Gigawatt AI...Zero Waste?

A single rack of next-generation GPUs can consume up to a megawatt of power. One gigawatt — the output of a nuclear power plant — now gets dedicated to a single AI deployment.

Those numbers stopped being hypothetical in May 2026, when Nebius broke ground on a gigawatt-scale AI factory in Independence, Missouri — a multi-building campus on 400 acres that will create 1,200 construction jobs and 130 permanent high-tech positions. It's the company's first gigawatt-class facility in the United States, and it's just one node in a global buildout that now spans 310 MW in Lappeenranta, Finland, 240 MW near Lille, France, and an expanding campus in Mäntsälä, Finland, where Nebius has operated for years.

The scale is staggering. So is the question it raises: how do you build at this pace without breaking the planet?

On the debut episode of Deep Geeks, Can AI Innovation Survive Its Own Energy Appetite, Dr. Serena Huang asked Nebius Head of Sustainability Daria Muktova and myself exactly that. When watching an early cut of the episode, something started clicking: a detailed, layer-by-layer engineering playbook — one that starts at the silicon and doesn't stop until a trained model is serving inference at the lowest possible energy cost per token.

That is what I’ll walk you through today.

Meet Nebius, the company behind the buildout

To understand why the Nebius sustainability story carries weight, you need to understand the scale of the company telling it. This isn't a research lab publishing a whitepaper about what's theoretically possible. It's an operator deploying infrastructure at a pace that rivals the major hyperscalers.

If you haven't been tracking Nebius, the numbers tell a compelling story. In Q1 2026, Nebius posted revenue of $399 million: a 684% year-over-year increase, with AI-specific revenue reaching $390 million (98% of total). The company raised full-year guidance to $3.0–3.4 billion. In March, NVIDIA invested $2 billion in Nebius to co-develop next-generation AI factory architecture, inference platforms, and fleet management. This is a partnership that will see Nebius deploy more than 5 gigawatts of NVIDIA systems by the end of 2030. Analysts at Futurum Group have called it one of the most strategically significant neocloud partnerships in the industry. And as The Motley Fool reported, Nebius has landed $46 billion in AI cloud deals which is a figure that signals this isn't a startup scaling hopefully, but a company scaling deliberately.

What makes Nebius different isn't just the capital or the contracts. It's the engineering philosophy. As Daria explained on DeepGeeks: "We don't treat sustainability as something that comes after…we treat it as a principle. For us, sustainability is basically a synonym for efficiency, a synonym for reliability, which makes it pretty much understandable how to treat it as an engineering approach."

Most cloud providers assemble infrastructure from off-the-shelf components. Nebius designs the full stack: custom servers, custom firmware, custom cooling, and the software orchestration layer that ties it all together. That vertical integration is what makes their sustainability story credible — because you can't optimize what you don't control.

Layer 1: Custom hardware that draws 20% less power

The foundation of the playbook is hardware designed for efficiency from the ground up. The Nebius custom-built servers draw roughly 20% less power than comparable commercially available hardware — not through exotic materials or bleeding-edge experimentation, but through disciplined thermal and electrical engineering.

The key design decision: Nebius servers are built to operate reliably at temperatures up to 45°C. That tolerance has cascading consequences. Higher operating temperatures mean less aggressive cooling is required, which means simpler infrastructure, lower energy overhead, and fewer points of failure. The company also develops its own firmware, reducing the troubleshooting overhead and failure rates that plague operators running vendor-standard stacks.

As Daria put it on DeepGeeks:

When you're deploying at gigawatt scale, that 20% isn't a line item — it's hundreds of megawatts of avoided demand.

Layer 2: Closed-loop liquid cooling with zero water intake

Cooling is where most data center sustainability stories begin and end. The standard approach — evaporative cooling towers — works, but it consumes staggering amounts of water. A typical hyperscale data center can use millions of gallons per day.

Nebius took a fundamentally different path. "At Nebius, we actually do not rely on water intake, even though we're introducing liquid cooling," Daria explained. "Our system is closed-loop — it doesn't include any evaporative components. We use outside air through dry coolers to bring down the temperature of the fluid that circulates within the same loop for thousands, millions of cycles." No evaporative towers. No water consumption from local supplies. It's a design choice enabled by the high-temperature tolerance of their custom servers: when your hardware runs comfortably at 45°C, you don't need to reject as much heat, and you can reject the heat you do produce without evaporating water to do it.

The Lappeenranta facility in Finland, announced in March 2026, will extend this approach to 310 MW of capacity. Like the Mäntsälä campus, it's designed to integrate a heat recovery system that donates excess server heat to the local district heating network. In Mäntsälä, this approach avoided approximately 4,000 tonnes of CO2 emissions in 2025 and reduced connected household heating costs by around 10%.

Daria framed the impact in human terms: "By reusing server heat that we capture — which is basically free — you're actually able to reduce the cost of producing heat for the municipality. Last year, households spent 10% less on heating because they were able to use free server heat as a resource." Between 2022 and 2024, over 50 GWh of captured heat was supplied to the local network — covering 65% of its needs in 2024 alone.

That's not a proof of concept. That's infrastructure doing double duty.

Layer 3: The software orchestration nobody talks about

This is where the DeepGeeks conversation got interesting (if I do say so myself) — and where Daria made a point that the broader industry needs to hear.

