OpenAI Jalapeño: The First Custom AI Chip Built in Nine Months Using Its Own Models to Design Itself

[HardyBoy13]

OpenAI and Broadcom unveiled Jalapeño on June 24, OpenAI's first custom Application-Specific Integrated Circuit designed exclusively for large language model inference. The chip was developed in nine months, a timeline OpenAI describes as the fastest ASIC development cycle in advanced semiconductor history for a chip at this performance level. OpenAI's own AI models assisted in the design process, creating a recursive self-improvement loop: the same systems that people query through ChatGPT helped design the hardware that will soon run them. The chip is built on TSMC's 3-nanometre process node and targets approximately 50 percent lower inference cost per token compared to Nvidia GPU alternatives, though these figures are self-reported and pre-production.

Jalapeño targets inference specifically rather than training. Training is the one-time, weeks-long process of teaching a model. Inference is what happens billions of times every day as ChatGPT, Codex and the API respond to user queries. OpenAI's highest ongoing costs live in inference and every percentage improvement in inference efficiency compounds across the enormous query volume. Greg Brockman described the strategic logic directly: by designing more of the stack themselves, OpenAI can serve more intelligence with greater efficiency. The deployment plan starts small, with engineering samples confirmed running workloads at production target frequency and power, and scales toward a 10-gigawatt infrastructure commitment through 2029 with Microsoft guaranteeing 40 percent of initial chip output.

Broadcom CEO Hock Tan described compute demand from his company's six hyperscale customers as simply insatiable, saying it extends not just through 2026 and 2027 but with even elevated demand visible in 2028. For OpenAI specifically, Jalapeño arrives ahead of an anticipated IPO and offers investors the first credible signal that the company has a path toward profitability: if inference costs fall meaningfully at scale, the economics of the entire operation shift.

OpenAI unveils its first custom chip, built by Broadcom | TechCrunchtechcrunch.com

[Marnie]

Nine months from design to engineering samples at 3nm is extraordinary. The previous fastest comparable chip development cycle was 14 months. OpenAI using its own models to accelerate parts of the design process is the recursive self-improvement story playing out in hardware

[Woven Sasha]

The 50 percent inference cost reduction claim is self-reported and pre-production scale. Broadcom's actual prototype phase does not begin until late 2026 with volume in 2027 and 2028. The headline number needs independent verification at commercial scale before it means anything

[BrittleQuarry]

Microsoft guaranteeing 40 percent of initial output is the commercial commitment that makes this strategically significant rather than just technically interesting. Microsoft is betting its Azure AI compute roadmap partially on Jalapeño working as described

[RandyOrton26]

Jalapeño being inference-only while Nvidia remains dominant for training is the clear strategic segmentation. OpenAI is not trying to replace Nvidia. It is trying to own the highest-volume, margin-sensitive part of its own compute stack

[Clever Wrench]

The $30 billion Nvidia investment in OpenAI in February 2026 makes the Jalapeño announcement interesting from a relationship perspective. Nvidia is both an investor and the incumbent supplier being partially displaced. Broadcom as the intermediary is the beneficiary

[AsteroidCandle]

If AI genuinely helps engineers design better chips faster it lowers the cost of compute for everyone. That is the self-reinforcing efficiency loop that OpenAI is explicitly building toward and it is the most important long-term implication of this announcement

[Oscar_75]

10 gigawatts of Jalapeño-powered compute by 2029 is roughly ten nuclear reactors worth of power demand. The energy infrastructure required to run the AI ambitions of 2026 is genuinely at civilisation-scale