OpenAI's New Custom AI Chip
OpenAI and Broadcom Unveil Jalapeño AI Chip, Shifting Towards In-House Hardware for LLMs
Key Points
Engineering samples of Jalapeño are already running ML workloads in OpenAI’s lab at production target frequency and power, including GPT-5.3-Codex-Spark.
Broadcom’s Tomahawk networking silicon is used as part of Jalapeño’s silicon implementation to enable high-performance, scalable data-center networking for the platform.
OpenAI and Broadcom plan to deploy Jalapeño at gigawatt scale with data-center partners across multiple generations.
OpenAI positions Jalapeño as a key step in its broader strategy to build the full stack behind its models, moving beyond reliance on Nvidia GPUs to run AI workloads.
Broadcom’s Hock Tan characterized Jalapeño as being on par with Nvidia’s Blackwell chips or Google’s TPUs in capability, underscoring its competitive positioning.
OpenAI and Broadcom officially unveiled "Jalapeño" on June 24, 2026 — OpenAI's first custom-built AI chip, designed to run large language models faster and at roughly half the cost of existing Nvidia GPUs. Financial Post reported that engineering samples are already running GPT-5.3-Codex-Spark in OpenAI's lab at full production power and speed.
Broadcom CEO Hock Tan called Jalapeño on par with Nvidia's Blackwell chips and Google's TPUs in raw capability, according to Reuters. OpenAI CEO Sam Altman described the chip as fulfilling a multi-year effort to secure the company's "compute destiny."
Nine Months From Blank Slate to Working Silicon
Jalapeño went from initial design to manufacturing tape-out in just nine months — an unusually fast cycle for a custom processor. GlobeNewswire reported the chip was built entirely from scratch, shaped by OpenAI's daily operational data from ChatGPT and Codex. Richard Ho, OpenAI's hardware lead and a former Google TPU engineer, said the chip was "designed from the ground up for LLM inference."
TSMC is fabricating the chip on a 3nm-class process, according to Investing.com. Celestica is handling board design and rack-scale manufacturing. Broadcom's Tomahawk networking silicon is built into the platform to support fast, scalable data-center connections.
50% Cost Cut Puts Nvidia's Pricing Power on Notice
Broadcom's Hock Tan predicts Jalapeño will cut per-query inference costs by roughly 50% compared to general-purpose GPUs. Leader Post noted early tests already show meaningful cost savings. Analysts at Startup Fortune said the move puts Nvidia's pricing power "on notice" in the high-margin inference market.
Inference — running a model to answer questions — now makes up about two-thirds of all AI compute spending, according to Startup Fortune. General-purpose GPUs burn extra power doing tasks they weren't built for. Jalapeño skips that waste by targeting only the specific math patterns LLMs actually use.
Gigawatt-Scale Rollout Targets 10GW Through 2029
OpenAI and Broadcom plan to deploy Jalapeño at gigawatt scale across multiple generations of data centers. The total planned power capacity reaches 10 gigawatts by 2029, according to Startup Fortune. Microsoft is reportedly set to receive 40% of the initial chip supply for its Azure-OpenAI clusters, per Global News.
Greg Brockman, OpenAI's president, said Jalapeño is "part of our long-term full-stack strategy to make compute more abundant," according to GlobeNewswire. The platform is designed to work with both current and future OpenAI models. OpenAI still relies on Nvidia for large-scale training clusters and has no plans to abandon them entirely.
Independent Benchmarks Still Missing, Safety Flags Raised
Analysts at Constellation Research called Jalapeño a genuine "blank-slate design" — not a retooled GPU. But critics warn that OpenAI has not yet released a full technical report with independent benchmarks. Self-reported lab numbers from engineering samples may not hold up at gigawatt-scale deployment in real data centers.
Broadcom shares rose 3.4% in pre-market trading after the announcement, pushing Broadcom's market cap to $1.81 trillion, according to Investing.com. Meanwhile, specialized AI safety newsletters flagged that GPT-5.3-Codex-Spark — the model running on Jalapeño — carries a "High" cybersecurity risk rating, raising questions about whether faster, cheaper hardware could accelerate deployment of risky autonomous agents.
Sentiment
Mostly PositivePublishers
73
Articles
166
Reach
239
