OpenAI and Broadcom unveiled Jalapeño, a custom ASIC for LLM inference, on June 24, 2026. The chip, designed to run large language models faster and cheaper, targets volume deployment by late 2026.
Key facts
- Jalapeño is an ASIC for LLM inference, not training.
- Volume deployment targeted by late 2026.
- OpenAI has raised over $40 billion in total funding.
- Chip developed in partnership with Broadcom.
- No performance metrics or cost data disclosed.
OpenAI and Broadcom have revealed a custom AI chip called Jalapeño, an Application-Specific Integrated Circuit (ASIC) built specifically for large language model inference. According to Bloomberg, the chip is part of OpenAI's bid to gain an edge by tailoring hardware to its AI products. The Verge reports that Jalapeño is designed to power current and future LLMs, with a focus on inference rather than training. The Decoder adds that volume deployment is targeted by late 2026.
Key Takeaways
- OpenAI and Broadcom unveiled Jalapeño, a custom ASIC for LLM inference, targeting volume deployment by late 2026.
- No performance metrics were disclosed.
Why an ASIC for Inference?
Jalapeño is an ASIC, meaning it is hardwired for a specific task — in this case, AI inference. This contrasts with GPUs, which are general-purpose and more flexible. By customizing the chip for inference, OpenAI and Broadcom aim to improve performance and energy efficiency, potentially lowering the cost of running models like GPT-5.3 and ChatGPT. The move follows a broader industry trend: Google has its TPU, Amazon has Trainium and Inferentia, and Microsoft has partnered with AMD and Intel for custom silicon. OpenAI, which has raised over $40 billion in total funding [per our knowledge graph], is now joining the custom-chip club.
Performance Claims and Missing Metrics
OpenAI and Broadcom have not disclosed specific performance metrics, power efficiency gains, or cost reductions for Jalapeño. The announcement per the OpenAI blog is light on technical detail, stating only that the chip will "improve performance, efficiency, and scale across AI systems." This lack of data makes it difficult to compare Jalapeño to existing inference chips like Nvidia's H100, Google's TPU v5p, or Amazon's Inferentia 2. The absence of benchmark numbers suggests the chip may still be in early production or that OpenAI is keeping competitive details close to the vest.
Strategic Context
This announcement comes amid a flurry of activity from OpenAI. In the past week alone, the company filed paperwork for an IPO [per our knowledge graph], partnered with 25+ security firms, and developed a technique to predict model failures. The Jalapeño chip could be a key enabler for OpenAI's scaling ambitions, especially as it pushes toward AGI. By controlling its own hardware, OpenAI reduces reliance on Nvidia, which has faced supply constraints and high prices for its H100 and B200 GPUs. The partnership with Broadcom, a major networking and ASIC designer, leverages Broadcom's expertise in high-volume chip production.
What's Missing
The announcement does not specify which foundry will manufacture Jalapeño (likely TSMC or Samsung), the chip's process node, memory bandwidth, or power consumption. Nor does it clarify whether the chip will be used exclusively for OpenAI's internal workloads or offered to third-party customers via Azure or OpenAI's API. The target of "late 2026" for volume deployment leaves a window for competitors to respond. Nvidia, for example, is expected to release its next-generation Blackwell Ultra GPU in 2026, which could offer competitive inference performance.
What to watch
Watch for specific performance benchmarks from OpenAI or Broadcom in Q3 2026, and for any announcements about which foundry and process node will manufacture Jalapeño. Also track whether Nvidia's Blackwell Ultra or Google's TPU v6 respond with competitive inference specs.
Source: bloomberg.com
