Rubin Ultra and Kyber NVL576: Planning the 2027 Flagship Rack

Overview

Kyber is the rack-scale generation after Vera Rubin, expected in 2027, and it is the point at which AI infrastructure stops being a data centre problem and becomes a building problem. NVIDIA has indicated Kyber racks will house 576 Rubin Ultra GPUs and that per-rack power will reach 600 kW to 1 MW, with the 800 VDC distribution architecture arriving in full-scale production alongside it. For anyone commissioning a facility in 2026, the practical question is whether the hall being built can host that generation, because retrofitting a building is far slower than replacing hardware.

What 576 GPUs in one domain changes

The scale-up domain — the set of GPUs connected by NVLink rather than by a scale-out network — has been the defining architectural parameter of each generation. GB300 NVL72 gives 72; Vera Rubin NVL144 gives 144; Kyber is expected to give 576. Each expansion moves more of a model’s parallelism inside the fast fabric and out of the slower network.

At 576, model-parallel strategies that today span many racks over InfiniBand sit inside a single NVLink domain. That matters most for two things: mixture-of-experts models, whose all-to-all expert routing is the most communication-intensive pattern in current architectures, and very large dense models requiring wide tensor parallelism. The practical effect is higher achieved utilisation on exactly the model families that are growing fastest — the MoE communication problem is described in fine-tuning mixture-of-experts models.

The power step is the discontinuity

The jump from roughly 120 kW to potentially a megawatt in a single rack is not an incremental step. It changes the electrical room, the busway, the UPS topology, the cooling plant and the structural loading of the floor. It also changes the failure blast radius: a single rack at a megawatt is a substantial fraction of a small data hall’s total capacity, so redundancy design must account for losing one.

Why 800 VDC is mandatory rather than optional

At conventional rack voltages, delivering a megawatt implies currents in the tens of thousands of amperes, which is not practically distributable. Raising distribution voltage to 800 V reduces that current by more than an order of magnitude, which is what makes the conductors and connectors physically feasible. NVIDIA reports the architecture carries over 150 percent more power through the same copper than 415/480 VAC distribution, with efficiency and maintenance benefits alongside.

This is why the 800 VDC transition and the Kyber generation are scheduled together — one enables the other. A facility planning for Kyber is planning for 800 VDC by definition, and the ecosystem work around power shelves, busways and protection devices is a supply chain that must be procured rather than improvised. The architecture detail is in 800 VDC power for megawatt AI racks.

What to provision in a 2026 build

The realistic advice for an Indian facility being designed now is provision, not install. Specifically: reserve physical space adjacent to the white space for power shelves and CDU capacity that current racks do not need; run conduit and cable tray paths sized for future high-voltage DC distribution; specify switchgear and transformer capacity with headroom rather than to current load; and design the cooling loop with the pipe sizing and heat-rejection capacity for a later density increase, since replacing pipework in a live hall is enormously disruptive.

Equally important is the utility interface. Grid capacity sanction in India has long lead times, often longer than any hardware procurement, and it is the item that cannot be accelerated by spending more later. Starting the capacity conversation with the utility during design, for a load the site will not draw for two years, is the single most valuable planning action available. The current-generation checklist is in hosting 100 kW racks in India.

Who this is actually for

It is worth being direct: Kyber-class infrastructure is not a general enterprise product. A megawatt rack requires a purpose-built AI factory, and in India the realistic hosts are national programmes, the largest cloud and neocloud operators, and a small number of purpose-built facilities. Most enterprises will consume this capacity as a service rather than own it.

That is the right outcome. The value of understanding the roadmap for a typical buyer is not to procure Kyber but to avoid two errors: over-building a facility for a generation you will never host, and under-building one that cannot take the generation you will host, which for most Indian enterprises is Blackwell- or Rubin-class at 120 to 200 kW. The scale-up versus scale-out decision for that tier is developed in one rack or nine nodes.

How to treat the roadmap in a buying decision

Three principles. Buy hardware for the workload you have funded, since a rack purchased today produces value across its depreciation life regardless of what follows. Build facilities for one generation beyond what you buy, because buildings outlive hardware by a factor of three or more. And do not let roadmap awareness become paralysis — there is always a better generation coming, and an organisation that waits for it never starts.

For most Indian buyers in 2026 that resolves to: deploy Blackwell-class or, if the schedule aligns, Rubin-class hardware; design the hall for higher density and an 800 VDC path; and plan to consume Kyber-class capacity from a provider rather than hosting it. The timing framework is in buy Blackwell Ultra or wait for Rubin.

Frequently asked questions