AI is becoming a utilities business
The most revealing AI announcement this week did not arrive as a chatbot, benchmark or clever new feature.
It arrived as an electricity problem.
OpenAI has agreed to secure approximately eight gigawatts of IT capacity at the PORTS-Pike Technology Campus in Ohio. SB Energy will build, own and operate the site under a 20-year lease. NVIDIA will supply the computing infrastructure and provide financial support for the initial 4.25 gigawatts. The first 800 megawatts are expected to become available in 2028, with construction continuing in phases through 2032.
Eight gigawatts is an extraordinary figure. For perspective, the largest US power station by generating capacity in the Energy Information Administration’s published 2022 ranking was Grand Coulee at just over seven gigawatts. IT capacity and power-station capacity are not identical measures, and the Ohio campus will not arrive all at once, but the comparison gives the scale some shape.
This is no longer a story about putting more servers in a warehouse. It is an industrial system involving new power generation, high-voltage transmission, cooling, buildings, finance, permits and a long supply chain.
The AI business is becoming a utilities business.
The new stack starts below the chip
The familiar AI stack begins with models, chips, cloud platforms and applications. PORTS-Pike exposes another layer underneath it.
NVIDIA calls that layer “land, power and shell”: the site, energy supply and physical building required before any accelerator can do useful work. Its account of the deal says frontier laboratories can grow faster than their balance sheets and credit profiles can support. They may have customers and rising revenue, but still struggle to sign the decades-long infrastructure commitments needed to turn demand into available computing capacity.
That is why NVIDIA is not merely selling chips here. It is helping secure the conditions in which generations of those chips can be installed, financed and replaced.
The commercial logic is striking. A data-centre building and its power connection can last through several hardware cycles. The computers inside can be upgraded repeatedly. NVIDIA estimates that each generation deployed at the initial site could represent about 1.5 million GPUs and $150 billion to $200 billion in NVIDIA revenue. Those are NVIDIA’s projections, not guaranteed outcomes, but they explain why a chip company would care about transmission lines and lease payments.
The scarce component is no longer only the processor. It is the powered place where the processor can run.
Cheap AI rests on expensive commitments
Users experience AI as a box on a screen and, increasingly, as an inexpensive API call. The physical reality underneath is moving in the opposite direction.
PORTS-Pike is expected to require 10 gigawatts of new power capacity, including 9.2 gigawatts of natural gas generation, alongside $4.2 billion of transmission investment. The developers say grid-upgrade costs will not be passed to Ohio ratepayers, and the site will use a closed-loop, air-cooled system designed to minimise ongoing water demand. Development still depends on permits, environmental review, finance and infrastructure being secured.
There is a useful tension here. AI services may become cheaper per unit while the systems providing them become more capital-intensive, more concentrated and more dependent on long-term energy decisions.
That does not mean the investment is necessarily excessive. Demand for AI inference may continue to rise as models are embedded in search, software, customer service, research and internal operations. Nor does a signed capacity agreement mean every proposed building will appear exactly as described. Large infrastructure projects are phased because forecasts, funding, regulation and engineering change.
The sensible conclusion is not “AI will consume everything” or “efficiency will solve everything”. It is that apparently weightless software now has material constraints, and somebody has to finance them years before demand is fully known.
What this means for ordinary businesses
Most companies will never negotiate a power contract for an AI cluster. They will still feel the consequences.
First, supplier resilience is becoming physical. When assessing an AI platform, uptime and continuity depend on more than model quality. Capacity allocation, regional availability, energy cost, data-centre delivery and financing all sit behind the service. A provider can have an excellent model and still face a constrained supply chain.
Second, low prices can be strategic rather than permanent. Providers may price aggressively while securing scale, filling capacity or building market share. Businesses should test the economics of a workflow at realistic volume and retain room for price, rate-limit and model changes.
Third, concentration risk is growing. The same partnerships increasingly connect model developers, chip suppliers, cloud operators, energy developers and financiers. That can accelerate deployment, but it also creates dependencies that are difficult for customers to see from an API dashboard.
A practical AI procurement check should now ask:
Can the workflow move to another model or provider without being rebuilt?
Which parts genuinely require a frontier model, and which can use a smaller or local system?
What happens if price, latency, quotas or regional availability change?
Are prompts, evaluations and business rules stored in a portable form?
Is the value measured per completed task, not merely per token?
Which provider claims about sustainability, reliability and future capacity are independently verifiable?
This is not an argument for avoiding the largest AI platforms. Scale can produce better reliability, lower unit costs and access to capabilities smaller providers cannot match. It is an argument for understanding what you are buying.
Intelligence has a balance sheet
The public conversation about AI still tends to float between magical software and alarming machines. PORTS-Pike is more concrete. It is a former uranium-enrichment site being redeveloped into a vast computing campus, supported by new generation, transmission, finance and a 20-year customer commitment.
That tells us where the competition is going.
Model improvements still matter. Chips still matter. But the frontier increasingly belongs to organisations able to assemble an entire industrial chain and carry the risk for long enough to make it productive.
For everyone else, the lesson is wonderfully unglamorous: treat AI like infrastructure. Know the dependency, test the economics, design for portability and read beyond the feature announcement.
The intelligence may appear in a browser. Its foundations are made of concrete, copper, gas turbines and contracts.
Key takeaway
AI services may feel weightless and become cheaper to use, but they increasingly depend on concentrated, capital-heavy physical infrastructure. Businesses should evaluate portability, continuity and task economics alongside model performance.
