The AI race is changing shape

Dr Mauricio Figueroa argues that the infrastructural and material changes of AI demand more environmental literacy in the law and tech community

For all the talk of chatbots and synthetic content, the AI revolution is becoming an extraordinarily physical undertaking. It requires legions of software engineers. It requires land and water. It requires electricity, and millions of highly sophisticated chips. And it’s also starting to require new ways of thinking about how to physically design the relevant devices and equipment, and where on Earth (or beyond) to put all that infrastructure. If AI is becoming more physical, technology law will have to become more environmentally literate.

Financial pressure doesn’t halt investment in AI infrastructure

To start with, there are financial concerns, but that doesn’t mean AI isn’t making money for technology giants. AI is good business; it just may not deliver the returns once imagined. MIT’s Daron Acemoglu contends that the AI impact on productivity within the next decade will be “quite modest”. Importantly, few consumers or AI users seem willing to pay for personal AI subscriptions. For instance, The Bank of America Institute, an in-house think tank established by the American financial institution of the same name, conducted a study and found that only 3% of households were paying for an AI service in 2026. The relatively low number of subscribers puts great pressure on enterprise sales and pushes for B2B deals. Some initiatives seem promising. For instance, Oxford University and the National University of Singapore, are two universities who have purchased institutional subscriptions for their staff and students to have access to ChatGPT Edu. More ambitious projects, like the OpenAI-Disney partnership faced greater difficulties, this particular project seems to have been called off, but AI providers keep exploring possibilities for expansion within existing pipelines of production across different industries.

The profits aren’t there yet, but neither is the willingness to risk missing the next breakthrough. The AI race is intense and unstoppable. As AI providers keep experimenting and pushing the frontiers of computational capabilities, the race demands substantial investment in the physical infrastructure that allows for those computational experiments to be conducted in the first place. As it is now well-documented, data centres consume substantial quantities of electricity and demand complex cooling systems. They produce noise and heat and can transform the communities in which they are built. For instance, during the heatwave in London in 2026, temperatures reached a little over 38 degrees at their maximum, but the exhaust air from  a data centre may match that easily. Academic studies of 2020 estimated discharged or exhausted air reaches 40 degrees for a data centre with an efficient air-cooling system, but with high-density equipment, it’s been reported that the exhaust air reaches hot aisle temperatures of over 50 degrees. Unsurprisingly, local resistance is growing. A 2025 Gallup survey found that seven in every ten Americans oppose the construction of data centres for AI in their local area, with only 7% strongly in favour. In Texas, the governor has announced an audit on data centres, halting approval requests currently under review.

Moving data centres elsewhere on Earth or beyond

The industry has two possible escape routes. One is outward, building in sparsely populated parts of America, or moving projects to other regions of the world. Last year I had the chance to discuss this topic with AI researcher and author Tamara Kneese and the award-winning journalist Diana Baptista in the SCL Podcast. The material ramifications of AI are expanding and reaching into the Global South. They are, however, equally found in Europe, where American technology companies are investing in data centre campuses. For example, in Spain, Amazon Web Services continues to invest in Aragon, where the AWS Europe Region is located, and is made up of three data centres in the provinces of Zaragoza and Huesca. The company has announced that it will increase its investments in Spain to 33.7 billion euros, and the first to benefit will be the Aragon data centre infrastructure, which it plans to expand and strengthen. Ten miles west of Heathrow, Slough has become one of the largest datacentre hubs in the world, hosting an estimated 40 huge facilities, many of them on a campus in the centre of town.

