Totally agree, Dan — inference is the real margin-killer coming.
Training gets the headlines, but the fabric (switches + optics) is quietly becoming a bigger line item than the GPUs themselves. That fabric = Broadcom’s kingdom.
Tomahawk 5 in 70-80% of new clusters, 1.6T optics monopoly, even NVIDIA’s Rubin needs their CPO in 2026.
Beautiful asymmetry: hyperscalers buy NVIDIA accelerators → still write $15-25B checks to Broadcom just to make them talk.
Modern gold rush, same winner: the guy selling the railroad.
Totally agree, Dan — inference is the real margin-killer coming.
Training gets the headlines, but the fabric (switches + optics) is quietly becoming a bigger line item than the GPUs themselves. That fabric = Broadcom’s kingdom.
Tomahawk 5 in 70-80% of new clusters, 1.6T optics monopoly, even NVIDIA’s Rubin needs their CPO in 2026.
Beautiful asymmetry: hyperscalers buy NVIDIA accelerators → still write $15-25B checks to Broadcom just to make them talk.
Modern gold rush, same winner: the guy selling the railroad.
Thanks for your thoughts🦉