OpenAI has introduced Jalapeño, its first custom AI processor designed for large language model inference. The new chip was developed with Broadcom and is intended to power products such as ChatGPT, Codex, the OpenAI API, and future AI agents.
Jalapeño is not a general-purpose accelerator built around older AI designs. OpenAI says the processor was created specifically for modern LLM inference, where speed, low response times, and efficiency matter most. The company plans to begin deploying Jalapeño systems before the end of 2026.
The launch shows that OpenAI is moving deeper into custom silicon as demand for AI computing continues to grow. Rather than relying entirely on Nvidia GPUs and other third-party hardware, OpenAI is building its own long-term hardware platform for AI services.
Jalapeño Is Designed for ChatGPT, Codex, and Future AI Agents
Inference is the stage where an AI model processes a request and generates a response. It is different from training, which involves building and improving a model using huge datasets.
OpenAI says Jalapeño is designed around the real workloads it already runs every day. That includes ChatGPT conversations, coding requests through Codex, API queries, and future agentic systems that may need to handle longer and more complex tasks.
The processor is intended to combine high throughput with lower latency. That could help OpenAI process more requests at once while keeping response times fast for interactive products.
| Detail | Jalapeño information |
|---|---|
| Developer | OpenAI |
| Production partner | Broadcom |
| Main purpose | LLM inference |
| Expected deployment | By the end of 2026 |
| Target products | ChatGPT, Codex, API, agentic tools |
| Memory design | Eight HBM memory locations |
| Platform plan | Multi-generation AI compute platform |
OpenAI says engineering samples are already running AI workloads at the company’s target power and performance levels. One of the models reportedly tested on the platform is GPT-5.3-Codex-Spark.
OpenAI Designed the Chip in Just Nine Months
OpenAI says Jalapeño was developed from early design through tape-out in around nine months. Broadcom helped turn that design into a production-ready chip platform, while Celestica is expected to support board, rack, networking, and system-level manufacturing work.
The company has not shared full technical specifications, including chip size, manufacturing process, total compute performance, or bandwidth details. However, images of the processor appear to show multiple high-bandwidth memory locations surrounding the central compute area.
HBM is important for AI accelerators because large language models need fast access to enormous amounts of data. Memory bandwidth can be just as important as raw compute power when handling large AI models at scale.
OpenAI is positioning Jalapeño as the first generation of a broader AI hardware effort rather than a one-time project.
Custom AI Hardware Could Reduce Dependence on Nvidia
Nvidia remains one of the dominant suppliers of AI hardware, but major technology companies are increasingly developing their own chips.
Google has its Tensor Processing Units, Amazon has Trainium and Inferentia chips, and other AI companies are exploring custom accelerator designs. OpenAI’s move follows the same wider trend.
Custom silicon gives companies more control over performance, power use, software optimisation, supply planning, and future data centre expansion. It can also reduce dependence on one supplier during periods of high demand and limited hardware availability.
OpenAI has already committed to major deployments of Nvidia infrastructure, so Jalapeño is unlikely to replace all of its existing GPU systems. Instead, the processor should become another part of a larger and more diverse compute strategy.
Broadcom Will Help Scale the Jalapeño Platform
Broadcom will play a major role in bringing Jalapeño from chip design to large AI data centre deployments. The companies are expected to combine custom accelerator hardware with Broadcom networking technologies for scale-up and scale-out systems.
This matters because modern AI infrastructure is not just about a single processor. Large models run across racks of servers, connected through high-speed networking, memory systems, storage, and power infrastructure.
OpenAI’s custom chip project also fits into its existing multi-year collaboration with Broadcom. That partnership is focused on building large-scale AI accelerator and networking systems for future OpenAI infrastructure.
Jalapeño Could Shape OpenAI’s Future AI Services
OpenAI’s first custom chip will not change ChatGPT overnight, but it could become important as the company expands its AI products.
A processor built specifically for LLM inference may help OpenAI serve more people, handle increasingly capable models, and lower the cost of running AI services at scale. It could also support faster responses for coding tools, APIs, enterprise products, and AI agents that need to reason through multiple steps.
Jalapeño is a major sign that OpenAI is no longer treating hardware as something supplied entirely by partners. The company is now designing part of the compute foundation that will support its next generation of AI products.
Discussion (0)
Be the first to comment.