That's the assembly time for a single compute tray, from Grace Blackwell to the new Vera Rubin platform. NVIDIA's CEO Jensen Huang says it's in full production now.

How they got there: 
a redesigned circuit-board midplane replaces the tangle of internal cable connections. 

No cables to route, no connectors to seat by hand. 
The fans are gone from the tray entirely, with liquid cooling handling the heat instead.

Fewer parts, fewer hands, fewer things to get wrong.

A 20x reduction in assembly time sounds like a manufacturing detail. It's closer to a capacity decision. 

Every data center build is constrained by how fast racks can be assembled, tested and serviced, and that constraint has nothing to do with chip supply.

Servicing improves for the same reason. A tray with no cables and no fans has far fewer failure points, and swapping one takes minutes rather than an afternoon.

Everyone watches the chip announcements. The thing that decides how fast this infrastructure actually goes up is whether a technician can put it together.

We're building open AI tooling at GenAI Works, and the round is open to everyone, from $1,000 with up to 20% bonus shares: https://lnkd.in/dJKiJvhK

Where do you think the real bottleneck sits now, in the silicon or on the data center floor?