OpenAI and Anthropic Are Developing Their Own Custom AI Chips: Here Are the Stocks That Could Benefit The Most

OpenAI disclosed its first custom AI chip (late June) while Anthropic is reportedly building a ground-up accelerator with Samsung, potentially easing Anthropic’s compute shortage. The article highlights Broadcom’s Jalapeño co-designed inference chip with OpenAI, targeting 10 GW of compute capacity by end-2029, and frames Broadcom’s post-earnings dip (driven by a weaker outlook) as a potential entry point. Overall, incremental progress in custom AI silicon and foundry partnerships is positioned as a cost-and-supply tailwind for hyperscalers, but near-term investor sentiment is likely mixed given recent earnings-related pullbacks in Broadcom and Samsung.

Analysis

The cleanest winner is the intermediary that turns model demand into a repeatable design franchise, not the lab that invents the silicon. That favors AVGO: each custom accelerator win deepens switching costs with hyperscalers and can compound into networking, memory, and packaging attach, while the market still appears to be pricing it like a cyclical semi rather than an AI infrastructure toll collector. GOOG/GOOGL benefit too, but more as a margin-defense story: custom silicon lowers unit inference cost and can widen cloud economics, though it also raises near-term capex intensity.

TSM is the better second-order beneficiary than the market usually recognizes because custom-chip ambitions are bottlenecked by advanced-node capacity, not ideas. Any incremental design win that reaches tape-out tightens wafer allocation and supports pricing, with the real upside arriving over several quarters as volumes ramp; SSNLF is a higher-beta way to express foundry share gain, but execution and yield risk make it a lower-conviction call. NVDA is not facing an immediate volume break, but custom inference ASICs are a gradual share leak in the most price-sensitive workloads, which could cap multiple expansion even if unit demand stays strong.

The contrarian mistake is assuming custom chips are purely substitutionary. Lower cost per token should expand usage, which means the aggregate compute pie likely keeps growing even if mix shifts away from merchant GPUs; that argues for being selective rather than outright bearish on the infrastructure complex. The key falsifiers are slower hyperscaler capex growth, delayed ramps from foundry constraints, or evidence on earnings calls that custom accelerators are taking share faster than expected in training as well as inference.

Key Decisions for Investors

• Buy AVGO on post-earnings weakness for a 6-12 month hold; risk/reward improves if management reaffirms AI revenue trajectory and custom-ASIC attach. Falsify if AI backlog or 2026 commentary disappoints again.
• Initiate a relative-value pair: long AVGO / short NVDA for 3-9 months to express mix shift toward custom inference silicon without betting on a total AI capex slowdown. Cover if NVDA reaccelerates gross margin or hyperscaler custom-chip adoption proves smaller than expected.
• Accumulate TSM on any sector pullback; advanced-node scarcity makes it the most direct bottleneck beneficiary over the next 12-18 months. Stop if foundry utilization or capex commentary signals demand digestion rather than scarcity.
• Watch GOOGL as a defensive AI-infrastructure long, especially if cloud margins expand while TPU usage rises. Best entry is on a broad tech selloff, since the market may underprice the margin lift from cheaper internal inference.
• Speculative only: consider a small SSNLF position as a foundry-share optionality trade if there is confirmed external logic-chip tape-out news. This is a lower-conviction idea until yield, packaging, and customer-qualification data are visible.