The Finished Chip

This is the demand signal that drives everything beneath it. NVIDIA's GPU roadmap — B200, B300, Rubin, and beyond — dictates how much CoWoS capacity TSMC needs (L05), how much HBM SK Hynix must produce (L04), how much EUV time ASML's machines must deliver (L02), and ultimately how much power the data center must supply (L14-L15).

The Custom Silicon Surge trend splits this layer into two distinct markets. The merchant market (NVIDIA, AMD) sells GPUs to anyone. The captive market (Google TPU, Amazon Trainium, Microsoft Maia, Meta MTIA) designs chips exclusively for internal use. Both markets are growing, but the competitive dynamics differ sharply. Merchant GPUs compete on software ecosystem (CUDA). Captive ASICs compete on total cost of ownership for specific workloads.

Broadcom sits at the center of the custom silicon trend — they design ASICs for Google, Meta, and others through their Custom Silicon division. Marvell does the same for Amazon and Microsoft. Both companies are evolving from networking chip companies into AI ASIC design houses, which is a higher-margin, higher-growth business.

The Inference Architecture Shift trend also flows through this layer. Training requires the largest, most expensive GPUs (B200/B300 class). Inference can potentially run on smaller, more efficient chips — creating an opening for Groq, Cerebras, and custom inference ASICs. If inference demand grows 10x faster than training (which many expect), the chip mix shifts away from NVIDIA's most expensive products toward a more diverse silicon ecosystem. That's the risk embedded in this layer.