Yeah, no, Jensen - we frame AI fabric three ways: scale-up within the rack, scale-out across racks, scale-across data centers, right? Scale-up's our 2027 entry - today that's NVLink from NVIDIA and some PCIe switching, Ethernet scale-up mostly '27, '28. At OFC we showed XPO: 12.8T a module, 204.8T an OCP rack, cold plate to 400W, hundred-plus vendors. CPO's still science experiments, proprietary - we'll embrace open CPO down the road, but XPO's got a decade run where liquid cooling's the ticket. Partners with OSFP at 400 and 800G, not a replacement. That's the way to look at it.
← stream@ANET solid