Goldman details 576‑GPU rack design
- Goldman Sachs highlighted NVIDIA’s next rack step — the Kyber system — as AI infrastructure shifts from server rooms to purpose-built “AI factories.” - The standout spec is 576 Rubin Ultra GPUs in one rack at roughly 600 kW, with 800 VDC power replacing bulky low-voltage copper. - That changes the bottleneck from chips alone to power, cooling, and construction timelines across whole data center campuses.
The thing to understand here is that this is not really a “better server” story. It is a power-and-physics story. Goldman Sachs used NVIDIA’s upcoming Kyber rack to show how fast AI infrastructure is leaving normal data-center design behind. One rack with 576 GPUs and around 600 kW of load is not a denser version of yesterday’s setup — it is basically a new building requirement. (goldmansachs.com) ### What is Kyber, exactly? Kyber is NVIDIA’s next rack-scale system for the Rubin Ultra generation. Goldman Sachs says it is due in 2027 and packs 576 GPUs into a single rack. NVIDIA has also framed Kyber as part of the move to gigawatt-scale “AI factories,” where the rack becomes the core unit of compute instead of the individual server. (goldmansachs.com) ### Why does 576 GPUs matter? Because the number is not just about more chips. It means more GPUs connected tightly enough to behave like one giant training machine. NVIDIA’s GB200-era reference designs already treat 576 GPUs as a full scalable unit in SuperPOD layouts. Kyber pushes that idea further by crammi(goldmansachs.com)er. (docs.nvidia.com) ### Why is the power number the real story? A 600 kW rack is absurd by traditional data-center standards. Goldman Sachs compares it to feeding enough electricity for roughly 500 U.S. homes into something the size of a filing cabinet. Older high-density racks were already stressing facilit(docs.nvidia.com) much energy safely?” (goldmansachs.com) ### Why does copper become a problem? Because low-voltage distribution gets physically ridiculous at this scale. NVIDIA says a 54 VDC approach for a roughly 1 MW rack would need up to 200 kg of copper busbar, and the power shelves could eat as much as 64U of rack space — basically the rack gets consumed by the(goldmansachs.com)ltage means lower current, less copper, less wasted space, and saner thermal behavior. (developer.nvidia.com) ### So what changes in the data center? Cooling, power delivery, and floor planning all change. Current DGX rack-scale systems already rely on liquid cooling, bus bars, manifolds, and passive copper backplanes inside the rack. Kyber-class systems push that further into facility design — p(developer.nvidia.com)sories. (docs.nvidia.com) ### Why does Goldman care? Because this shifts where money gets made and where delays show up. If AI demand keeps scaling, the scarce thing is no longer just advanced silicon. It is substations, transformers, liquid-cooling gear, high-voltage distribution, and the contractors who can build custom halls around these racks. That widens the AI su(docs.nvidia.com)d data-center engineering. (goldmansachs.com) ### Is this happening now or later? The exact Kyber configuration Goldman flagged is a 2027 story, not a 2026 deployment story. But the transition is already underway. NVIDIA’s current GB200 and GB300 rack-scale systems, SuperPOD reference designs, and 800 VDC roadmap all point the same way — toward fewer, denser, liquid-cooled racks that demand purpose-built facilities. (goldmansachs.com) ### Bottom line? Goldman’s point is simple — AI infrastructure is turning into heavy industry. The headline number is 576 GPUs, but the real takeaway is that future AI growth depends as much on electricity, cooling, and construction as on chips. (goldmansachs.com)
