OpenAI's Jalapeno Chip Outperforms GB300 on Power and Speed
According to OpenAI, its Jalapeno chip, developed with Broadcom, outperforms Nvidia’s GB300 in terms of AI work per unit of power and response speed at 700 watts. However, it was not tested against Nvidia’s newer Vera Rubin hardware and is not designed for model training.
Jalapeno achieved its results while running a small open model from OpenAI, as well as third-party models from DeepSeek and Moonshot, with the most significant advantage demonstrated on Kimi, the largest model tested.
Key Takeaways:
- Performance: Jalapeno beat GB300 in AI work per watt and response speed.
- Limitations: Not tested against Vera Rubin, cannot train models.
- Future: A second version is coming soon, with a third generation already conceptualized.
- Competition: Other large buyers like Anthropic and Google are also developing custom inference chips.
- Market Impact: Europe’s AI chip spending is significant, driven by initiatives like the EU’s commitment of €20bn to AI gigafactories.