OpenAI’s Jalapeno chip outperformed the GB300 on power and speed, according to OpenAI

OpenAI claims its Jalapeno inference chip (built with Broadcom) outperformed Nvidia’s GB300 on AI work per unit of power and on response speed, with OpenAI publishing utilization/efficiency figures suggesting it can “need Nvidia less.” The comparison is limited to GB300 (no test vs Nvidia’s newer Vera Rubin hardware) and Jalapeno is positioned for inference rather than training, making the headline incrementally positive but still somewhat qualified.

Analysis

The market read-through is less about a single benchmark and more about the accelerating bifurcation between training and inference. If hyperscalers believe they can offload a meaningful share of inference onto custom ASICs, the pressure lands first on NVDA’s pricing power and mix in lower-complexity workloads, not on its data-center growth line immediately. That said, this does not threaten the training stack; it mainly raises the bar for NVDA to defend inference attach rates with newer platforms, software stickiness, and systems-level performance.

AVGO is the cleaner beneficiary because the economic value is in design wins, packaging, and long-cycle platform lock-in rather than unit shipment hype. The second-order effect is that every credible custom-chip announcement encourages other large buyers to revisit in-house silicon, which can compress merchant GPU share over 12-18 months even if total AI capex keeps rising. The likely near-term loser is NVDA sentiment, but the fundamental damage is limited unless we see broader hyperscaler procurement shifts or evidence that Vera Rubin fails to reclaim the performance-per-watt lead.

Contrarian view: the move may be overinterpreting a non-comparable test. OpenAI’s inference workload is not the full AI stack, and the lack of training relevance means NVDA’s moat is still intact where the spending is largest. The key falsifier is whether subsequent hyperscaler comments and procurement data show custom inference silicon moving from pilot to production at scale over the next 1-3 months; absent that, this is more a narrative headwind than an earnings revision event.

Key Decisions for Investors

• Tactically long AVGO / short NVDA as a 1-3 month relative-value pair, sized modestly; thesis is inference custom-silicon share gains and sentiment pressure on NVDA, with the pair invalidated if NVDA’s next platform commentary reasserts a clear performance lead.
• Use NVDA downside expressions via put spreads rather than outright shorting; best risk/reward is on a 1-2 quarter window if hyperscaler commentary broadens beyond a single customer, but the trade should be cut if data-center guidance or backlog accelerates.
• Add AVGO on weakness only if the tape confirms no broader AI multiple de-rating; the durable upside is in custom silicon and networking attach, while the risk is that investors treat this as a one-off PR item and fade it quickly.
• Watch for confirmation in upcoming capex/buying signals from large cloud customers; if multiple buyers mention inference custom silicon, rotate toward semiconductor enablers of bespoke AI stacks and away from pure GPU beta.