Broadcom (NASDAQ: AVGO) & Marvell (NASDAQ: MRVL) Custom AI XPU ASICs vs. Nvidia GPUs: Who Wins TSMC 3nm and CoWoS-L Wafer Capacity?

As Taiwan Semiconductor Manufacturing Co. (TSMC) expands advanced CoWoS packaging past 80,000 wafers per month, Nvidia has locked up over 60% of CoWoS-L capacity for Blackwell and Rubin GPUs, forcing custom XPU designers Broadcom and Marvell to battle for 3nm wafers and HBM4 stacks.

In SEC Form 10-K segment disclosures and supply-chain filings monitored under U.S. Department of Commerce Bureau of Industry and Security (BIS) export controls and CHIPS Act packaging grants, Broadcom (NASDAQ: AVGO) is ramping custom 3nm (N3P) AI accelerators for Alphabet's (NASDAQ: GOOGL) TPU v6p and Meta's MTIA, while Marvell Technology (NASDAQ: MRVL) scales Amazon's (NASDAQ: AMZN) AWS Trainium2.

While Nvidia (NASDAQ: NVDA) commands ~75% gross margins on merchant GPU racks fabricated at TSMC (NYSE: TSM), hyperscalers are shifting internal inference workloads toward Broadcom and Marvell custom ASICs operating at 55%–60% gross margins to lower Total Cost of Ownership (TCO). Will custom XPUs capture a growing share of TSMC CoWoS-L allocations, or will Nvidia's merchant dominance keep supply locked up?