Can the AI Compute Boom Keep Growing? Power, Chips and Partnerships

[Emily87]

Behind every AI chatbot, coding agent and image generator sits an enormous amount of computing power. Demand for that power has grown so fast that it now shapes deals between some of the biggest companies in technology. Data Center Dynamics reported in September that Anthropic alone had signed around $517 billion in compute agreements over roughly eleven months, adding up to about 14.8 gigawatts of capacity. Rivals are renting capacity from each other, as the Wall Street Journal noted this week. So can this boom keep growing, and what will it take?

The scale is remarkable. Training a frontier model can use tens of thousands of specialised chips running for months, and serving millions of users requires even more capacity over time. AI agents add to the load, because they run longer tasks with many steps. GitHub reported that activity on its platform more than doubled in a year, partly because of agents. Every part of the computing stack, from chips to storage to networking, is being pushed to grow faster than ever before

Electricity is one of the biggest constraints. The International Energy Agency estimated in 2025 that data centres used around 415 terawatt hours of electricity in 2024, about 1.5% of global consumption. It projected that figure could roughly double by 2030, with AI the main driver. That is a lot of power, but it is also important to keep in perspective. Data centres remain a relatively small share of global electricity use compared with industry, transport and buildings

The positive side is that AI demand is driving huge investment in new energy sources. Technology companies have signed deals for solar, wind, geothermal and nuclear power, including agreements to restart or extend existing nuclear plants and to buy power from future small modular reactors. Some of these projects would have struggled to find funding without AI demand. If managed well, the AI boom could speed up the shift to cleaner electricity rather than slow it down

Chips are the second constraint. Advanced AI processors are made by a small number of companies, and manufacturing capacity takes years to expand. That scarcity explains the partnerships between rivals. When nobody can build fast enough, renting spare capacity from a competitor makes sense for both sides. Anthropic renting xAI's Colossus 1 facility in Memphis, reported by Forbes, is one striking example. Google's custom TPU chips, Amazon's Trainium and other in house designs are also helping to widen supply beyond a single dominant supplier

Efficiency gains are an encouraging counterweight. Each new generation of chips does more work per watt. Software techniques such as better model architectures, smaller specialised models and more efficient training methods reduce the computing needed for a given result. Meta's CRAM memory compression work, which we discussed this week, is a small example of the kind of engineering that squeezes more out of existing hardware. Historically, efficiency improvements have been enormous in computing, and there is no sign that trend is over

There are reasonable questions about sustainability in the financial sense too. Commitments worth hundreds of billions of dollars assume that AI revenue will grow quickly enough to pay for them. If demand slows, some of those contracts could become a burden. The 90 day exit clauses reported in some deals show that companies want flexibility. Careful planning, rather than unlimited spending, will help the industry avoid painful corrections

Local communities matter as well. Data centres bring investment, construction jobs and tax revenue, but they also need land, water for cooling in some designs and grid connections. In some places, residents have raised concerns about noise, water use and electricity prices. The best projects work closely with local authorities, invest in grid upgrades and use efficient cooling. The UK is seeing interest in AI growth zones, and getting these relationships right will be important here

My overall view is that the compute boom can keep growing for some years, but its shape will change. Efficiency gains will matter more, partnerships and shared infrastructure will become normal and energy investment will be a central part of every plan. The most successful companies will be those that grow capacity responsibly, working with communities and energy providers rather than against them. If that happens, AI could leave behind a stronger, cleaner energy system as well as smarter software

The UK has some real opportunities in this area. It has strong research universities, a growing number of AI companies and plans for AI growth zones to speed up data centre development. Grid connections and planning delays are the main obstacles, and both are being worked on. Cooler weather helps with cooling efficiency, and offshore wind provides a large source of clean power. Getting this right could bring investment and skilled jobs to many parts of the country

There is also a role for smarter scheduling. Some AI workloads, such as training runs and batch processing, do not need to happen at a specific time. Running them when renewable power is plentiful and demand is low can reduce both costs and emissions. Several companies are already experimenting with this kind of flexible computing. It is a simple idea with a lot of potential as grids become more dependent on wind and solar

Do you think the AI compute boom is sustainable? Is electricity or chip supply the bigger challenge? And would you welcome a large data centre near where you live?

Anthropic Signed 517Bn In Compute Agreements In Past 11 Monthswww.datacenterdynamics.com