AI datacenters putting zero emissions promises out of reach

The datacenter industry looks set for a turbulent 2025 as AI growth threatens to trump sustainability commitments and authorities are likely to see growing public hostility to new projects.

In addition, an overhaul of the way power is supplied and distributed to server farms seems increasingly necessary.

These findings come via the latest report from Uptime Institute, which offers predictions for the year ahead based on its research into the latest developments and challenges shaping the industry.

Perhaps not surprisingly, all five predictions relate to the huge growth in demand for more and more datacenters to support the anticipated boom in AI and associated cloud services.

Uptime's first forecast is that server farms will become more controversial, as their resource use and greenhouse gas (GHG) emissions lead to greater local opposition over new builds, while governments are more concerned with reaping any economic benefits.

The Reg has extensively covered the dire warnings for datacenter energy consumption resulting from AI, with analyst Gartner recently saying this could expand 160 percent over the next two years. Datacenters' use of water and land are other bones of contention, which in combination with their reliance on tax breaks and the limited number of local jobs they deliver, will see them face growing opposition from local residents and environmental groups.

Uptime highlights that many governments have set targets for GHG emissions to become net-zero by a set date, but warns that because the AI boom look set to test power availability, it will almost certainly put these pledges out of reach.

Governments say yes, residents say no

Many governments seem convinced of the economic benefits promised by AI at the expense of other concerns, the report notes. The UK is a prime example, this week publishing the AI Opportunities Action Plan and vowing to relax planning rules to prioritize datacenter builds.

Uptime also forecast that datacenter operators will have to become active participants in managing energy grids, because of their large and growing demand. It foresees that operators will face difficulties in buying or generating power without closer collaboration with utility companies, and suspects that server farms may need to provide or store power, and be willing to shed loads when required.

Microsoft, for example, has deployed "grid-interactive UPS technology" at its Dublin campus which allows the energy storage systems installed for backup power to feed energy back to the grid if required. This is intended to help smooth out any variability in the power supply due to the variable nature of renewable energy sources.

The cloud and software biz is also constructing a private gas-powered power plant at the same campus, so its infrastructure can keep up and running during times of peak power demand. Uptime says this trend is likely to see broader adoption as operators try to offset some of their investment by charging for excess power sent to the local grid.

Time to rebuild

This feeds to Uptime's next forecast: datacenters will need a radical overhaul internally because of the growing energy demands of AI infrastructure. AI training is already pushing rack densities towards a level normally seen in supercomputing facilities, and the report says kit built around Nvidia H-series GPUs is hitting 40 kW per rack.

Increasing rack power presents several challenges, the report warns, including the sheer space taken up by power distribution infrastructure such as switchboards, UPS systems, distribution boards, and batteries. Without changes to the power architecture, many datacenters risk becoming an electrical plant built around a relatively small IT room.

Solving this will call for changes such as medium-voltage (over 1 kV) distribution to the IT space and novel power distribution topologies. However, this overhaul will take time to unfold, with 2025 potentially a pivotal year for investment to make this possible.

Another prediction is that AI models will be trained in the cloud, with enterprises using public cloud resources rather than procure and deploy their own dedicated GPU server clusters.

This view contradicts Canalys chief analyst Alastair Edwards, who last year claimed that deploying large-scale AI models in the cloud would quickly become "unsustainable from a cost perspective." He believes enterprises will instead turn to colocation and specialized hosting providers such as Coreweave and Foundry, which are also mentioned by Uptime.

• AI frenzy continues as Macquarie commits up to $5B for Applied Digital datacenters
• Amazon splashes $11B on AI datacenters in Georgia
• AI spending spree continues as Microsoft commits $80B for 2025
• AI's thirst for water is alarming, but may solve itself

The report's final forecast is that datacenter operators will start to give greater consideration to alternatives to Nvidia's energy guzzling GPUs for AI processing. There are signs that the AI hardware market will become more diverse in 2025, Uptime claims.

Inferencing requires a fraction of the compute necessary for model training, and GPUs may not be necessary here. Alternative AI hardware such as that from Cerebras and SambaNova could see greater traction, the report adds.

2025 will test datacenter operators' ability to navigate the opportunities and uncertainties presented by AI, Uptime concludes. They will need to balance this and still maintain existing service level agreements, achieve sustainability goals, and meet financial objectives. ®