NVIDIA Blackwell Ultra GB300: Liquid cooling or meltdown?

It is no surprise that the performance of modern AI accelerators is constantly increasing. Nor is it a surprise that they are consuming ever more absurd amounts of energy and mercilessly pushing the thermal limits of the hardware. The solution? For NVIDIA, it’s “liquid cooling or nothing”. With the Blackwell Ultra GB300 AI servers, which are to be presented at GTC 2025, NVIDIA is saying goodbye to air cooling and switching completely to water-based cooling systems. And there is a simple reason for this: with up to 1400 watts TDP per accelerator, air cooling would make about as much sense as a fan in a blast furnace. But what does this mean for data centers? And how much will NVIDIA tighten the price screw when even the GB200 models already cost absurd sums?

Performance with side effects: The GB300 series pushes the limit

The details known so far about Blackwell Ultra indicate that NVIDIA will not be satisfied with minor updates. Instead, there is a brute leap in performance – with the corresponding consequences:

• TDP of 1400W – almost 50% more than the already heathy GB200
• More memory: from 192 GB HBM3E to 288 GB, thanks to new 12 Hi stacks
• 1.4 times the FP4 performance of its predecessor

All well and good, but with such massive power consumption comes massive heat development. While smaller GPUs can still be fobbed off with air coolers, the GB300 will only work with liquid cooling – whether you like it or not.

Data centers must retool – or stay outside

According to Taiwan Economic Daily, suppliers have already received a flood of orders for liquid cooling systems. So NVIDIA doesn’t seem to be doing things by halves and is moving away from hybrid air-liquid cooling solutions completely. This is a bold move, as not every AI farm is designed for a pure liquid cooling system. So if you want to invest in Blackwell Ultra, you have to upgrade first – and that costs money. Speaking of costs, the GB200 NVL72 cluster already costs around 3 million dollars, and the new generation will certainly not be any cheaper. Prices are rising, but what can you do when there is no alternative?

Market dominance: If you need AI performance, stay with NVIDIA

Of course there is competition. AMD has the MI300X, Intel has the Gaudi 3, and some start-ups are trying to gain a foothold in the AI segment. But the reality is this:

• CUDA is the standard. Anyone who is serious about AI relies on NVIDIA – not because they really want to, but because there are hardly any alternatives.
• Hyperscalers have no choice. Google, Microsoft, Amazon and co. need the best hardware – no matter what it costs.
• Demand is exploding. Despite the high prices, bottlenecks are already foreseeable.

Even if cooling remains a challenge, it is unlikely that NVIDIA will lose customers. Because if you want to be a player in AI, there is no way around NVIDIA.

Revolution or just hotter?

The Blackwell Ultra GB300 series is a turning point. Full liquid cooling, record TDP, even more memory – and probably even higher prices.

Questions that remain unanswered:

• How many data centers will be able to afford the switch to liquid cooling?
• Will the extreme power consumption become a problem?
• And how long will NVIDIA maintain its monopoly before someone starts to seriously compete?

One thing is certain: NVIDIA will remain the measure of all things in the AI sector for the time being. And if it means that entire data centers have to be upgraded – then it will be done.

Source: Taiwan Economic Daily