ASIC programs face high failure rates ahead
The guest argued that most custom ASIC programs at major tech companies are likely to fail or be canceled over the next three years, leaving NVIDIA and Google's TPU as the dominant AI compute architectures.
The argument
The guest noted that building custom silicon is incredibly difficult, pointing out that it took Google three generations to get its TPU working correctly. While Broadcom is helping companies like Meta build custom fabrics and ASICs, the guest believes these programs will struggle to compete with NVIDIA's integrated systems or Google's mature TPU ecosystem.
The thesis, stress-tested
✓ What validates it
- ✓High-profile cancellation of custom ASIC programs by hyperscalers within the next three years
- ✓Google officially offering TPU cloud instances to external AI startups on a large scale
▸ Risks discussed
- ▸Amazon's Annapurna team could succeed with Trainium 3
- ▸Google could begin selling TPUs externally to third parties like Anthropic
- ▸Broadcom could successfully deliver open-standard Ethernet fabrics that rival NVLink
Hear it yourself
"And what Broadcom is saying to companies like Meta is, hey, we will build you a fabric that can theoretically compete with NVIDIA's fabric, which is a mixture of NVLink and either InfiniBand or Ethernet."
00:00 / 00:14
AFFILIATE LINK · ZORTIX MAY EARN A COMMISSION · NEVER A RECOMMENDATION TO TRADE