OpenAI Is Building Its Own AI Chip With Broadcom. Should Nvidia Investors Be Worried?

OpenAI and Broadcom launched Jalapeño, OpenAI's first custom AI chip, as part of a plan to deploy 10 gigawatts of OpenAI-designed accelerators from 2H 2026 through 2029. Broadcom's AI semiconductor revenue jumped 143% year over year to $10.8 billion in fiscal Q2, and management said AI orders topped $30 billion while guiding AI chip revenue above $100 billion in fiscal 2027. The article argues Nvidia's moat remains intact for now, with fiscal Q1 2027 revenue of $81.6 billion and data center revenue of $75.2 billion far exceeding Broadcom's current AI chip scale.

Analysis

The market is likely misreading this as a binary NVDA versus AVGO event, when the more durable second-order effect is margin migration across the AI stack. Custom silicon does not need to displace Nvidia in aggregate to matter; it only needs to capture the lowest-cost inference workloads and commoditize enough of the demand curve to compress pricing power over the next 12-24 months. That is most relevant once hyperscalers and model providers move from “can we get capacity?” to “can we lower cost per token?”—a transition that usually arrives after the first wave of buildout, not during it.

Broadcom is the cleaner relative winner because it monetizes the trend without relying on one customer’s model success, and the networking attach is the underappreciated kicker. The real economic prize in custom AI deployments is often not the accelerator alone, but the surrounding interconnect, packaging, and system-level integration; that raises AVGO’s revenue durability and makes its AI mix less cyclical than the market may assume. The risk, however, is concentration: if just one or two flagship customers delay deployment or reset spending timelines, sentiment can reverse quickly even if the structural thesis remains intact.

For Nvidia, the near-term earnings path is still supported by overall market expansion, but the stock’s multiple is now more exposed to narrative deterioration than to actual unit losses. The consensus seems to be overestimating immediate share loss and underestimating medium-term gross margin pressure as custom chips carve out inference workloads, especially in 2027-2028. The cleanest bear case is not volume collapse; it is a slower mix shift that reduces incremental pricing power and caps upside surprises. In other words, the stock can keep growing while the multiple compresses.

Key Decisions for Investors

• Long AVGO vs NVDA as a 6-12 month relative-value trade: own the custom-silicon winner with networking exposure while shorting the name most exposed to future inference commoditization; target 10-15% spread capture if AI mix shift continues.
• Buy NVDA on 3-6 month pullbacks rather than chasing strength; use the trade as a momentum-plus-quality long, but keep size smaller than prior cycles because the multiple is now vulnerable to margin narrative risk.
• Add GOOGL and META on weakness as indirect beneficiaries of custom silicon adoption: both can reduce unit inference cost and improve AI product economics without needing to win the chip race directly.
• For hedged exposure, pair long AVGO with a short basket of semiconductor hardware names that lack custom-design leverage; this isolates the more durable IP/networking economics from generic GPU beta.
• Consider NVDA downside protection via 6-9 month puts if the stock re-rates on any AI sentiment cooling; the key risk to the bull case is not this quarter’s demand, but a 2027-2028 market share and margin reset.