manufacture and acquire, but once the equipment reaches a prepared facility, installation can proceed within an established deployment process.

Building new electricity infrastructure follows a different schedule.

Power plants require financing, equipment, construction, fuel or resource arrangements, and regulatory approvals. Grid connections may involve new substations or transmission infrastructure. Large transformers and switchgear can have lengthy manufacturing lead times, while proposed projects compete for limited construction and engineering capacity.

According to the International Energy Agency’s 2025 Energy and AI report, new transmission lines in advanced economies can take roughly four to eight years to build. Waiting times for some critical grid components have also increased substantially.

Meanwhile, electricity demand from data centers is expected to rise sharply.

Projected Global Data-Center Electricity Consumption

According to the International Energy Agency (IEA), global data-center electricity consumption is expected to increase significantly by 2030.

• 2025: 485 TWh Estimated annual electricity consumption.
• 2030: 950 TWh Projected annual electricity consumption.

That’s nearly a 96% increase in just five years.

These figures cover all data centers, not exclusively AI workloads.

Source: International Energy Agency Key Questions on Energy and AI, April 2026.

The IEA’s April 2026 update projects global data-center electricity consumption increasing from approximately 485 terawatt-hours in 2025 to 950 TWh by 2030. Electricity use at AI-focused facilities is expected to grow even faster than the overall data-center sector.

These are projections, not guaranteed outcomes. Investment conditions, efficiency improvements, equipment shortages, and the pace of AI adoption could change the trajectory.

Still, they illustrate why electricity has become a strategic constraint.

The industry is facing a mismatch between how rapidly companies want to expand computing capacity and how quickly the energy system can deliver the infrastructure required to support it.

That also explains the economic logic behind securing generation equipment directly. If the normal procurement process means waiting years for the necessary power capacity, access to equipment arriving sooner may have enormous strategic value.

A company making such an arrangement is not merely buying future electricity. It may also be buying a shorter path to operational computing capacity.

The Cloud Abstraction Is Meeting a Physical Limit

Software engineering has spent decades developing abstractions that allow programmers to ignore underlying infrastructure.

Developers no longer need to manage individual machine instructions for ordinary application work. Cloud platforms let teams deploy applications without purchasing servers, and serverless systems make much of the execution environment disappear from the programming model.

These abstractions work because someone else manages the infrastructure underneath them.

AI scaling is revealing what happens when an underlying resource is no longer sufficiently elastic.

The problem resembles familiar situations in software architecture. A database abstraction is convenient until query patterns reveal an expensive access path. A message queue appears simple until ordering and duplicate delivery become important. A cloud region feels interchangeable with another until latency, compliance, or capacity constraints make location significant.

The abstraction hasn’t necessarily failed. The resource underneath it has become important enough that engineers can no longer ignore its behavior.

Something similar is happening with AI compute.

A company building a large model needs to consider accelerator utilization, rack power density, cooling capacity, and facility reliability. At sufficient scale, those discussions extend to substations, grid connections, equipment manufacturers, and electricity-generation contracts.

This is a different kind of full-stack engineering from what software developers usually mean by the term.

The stack extends all the way to the energy required to make an instruction execute.

And unlike software resources, electrical infrastructure cannot always be expanded simply by allocating more money to a cloud account.

At Hyperscale, Inefficient Software Becomes Physical Infrastructure

The most interesting consequence for software engineers is how these physical constraints change the value of efficiency.

Imagine an AI application using a large model for every request, even when many tasks could be completed by a smaller model. At modest scale, that decision might result in an unnecessarily expensive API bill.

At enormous scale, the same design choice can propagate throughout the infrastructure.

More computation requires additional accelerator time. Higher sustained utilization may require more accelerator capacity, networking equipment, electrical power, and cooling. If the existing facilities cannot accommodate that demand, expansion may involve new buildings or additional energy infrastructure.

Software inefficiency can eventually become a physical construction problem.

