What the Nvidia Vera Rubin Era Means for Data Center Providers

Nvidia Vera Rubin is the GPU platform succeeding Blackwell, in full production as of January 2026 with volume shipments in H2 2026. The VR200 NVL72 rack draws roughly 190–230 kW (up from 120–130 kW for Blackwell and ~40 kW for Hopper), and the 2027 Rubin Ultra "Kyber" rack is specified at ~600 kW, with 1 MW-class racks behind it. For data center providers, Rubin isn't a chip refresh. It's a forced redesign of power delivery, cooling, and the economics of building anything custom.

This post covers the Rubin roadmap as it stands in mid-2026 → the rack power trajectory to 1 MW → what changes for providers (800 VDC, all-liquid cooling, structure) → the stranded-asset problem → who's committed → why the annual cadence breaks the traditional build model.

The Rubin roadmap, as of June 2026

At CES in January 2026, Jensen Huang announced the Vera Rubin NVL72 is in full production, with volume availability in the second half of the year. One naming note worth getting right: the rack originally announced as "NVL144" (counting GPU dies) is now officially the VR200 NVL72, counting 72 GPU packages of two dies each.

The platform is six new chips shipping as one system. The Rubin GPU delivers 50 PFLOPS of NVFP4 inference per package (five times Blackwell) with 288 GB of HBM4 at 22 TB/s, a bandwidth figure Nvidia raised from its original 13 TB/s target. The Vera CPU brings 88 custom Arm "Olympus" cores. NVLink 6 doubles per-GPU bandwidth to 3.6 TB/s, ConnectX-9 doubles the scale-out NIC to 1.6 Tb/s, BlueField-4 handles the data plane, and a GDDR7-based Rubin CPX accelerator targets long-context prefill in a separate ~370 kW rack variant.

Per-GPU power: roughly 1.8 kW baseline, up to ~2.3 kW in the "Max P" configuration, per supply-chain analyst Ming-Chi Kuo (2026). Kuo estimates 5,000–7,000 VR200 racks shipping in H2 2026 alone.

And the cadence is now explicit: Rubin in 2026, Rubin Ultra in the 600 kW Kyber rack in 2027, Feynman in 2028 with co-packaged optics. Huang opened GTC 2026 citing roughly $1 trillion in orders through 2027. A new rack-scale platform, with a materially different facility envelope, every twelve months. Hold that thought.

The rack power trajectory: 40 kW to 1 MW in five years

Sources: Nvidia (2025–26); Ming-Chi Kuo (2026); Hashrate Index (2026); DCD (2025).

Read the table as a provider, not a chip enthusiast. In five years, the unit of deployment went from "a rack your facility could power" to "a rack that consumes a small factory's worth of electricity." Each step invalidated the previous step's electrical design.

Nvidia's answer to the physics is 800-volt DC power architecture: facility-to-rack DC distribution that moves over 150% more power through the same copper than 415/480 VAC, eliminating roughly 200 kg of busbar per rack. Some twenty ecosystem partners (Vertiv, Schneider Electric, Siemens, ABB, Eaton, and Hitachi Energy among them) are building the silicon, sidecars, and switchgear. SemiAnalysis sizes the shift at ~39 GW of new capacity on 800 VDC by 2030, with an ~$11B power-sidecar market and ~$13B in solid-state transformers behind it. Schneider Electric's own engineers put it flatly in October 2025: the 1 MW rack is coming, and it needs 800 VDC.

What actually changes for data center providers

Power delivery becomes the product. Per-rack feeds jump from 60 A whips on GB300 to 100 A on VR200 (Kuo, 2026). At Kyber densities, the familiar AC chain (UPS, PDUs, busways) gives way to 800 VDC power rooms and sidecar racks sitting next to the compute. If you want to host Kyber on day one in H2 2027, the electrical redesign starts now. How redundancy topologies translate into AI-era designs is covered in our power architecture guide.

Cooling goes all-liquid, and then some. VR200 compute and switch trays are fanless. Coolant flow per rack roughly doubles versus GB300, while rack airflow requirements drop about 80% (Kuo, 2026). Even Nvidia's 8-GPU HGX Rubin boxes now require liquid cooling. Dell'Oro forecasts the data center liquid cooling market approaching $7B by 2029. For the facility, this means warm-water direct-to-chip loops, CDU capacity as a first-order design parameter, and water plant sized in hundreds of kW per rack position. The full stack is in our liquid cooling guide.

Structure follows. A 600 kW Kyber rack with liquid-filled manifolds concentrates several tonnes per rack position. Overhead prefabricated power-and-cooling busways replace underfloor everything. The legacy raised-floor hall isn't just underpowered for this era; it's the wrong shape.

And the facility becomes part of the product roadmap. Nvidia now ships a facility-level reference design: Vera Rubin DSX, with an Omniverse digital-twin blueprint. Schneider, Siemens, Trane, and Vertiv provide simulation-ready models so a gigawatt plant can be validated before groundbreaking. The data center is being designed top-down from the rack outward, on Nvidia's annual clock. As The Register put it: enterprises don't buy products, they buy roadmaps, and the hard part is having somewhere to put them.

