SpaceX Is Reportedly Building Its Own AI Rack, Buying Nvidia's Chips but Not Its Boards

Nvidia hasn't sold its biggest customers just chips for years. It sells fully built circuit boards, and increasingly entire racks. Now SpaceX reportedly wants to take that layer back: keep buying Nvidia's chips, but build everything around them itself.

What happened

On Nov. 11, chip research firm SemiAnalysis said in a thread on X that SpaceX is developing a custom AI rack codenamed "Kessel." It's a reworked version of Nvidia's next-generation Vera Rubin NVL72 rack that keeps 72 GPU packages, 18 compute trays and 9 switch trays, but swaps the internal NVLink copper cable cartridges for a single passive printed circuit board in the middle, called a midplane. Trays plug in back to back on either side of it.

The bigger change is in purchasing. In SemiAnalysis's words: "SpaceX plans to purchase chip packages from Nvidia and procure the rest directly, bypassing Nvidia's L6 compute boards. That would give SpaceX more control over supplier selection and component sourcing." Chinese financial outlets including Zhitong Finance picked up the report later that day. To be clear, this comes from a research firm. Neither SpaceX nor Nvidia has publicly confirmed it.

Some background

What's "L6"? In server manufacturing, L6 usually means a finished board with the chips mounted; L10 and L11 are complete servers and racks. Over the past few years Nvidia moved up that stack, from selling GPUs to selling HGX boards to selling whole NVL72 racks. With each step, it gained more say over whose boards, connectors, power and cooling go inside.

What's a midplane? Think of it as one giant backbone circuit board inside the rack. Every compute and switch tray plugs into it, and traces on the board replace bundles of copper cable. Nvidia's own next-generation Kyber rack uses the same idea, but SemiAnalysis previously reported that Kyber needs a 156-layer board so hard to make that the rack slipped to 2028 (Nvidia disputed that its roadmap had changed). Kessel shrinks the design so it needs only 84 layers, but on a higher-grade M10 laminate material. The trade is less density for a better chance of actually building it. Even so, an 84-layer board is extremely hard to produce.

Why it matters

This may be the first time a customer has pushed back against Nvidia at the rack level. If it works, the supplier list for boards, connectors, power and cooling becomes SpaceX's own rather than Nvidia's approved list. And a rack SpaceX designs itself would make swapping in other chips easier later. On the page for its Starmind orbital computing project, SpaceX says it is "agnostic" about AI chip suppliers. SemiAnalysis says Kessel is meant for both ground data centers and computing in orbit.

The timing is awkward for Nvidia, which is in the middle of a board investigation tied to reports about CEO Jensen Huang, and whose stock has fallen sharply. One big customer buying chips but not boards won't dent Nvidia much. If other large buyers follow, though, its model of earning a premium on fully built racks could come under pressure.

What we don't know

SemiAnalysis didn't say how far along Kessel is, how many will be built or when they'll be deployed, and it isn't clear yet whether an 84-layer M10 board can be made at volume. SemiAnalysis itself says the real bottleneck may be electricity: its earlier estimates of available power fall short of what SpaceX's planned data centers need.

What it means for you

You won't notice anything soon. But much of what AI services cost comes down to how expensive the hardware is. If the biggest buyers can skip Nvidia's bundled pricing and assemble cheaper racks themselves, AI computing could get cheaper and less dependent on one supplier over time. That would eventually show up in the price of the chatbots, AI search and other AI apps you use.

Original sources published: SemiAnalysis thread on X, Nov. 11, 2026, 11:03 Beijing time (UTC+8); Zhitong Finance, Nov. 11, 2026, 16:06 Beijing time