Morgan Stanley Greater China Semiconductors Research published a new report on May 8, 2026

-> NVIDIA racks and servers are still the biggest capex driver, representing around 58–63% of total 1GW data center cost, excluding HBM and CPU

-> For custom ASICs, racks and servers are much lower, around 39% of total cost. That is the clearest reason why hyperscalers keep pushing internal silicon

-> Morgan Stanley estimates that a 1GW NVIDIA GPU data center costs:

• Vera Rubin: ~$41B

• GB300: ~$33B

• B200: ~$24B

• H100: ~$23B

Compared to custom ASICs:

• TPUv7: ~$27B

• Trainium3: ~$15B

-> Current-gen NVIDIA systems can cost up to ~2x more per GW than custom ASICs. Vera Rubin is almost 3x Trainium3 on this estimate.

-> This does not mean custom ASICs are automatically better. NVIDIA still has the strongest full-stack ecosystem, software, networking, availability, and model support. But it shows why hyperscalers have a massive economic incentive to scale ASICs.

-> The second-biggest cost bucket is networking, around 19–23% across most systems.

-> Power shell, cooling, DRAM, HBM, and CPU are meaningful, but secondary compared to racks, servers, and networking.

-> This is also bullish for rack-scale infrastructure suppliers, because the AI capex debate is not just GPU vs ASIC. It is racks, networking, power delivery, cooling, and the full data center buildout.

-> The hyperscaler world will split between premium NVIDIA clusters for maximum performance and flexibility, and custom ASIC clusters for lower-cost, more optimized workloads.

-> NVIDIA still captures the richest part of AI infrastructure capex, but custom ASICs are becoming the cost-pressure mechanism hyperscalers will use to reduce long-term dependency.