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

OpenAI unveiled its first custom AI chip (Jalapeño) and plans broader inference deployments, while Anthropic is reportedly pursuing a ground-up accelerator with Samsung as TSMC faces capacity constraints. Broadcom shares fell after a disappointing AI-related earnings outlook, but the article frames the setup as an opportunity tied to custom compute demand and potential easing of token costs via in-house silicon. Overall, the news is constructive for AI compute infrastructure prospects but highlights near-term volatility in key suppliers (Broadcom/Samsung).

Analysis

This is less a “winner takes all” AI-chip shift than a redistribution of margin from merchant GPUs toward vertically integrated stacks. The first-order beneficiaries are the firms that own the design relationship and the manufacturing bottleneck: they get sticky, multi-year programs and more pricing power than a one-time chip sale suggests. The hidden loser is GPU mix at the margin — especially inference workloads where economics matter more than absolute performance and buyers can justify ASIC-like customization faster than consensus expects.

The second-order effect is that lower per-token cost can expand total usage, so the real demand destruction risk for NVIDIA is narrower than bears assume. In the near term, the biggest earnings leverage is likely in cloud economics, not chip revenue: better custom silicon utilization should improve gross margin and differentiate the cloud stack, while any foundry gains for Samsung are more of an execution option than an immediate profit driver. TSM still looks structurally better positioned than the headline would imply because leading-edge capacity, packaging, and ecosystem lock-in matter more than nominal customer wins.

Contrarian view: the market may be too eager to capitalize Samsung’s potential foundry share gain and too slow to discount how long it takes to move from announcement to repeatable yield. If model demand softens, if custom-chip ROI is delayed, or if NVIDIA resets pricing aggressively, the whole custom-silicon thesis gets pushed out by quarters, not days. The real falsifier is not talk of partnership but evidence of production ramps, utilization, and AI capex conversion in the next 1-2 earnings cycles.

Key Decisions for Investors

• Buy AVGO on post-earnings weakness, targeting a 3-6 month rebound as custom-design revenue visibility improves; stop if the next AI revenue guide fails to reaccelerate or customer concentration worsens.
• Long GOOGL/GOOG vs. short NVDA as a 6-12 month relative-value trade: TPU-driven inference economics should help cloud margins while merchant-GPU pricing faces mix pressure; cover if NVDA re-accelerates datacenter growth or custom chip ramps slip.
• No aggressive chase in SSNLF yet; treat it as an option on foundry share, not a core long. Revisit only after evidence of AI accelerator yield and packaging execution, otherwise the market is likely ahead of fundamentals.
• Stay constructive on TSM on dips rather than trading the Samsung headline; the moat is in constrained leading-edge capacity and packaging, and that should keep volume tight even if share headlines wobble.
• Watch next quarterly capex and commentary from OpenAI/Anthropic and Google Cloud as the key catalyst set; if capex expands faster than expected, buy the ecosystem, but if spend pauses, reduce exposure to AVGO and the foundry names first.