Market predictions of rising power requirements from new data centre construction mean that all sectors of the energy industry are eyeing up opportunities to satisfy demand. The renewables sector is no exception, with the industry’s shorter time-to-production metrics potentially helping to fill the gap left by the high demand for and increasing costs of ‘traditional’ data centre power sources, such as gas turbines.
Much of the latest boom in data centre development is created by technology ‘hyperscalers’ (Google, Meta, Oracle, X and Microsoft) commissioning new facilities to satisfy the needs of two main customers: Anthropic – perhaps best known for its Claude models – and the company that set the world on its AI course, OpenAI, whose ChatGPT models made global headlines in 2022. The hyperscalers have promised up to $2 trillion for data centre development in the next four years.
Despite the existence of seemingly hundreds of thousands of AI companies that have sprung into existence in the last four years, the vast majority of them comprise of services based on the biggest large language models as developed and run by Anthropic and OpenAI (Elon Musk’s Grok, Google Gemini and Meta’s models are minor players in the AI market). Even relatively well-known names such as AI legal services outfit Harvey.ai are, at their core, providing a specialist ‘wrapper’ around the LLMs from Anthropic, Google and OpenAI.
That concentration of the data centre market around the computing needs of two primary customers should give every sector considering investment in the AI industry serious reasons to consider any commitment. In addition to the financial condition of the “big two” – which I cover in more detail below – there is a host of detail to the current data centre construction boom that throws a quite different light on the long-term prospects for AI data centres, and by proxy, the AI industry’s suppliers.
Construction pains
It’s important to note that building a data centre to be used for AI requires what are very much a new set of skills, methods and equipment. These are different from those needed for ‘traditional’ data centres like those that power cloud computing services. As a result, AI data centre construction is generally slower than the predictions made at a project’s outset. The larger the facility, the longer it takes to build. For example, the 1.1GW Stargate Abilene Texas facility was declared open at the end of 2025, yet only two of the eight buildings were complete.
Indeed, ‘under construction’ in the data centre industry can describe a variety of states, from plot acquisition and planning permission, through to land clearance and preparation, to the completion of one part of many (larger data centres comprise of several discrete data centres joined into a campus). Rather than the optimistic predictions of 18 month-long projects perhaps made by commissioning hyperscalers, a two-and-a-half to three-and-a-half year timescale seems to be more realistic.
With specialist contractors and equipment in short supply, construction costs have tended to rise steeply, and the final equipping of the nascent data centre with racks of GPUs, cooling and networking equipment brings further costs that have risen over the course of construction. This is most apparent in the cost of silicon-based products such as computer memory, which have doubled in price thanks to the insatiable demand of the AI industry. The beating heart of AI data centres, the GPU hardware – supplied largely by the world’s biggest company, NVIDIA – has been doubling in price with each annual iteration, rising from 2022’s DGX A100 costing around $200,000 per unit, to $500,00 per Blackwell SuperPod in 2024.
In short, building an AI data centre is expensive and slow, and the slower the construction, the higher the final cost for a finished facility.
There’s also the issue of local opposition to any new data centre construction project. While a local area may see a temporary uptick in construction job opportunities during the years of building, once running, finished facilities need few long-term staff (perhaps in the 100s for even the largest campus). Protests about drinking water use by a proposed data centre, the potential for higher electricity bills in the area, and the noise huge data centre campuses make mean that many proposed sites are stuck in legal challenges, political wrangling and planning permission delays.
The two customer problem
Anthropic and OpenAI are not only unprofitable, they run at a huge loss. Without third-party funding (private equity, venture capital, and dubious circular financing deals from the hyperscalers themselves), both companies would not be able to function. Their commitments to use the computing power housed in data centres mean that between them, they will need to generate around $1tn of income by 2030 to pay the hyperscalers, cover their running costs (which include hugely expensive training runs for new models and on-going, so-called ‘post-training’ to tweak existing models) and, of course, pay their many creditors.
This situation could be passed off as how markets work: companies borrow money to establish a business, and over time, demand for their services grows. Charged prices cover costs and any financial commitments, and hopefully make their owners some profit, somewhere down the line.
The problem with a boilerplate financial summary in the particular instance of AI is twofold. Firstly, although there are billions of dollars in monthly demand for ‘AI’ (and by this I mean the large language models operated by Anthropic and OpenAI), it is insufficient at present to make a viable business, and the signs are that demand for AI is slowing.
The second fly in the ointment is the likely cause of the current slowdown in use of AI: a necessary change in the way that users were made to pay for AI. From 2022, both companies offered AI on a subscription basis, one that will be familiar to any user of modern, cloud-based business software. For as little as $20 per user, per month (up to and beyond $200 a month for a business plan) users of AI could use much more computing power than they were paying for. A $200 per month business subscriber could use many thousands of dollars-worth of computing (as much as $4,000 according to some reports), with Anthropic or OpenAI subsidising the difference.
When, at the end of 2025, both companies began to charge their heaviest AI users on a ‘per-token’ basis, organisations went from encouraging their employees to use as much AI as they could (so-called token-maxxing) to attempting to limit the type and frequency of workers’ queries to ChatGPT or Claude. The rate of rise in demand for AI fell as every token spent had an attributable cost. (A token is around two-thirds of a single word, either input or output from an AI). Agentic AI and ‘looping’, in which software agents repeatedly query the large language model to get to some outcome, were particularly blamed. Without meaningful measures of increased productivity that could be placed at the door of AI use, many organisations are beginning to analyse the costs and benefits of new, AI-powered software.
The bright future
Despite these growing pains, many in the technology industry remain optimistic about the future of what we term AI – although large language models would be a more accurate description of what’s commonly on offer. i) Anthropic and OpenAI will increase their customer base and those customers will be willing to pay at least six times more than they do currently for AI. ii) The investments made in data centres will pay off despite the rising costs and delays, and the hyperscalers building data centres will get paid by the ‘big two’ to use the racks of GPUs. iii) The large language models developed by Anthropic and OpenAI will continue to evolve and improve, increasing the number of use-cases for AI that every industry will soon discover. This is the current ethos among Silicon Valley and its many investors. It’s also the message promoted by marketing departments with skin in the AI game, not least those operating out of OpenAI and Anthropic.
If an organisation is considering investing its time and money in AI data centres and their surrounding infrastructure (water, power, supply chain, tertiary industries, etc.) then it needs to trust that two companies will make good on a business model that runs at a massive loss and relies on huge loans to keep afloat. Over the course of the next few years, the ‘big two’ will need to bring in trillions of dollars of revenue (note, the current non-AI software market is worth around $800bn annually), and, therefore, the slow and expensive process of building new data centres will be worthwhile.
The latest generation of data centres, designed for what we now call AI, are not suitable for anything other than running LLMs. Unlike the dot-com bubble, which at least led to the widespread rollout of digital fibre infrastructure, AI data centres have only one use. Organisations in the utilities sector or data centre supply chain need to be absolutely sure that AI data centres and the companies that will use them are viable businesses in the long term.