Blog

GPU racks now draw well over 100 kW, far beyond what most air-cooled rooms were built for. Before your first liquid-cooled rack arrives, check heat load, heat rejection, CDUs, piping, leak detection and operations — here is the practical checklist.

GPUs only earn their money while they work — and too often they wait for data. NetApp and Supermicro now offer jointly validated AI infrastructure: Novus, AIPod and a reference design with Supermicro HGX B300, HGX B200 and RTX PRO 6000 servers. Here is what it includes and how to size it.

Book a real Supermicro GPU or CPU server, log in from your browser and run your own workloads on it — free. Here is how JumpStart works, which systems you can test today, and how to turn a good test into a delivered system.

Ten rack-ready Supermicro building blocks share one GPU generation, but not the same rack, power or cooling needs. We compare HGX B300 in 4U, 8U and 2-OU, GB300 and GB200 NVL72, GB200 NVL4 and HGX B200 — and show which one fits your room.

Stands in Tallinn and Helsinki, a data centre opening in Riga, a robot that stopped to wave at us, a trail e-bike ride in the rain and NVIDIA Vera Rubin NVL72 up close. Here’s where we’ve been in 2026 — and where to meet us next.

In a dense liquid-cooled node the back of the chassis is full: coolant hoses with quick disconnects, 6,600 W power supplies and fan modules. Supermicro moved networking, storage and management to the front, so NICs, DPUs, E1.S drives and IPMI are all reachable from the cold aisle — no work next to live coolant lines.

The opposite answer to last week’s four-GPU edge server: up to 84 AMD EPYC cores in a short-depth 1U box, and no accelerator cards at all.

Four double-width GPUs in 2U — twice the accelerator capacity of anything else in Supermicro’s Intel edge line-up. Built for sites where inference has outgrown small boxes but still cannot move to a public cloud region.

If your grid connection is fixed, megawatts are your scarcest resource — and throughput per megawatt is the metric that decides how much AI business your existing envelope can carry. NVIDIA measured Vera Rubin NVL72 against GB300 NVL72 on recorded agentic coding sessions, not single-prompt benchmarks. But the 30× is not one number: it widens from roughly 2× to 30× as interactivity rises, and where your product sits on that curve changes which platform is economically correct.

NVIDIA's 800 VDC architecture is usually read as a greenfield story. The commercially interesting part is the opposite: the MGX-compatible power rack shipping in the second half of 2026 is designed to drop into halls that are already built, delivering 800 VDC inside the row without touching the building's electrical system. Higher density becomes a row-level equipment decision instead of a multi-year rebuild.