The other possibility is rather more extraordinary, making the projects in Aragon and Slough come across as out of fashion. Look up: SpaceX is developing “AI1” satellites designed to process data in orbit. The idea is to position computing infrastructure where satellites can remain in sunlight for almost the entire orbit, harvesting solar power, using large radiators to dispose of heat. Optical links between satellites could, in principle, turn a constellation into something resembling a data centre circling the Earth. This has made the scientific community to take a critical stance, because the satellites are likely to cause negative impact on the night sky if the project is executed without careful planning to reduce brightness. It is, at the same time, a defiant project that reminds us of the administrative law component of space wherein American institutions, particularly the Federal Communications Commission (FCC), play a major role in defining and allocating rights related to outer space whose impact goes far beyond American interests. Of course, having no regulator would be worse than this, albeit I’m sure that public international law scholars or space lawyers may have a wider view on the topic, as this is a domestic communications regulator making decisions that effectively alter the global commons, externalising the costs to  the rest of the planet.

The SpaceX project to build space data centres is attractive but not necessarily cheap. It’s attractive because a data centre in space does not have neighbours complaining about the noise, impact on water or electricity costs. Solar energy is plentiful. It is expensive, however, because of the launch costs. If this is an idea SpaceX wants to entertain further, they will need to invest heavily in research on the reduction in those  costs or to attain fully reusable rockets fit to take those satellites to the Earth orbit. Until that  happens, orbital computing must be understood more of an idea than a business strategy.

Changing shapes within the computers

Even if tech companies solve the problem of “space” as in where to put all its computers, it confronts another constraint inside the computers themselves. For roughly half a century, engineers made transistors smaller, fitted more of them onto a chip and, one generation after the other, computers became faster and more efficient. Or so the story goes. Then physics began sending the bill. Transistors had become so small that electrical current increasingly leaked through components even when they were meant to be switched off. This results in both wasted energy and failed devices. Each subsequent reduction in transistor has become harder and more expensive. So, chipmakers are changing direction.

Instead of squeezing ever more components next to one another on a flat piece of silicon, engineers are beginning to stack computing components vertically. IBM has demonstrated versions of this approach. Samsung also has pointed to that direction. But building upwards introduces its own engineering problems. Heat is one. Packing computing components on top of one another results in enormous quantities of heat in a tiny space. Engineers will need to think of effective methods to ferry heat out of the chip. Manufacturing is another. As IBM licenses the underlying IP rights to the relevant manufacturing partners, these will require the most advance fabrication plants and machinery to meet the extraordinary precision these vertical chips require, and only after that happens, would commercialisation  be feasible.

A change in the legal profession?

The AI race, then, is not slowing down so much as changing shape. The race for better and more profitable AI is increasingly becoming a race to redesign the physical world that makes AI possible. For technology lawyers, understanding AI will no longer mean solely understanding data flows, terms of service, IP rights or platform regulation debates. These are crucial topics, of course. But an effective legal interrogation of computational systems will increasingly demand environmental literacy, including a basic, but solid, understanding of electricity, water, land and the other material resources that make computation possible.

There is precedent for this kind of intellectual adjustment. Over the previous decade, legal scholars and practitioners invested enormous effort in understanding machine learning (ML) itself. Lawyers who had never been trained as computer scientists learned how computational systems make inferences from datasets and what it means to train or fine-tune a model. In fact, without that understanding, it would be hard to advance solutions or critiques on how ML processes may implicate existing rights. Some have gone  further, asking what happens when rights that appear relatively straightforward in legal doctrine become extraordinarily difficult to exercise against an algorithmic system, such as the right to deletion against a deployed ML system.

As technology lawyers have become keen to apprehend what happens inside computational systems, we have to start paying closer attention to the environmental and material systems that make computation possible in the first place. But we do not have to invent an entirely new field to do this. Quite the opposite. Environmental law has decades of scholarship and practice dealing with questions of energy, land use, pollution and environmental impact. The task is to bring that body of knowledge into much closer dialogue with technology law.

This should also change how we train the next generation of technology lawyers. They need not become environmental scientists or electrical engineers, just as the previous generation did not need to become computer scientists to understand ML. But they should acquire enough environmental and technical literacy to recognise the material consequences of the systems they advise on. In other words, the next generation of technology lawyers will need to be trained to cross the boundaries that, increasingly, AI itself is crossing.