At Rise of AI 2026 in Berlin, Christian Winterfeldt of Dell Technologies, who leads Dell's data center business for Germany, Austria and Switzerland, argues that AI can become more sustainable through disciplined engineering rather than by avoiding the question. He opens with the scale of the problem: data centers will consume more than 2,000 terawatt-hours of energy by the end of the year, according to one study, roughly half of North America's total energy use and double the figure from three years ago. He explains that unlike CPU infrastructure, GPUs cannot be oversubscribed because their memory and bandwidth have hard limits, so the old cloud economics of 'more usage, lower cost' break down and orchestration replaces oversubscription. He presents the 'AI factory' as a full-stack production system optimized for useful output per watt, covering rack-scale design, liquid cooling, multi-tenant GPU scheduling, granular energy measurement, heat reuse, model-size choices and storage tiering. He describes a customer case where software optimization cut a GPU footprint from twelve GPUs to one, and closes with a five-step plan and the view that AI will remain a major energy consumer but can be built intelligently.
Key takeaways
- Data centers will consume more than 2,000 terawatt-hours of energy this year, about half of North America's total use, according to Winterfeldt.
- GPUs cannot be oversubscribed like CPUs because of hard memory and bandwidth limits, breaking traditional cloud economics.
- Moving from static GPU pools to multi-tenant scheduling can raise utilization from around 30-90 percent to 90 percent or higher.
- A customer cut its GPU footprint from twelve GPUs to one by optimizing caching, model serving and model size rather than adding hardware.
- A triple-digit number of German enterprises move AI workloads on-premise every quarter, mainly for governance and sovereignty reasons.
Detailed notes
The scale of AI's energy demand
Winterfeldt opens by describing data centers being built today that consume more energy than a small town, not for hospitals or transport but to generate tokens from neural networks. Citing one study, he says data centers will consume more than 2,000 terawatt hours of energy by the end of the year, a figure that has doubled over the last three years and equals roughly half of North America's total energy consumption. He frames this as the backdrop to a question he hears constantly from executives: how to scale AI within an organization without destroying the economics, since consuming AI purely from the cloud can become very expensive as usage grows.
Why old cloud economics break down for GPUs
He explains that traditional data centers thrived on oversubscribing CPUs and sharing memory across statistical averages, which let cloud providers offer aggressive pricing. GPUs break this playbook because their memory and bandwidth have hard boundaries: oversubscribing a GPU immediately shows up as rising latency or dropping throughput. As a result, token bills explode as AI usage and sophistication grow, the opposite of typical cloud economics, and orchestration replaces oversubscription as the core design principle.
The AI factory: rack design, liquid cooling and on-premise sovereignty
Winterfeldt describes an 'AI factory' as a full-stack production system spanning compute, storage, networking, scheduling and telemetry, optimized for useful output per watt rather than raw GPU count. He notes that a modern high-density AI rack can draw far more power than expected a few years ago: 30 to 170 kilowatts, figures he calls already outdated, compared with Nvidia's newly announced Vera Rubin racks reaching several hundred kilowatts. While 99 percent of data centers are still air cooled, he sees about 20 percent growth in liquid cooling, which removes heat more efficiently and reduces chiller and fan load. He reports that a triple-digit number of German enterprise and mid-market customers are moving AI on-premise every quarter, mainly for governance and sovereignty reasons, typically investing in smaller increments than the large sums spent by US and APJ companies.
Fixing GPU utilization through multi-tenant scheduling
He identifies idle GPUs, not cooling, as the biggest practical problem, describing organizations that split GPU capacity into static pools per team or use case, leaving training clusters at around 30 percent utilization, inference pools at 90 percent, and some capacity unused entirely. His proposed fix is moving from static pools to multi-tenant scheduling with a centralized queue, such as Nvidia's Run:AI, which Dell has built AI factories around together with Nvidia since creating their first joint AI factory three years ago. With such scheduling, he says utilization can rise to 90 percent or higher, and he stresses that sustainability and economics move together: better utilization improves both at once.
Measuring, reusing heat and optimizing the software stack
Winterfeldt argues every system should be measured for energy use and CO2 footprint per workload, using tools such as Dell's iDRAC and OpenManage Enterprise, to find 'zombie racks' that consume power while doing nothing useful. He argues waste heat should be treated as a resource, reused for residential or district heating, industrial or agricultural production, or desalination in warmer climates, rather than vented from isolated chiller boxes. On the software side, he says the biggest energy saver is avoiding unnecessary compute: choosing smaller models where a large model is not needed, efficient retrieval-augmented generation, adaptive batching, right-sized context windows and flexible model serving. He cites a customer that moved from cloud to on-premise where optimizing caching and model serving cut its GPU footprint from twelve GPUs to one, and says Dell has more than 800 engineers working on such model optimization and shares findings publicly. He also describes storage tiering, using fast flash storage for active inferencing and KV caches and the cheapest tier for archiving and logging, as an underused lever. He closes with a five-step, roughly 18-month plan: measure everything and set clear KPIs, build a water-cooled, rack-scale reference pod, move to dynamic scheduling, assign top software engineers to optimize the full stack, and align infrastructure and sustainability teams around shared goals.
Q&A highlights
Asked why he did not mention immersion cooling, Winterfeldt says Dell is watching it closely but sees better near-term return on direct liquid cooling, since immersion cooling is currently more expensive to deploy; he adds that Dell would move quickly if Michael Dell saw a clear opportunity there. Asked why more organizations are not already doing this if it is 'easy', he says most companies still lack an AI strategy: the last three years were mostly pilots using cloud resources because they are fast and easy, and scaling requires a deliberate shift, with a triple-digit number of German customers joining this on-premise strategy every quarter while hybrid cloud and on-premise use will coexist. Asked how organizations without in-house expertise can improve GPU utilization, he recommends starting from a proven AI factory reference architecture built with Nvidia, using colocation where needed, and working with an experienced partner; he notes Dell has sold more than 3,000 AI factories globally over the past two years. Asked about his personal motivation beyond cost and environmental benefits, he points to data sovereignty and governance: keeping critical data under an organization's own control rather than copying it into a third-party cloud provider's data center.
Quotes
In old times, oversubscription was your friend. Today, orchestration is your friend.
Christian Winterfeldt
You want to have a useful output per watt, and this is the key element here, a useful output per watt and not what per workload.
Christian Winterfeldt
The question is not whether AI data centers are going to be built. The question is, are we building them intelligently.
Christian Winterfeldt
Questions answered in this talk
How much energy do data centers use for AI?
According to Christian Winterfeldt at Rise of AI 2026, one study projects data centers will consume more than 2,000 terawatt-hours by the end of the year, a figure that has doubled over three years and is roughly half of North America's total energy consumption.
Why can't GPUs be oversubscribed the way CPUs are in a data center?
According to Christian Winterfeldt at Rise of AI 2026, GPU memory and bandwidth have hard boundaries, so oversubscribing them immediately causes rising latency or dropping throughput; this is why multi-tenant orchestration and scheduling, rather than oversubscription, are needed to run AI infrastructure efficiently.
How can a company make its AI infrastructure more sustainable?
According to Christian Winterfeldt at Rise of AI 2026, companies should measure energy use and CO2 footprint across the stack, build a water-cooled, rack-scale reference architecture, use multi-tenant scheduling to raise GPU utilization, optimize model size and the software stack, reuse waste heat, and align infrastructure and sustainability teams around shared KPIs.