Energy per token, model right-sizing, inference optimization, and carbon-aware computing.
The energy per AI query has fallen more than 30-fold in a year, and total AI emissions are still climbing. For CIOs and chief AI officers, the levers that matter most are model choice, workload design, and procurement.
Microsoft says Azure Maia 200 is in production across the US, Europe and Australia, supporting Copilot, OpenAI models and its MAI models. Maia 300, slated for deployment in 2027, uses 96-accelerator racks with a 167kW power envelope and 27.6TB of HBM. Microsoft says Maia 300 is expected to deliver up to 40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale token throughput of Maia 200 (company claims, not independently verified).
Why it matters: A 167kW rack envelope and a stated plan for multi-gigawatt inference deployments give builders concrete power and cooling densities to plan for, while the tokens-per-megawatt claim shows how hyperscalers are framing efficiency for AI strategy decisions.
Source: Data Center Dynamics
VivoPower has begun the first phase of construction at its 42MW Mo i Rana data center in northern Norway, a former crypto mining site acquired from Cowa in April. The company has selected an undisclosed AI tenant and has not given a completion date. It also plans to spin off a Singapore-headquartered, separately listed platform for a non-Nordic AI data center portfolio that it says exceeds 2.2GW.
Why it matters: Converting a former crypto mining site with planned waste heat reuse via district heating shows one route to bring AI capacity online on existing sites, though the heat offtake arrangements are still under evaluation.
Source: Data Center Dynamics
The Karlsruhe Institute of Technology has announced an AI data center at its Campus North in Baden-Württemberg, Germany, with an estimated cost of €24 million ($27.2m). Construction is due to finish by the end of 2029, with opening the following year. Power consumption is predicted at 2MW, with an option to scale to 10MW if demand stays high, and IT compute capacity has not been determined.
Why it matters: A phased build with a defined path from 2MW to 10MW, plus waste heat feeding a district heating network, offers a reference design for builders pairing AI capacity with heat reuse and demand-matched workloads.
Source: Data Center Dynamics
CNBC reports both labs are looking for smaller 20 to 30 MW deployments in the US, UK, and Nordics to serve inference, alongside their gigawatt-scale training campuses.
Why it matters: Inference capacity is dispersing closer to users, which opens the market to mid-sized sites that can deliver power quickly.
Source: Data Center Dynamics
