Blackwell, decoded: GB200, GB300 and the 120 kW rack
Most coverage of Blackwell counts FLOPS. We'd rather count kilowatts, because the most important thing Blackwell did wasn't go faster. It made the room impossible to cool with air, and quietly retired a decade of data-center assumptions in the process.
One rack, the size of a small substation
The GB200 NVL72 binds 72 Blackwell GPUs and 36 Grace CPUs into one liquid-cooled rack carrying 13.5 TB of HBM3e and around 1,440 PFLOPS of FP4 tensor performance. It draws roughly 120 kW and lands somewhere near $2-3M per rack, with shipments that began in early 2025 (Awesome Agents, Spheron).
Hold that 120 kW number against your intuition. A "dense" data center rack a few years ago was 8-15 kW. Blackwell is an order of magnitude past that, in the same footprint. The power didn't trickle up. It stepped.
GB300: more memory, more heat
The GB300 NVL72, Blackwell Ultra, pushes per-GPU memory from 192GB to 288GB of HBM3e, lifts FP4 inference toward 1,100 PFLOPS, and runs each GPU near 1.4 kW, for roughly 50% more performance than GB200 (Tech Insider, Spheron). DGX Station systems in both GB200 and GB300 flavours began shipping to customers in March 2026, and GB200 NVL72 is now broadly available across CoreWeave, Oracle Cloud, Azure and Google Cloud (Introl).
Why air cooling quietly died
At 120 kW in a single rack, you are no longer cooling electronics, you're managing a concentrated heat source that a fan cannot keep up with. That's why GB200 ships as a liquid-cooled design, with coolant brought directly to the chips (Introl). Direct-to-chip liquid cooling stopped being an exotic option and became table stakes.
This is the part that catches operators out. You can buy the GPUs. You can even get them delivered. But if your hall was designed for 15 kW racks and chilled air, a Blackwell cabinet doesn't fit your facility, it fits a different category of building entirely.
The asset is only as good as the room
A $2-3M rack earns its keep by running near-continuously at peak. Two things stop that: thermal throttling and expensive power. A Blackwell rack that throttles because the loop can't carry the heat away is a multi-million-dollar machine performing like a cheaper one. The building isn't a backdrop to the silicon. It's a multiplier, or a tax, on every dollar of it.
Liwa is built for exactly this class of hardware: direct-to-chip liquid cooling, racks rated to 150 kW (headroom above today's 120 kW Blackwell, and ready for Rubin), power at $0.10/kWh, PUE ≤1.2, in a UAE free zone. You bring Blackwell, or whatever ships next, and run it under your own brand. We keep it from throttling.
Questions we're sitting with
- If your current hall tops out at 15-30 kW per rack, is a "GPU upgrade" really a hardware purchase, or a facility decision in disguise?
- What's the true cost of one throttled Blackwell rack over a year, and how much power price would you trade to never see it?
- Blackwell is 120 kW; Rubin will ask for more. Do you design for today's chip, or for the one two years out?
Got Blackwell racks and nowhere dense enough to put them?
Reserve 150 kW-ready, liquid-cooled space at $0.10/kWh, your hardware, your brand.
Sources
- Awesome Agents, GB200 NVL72 rack-scale Blackwell
- Spheron, GB200 NVL72 specs, pricing & architecture
- Tech Insider, Blackwell GPU pricing & GB300
- Introl, GB200 NVL72 liquid-cooled deployment
Specs and pricing are vendor statements and market estimates as of May 2026 and vary by configuration.