- Cooling costs surge as Nvidia pushes power limits across successive rack generations
- Compute trays dominate expenses due to rising cold plate requirements and thermal density
- Switch tray savings cannot offset escalating demands from high-power GPU trays
The cost of keeping Nvidia’s high-end rack systems cool continues to climb as each generation pushes deeper into extreme power levels.
A Morgan Stanley report obtained by @Jukanlosreve reveals the liquid cooling hardware inside the GB300 NVL72 costs $49,860, roughly enough to buy a new Tesla Model Y.
The report further estimates the liquid cooling system required for the newer Vera Rubin NVL144 configuration will approach $55,710, a 17% increase.
The economics of tray-level cooling
This platform depends on hotter Rubin GPUs rated at up to 1,800W per unit, along with next-generation NVSwitch 6.0 components.
The cost for cooling this system is linked to the individual compute trays, and each compute tray will need higher-capacity cold plates.
The cost per compute tray is expected to rise by 18% to about $2,660 – and since the Vera Rubin NVL144 system has 18 trays, the total compute-side cooling expense reaches approximately $47,880.
The increase comes from higher-capacity cold plates, which climb to $400 per unit as CPUs and GPUs push thermal limits.
Meanwhile, switch tray cooling appears less burdensome, falling to $870 per tray and totalling $7,830 per rack.
However, this reduction is overshadowed by the much larger jump on the compute side, as the cost trajectory follows a pattern: the transition from GB200 NVL72 saw a 20% rise in cooling demands for GB300 NVL72.
Similarly, the move from GB300 NVL72 to the Vera Rubin NVL144 adds another 17%. The power levels explain this trend.
Each Blackwell Ultra data center GPU draws 1,400W, a Grace CPU takes 300W, and memory contributes 200W per socket.
As workloads escalate, the value of precision cooling grows just as fast, but future systems will further compound this. Nvidia plans a shift to Rubin Ultra GPUs that may reach a thermal design power of 3,600W per package, and meeting that requirement may force new types of cold plates or more aggressive cooling techniques.
Nvidia is also preparing the liquid-cooled NVL576 “Kyber” system, which will include 144 GPU packages, and will deliver higher performance than the Vera Rubin NVL144, while carrying an even greater cooling bill.
Although the final dollar amount is not confirmed, high-capacity plates capable of removing 3.6kW of heat will clearly exceed the current $400 per unit.
This sends a signal that future data center installations will face even steeper thermal expenses.
Via Toms Hardware
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
https://cdn.mos.cms.futurecdn.net/XDJuiMvqGqqEFZZrYzhqyL-1920-80.jpg
Source link
