Archive · 26 Aug 2026
OpenAI Says Jalapeno Chips Beat Nvidia in Key Tests
OpenAI built a new computer chip called Jalapeno to help its AI answer questions.
The company says Jalapeno used less power and responded faster than one Nvidia chip in tests.
It is meant to run AI after the models have already been trained.
It is not designed to train those models.
OpenAI hopes to start using Jalapeno later this year.
The chip was made with help from Broadcom and uses 700 watts.
OpenAI says this could lower the cost of running its data centers.
However, the tests did not compare Jalapeno with Nvidia's newest Vera Rubin chips.
OpenAI also says it will keep using Nvidia and other computing providers.
OpenAI says its Jalapeno processor outperformed Nvidia's GB300 in power efficiency and response speed tests.
Jalapeno is designed for AI inference, not for training artificial-intelligence models.
OpenAI plans to begin using the chip to support its models later this year.
The chip was developed with Broadcom and operates at 700 watts, which OpenAI says could reduce data-center costs.
Jalapeno was not tested against Nvidia's newly shipping Vera Rubin chips, and OpenAI says it will continue using Nvidia and other providers.
- Who
- OpenAI, led in the effort by chip chief Richard Ho, developed Jalapeno with Broadcom.
- What
- OpenAI claims Jalapeno outperformed Nvidia's GB300 in AI work per unit of power and response speed.
- When
- OpenAI plans to deploy Jalapeno later this year; a second version could reach final design in the coming months.
- Why
- OpenAI wants to reduce the cost and power demands of its expanding AI infrastructure and handle more workloads with its own chips.
OpenAI's Performance Claims
Testing Limits and Continued Nvidia Reliance
Performance
OpenAI's Performance Claims
OpenAI says Jalapeno led Nvidia's GB300 in power efficiency and response speed, combining high throughput with low latency.
Testing Limits and Continued Nvidia Reliance
The comparison did not include Nvidia's newer Vera Rubin chips, which had begun shipping, so the results do not cover Nvidia's latest generation.
Cost savings
OpenAI's Performance Claims
OpenAI says Jalapeno's 700-watt operation and low-voltage performance could significantly reduce data-center costs.
Testing Limits and Continued Nvidia Reliance
The expected savings depend on OpenAI increasing Jalapeno production and achieving the anticipated cost benefits.
Replacement of other providers
OpenAI's Performance Claims
OpenAI says Jalapeno can handle larger models and could let it perform more AI work internally.
Testing Limits and Continued Nvidia Reliance
Jalapeno is not designed for model training, will not soon replace providers such as Cerebras, and OpenAI says it will continue needing many Nvidia chips and other providers.
- Chip
- Jalapeno
- Comparison chip
- Nvidia GB300
- Primary use
- AI inference, including responding to prompts and handling tasks
- Reported advantages
- Higher AI work per unit of power and faster response times in testing
- Power level
- 700 watts
- Development partner
- Broadcom
- Planned deployment
- OpenAI expects to use Jalapeno later this year
- Testing limitation
- Jalapeno was not tested against Nvidia's Vera Rubin generation
