Broadcom's AI semiconductor revenue exploded 221% while TSMC absorbs 2nm margin compression — and circular financing between hyperscalers, ASIC designers, and foundries is becoming a systemic risk. Today's briefing unpacks the widening gap between fabless economics and fab-heavy balance sheets.
Audio is available on Spreaker — see link below.
Broadcom just made the most compelling case yet that in AI hardware, you don't need to own a single fab to win. The numbers are stark.
TSMC's position is not weak. Its October revenue surged nearly sixty percent year-on-year to a record five hundred and eighty-six point eight nine billion New Taiwan dollars.
Broadcom's AI semiconductor revenue hit sixteen point seven billion dollars, up two hundred and twenty-one percent year on year. Seventy-three percent of that came from custom accelerators: Google's TPUs, Meta's MTIA, and others in that class.
There's a less visible part of this dynamic that matters. A report out of Taiwan suggests TSMC is now enforcing balance-sheet-based credit limits on ASIC customers, demanding prepayment and minimum cash thresholds before committing capacity.
That gap is pushing ASIC designers toward convertible debt. MediaTek raised three point nine billion dollars in convertible bonds, with Nvidia taking three point five billion of that.
One more shift worth tracking. Memory is now the binding constraint in AI inference, not compute.
The two metrics worth watching closely from here: TSMC's gross margin trajectory through the two-nanometre yield ramp, and whether ASIC designers start showing convertible debt stress if hyperscaler commitments slow. Both are live tests of how durable this AI hardware boom actually is at the infrastructure layer.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.