OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-designed AI inference processor engineered for large language model workloads, set to deploy at scale by late 2026.

OpenAI announced the OpenAI Jalapeño Intelligence Platform, developed in partnership with Broadcom, marking a significant step in the company's efforts to build custom silicon. The partnership represents a major win for Broadcom's design capabilities as OpenAI addresses massive demand for model serving capacity by investing in proprietary accelerator chips, following the established pattern of traditional hyperscalers.

The accelerator design achieved a noteworthy milestone: OpenAI completed tape-out in just nine months. In a photograph showing Sam Altman and Broadcom CEO Hock Tan with the accelerator, the design features HBM stacks positioned on either side of the silicon, with what appears to be six spaces allocated for high-bandwidth memory modules.

The silicon is already operational in OpenAI's labs, running GPT-5.3-Codex-Spark. The platform incorporates Broadcom Tomahawk silicon for networking and relies on Celestica for board and rack-scale infrastructure. OpenAI indicated that architectural details and performance data will be disclosed over the coming months.

The development reflects a significant shift in the industry landscape. Rather than seeing primarily startups attempt to build next-generation NVIDIA competitors, the cutting-edge custom silicon efforts are now dominated by large companies, hyperscalers, and major AI serving enterprises. OpenAI's announcement represents a representative example of this broader trend reshaping competitive dynamics in AI infrastructure.