TL;DR
NVIDIA’s Rubin generation pushes 8-GPU servers past 20 kW per chassis, beyond what air cooling can handle — and most enterprise data centers aren’t built for liquid. That leaves three options: rebuild your facility, hope AMD, Intel, or Google ships 4- or 6-GPU air-cooled boxes to buy a few more years, or rent GPUs from hyperscalers forever. Pick your constraint now, or the market picks it for you.
Key takeaways:
- Rubin-class 8-GPU servers are heading toward 20 kW per chassis, past the practical limit of air cooling
- Rack-level liquid cooling only pencils out when racks are full, enterprise power limits prevent that
- Hyperscalers are absorbing early supply of liquid-only GPU modules, leaving enterprises waiting quarters
- A 4- or 6-GPU air-cooled box from AMD, Intel, or Google could extend on-prem life without a full data-center rebuild
- Defaulting to cloud means renting compute forever at prices hyperscalers control and are likely going to increase
Every few years, the IT industry hits a moment where the pace of innovation outruns the physical world built to house it. I believe we’re in that moment right now. NVIDIA is pushing GPU performance so far, so fast, that the data center infrastructure most enterprises rely on simply cannot keep up.
The next generation of NVIDIA product, Rubin, pushes beyond the limits of what air cooling can support in an 8-GPU server. And once that happens, the entire enterprise market must confront a choice it has been trying to postpone.
We are arriving at a fork in the road, and neither path is easy.
Why Air Cooling Can’t Handle NVIDIA Rubin
For the last few years, 8-GPU servers made financial and architectural sense. They delivered density, throughput, and manageable thermals. But the wattage curve has spiked hard. A server that just a few years ago hit 3-4 kW now pushes over 10 kW, and we’ll soon see configurations upwards of 20 kW per 8-GPU chassis according to NVIDIA’s GTC 2026 disclosures. No matter how clever the shrouds or airflow designs become, air cooling simply won’t carry systems like this forward. And the Rubin redesign suggests NVIDIA is going all-in with direct liquid cooling (DLC).
This is the moment when enterprises must choose between massive data center upgrades or relinquishing long-term control of their AI workloads to the cloud. The physics leaves no third option, at least not a comfortable one.
Why Direct Liquid Cooling Is Hard for Mid-Market Enterprises
Liquid cooling works beautifully in hyperscale environments. Those facilities were engineered for it. They have the water infrastructure, the mechanical teams, the density, and the purchasing power to make it economical. A mid-market enterprise does not. When a site can only power one or two 8-GPU servers per rack, a rack-level cooling distribution unit (CDU) becomes disproportionately expensive. These cooling systems only make economic sense when the rack is full enough to consume most of the cooling budget; power limits prevent that in many traditional data centers. So liquid cooling isn’t just a technical hurdle, it’s a financial one.
And even if an enterprise wants to move to liquid, there’s another problem: hyperscalers are already consuming nearly all the early manufacturing supply for the liquid-only GPU modules. Enterprises are often left waiting months, sometimes quarters, before hardware becomes available.
Will AMD, Intel, or Google Offer Air-Cooled 4-GPU and 6-GPU Servers?
This reality may push server and GPU manufacturers into an uncomfortable shift: reducing density. If the next wave of NVIDIA GPUs can’t be air-cooled, is it possible one of the smaller players like AMD, Intel, or Google will come out with more air-cooled options in 6-GPU or 4-GPU air-cooled configurations to capture a few more years in the enterprise space? It’s not an elegant solution, but it would allow organizations to avoid rebuilding their data centers from scratch.
It also stretches the life of thousands of facilities that were never designed for the thermals of modern AI workloads. For many enterprises, this will be the only path that preserves on-prem compute autonomy without triggering a full mechanical overhaul.
Should Enterprises Run AI Workloads On-Prem or in the Cloud?
If air-cooled options disappear, the cloud becomes the fallback. But anyone who has taken a pilot AI workload into production already knows the danger here: cloud GPU costs balloon quickly, especially with model training and agentic applications that run nonstop. Renting compute forever is not a strategy, it’s an inevitability created by not investing early enough in on-prem capability. Either enterprises take steps now to preserve optionality, or they risk locking themselves into a long-term dependency on cloud providers who control both supply and price.
How Enterprises Should Plan AI Infrastructure for 2026–2028
The real decision isn’t air vs. liquid or on-prem vs. cloud. It’s whether you choose your constraints today or inherit them later from the market. Enterprises who wait will find themselves pushed into a single outcome: renting GPUs at whatever price the cloud dictates.
The fork in the road is already here—you only get to decide whether you’re the one steering when you reach it.
