OpenAI Tests Custom Jalapeno Chip With 50% Cost Savings

OpenAI has received initial samples of its first custom AI chip and begun testing the hardware designed to run its models more efficiently than off-the-shelf alternatives.

The chip, named Jalapeno, emerged from a partnership with semiconductor company Broadcom and represents OpenAI's push to control more of its technology stack. Early testing shows the accelerator delivering approximately 50% cost savings compared to standard AI graphics processing units, according to Broadcom CEO Hock Tan.

Custom silicon for inference optimization

The companies announced the milestone Wednesday as OpenAI evaluates how the silicon handles AI workloads in real-world conditions. Custom chips allow AI companies to optimize for their specific model architectures and usage patterns rather than relying on general-purpose GPUs designed for broader markets.

OpenAI's move follows similar efforts by other major AI players to develop proprietary hardware. Google has deployed custom Tensor Processing Units for years, while Amazon Web Services offers its own Trainium and Inferentia chips for training and running AI models.

Why it matters

Inference costs—the expense of actually running AI models to generate responses—represent a growing financial challenge as AI applications scale. A 50% reduction in compute costs would significantly improve OpenAI's unit economics, especially for high-volume products like ChatGPT. Custom chips also reduce dependence on Nvidia, whose GPUs currently dominate AI infrastructure but face persistent supply constraints and premium pricing. If Jalapeno performs as early tests suggest, OpenAI could gain both a cost advantage and greater supply chain control as competition in generative AI intensifies.

From samples to production

Receiving first samples marks an early stage in chip development. OpenAI will need to validate performance across diverse workloads, refine the design based on testing results, and eventually scale manufacturing before the chip can power production systems. The timeline from samples to deployment typically spans months to over a year, depending on complexity and any necessary design iterations.

Broadcom brings deep expertise in custom chip development, having worked with major technology companies on application-specific integrated circuits. The partnership allows OpenAI to leverage Broadcom's manufacturing relationships and design capabilities without building an entire semiconductor operation in-house.

The cost savings Tan cited compare Jalapeno's performance against "typical AI graphics processing units," though the companies did not specify which GPU models served as the baseline or detail the workloads used in testing.

Bloomberg first reported these details on June 24, 2026.