That makes techniques such as model routing, batching, caching, quantization, and efficient inference serving more strategically important.

Consider an application that repeatedly submits the same large context to a model. If the serving system can reuse cached computation, it may avoid performing some of the same work again. Similarly, routing straightforward requests to a smaller model can reduce resource consumption while preserving a more capable model for tasks that genuinely require it.

Batching can improve hardware efficiency by processing compatible work together, although it must be balanced against latency requirements. Quantization may reduce memory use and improve serving efficiency, but it requires evaluation of the tradeoffs in model quality and hardware performance.

None of these optimizations is universally beneficial. Their value depends on workload characteristics, service-level objectives, model behavior, and the infrastructure being used.

The engineering goal is not simply to use fewer GPUs or minimize electricity consumption regardless of product quality. It is to deliver useful work reliably while making efficient use of compute, power, and capital.

This becomes even more important as models support richer and more computationally expensive tasks. The IEA reports that AI-server power density increased elevenfold between 2020 and 2025, with further rapid growth expected.

Those trends mean software architecture and physical infrastructure planning are becoming more tightly connected.

At sufficient scale, reducing unnecessary token generation or improving inference efficiency can influence how much hardware a company needs to deploy in the first place.

The Geography of AI Is Becoming an Energy Decision

Electricity availability also affects where future AI infrastructure can be built.

Data-center operators have always considered land, networking, electricity prices, climate, latency, and regulatory requirements when choosing locations. But a site with excellent connectivity and favorable economics is less attractive if the grid cannot provide the additional power required for years.

This creates an unusual possibility: rather than expecting energy infrastructure to expand wherever computing companies prefer to build, some computing capacity may increasingly move toward regions where sufficient energy is already available or can be developed sooner.

That decision has technical consequences. Moving computation affects latency, data-transfer costs, regional resilience, data residency, and the relationship between training clusters and inference services.

The solutions are also becoming more diverse.

In January 2026, Meta announced agreements intended to support up to 6.6 GW of existing and future nuclear generation by 2035. In October, Google and Constellation announced an agreement targeting 890 megawatts of additional nuclear capacity through improvements at existing plants.

These efforts are different from Anthropic’s reported gas-generator agreement, but they reflect the same underlying pressure. Companies are making long-term electricity commitments because energy availability is increasingly tied to AI infrastructure expansion.

Natural gas, nuclear, renewables, storage, grid expansion, and demand flexibility are all part of the possible response. Each brings different costs, construction timelines, reliability characteristics, environmental consequences, and regulatory constraints.

No single energy technology explains the entire AI infrastructure race.

The deeper issue is the scale of capacity companies are trying to make available — and the difficulty of delivering it quickly enough.

AI Has Reached the Physical World

The reported Anthropic deal sounds unusual only if AI is still imagined primarily as a software industry.

At frontier scale, the industry is increasingly defined by the physical resources required to operate its models. Accelerators need networking, racks need cooling, facilities need electrical distribution, and the entire system needs dependable generation capacity.

The GPU race hasn’t ended. Advanced chips, high-bandwidth memory, networking, and capital remain crucial constraints.

But the race has expanded beneath the chips.

A company that secures enough accelerators without securing the infrastructure to operate them hasn’t completed its capacity plan. And a company that builds enormous facilities without sufficient utilization or economic returns hasn’t necessarily created a successful AI business.

The next phase will reward organizations that can coordinate all these layers: hardware, energy, infrastructure, software efficiency, location, and economics.

The AI race began by asking who could secure the most compute. Increasingly, it asks who can turn enough electricity into useful compute at the right cost, in the right place, and at the right time.

Once competition reaches the scale of gigawatts, AI starts looking less like software running on expensive servers and more like an industrial system whose final product happens to be intelligence.

The AI Infrastructure Race Has Moved From GPUs to Gigawatts was originally published in Skill Stuff on Medium, where people are continuing the conversation by highlighting and responding to this story.