The stranded-asset problem

Here's the uncomfortable arithmetic nobody puts in the brochure. A facility designed in 2024 around 120 kW Blackwell racks cannot, by spec, host 2027's 600 kW Kyber racks without rebuilding its power distribution, cooling plant, and possibly its structure. The Next Platform headlined it precisely in January 2026: Vera Rubin "obsoletes current AI iron six months ahead of launch." Industry analyst Tony Grayson warns of "billion-dollar concrete husks": shells built for one density generation that can't adapt to the next. Uptime Institute notes operators now design conservatively because the financial risk of stranded power and cooling capacity outweighs the risk of unused space.

One honest caveat, because the panic is overdone in places: the silicon doesn't strand. Hopper-class GPUs are still earning in 2026, cascading into inference as new generations take frontier training. What strands is the facility that can't accept the next generation. The asset at risk isn't the chip. It's the building.

Who's committed to Rubin

The demand side is not speculative. Vera Rubin powers the first ~1 GW of Stargate capacity coming online in H2 2026, within OpenAI's ~10 GW program. Microsoft's Fairwater sites will run VR200 at hundreds-of-thousands-of-superchips scale, and reporting on the revised Microsoft–OpenAI agreement cites roughly 5 GW of Vera Rubin capacity. First Rubin instances arrive from AWS, Google Cloud, Azure, OCI, and CoreWeave in H2 2026; Nebius takes VR200 from H2 2026 against Microsoft's up-to-$19.4B commitment. The buyers' business model, and why they care about every week of energization delay, is laid out in how the neocloud business works.

Why the annual cadence breaks the custom build

Now connect the two clocks. Nvidia ships a new rack-scale platform with a materially different power and cooling envelope every 12 months. A custom greenfield data center takes 18–36 months from design freeze to energization. Do the subtraction: a facility specified around GB200's 120 kW racks was two generations stale at handover.

There are only two rational responses. The first is to over-provision wildly and eat the stranded-capacity cost. The second is to stop treating the data center as a construction project and start treating it as a manufactured, upgradeable product. That is exactly where the ecosystem is heading. Nvidia's DSX reference design and MGX/Kyber rack standards give factory builders a fixed interface to build against. Vertiv reports its pre-configured, factory-integrated modules deploy up to 50% faster than traditional on-site builds. With sidecar power racks, skid-mounted CDUs, and modular 800 VDC power rooms, the generation-to-generation delta becomes a module swap instead of a facility rebuild.

That's the factory-built modular model applied to the Rubin era: deploy capacity engineered for today's 130–230 kW racks, pre-provisioned (pipe diameters, busway ratings, floor loading, white-space geometry) for what the roadmap already tells you is coming. European providers have it harder still: grid queues of 7–10 years and a power problem that increasingly means generating your own make the 24-month custom build slower yet, and the broader resilience case for modular infrastructure stronger. The full build sequence is in our AI data center guide.

The Vera Rubin era's message to data center providers fits in one line: Nvidia told you the next three years of facility requirements in advance, in public, with spec sheets. The providers who win are the ones who believe the roadmap and build for it — in a factory, not a field.

FAQ

What is Nvidia Vera Rubin?

Vera Rubin is Nvidia's GPU platform succeeding Blackwell, named after the astronomer who confirmed dark matter. It pairs the Rubin GPU (50 PFLOPS NVFP4, 288 GB HBM4) with the Vera CPU (88 Arm cores) in the rack-scale VR200 NVL72. It entered full production in January 2026, with volume shipments in H2 2026.

How much power does a Vera Rubin rack use?

The VR200 NVL72 draws roughly 190–230 kW per rack depending on configuration, with the CPX long-context variant around 370 kW. That compares with 120–130 kW for Blackwell GB200 NVL72 and roughly 40 kW for Hopper-era racks.

What is the Kyber rack?

Kyber is Nvidia's rack architecture for Rubin Ultra NVL576, arriving H2 2027: 576 GPU dies per rack, about 15 EFLOPS of FP4 inference, and roughly 600 kW of power per rack, paired with separate power and cooling sidecar racks and 800 VDC distribution.

What is 800 VDC and why does it matter?

800 VDC is a facility-to-rack direct-current power architecture Nvidia announced in 2025 to make 600 kW–1 MW racks feasible. It carries over 150% more power through the same copper than 415/480 VAC distribution. Around 20 partners including Vertiv, Schneider Electric, Siemens, and ABB are building the ecosystem, and SemiAnalysis projects ~39 GW of new capacity on 800 VDC by 2030.

Does Vera Rubin require liquid cooling?

Yes, entirely. VR200 compute and NVSwitch trays are fanless, coolant flow roughly doubles versus GB300, and even 8-GPU HGX Rubin servers require liquid cooling. Facilities need warm-water direct-to-chip loops, expanded CDU capacity, and water plant sized for hundreds of kW per rack.

Will Vera Rubin make existing data centers obsolete?

Facilities designed for earlier density generations cannot host Rubin-class racks without rebuilding power and cooling; analysts warn of stranded facilities, though older GPUs keep earning on inference workloads. Designs that modularize power rooms, CDUs, and sidecars can absorb generation jumps as module swaps rather than rebuilds.

When should providers start preparing for Rubin Ultra?

Now. Kyber's ~600 kW racks and 800 VDC distribution arrive in H2 2027, and electrical redesign, water plant sizing, and structural verification take longer than one GPU cycle. Providers planning day-one Kyber capacity need 800 VDC and high-capacity liquid cooling in current designs.

‍