There's a hard truth that the GPU infrastructure industry needs to confront: air cooling is finished. Not in five years, not gradually — it's already inadequate for the workloads that matter most. If you're building GPU infrastructure in 2026 and you're still planning around air-cooled racks, you're building yesterday's data center.
I don't say this lightly. I've been designing and operating data centers since 2009, when I founded Datacell in Reykjavik. I've watched cooling technology evolve through every generation. But the jump from traditional CPU-centric workloads to modern GPU clusters represents a discontinuity — not an incremental change — in thermal management requirements.
The Numbers Don't Lie
A single NVIDIA H100 GPU draws roughly 700 watts of power. A DGX H100 system with eight GPUs consumes around 10.2 kW. Pack a rack with four of these systems, and you're looking at 40+ kW per rack — before you even account for networking, storage, and management overhead. The latest B200 generation pushes these numbers even higher.
Traditional air-cooled data centers are designed for 8 to 15 kW per rack. Some modern facilities stretch to 25 kW with hot/cold aisle containment and precision cooling. But 40, 60, or 100 kW per rack? Air simply cannot move enough thermal energy at those densities. You'd need hurricane-force airflow, and even then the temperature gradients across the chip would cause throttling and reliability issues.
At AI Green Bytes, we design our immersion-cooled facilities for 100+ kW per rack from day one. The dielectric fluid makes direct contact with every heat-generating component, eliminating the thermal resistance that limits air cooling. GPU junction temperatures stay lower and more uniform, which means consistent boost clocks, longer hardware life, and fewer thermal throttling events.
The Economics of GPU-as-a-Service
The GPU rental business is fundamentally a utilization game. Every hour a GPU sits idle or thermally throttled is lost revenue. Every GPU that fails prematurely is a capital expenditure that didn't earn its return. The economics are unforgiving.
Immersion cooling directly improves both metrics. Lower, more stable operating temperatures mean GPUs maintain peak performance for longer sustained periods — critical for the multi-hour training runs that enterprise customers demand. The elimination of thermal cycling stress extends hardware lifespan by an estimated 2 to 3x, which fundamentally changes the depreciation math.
We work with Midas for our immersion tanks and Oleon for the dielectric cooling fluid. The upfront cost of immersion infrastructure is higher than traditional air cooling — roughly 15 to 20% more per rack. But when you factor in the elimination of CRAC units, raised floors, and complex air handling systems, plus the reduced energy costs and extended hardware life, the total cost of ownership over five years is significantly lower.
For our GPU rental customers, this translates directly to competitive pricing. We can offer H100 and B200 GPU hours at rates that air-cooled competitors struggle to match, because our operational costs are structurally lower.
Density Is Destiny
There's a spatial dimension to this that often gets overlooked. An air-cooled facility needs roughly 3x the floor space of an equivalent immersion-cooled facility to deliver the same compute capacity. That's 3x the real estate cost, 3x the construction footprint, and 3x the physical security perimeter.
In our edge data center model — smaller facilities distributed across Iceland, Norway, Sweden, France, and Portugal — space efficiency is critical. We're not building hyperscale campuses with unlimited room to expand. We're building targeted, high-density facilities close to where the compute is needed. Immersion cooling makes this possible in a way that air cooling simply cannot.
Hypertec supplies our server hardware, and Cloudification handles the cloud orchestration layer. Together with Midas immersion tanks, we've built a stack that delivers hyperscale-grade GPU compute in edge-scale footprints. That's the competitive advantage that matters.
The Reliability Imperative
When you're running a GPU-as-a-Service business, uptime isn't just a metric — it's your reputation. Enterprise customers running multi-day training jobs need absolute confidence that the infrastructure won't fail mid-run. A thermal event that takes down a GPU node during a 72-hour training run doesn't just cost you the compute time; it costs you the customer.
Immersion cooling eliminates the most common failure modes in traditional data centers. No fans to fail. No air filters to clog. No hot spots from uneven airflow distribution. The thermal environment is inherently more stable and predictable than anything air cooling can achieve.
In our facilities, we've seen hardware failure rates drop by over 60% compared to industry averages for air-cooled GPU deployments. That reliability is what lets us offer the SLAs that enterprise customers require.
The Market Is Moving
The global GPU-as-a-Service market is projected to reach $50 billion by 2030. Every major cloud provider is scrambling to add GPU capacity. But the providers who will win this market aren't the ones with the most GPUs — they're the ones who can operate those GPUs most efficiently, most reliably, and most sustainably.
Immersion cooling is the enabling technology for all three. If you're in the GPU business and you're not investing in immersion cooling today, you're building on a foundation that's already cracking.
