For years, OpenAI has run its models on other companies' hardware — primarily NVIDIA GPUs. That changes now. On June 24, 2026, OpenAI and Broadcom announced Jalapeño, a custom AI inference chip designed from the ground up to run large language models at scale. It is OpenAI's first custom silicon — and it was built in under a year.
What Jalapeño Is
Jalapeño is a reticle-sized ASIC (application-specific integrated circuit), meaning it uses the maximum silicon area physically allowed on a single wafer exposure. That design choice reflects a deliberate priority: peak per-chip throughput over conservative die sizing.
Unlike a general-purpose GPU, Jalapeño was architected around the specific patterns of LLM inference — memory movement, kernel execution, network communication between chips, and the serving patterns that matter at scale. Broadcom built the chip based on detailed input from OpenAI's research teams, meaning the hardware encodes years of operational knowledge about how frontier AI models actually run in production.
Nine Months From Design to Tape-Out
The development speed is arguably the most striking detail. Tom's Hardware reports that Jalapeño went from initial design to manufacturing tape-out in nine months — described as the fastest ASIC development cycle achieved for a high-performance advanced semiconductor at this complexity. Traditional custom chip programs at this scale typically take two to four years.
Engineering samples are already in the lab, running ML workloads including GPT-5.3-Codex-Spark, at production target frequency and power. Broadcom's investor announcement states that early testing shows performance per watt substantially better than current state-of-the-art offerings.
The Gigawatt-Scale Ambition
The stated goal is gigawatt-scale infrastructure — deployments measured in whole power stations, not data center racks. OpenAI and Broadcom are planning these deployments with Microsoft and other partners beginning in 2026. This reflects the sheer compute appetite of frontier AI: running GPT-class models for hundreds of millions of users is an infrastructure problem that dwarfs anything the cloud computing industry has previously built.
Why This Shifts the Industry
Jalapeño is the first chip in a planned multi-generation compute platform, not a one-off experiment. That matters because it signals OpenAI committing to a long-term custom silicon strategy — following a path that Google (TPUs), Amazon (Trainium), and Microsoft (Maia) have already taken.
The strategic logic is straightforward: inference — where the day-to-day compute actually runs, serving real user requests — is where custom ASICs have a structural cost and efficiency advantage over general-purpose GPUs. Training still favors NVIDIA's ecosystem, but inference can be won by specialized silicon. If Jalapeño performs as claimed in volume production, OpenAI's cost per query could drop substantially — a significant competitive lever as the AI market matures.
The era of OpenAI running entirely on borrowed hardware is over.