OpenAI and Broadcom have unveiled Jalapeño, the ChatGPT maker's first custom AI accelerator chip — a purpose-built ASIC designed specifically for large language model inference rather than the general-purpose GPUs that have dominated the market. Source: VentureBeat
The chip reportedly cuts inference costs by about 50% compared to existing solutions, and its development timeline was extraordinary: just nine months from schematics to fabrication readiness, accelerated by using OpenAI's own models in the chip design process. Greg Brockman, OpenAI's president, described the performance-per-watt as "incredible" in an appearance alongside Broadcom CEO Hock Tan on CNBC.
Jalapeño is already being tested with GPT-5.3-Codex-Spark in production workloads, with plans to begin data center rollout by the end of 2026. Broadcom contributed core silicon and Tomahawk networking technology, while Celestica is handling board and system integration.
The push into silicon comes as OpenAI's audited financials reveal the brutal economics driving the decision: the company generated $13.07 billion in revenue in 2025 but spent $34 billion, losing nearly $21 billion — with $10.59 billion of that going to Microsoft for compute infrastructure alone.
Analysis
The Jalapeño chip is not a technical flex — it's a survival move. OpenAI lost $21 billion last year, and the single biggest line item was paying Microsoft for GPU compute. Every percentage point of inference cost reduction doesn't just improve margins; it directly determines whether the company can continue operating at its current scale without another round of dilutive fundraising. The nine-month development cycle is genuinely impressive, but speed in silicon often means corners cut in validation. If Jalapeño ships with errata that affect model accuracy — even subtly — the reputational damage to OpenAI's brand as the quality leader could outweigh the cost savings. The real question isn't whether OpenAI can build a chip, but whether it can build a reliable one at scale while simultaneously fighting a multi-front war in models, products, and geopolitics.
