OpenAI and Broadcom unveiled Jalapeño, a custom ASIC for AI inference optimized for cost (~50% reduction) and enabling gigawatt-scale deployment from 2026.

OpenAI and Broadcom have unveiled Jalapeño, a custom ASIC chip designed specifically for AI inference workloads. The announcement signals a strategic shift toward custom silicon as major AI builders look beyond traditional GPU suppliers like Nvidia to optimize their infrastructure costs and scaling capabilities.

The Jalapeño chip delivers approximately 50% cost reduction for inference operations, addressing one of the largest operational expenses in AI deployment. The partnership enables gigawatt-scale deployments, with OpenAI targeting deployment from 2026 onward, indicating the company's readiness to transition significant inference workloads off standard processors onto purpose-built silicon.

This move underscores the broader industry transition in AI infrastructure. As inference becomes a larger fraction of AI compute demand and custom silicon matures, major cloud builders and AI labs are reducing dependency on general-purpose GPU suppliers. For the AI buildout, this diversification of chip architectures and supply chains signals accelerating infrastructure verticalization and a shift toward purpose-optimized silicon architectures for specific AI workload categories.