The Future of AI: ASICs vs GPUs

Ben and Jay unpack why Broadcom’s “fourth customer” (~$10B) custom-ASIC win reset sentiment even after a modest beat/raise, and how that squares with hyperscalers second-sourcing away from NVIDIA in the near term. They frame the true battleground as networking—Ethernet’s ubiquity vs. NVLink’s tight integration—then differentiate GPUs’ performance-per-watt advantages from custom ASIC cost calculus, arguing that “lumpiness” (program outcomes) is not “cyclicality” (inventory swings). They stress TAM realism: it’s easy to total up CapEx, but the ROI numerator (revenue/profit) is still unknowable. Structurally, TSMC remains the default winner, with a plausible Intel Foundry financing path in the wings, while Google looks more likely to “sell capacity” for TPUs than chips. Net: GPUs keep the bulk of spend through 2030 even as select first-party silicon scales, and the market should judge claims against networking choices and workload fit—not headlines.