APREX · EST. 2026
OpenAI and Broadcom have unveiled their first joint AI chip: a custom-built Intelligence Processor codenamed Jalapeño. Unlike general-purpose AI accelerators, the chip has been built from the ground up for a single task — inference with large language models (LLMs). The announcement was made on OpenAI's official blog.
Nine Months from Concept to Production
One of the most remarkable aspects of Jalapeño is the speed of its development. According to OpenAI, just nine months elapsed from the start of design to tape-out — the stage at which a chip is sent for fabrication. This is described as one of the fastest ASIC development cycles in the history of high-performance semiconductors.
OpenAI states that the company actively used its own AI models during parts of the design and optimization process, which helped significantly compress the timeline.
Jalapeño was designed from the ground up with a deep understanding of how large language models actually work — not as a generic accelerator retrofitted after the fact.
What Sets Jalapeño Apart from the Competition?
Whereas Nvidia's GPUs and Google's TPUs are built to handle a broad portfolio of AI workloads, Jalapeño is exclusively focused on inference. This specialization makes it possible to optimize the architecture to reduce data movement and balance compute, memory, and network resources more precisely — with the goal of operating closer to theoretical peak performance in practice.
Broadcom is responsible for silicon implementation and high-performance networking components, including Tomahawk network chips. Celestica handles board integration and scalable manufacturing.
Claims Parity with Nvidia Blackwell
Broadcom CEO Hock Tan is reported in research materials to have stated that Jalapeño is "just as good" as Nvidia's Blackwell GPUs and Google's TPUs in data center settings. It is a bold claim, but one that has not yet been substantiated by independent, publicly available benchmarks.
OpenAI reports that technical prototypes are already running ML workloads in the company's own labs, including GPT-5.3-Codex-Spark. Outside observers cannot yet verify the performance figures.
What Are the Competitors Saying?
Today, Nvidia dominates the LLM inference market with its Blackwell architecture (B200, GB200). Benchmarks show that the B200 can deliver up to four times higher throughput than the previous-generation Hopper GPUs on models such as Llama 3.3 70B. Google's TPU v5p, meanwhile, is optimized for large batch jobs in JAX-based environments.
Jalapeño enters a market where demand for specialized inference hardware is growing sharply as companies scale up model services. The cost argument — 50 percent cheaper per token — would, if it holds up in production, represent a significant competitive advantage.
Deployment by End of 2026
The first production rollout is planned for the second half of 2026, with data centers such as Microsoft among the early partners. OpenAI describes Jalapeño as the first in a multi-generation platform, with the next chip expected in 2028 and annual updates to follow thereafter.
How quickly and to what extent Jalapeño will actually replace purchased Nvidia GPUs in OpenAI's infrastructure remains to be seen. The company remains in a close partnership with Microsoft, which uses Nvidia hardware at massive scale. Nevertheless, it is clear that OpenAI, like Google and Amazon, is looking to reduce its long-term dependence on third-party AI hardware.