As Daria put it: "While the discussion is largely focused on hardware — how the chips are performing, how the servers are performing, what the cooling systems are — the big question is about software. Software plays an important role in orchestrating all of that hardware and infrastructure, allocating workloads in a way that uses all available capacity and ensures servers aren't staying idle. Idling is actually a killer to efficiency."

Consider the idle tax: GPUs that aren't actively processing still draw 12–25% of their peak power. Across a fleet of tens of thousands of accelerators, that idle draw adds up fast. Nebius addresses this through autoscaling that dynamically matches cluster size to real-time demand — rather than statically allocating GPUs to individual customers regardless of utilization.

Their virtualized fabric architecture takes this further. Traditional bare-metal providers leave small GPU slices unused because they can't be allocated to other workloads. Nebius's software captures those fragments, cutting buffer waste from the industry-standard 10–20% down to 4–5%. At gigawatt scale, recovering that 15 percentage points of utilization is the equivalent of building an entirely new mid-size data center — without pouring a single foundation.

On cluster reliability, the numbers are equally striking. One Nebius customer recorded 56.6 hours of stable operation on a 3,000-GPU production cluster, compared to industry averages of roughly 9.8 hours between interruptions for clusters of that size. Every avoided restart means a training job that doesn't have to re-consume the energy it already spent. Failure recovery isn't just an uptime metric. It's an energy metric.

Layer 4: Post-training optimization and the Token Factory

At the model layer, Nebius's Token Factory applies runtime optimizations that boost throughput without changing the model itself or the underlying hardware. Their post-training tools tune models for the specific GPU architecture hosting them — meaning faster inference, fewer wasted cycles, and less total energy per completion.

This is where efficiency gets personal for AI practitioners. You don't have to redesign your model or switch providers. You run your model through optimization tooling that adapts it to the exact hardware it will serve on, and you get more tokens per watt out the other side.

It's the kind of capability that emerges naturally from vertical integration. Because Nebius controls the hardware, the cooling, the cluster software, and the model-serving layer, they can execute multi-layered optimizations that pure infrastructure or pure software providers simply cannot.

The missing puzzle piece: Memory tiering and the storage layer

There's one more layer in this story — and it's the one that connects Nebius's infrastructure vision to the broader AI systems conversation.

GPU compute and model optimization get most of the attention when people talk about AI efficiency. But data movement — how training data, checkpoints, and model weights flow between storage tiers and accelerators — is one of the largest hidden energy costs in the stack. Every unnecessary data fetch, every redundant checkpoint write, every time a GPU stalls waiting for storage I/O is wasted energy and wasted time.

This is where WEKA’s NeuralMesh fits into the picture. NeuralMesh enables intelligent memory tiering that moves data automatically between hot, warm, and cold storage based on access patterns and workload demands. Hot data lives close to the GPUs. Cold data drops to lower-cost, lower-power tiers. The result: GPUs spend more time computing and less time waiting, which means more useful output per watt — the metric that matters most at gigawatt scale.

It's the storage layer that completes the Nebius full-stack vision. Custom hardware, zero-water cooling, software orchestration, and model optimization get you most of the way there. Intelligent data movement gets you the rest.

What this means for AI teams evaluating infrastructure

If you're an AI team choosing where to run your workloads, the Nebius story reframes what you should be asking about. Cost per GPU-hour and uptime SLAs are table stakes. The question that matters now — especially as workloads scale from chat to reasoning to agentic AI, with energy demands growing 100x along the way — is: how much useful output does your infrastructure produce per watt?

That's a full-stack question. It touches model optimization (model & token routing, KV cache transfer scheduling), cluster reliability, fabric utilization, cooling architecture, storage I/O, and software orchestration. If your provider can't give you specifics at each layer, it's worth asking why.

The Nebius white paper The Energy Behind AI is one of the most transparent examinations of these tradeoffs published by any infrastructure provider. It's worth reading alongside the DeepGeeks episode for the engineering detail that doesn't fit in a press release.

Why this conversation matters now

The AI infrastructure buildout happening right now is the largest construction of compute capacity in history. Nebius alone is targeting more than 3 GW of contracted power by the end of 2026, with a path to 5+ GW by 2030. The Missouri AI factory, the Finland expansions, the NVIDIA partnership — these aren't announcements. They're concrete being poured.

The question isn't whether we'll build gigawatt-scale AI. We already are. The question is whether we'll build it smartly — with efficiency engineered into every layer, rather than bolted on as an afterthought.

Nebius's playbook shows what's possible when a company controls the full stack and treats energy efficiency as a design constraint rather than a marketing claim: 20% lower server power draw, PUE of 1.15, zero water intake, 50 GWh of avoided consumption in a single year, thousands of homes heated by recovered waste, and a software layer that wrings utilization out of every GPU cycle.

Daria's closing words on the episode said it best:

"What will define the future of AI is how well we engineer it. If we think about this as an ecosystem of different decisions made at different levels — from chip providers to infrastructure providers and how they design the system, to software and how different tools are built — and if we all work together, this is how I think we can achieve a sustainable AI that produces maximum value but is also mindful and conscious about the resources it uses."

That's not sustainability theater. That's engineering.

Watch the full conversation: Deep Geeks Episode 1 — "Can AI Innovation Outpace Its Own Energy Demands?" features Daria Mukhortova and myself going deep on everything from tokens per watt to the 100x energy jump from chat to agents.

Download the resource: The AI Energy Metrics Cheatsheet covers what to measure and what to fix when it comes to building efficient AI infrastructure.

Deep Geeks is a podcast sponsored by WEKA.