What happened
Broadcom's latest earnings showed AI-chip revenue up 221% year over year, making it one of the biggest direct beneficiaries of the AI compute buildout.
Meanwhile, Broadcom has begun mass production of Google's TPU 8i, which the company says can match the performance of Nvidia's Vera Rubin platform.
The memory squeeze
Memory accounts for roughly 75% of AI server bill-of-materials costs, and Google has reportedly been disassembling old servers to recover DDR4 memory to ease shortages.
Why it matters
Custom ASICs (such as Google TPUs) keep siphoning demand away from Nvidia's general-purpose GPUs, pushing AI compute supply into a dual-track era of general-purpose GPUs + custom ASICs:
- For cloud providers: custom silicon cuts cost and reduces single-vendor dependence
- For Broadcom: custom-ASIC orders from hyperscalers are its AI growth engine
- For Nvidia: forced to enter custom-chip ecosystems via interconnect tech like NVLink Fusion
Source: Daily Tech News / Unite.AI (Broadcom results)
No comments yet.