Custom Silicon: Why ASICs Matter Even if NVIDIA Wins
Custom AI silicon — Google’s TPU, Amazon’s Trainium and Inferentia, Meta’s MTIA, Microsoft’s Maia, and a wave of customer-specific chips designed by Broadcom and Marvell — is often framed as a threat to NVIDIA. The framing is wrong. Custom silicon is not winner-take-all; it is workload-specific. NVIDIA wins the training and general-purpose inference market because of CUDA and ecosystem lock-in. ASICs win specific inference workloads where hyperscalers can justify the design cost. Both trends grow simultaneously, and Broadcom and Marvell collect design fees from the ASIC side regardless of which hyperscaler wins.
The Three Players in Custom Silicon
- Hyperscalers designing in-house: Google (TPU v5), Amazon (Trainium 2, Inferentia 3), Meta (MTIA v2), Microsoft (Maia). Multi-year programs, billions in design spend.
- Custom silicon foundries: Broadcom (the largest, with Google TPU among customers), Marvell (AWS, others). They design and tape out chips to customer specifications.
- IP and design tools: Arm, Cadence, Synopsys. Every custom chip uses licensed IP and design software — and these vendors are paid regardless of which chip wins.
Why ASICs Make Sense for Specific Workloads
A workload running at scale 24/7 — search ranking, ad inference, recommendation systems — can amortize a $500M chip design cost across years of operation. ASICs deliver 3–10x better performance-per-watt versus GPUs on specific narrow workloads. Once a hyperscaler reaches sufficient internal volume, in-house silicon becomes cost-justified on a clean ROI basis. The math is straightforward, and it does not depend on any particular AI trend continuing.
Why ASICs Do Not Kill NVIDIA
Training frontier models requires general-purpose programmability that CUDA + GPUs provide. New model architectures emerge faster than ASIC design cycles can follow — a 2-year chip design cycle versus 6-month architectural shifts. CUDA ecosystem effects (developer mindshare, library maturity, compatibility) compound annually. Customers who run mixed workloads — most enterprises — cannot justify ASIC design; they buy GPUs. The “NVIDIA killer” framing has been wrong for a decade because it misreads the market structure.
Where the Money Flows
NVIDIA captures training, general inference, and the long tail of enterprise AI workloads. Hyperscalers capture cost savings on their internal high-volume inference. Broadcom and Marvell capture design-services revenue regardless of which hyperscaler wins. TSMC captures wafer share from both NVIDIA and the ASIC ecosystem. The total AI silicon pie grows; the slices are different.
The Investment Implication
The barbell — long NVIDIA for training and general-purpose, long Broadcom/Marvell/Arm for the ASIC ecosystem — captures both legs of the trend. Betting against NVIDIA on ASIC threat alone has been a losing trade since 2017. Ignoring custom silicon misses one of the cleanest design-services growth stories in semiconductors.