The Kyber Midplane

AI supercomputers are built as refrigerator-sized cabinets ("racks") packed with GPU chips. The chips inside one rack must talk to each other constantly, at absurd speed. Today NVIDIA wires them together with about 5,000 individual copper cables per rack — over two miles of wire.

Kyber is NVIDIA's next design: throw away the cables and plug all the computers into one giant printed circuit board — the midplane — like game cartridges into a huge console. It packs twice as many GPUs into one rack. The catch: that board may be the hardest circuit board anyone has ever tried to mass-produce, and in July 2026 a fight broke out over whether it can ship on time.

Think of the rack as a bookshelf. The "books" are compute blades — each one a complete slab of computer with GPUs, CPUs, and its own water-cooling plumbing. You slide a book in, and connectors on its spine snap into the big board in the middle. Behind that board sit the "switches" — the traffic routers that let any GPU talk to any other. A second cabinet next door (the "sidecar") holds the power converters and cooling gear, so the main cabinet is almost pure computer.

A circuit board is copper wiring printed onto layers of fiberglass, like a many-storey road map. Your laptop's board has maybe 10 layers and is smaller than this page. The Kyber midplane is reportedly about a square metre with ~78 layers — a skyscraper of printed wiring — made from exotic materials so signals don't fade. Printing it perfectly is the problem: one bad via out of tens of thousands and the whole expensive board is scrap. That yield question is the heart of the entire controversy.

The GPUs shout data at each other so fast that each "bit" of information occupies a sliver of time almost too small to imagine — trillionths of a second. At that speed, copper wire behaves less like a wire and more like a bad phone line: signals blur, echo, and leak into their neighbors. Every plug, every extra centimetre, every manufacturing wobble eats into the tiny margin. Kyber's answer: stop hand-wiring and print the wiring instead, so every path is identical and machine-perfect. But that just moves the difficulty from assembly into the printing press.

Each "GPU" NVIDIA sells is actually a package containing several chips glued together — like a chocolate bar made of squares. In 2025, NVIDIA marketed Kyber by counting the squares (576). In 2026 it went back to counting the bars (144). Nothing physical changed — only the marketing arithmetic. So when you see "NVL576" in a headline, check the date: it means one futuristic rack in 2025 coverage, but eight of today's racks in 2026 coverage.

Power = voltage × current. Push the same power at low voltage and you need enormous current — which needs enormous copper bars, like needing a fire hose instead of a garden hose. A Kyber rack drinks roughly the power of 500 homes. At today's 54 volts that would mean ~11,000 amps and hundreds of kilos of copper bar; at 800 volts it's a manageable ~750 amps. That's why NVIDIA is pushing the whole industry — power companies, converter makers, data-center builders — to rewire around 800-volt DC. Try the slider below.

A 600 kW rack gives off as much heat as a few hundred space heaters running at once — no amount of fans can blow that away. So Kyber has no fans at all. Instead, warm water flows through metal plates bolted directly onto every hot chip, carries the heat out to the sidecar, and returns cooled. Removing fans is also what frees the big board to fill the whole cabinet — no holes needed for airflow. The flip side: if the water stops, everything stops.

On July 5, a respected research firm reported that the giant board can't yet be manufactured reliably and the whole rack slips a year, to 2028. Stocks of the board's Asian suppliers dropped double digits. NVIDIA responded with a carefully-worded "our roadmap is intact" — notably not "the report is wrong." Five days later Morgan Stanley relayed the real story: the chips still ship on time, but the rack design itself is being replaced with a "better solution." Translation: the report was probably right about the board being too hard — and NVIDIA had already changed course.

Reading the room: one research firm (SemiAnalysis) drove the whole story; Wall Street banks then interpreted it; NVIDIA said little. Retail investors called it fake news, engineers mostly said "the board really is that hard," and by July 10 the argument settled into "the design is changing, the schedule maybe isn't." Notably absent: any actual NVIDIA engineer or big-cloud insider speaking publicly.

The short version: can the board be printed reliably, does it ship in 2027 or 2028 (and in what shape), and can any existing building even host a rack this hungry? Nobody outside NVIDIA can answer these today.