23 July 2026
|
8:10:45
OpenAI and Broadcom have officially unveiled Jalapeño, a custom AI chip purpose-built for large language model inference. Announced on June 24, 2026, the chip marks OpenAI's first move into custom silicon and represents a significant step in the company's push to control more of the infrastructure powering its models, from ChatGPT to Codex and its developer API.
Jalapeño is described by OpenAI as its first "Intelligence Processor" an accelerator built entirely around the company's understanding of how large language models actually run in production. Rather than adapting an existing general-purpose GPU architecture, OpenAI designed the chip from the ground up, informed by its own roadmap of models, kernels, serving systems, and product needs.
The chip was delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom President and CEO Hock Tan and Semiconductor Solutions President Charlie Kawwas, symbolizing the formal debut of a partnership that OpenAI says extends its full-stack strategy from products and models all the way down to chip architecture.
OpenAI's hardware program lead, Richard Ho, explained that Jalapeño was optimized around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models. The goal is to execute OpenAI's most important workloads as close as possible to the hardware's theoretical performance limits.
This matters because inference the process of actually running a trained model to generate responses is where most of the ongoing computing cost of AI products lives. As demand for AI services scales, even small gains in efficiency translate into significant savings in energy, hardware, and operating costs.
One of the more striking details of Jalapeño's development is its speed. OpenAI reportedly compressed the chip's design cycle to roughly nine months, in part by using its own AI models during the design and optimization process a detail that hints at how AI tools are beginning to accelerate the very hardware development that powers them.
Jalapeño's core design philosophy centers on reducing data movement and balancing compute, memory, and networking resources, aiming for realized utilization much closer to theoretical peak performance than most current systems achieve. Engineering samples are already running machine learning workloads in the lab at production target frequency and power, including OpenAI's GPT-5.3-Codex-Spark model.
Broadcom's contribution includes its silicon implementation expertise and networking technologies, notably its Tomahawk 6 networking silicon, which delivers 1.6 terabits per second of throughput and is integrated directly into the inference stack alongside the compute die. Celestica is also involved, providing system integration support to help bring the platform to large-scale production.
Early testing suggests Jalapeño will deliver performance per watt substantially better than current state-of-the-art AI accelerators, according to OpenAI. Broadcom CEO Hock Tan has also pointed to potential cost savings of roughly 50% compared with current AI GPUs.
It's worth noting these figures are self-reported by OpenAI and Broadcom, and no independently verified benchmarks have been published yet. Both companies say a detailed technical performance report will follow in the coming months, which will offer the first real opportunity for outside experts to evaluate these claims.
Jalapeño positions OpenAI and Broadcom as a new competitive force in a market historically dominated by Nvidia's GPUs. Unlike a general-purpose accelerator, Jalapeño was engineered specifically for LLM inference workloads, a narrower but increasingly critical use case as more companies deploy large models into everyday products.
Jalapeño is also designed with enough flexibility to support current and future large language models across the broader AI industry, not just OpenAI's own models positioning it as a potential inference option for other AI developers down the line, not solely an internal tool.
OpenAI and Broadcom plan to deploy Jalapeño at gigawatt scale with data center partners, with Microsoft reportedly targeting deployment by the end of 2026. This is described as just the first chip in a multi-generation compute platform the two companies are building together, suggesting Jalapeño is the starting point for a longer-term hardware roadmap rather than a one-off product.
For enterprises and developers building on OpenAI's platform, Jalapeño's promised efficiency gains could eventually translate into lower latency for interactive AI products, higher throughput, and reduced operating costs passed down through pricing. However, since final performance data hasn't been published, businesses evaluating AI infrastructure decisions today should treat these benefits as directional rather than confirmed, and watch for OpenAI's forthcoming technical report before making infrastructure commitments based on Jalapeño specifically.
Comments
There are no comments for this Article.