OpenAI Jalapeño: The Custom Inference Chip That Puts OpenAI in the Silicon Business

OpenAI unveiled Jalapeño, its first custom AI accelerator (the “Intelligence Processor”), on June 24, 2026 — co-designed with Broadcom and Celestica. The chip marks OpenAI’s transformation from a pure model company into a vertically integrated AI infrastructure builder.

What Makes the Jalapeño Chip Different

Jalapeño is designed from the ground up for LLM inference — not a general-purpose accelerator adapted from other AI workloads. OpenAI’s deep understanding of its own transformer architectures informed every aspect of the chip design. The result, by OpenAI’s account, is performance-per-watt that is “substantially better than current state-of-the-art” in early testing.

The chip’s architectural focus is on reducing data movement. Unlike general-purpose GPUs that must handle diverse workloads, Jalapeño balances compute, memory, and networking specifically for transformer inference — aiming for “realized utilization much closer to theoretical peak.”

The Nine-Month Tape-Out

The chip went from design to manufacturing (tape-out) in nine months — one of the fastest ASIC development cycles in high-performance semiconductor history. For context, typical custom silicon cycles run 18 to 36 months. OpenAI achieved this in part by using its own models to accelerate the chip design process, creating a feedback loop between the software and hardware teams.

Engineering samples are currently running GPT-5.3-Codex-Spark at production target frequency and power levels. Deployment targets the end of 2026, with gigawatt-scale data centers.

Strategic Context: OpenAI’s Full-Stack Ambition

Jalapeño is part of OpenAI’s end-to-end strategy: products serving millions of users, models optimized for frontier performance, custom silicon tailored to those models’ specific inference patterns, and the infrastructure to deploy them at scale.

Broadcom CEO Hock Tan described the chip program as “just the beginning of a multi-generation roadmap.” The implication is clear: OpenAI is not treating this as a one-off custom chip but as the foundation of a long-term hardware platform.

What This Means for the AI Industry

Jalapeño signals several important trends. First, the cost of AI inference at scale is low enough — and margins high enough — that model companies find it economically rational to invest in custom silicon. Second, vertical integration is becoming the defining strategy in AI: every major player (OpenAI, SpaceXAI, Anthropic via partnerships, Google via TPU) is building its own hardware-software stack.

For the rest of the industry, Jalapeño means more competition in the inference hardware market, downward pressure on inference pricing as OpenAI reduces its dependency on NVIDIA GPUs, and a potential inflection point in the cost per token of state-of-the-art models. Cheaper inference means more accessible AI — and that is the outcome that matters most for end users.