OpenAI and Broadcom have unveiled Jalapeño, a purpose-built inference processor — the first chip OpenAI calls its own.
Jalapeño is a large-compute ASIC designed specifically for large language model inference and agentic workloads, not a repurposed training chip. The compute chiplet is estimated at roughly 840 mm² — near the physical limit of EUV lithography — and is paired with six HBM memory modules to prioritize low latency over cost. Engineering samples are already running workloads including GPT-5.3-Codex-Spark. Broadcom CEO Hock Tan says the chip will be deployed at gigawatt-scale data centers with Microsoft and other partners before the end of 2026, with tape-out reached in just nine months — roughly half the typical ASIC design cycle.
The compressed timeline matters because it signals that AI-assisted chip design is compressing hardware iteration in ways that could shift competitive dynamics fast. OpenAI also says Jalapeño is built to serve third-party LLMs, not just its own — a hint that silicon could become a revenue line, not just an infrastructure cost. That puts it on a collision course with Nvidia and AMD for data center mindshare, not just compute efficiency.
The performance-per-watt claims against AMD's Instinct MI350 and Nvidia's Blackwell accelerators arrive without a single published benchmark, which makes them marketing until proven otherwise — and the real test will come when AMD's MI400 and Nvidia's Rubin-based chips are in the field.