Why everyone from OpenAI to SpaceX is building their own chips (and turning up the heat on Nvidia)

OpenAI is moving to reduce dependence on Nvidia by developing a custom inference chip, Jalapeño, with Broadcom. The article frames this as part of a broader shift among major tech companies, including Google, Apple, and SpaceX, toward multi-sourcing AI silicon and lowering single-supplier risk. The news is modestly positive for Broadcom and slightly negative for Nvidia’s long-term competitive positioning, though the immediate market impact is likely limited.

Analysis

The market is slowly moving from a pure monopoly narrative in AI inference toward a multi-vendor procurement cycle, and that matters more for margin structure than headline unit demand. If hyperscalers and frontier labs can credibly dual-source custom silicon, NVDA’s pricing power should compress first in inference, then eventually in adjacent networking and software attach rates as customers bargain harder across the stack. The second-order winner is AVGO: even if one design win displaces less merchant GPU volume than investors fear, custom ASIC content can scale with the same capex budgets that would otherwise have gone to NVDA, making AVGO a beneficiary of silicon budget reallocation rather than just share capture.

The near-term reaction can still overstate NVDA downside because custom chips are a multiyear transition and they solve a narrower workload set than GPUs. The practical bottleneck is not chip design, it is software migration, compiler maturity, and operational risk tolerance; that means NVDA can retain the default choice for model training and bleeding-edge inference while losing only the most cost-sensitive, steady-state workloads over 12-36 months. That argues for a slower multiple de-rating than a fundamentals collapse, especially if NVDA can keep monetizing networking and full-stack deployment.

The contrarian read is that this is not primarily a moat-break story, but a supply-chain bargaining story. The more customers publicize custom chips, the more they strengthen their negotiating leverage on lead times, allocation priority, and bundle pricing with NVDA and its ecosystem suppliers. For AVGO, this broadens the TAM and improves strategic relevance; for GOOGL and AAPL, the signal is that vertical integration remains a margin-defense tool, not just an engineering flex, which could encourage peers to emulate the playbook in edge AI and on-device inference.

Key Decisions for Investors

• Reduce NVDA exposure on rallies; use a 1-3 month horizon to fade strength via bearish call spreads rather than outright shorting, since the selloff should be capped until the market prices a real migration curve. Best risk/reward is to short-dated upside convexity, not the underlying, because the thesis is multiple compression, not an imminent demand air pocket.
• Add AVGO on weakness as the cleanest relative winner from custom ASIC proliferation. Prefer a 6-12 month time frame; upside comes from broader design-win optionality and higher strategic value to hyperscalers, while downside is limited unless cloud capex rolls over materially.
• Pair trade long AVGO / short NVDA for a 3-6 month window to express the view that custom silicon narrows NVIDIA’s pricing power faster than it shrinks AI capex. Risk to the pair is an AI spend acceleration phase that re-ignites GPU scarcity and lifts NVDA gross margin more than expected.
• For event-driven traders, buy NVDA downside protection into any vendor-announcement cycle with 2-4 month tenor. The catalyst is not one customer switch, but a series of procurement disclosures that incrementally normalize multi-sourcing and pressure forward estimates.
• Keep GOOGL and AAPL on the watchlist as structural beneficiaries of vertical integration, but treat them as second-order trades only if the market starts rewarding in-house silicon as a margin defense mechanism. The cleaner expression is relative underperformance avoidance, not outright longs.