Huge grid and heat challenges ahead as Nvidia set for 1MW rack

The impending growth of datacentre rack power draw to 1MW within two years will bring a step change in energy use and potential to waste massive amounts of heat.

This means datacentre operators must become responsible partners to the electricity grid, and with just one 1MW rack set to produce as much heat as 200 5kW ovens, the datacentre industry, government and local authorities must remove obstacles to heat re-use.

Those are the views of Schneider Electric vice-president, secure power and datacentre division, UK and Ireland, Matthew Baynes, who spoke this week at the Datacentre World event in London about the energy issues around so-called artificial intelligence (AI) factories.

“We’re seeing a major shift into very, very extreme high-density applications with Nvidia GPUs [graphics processing units]. This poses extreme challenges for us as an industry. But the number one challenge is energy production, access to energy and being responsible with this critical asset of electricity,” said Baynes.

“Recently, we were designing up to 10, 15, 20, 40kW per rack in traditional datacentres housing cloud applications, but they are moving to becoming AI factories, so that predictable nature we had is harder for us to understand. The pace at which Nvidia is moving with the GPU technology makes it very, very difficult,” he added.

Here, the context is Nvidia’s roadmap for GPU products. While right now the densest datacentre racks running Nvidia Blackwell GPUs draw well under 200kW, that’s set to multiply by five or six times from around 2028.

Firstly, Nvidia’s Rubin and Rubin-plus GPUs are set to take rack power draw from around 240kW this year to 600kW-plus in 2027. Then will come Nvidia’s Feynman GPU hardware, which will attain 1MW in a rack. Their power draw will be up to 2kW, each, with 576 per rack.

From a technical point of view, this will mandate direct-to-chip liquid cooling, where air cooling has sufficed until now. It will also require a move to 800V DC power distribution to the rack, which is where Schneider’s product set intersects with the issue. 

At the high level, this all equates to massive and rapidly increasing demand for power, with new datacentre capacity projected to reach 240GW globally by 2030. 

Datacentres need to be a stable asset on the grid, not one that is causing disruption
Matthew Baynes, Schneider Electric

For the datacentre industry, Baynes said, that translates to the need to mitigate energy supply constraints, which include flexible use of the grid and making best use of the heat produced in datacentre operations.

“I was at a conference yesterday with ministers, and the grid situation is challenging, with something like 8GW of applications on the London grid for datacentres alone,” said Baynes. 

“Datacentres need to be a stable asset on the grid, not one that is causing disruption, not one that is having dynamic heat loads on and off. You’ve got to offer technology and be an active asset to the grid so that datacentres can run at their most efficient and be the least disruptive they can be to the national grid,” he added.

“We also need to innovate on heat re-use,” said Baynes. “We need to harness some of that hybrid waste heat. We need to work with communities and see how we can leverage that and actually be part of society as well as an industry.

“The challenges are government and regulation, [the lack of] district heat networks, the planning permission for that, and then having off-takers for it. It’s not as simple as grabbing the heat and putting it into a swimming pool. There’s a lot more in between those two equations that need to happen, a lot more technology that needs to be implemented.”

Related posts

Microsoft Azure Unveils Nvidia GB300 NVL72 Cluster Built for OpenAI’s AI Workloads

Microsoft Azure Unveils Nvidia GB300 NVL72 Cluster Built for OpenAI’s AI Workloads

Microsoft Azure Unveils Nvidia GB300 NVL72 Cluster Built for OpenAI’s AI Workloads