NVIDIA’s next few AI GPU generations are not just about faster chips. The real battle may be power delivery — and the numbers are starting to look properly wild.
According to figures highlighted from Morgan Stanley Research, NVIDIA’s future Feynman AI rack generation could carry more than US$191,000 in power semiconductor content per rack. For Malaysian readers, that is roughly around RM900,000 before taxes, shipping, integration and exchange-rate swings — just for the power-related semiconductor side, not the full rack.
That is the part that should make datacentre operators sit up. Compared with Blackwell, the Feynman figure is estimated to be around 17 times higher.
From Blackwell To Feynman: Power Becomes The Expensive Part
The current Blackwell B200 baseline is estimated at about US$11,234 in power semiconductor content. GB200 adds roughly another US$4,000, while GB300 pushes it further by about US$3,500. Across the Blackwell generation, the estimate rises to US$17,761.
Then things start scaling much harder.
Rubin, expected after Blackwell, is estimated to pass US$33,000 in power semiconductor cost. Rubin Ultra goes even higher, with the reported estimate around US$95,000. Feynman, planned for 2028 after Rubin, then doubles that Rubin Ultra figure to more than US$191,000.
In simple terms: NVIDIA’s AI racks are becoming less like “GPU boxes” and more like full power engineering projects. The chips are still the stars, sure, but feeding them electricity safely and efficiently is becoming a massive cost centre.
What Parts Are Driving The Cost?
The biggest chunks reportedly come from the power conversion system and second-stage voltage regulation modules. The PCS accounts for about 27%, while VRM-related components take around 26%.
Power supply units make up another 19%. Lateral VRMs sit at around 15%, while intermediate bus converters, battery backup, UPS parts and other supporting components take smaller slices. Switches, NICs and eFuses are also part of the mix.
Basically, every step of moving power from the facility into hundreds of hungry GPUs is getting more complicated.
Why NVIDIA Is Moving To 800V DC
NVIDIA has already outlined a shift toward 800V DC architecture for future AI datacentres. This would move beyond older 48V or 54V-style systems, which are becoming increasingly awkward for megawatt-scale AI racks.
The reason is physics. Higher voltage means lower current for the same power level, which can reduce cable thickness, copper requirements and energy loss. Wccftech’s source material notes that a 1MW rack using 54V DC could require up to 200kg of copper busbar. Scale that to a 1GW datacentre, and the copper requirement becomes ridiculous.
800V DC also helps reduce repeated power conversions, which can waste energy and introduce more failure points. The architecture leans on advanced power electronics such as gallium nitride and silicon carbide components, both important for efficient high-voltage switching.
NVIDIA’s Kyber racks, expected in 2027, are planned to introduce 800V DC with Rubin Ultra GPUs. The setup is described as a dense, liquid-cooled 600kW rack design with 576 Rubin Ultra chips.
Why Malaysia And SEA Should Care
This sounds like hyperscaler stuff, but it matters locally too. Malaysia is already attracting serious datacentre investment, especially around Johor and other strategic locations in the region. If AI infrastructure keeps moving toward megawatt-class racks, power availability, cooling, grid planning and component supply will become even more important.
For gamers and PC builders, this does not mean your next GeForce card will suddenly cost RM900k, relax bro. But it does show where NVIDIA’s highest-end engineering focus is going: AI factories, not just gaming GPUs.
For SEA, the bigger question is whether regional datacentres can support this kind of power density without energy costs and infrastructure limits becoming the real bottleneck. Faster AI needs stronger power systems — and that part is getting expensive fast.
Source: Wccftech Gaming