OpenAI Unveils Custom 'Jalapeño' AI Chip Developed With Broadcom

OpenAI Unveils Custom ‘Jalapeño’ AI Chip Developed With Broadcom OpenAI’s decision to build its own AI chip, Jalapeño, marks a pivotal shift in the race for computing power, pitting Big Tech’s custom silicon bets against Nvidia’s long-dominant GPUs.

In June 2025, OpenAI quietly began a nine‑month sprint to design an application‑specific integrated circuit (ASIC) optimized purely for large language model (LLM) inference, working closely with Broadcom on implementation and manufacturing. The company now describes Jalapeño as its first “Intelligence Processor,” architected around its roadmap of models, kernels, and serving systems to make advanced AI “faster, more reliable, and more accessible to more people.”

By late June 2026, engineering samples were already running machine‑learning workloads, including next‑generation systems like GPT‑5.3‑Codex‑Spark, at production target frequency and power. OpenAI says early tests show Jalapeño will deliver “performance per watt substantially better than current state‑of‑the‑art,” with a full technical report promised in the coming months. The chip is the first step in a “multi‑generation compute platform” the firm expects to begin deploying widely by the end of 2026.

Public rollout details followed quickly. OpenAI confirmed it had begun testing Jalapeño in its labs on tasks similar to answering Codex queries, reporting “even better thermal performance than anticipated,” with commercial use at Microsoft and other partners expected by year’s end and volume ramping in 2027. The company ultimately wants its custom chips to power 10 gigawatts of compute by 2029, a bid to control costs and reduce dependence on Nvidia.

Industry and investor reactions frame Jalapeño inside a broader realignment. The Verge highlighted Broadcom CEO Hock Tan’s claim that Jalapeño matches Nvidia’s Blackwell GPUs and Google’s Tensor Processing Units, underscoring how custom ASICs are becoming a credible alternative for inference workloads. Ars Technica similarly emphasized that Jalapeño is designed from scratch for data‑center LLM inference, with both companies presenting it as the opening move in a long‑term hardware refinement effort.

TechCrunch cast the chip as “Big Tech’s spiciest move away from Nvidia,” noting that OpenAI now joins Google, Apple, Amazon, Meta and even SpaceX in building custom silicon to escape single‑supplier risk while tuning hardware to their own AI stacks. As one analysis put it, custom chips are “less of a clean break and more of a hedge,” analogous to the performance gains Apple achieved when it moved from Intel to its own processors.

Inside OpenAI, leaders have framed Jalapeño as both a technical and strategic milestone. President Greg Brockman described it as “designed from scratch for LLM inference over nine months,” adding that performance per watt looks “incredible.” CEO Sam Altman reacted more tersely but enthusiastically: “team cooked, spicily.”

Beyond the corporate narrative, open‑source advocates see the move as validation of a different trend. Hugging Face CEO Clément Delangue amplified a view that “Local and Opensource AI are going to win,” suggesting that as hyperscalers lock in efficiency with custom chips, independent developers will push harder on decentralized and local alternatives.

Together, these perspectives sketch a turning point: Nvidia remains central for training, but the economics and control of inference are shifting. Jalapeño is less about abandoning GPUs outright and more about reshaping the AI stack—from model to data center rack—around the specific demands of large language models.

Continue reading https://foxvector.com/stories/019f0979-e85a-360d-7079-158489ad8279