OpenAI's new AI chip outperforms Nvidia's GB300 in efficiency tests, company says

OpenAI said its custom-built Jalapeño inference chip (with Broadcom) showed better throughput per watt and lower response latency than Nvidia’s GB300 in internal testing. The inference-focused chip runs at ~700 watts and is optimized for inference rather than model training, supporting claims of improved cost/performance for AI serving.

Analysis

This is less about one benchmark win and more about the market finally pricing inference as a separate silicon profit pool. If large cloud buyers conclude that custom chips can meet latency targets at materially better unit economics, the near-term winner is Broadcom: every incremental design win increases the stickiness of its custom compute franchise and improves mix toward high-margin, high-visibility revenue. The second-order effect is that the “AI spend = NVIDIA spend” trade is becoming less clean; a growing share of capex may shift from merchant accelerators to bespoke ASICs without reducing overall AI capex intensity.

For NVIDIA, the risk is not a sudden loss of share in the near term but a slower compression of pricing power and wallet share in inference-heavy workloads over 6-18 months. That matters because inference is the part of the stack most exposed to production scale, so even modest penetration by custom silicon can pressure refresh cycles and make the market pay more attention to competitive alternatives from AVGO, hyperscaler in-house silicon, and potentially AMD. The immediate read-through is negative for NVDA sentiment, but the real question is whether this changes 2025-2026 capex allocation or merely confirms a structural diversification already underway.

The contrarian view is that internal testing is easy to over-interpret: shipping silicon economics, software tooling, uptime, supply assurance, and deployment friction usually decide adoption, not lab throughput. If OpenAI’s own workload mix remains training-dominant or if integration takes longer than expected, NVDA’s ecosystem advantage can keep it the default standard and blunt the downside. What would falsify the bearish NVDA read is continued hyperscaler commentary showing no cannibalization of merchant GPU budgets, or AVGO failing to translate custom wins into visible backlog/revenue acceleration over the next 1-2 quarters.

Key Decisions for Investors

• Initiate a modest long AVGO / short NVDA pair for 1-3 months, sized for catalyst risk rather than conviction: thesis is gradual inference share shift and better multiple support for AVGO if custom silicon wins become more visible; stop if NVDA re-accelerates orders or AVGO commentary does not confirm design-win momentum.
• For outright exposure, prefer AVGO on pullbacks over NVDA into the next earnings cycle: the reward/risk is better if management can show backlog and hyperscaler concentration is not impairing margin quality; watch for guide-up in custom ASIC revenue or mix.
• If already long NVDA, hedge with short-dated call spreads or partial stock reduction ahead of the next hyperscaler capex print; downside is more about multiple compression than fundamental collapse, but the stock is vulnerable to any narrative shift in inference economics.
• Set a catalyst watch on cloud capex disclosures over the next 1-2 quarters: if major buyers start specifying custom inference silicon adoption, increase AVGO exposure and reduce NVDA beta; if not, treat this as a sentiment event only.
• Use a relative-value lens versus SMH/XLK rather than a standalone directional call: the market may overreact to one benchmark, but the durable trade is only if custom silicon adoption becomes a broader industry pattern.