600 kW Racks Are a 2027 Story

Walk any 2026 infrastructure conversation and you will hear about 600 kW racks. What you will hear less often is that the 600 kW rack, NVIDIA's Kyber housing the Rubin Ultra NVL576, is on the roadmap for the second half of 2027, not for the facility you are specifying now. The rack that is actually going into production this year is Vera Rubin, and it lands closer to 190 kW. Conflating the two is the single most common way a 2026 design decision gets anchored to the wrong number.

The Actual Roadmap, Dated

The density curve is steep and public. Hopper-generation racks ran around 40 kW. Blackwell, the GB200 NVL72, runs roughly 120 to 130 kW. Vera Rubin, announced in full production at CES with volume availability in the second half of 2026, draws roughly 190 kW at the NVL72 configuration, with denser variants higher (The Register, Tom's Hardware, 2026). Then comes the jump everyone quotes: Rubin Ultra in the Kyber rack, at 600 kW, in the second half of 2027, followed by Feynman in 2028 (Tom's Hardware, DCD, 2026).

So the number that matters for a facility being designed in mid-2026 is 190 kW, not 600. And even the 600 kW figure carries a footnote most decks omit: the Kyber design, as NVIDIA has shown it, needs roughly two racks' worth of physical footprint because of the power and cooling sidecar it requires. A 600 kW rack is not a 600 kW slot in your existing floor plan. It is a new mechanical and electrical envelope.

Generation	Rack Power	Timing
Hopper	~40 kW	Prior generation
Blackwell (GB200 NVL72)	~120 to 130 kW	Deployed now
Vera Rubin (NVL72)	~190 kW	Volume H2 2026
Rubin Ultra (Kyber NVL576)	~600 kW (roughly 2-rack footprint)	H2 2027
Feynman	Higher still	2028

Roadmap per NVIDIA CES/GTC disclosures as reported by The Register, Tom's Hardware, and DCD, 2026. Figures are approximate and generation-dependent.

The Number That Matters

190 kW is what you deploy in 2026. 600 kW is what you read about for 2027, and it needs roughly two racks of footprint when it arrives. Designing a 2026 facility around 600 kW is designing around a product that is not shipping, on a floor plan that does not fit it.

The Obsolescence-by-Design Trap

Here is the uncomfortable engineering truth the density curve creates. 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, its cooling plant, and possibly its structure (Introl, 2026). That is not a knock on the 2024 design. It is the nature of a curve that has moved from 40 kW to a projected 600 kW in the span of a few product generations. Every fixed building is a bet on which density generation it can hold, and the useful life of that bet is getting shorter as the curve steepens.

This is why "future-proof" has quietly become one of the least honest words in a data center sales conversation. You cannot economically build a hall today whose power and cooling envelope will absorb a 5x density increase you cannot yet fully specify. Anyone who tells you their fixed facility is future-proof against the Kyber generation is either over-provisioning at enormous cost for hardware that is more than a year out, or they are hoping the curve stalls. The curve is not stalling.

The Alternative: Make the Unit Replaceable

If you cannot future-proof the building, the design goal changes. Instead of trying to make one fixed envelope absorb every future generation, you make the compute unit itself the thing that gets swapped or added when the generation turns. A purpose-built block designed for today's density, deployed against power you actually have, and replaceable when the next generation demands a different power and cooling envelope, is a fundamentally different bet than a monolithic hall you cannot economically retrofit.

This is the modular argument stated honestly, without the future-proofing spin. A modular block does not defy the density curve either. What it does is contain the obsolescence to the unit rather than the whole facility. When 600 kW Kyber actually arrives in 2027 and needs a new power and cooling envelope, replacing or adding a purpose-built block is a far smaller decision than rebuilding a hall that was spec'd around 120 kW. You are not future-proofing the building. You are making the part that goes obsolete the part that is designed to be replaced.

The Design Principle

Design for the density that ships now (roughly 190 kW in 2026), not the density in the marketing deck.
Accept that no fixed building absorbs a 5x jump economically. Future-proofing the hall is the wrong goal.
Put the obsolescence in the unit. Make the block the thing that gets swapped when the generation turns, so a rebuild is an addition, not a demolition.

The Practical Read

When someone pitches you a facility built for 600 kW racks in 2026, ask two questions. First, which shipping hardware needs 600 kW today, because the answer is none until the second half of 2027. Second, what does the footprint and cost look like when the real Kyber generation arrives with its two-rack envelope, because that is the moment the "future-proof" claim gets tested. The gap between those answers and the pitch is where the money gets wasted.

The density roadmap is not the problem. The problem is treating a 2027 number as a 2026 requirement, and treating a fixed building as if it can absorb a curve that has moved 15x in a few generations. Design for what ships, make the unit replaceable, and let the block, not the building, carry the obsolescence. That is the honest version of building for the future.

Design for the Density That Ships

The right envelope for a 2026 design is the density that is shipping, not the 2027 number. The cooling and power consequences of that density are a product question as much as a building question.

See modular data centers