# OpenAI and Broadcom Unveil Jalape1o Inference Processor > Source: > Published: 2026-06-24 15:45:21.195134+00:00 Photo: platform.theverge.com · rights & takedowns OpenAI and Broadcom unveiled Jalapeno , described as OpenAI's first "Intelligence Processor," an ASIC built for LLM inference, according to OpenAI's announcement and Broadcom's investor release. Engineering samples are running ML workloads at production target frequency and power, including GPT-5.3-Codex-Spark , per OpenAI. The chip was co-developed from design to tape-out in nine months - what OpenAI calls the fastest ASIC development cycle for high-performance semiconductors. Broadcom CEO Hock Tan and President Charlie Kawwas handed the first wafer to Sam Altman and Greg Brockman. Deployment is planned for late 2026 at gigawatt scale with Microsoft and other partners. Market reaction included a Broadcom share rise of about 3.4% per Yahoo Finance. What happened OpenAI and Broadcom publicly unveiled Jalapeno , billed as OpenAI's first "Intelligence Processor," in a joint announcement and Broadcom investor release on June 24, 2026 (OpenAI; Broadcom IR). The chip is described by the companies as an ASIC accelerator optimized for LLM inference, with engineering samples reportedly running ML workloads at production target frequency and power, including GPT-5.3-Codex-Spark (OpenAI). The design-to-tape-out cycle took nine months - what OpenAI calls the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors; OpenAI's own models helped accelerate parts of the design and optimization process (OpenAI; The Decoder). Early testing shows the first-generation accelerator delivers performance per watt "substantially better" than current state-of-the-art; a detailed technical report is promised in the coming months (OpenAI; Broadcom IR). Technical details Per OpenAI's announcement, the Jalapeno architecture focuses on reducing data movement and balancing compute, memory, and networking to increase realized utilization closer to theoretical peak. The chip is a blank-slate design for modern LLM inference, not a general-purpose accelerator adapted from earlier AI workloads. Broadcom's contribution includes silicon implementation and networking technologies, notably Tomahawk networking silicon, while Celestica is named for board, rack, and system integration (OpenAI; Broadcom IR; AA). The chip is positioned for inference workloads rather than training, i.e. serving model responses in products such as ChatGPT and Codex (The Verge; AA). Industry context Editorial analysis: Industry reporting frames Jalapeno as part of a broader wave of custom silicon efforts by major AI players seeking greater cost and energy efficiency for inference. Observers note that ASICs, while less flexible than general-purpose GPUs, can offer improved performance per dollar and per watt for targeted workloads (The Verge; CNBC). Public coverage compares the new chip to existing offerings from Nvidia and Google, noting that ecosystem maturity and overall peak performance remain important factors in adoption (The Verge). The Decoder also notes that GPT-5.3-Codex-Spark currently runs on Cerebras hardware, which also specializes in inference. What to watch Editorial analysis: Practitioners and infrastructure buyers will look for the forthcoming technical report and independent benchmarks that verify the companies' performance-per-watt claims. The Decoder notes that performance figures are self-reported and untested against specified competitors under specified conditions. Other near-term indicators include compatibility with existing model runtimes, compiler and tooling support for porting models to the new accelerator, the pace of production ramp and supply commitments, and Broadcom's reported requirement that Microsoft commit to buying 40% of chips to secure the first phase (The Decoder; OpenAI; Broadcom IR; CNBC). For practitioners Editorial analysis: If independent benchmarks confirm meaningful efficiency gains for inference, operators of high-volume serving fleets could reassess cost models for at-scale LLM deployments. Companies evaluating heterogeneous inference stacks will need to consider model-porting costs, potential quantization and kernel changes, and how orchestration and monitoring will handle new accelerator types. The broader pattern of custom inference silicon increases the importance of portable runtimes and compiler tooling in production ML pipelines. Market and reaction Reporting: CNBC reported that Broadcom shares rose about 2% following the announcement, while Yahoo Finance reported a roughly 3.4% rise; coverage reflected investor interest in Broadcom's expanded role in AI infrastructure. The Verge and other outlets contextualized the move as another example of AI firms reducing dependence on GPU suppliers by developing specialized inference hardware (CNBC; Yahoo Finance; The Verge). Bottom line Editorial analysis: Jalapeno is a concrete statement of intent by OpenAI and Broadcom to pursue custom inference silicon at scale. Confirmation will hinge on the promised technical report, third-party benchmarks, and the practicalities of integrating a new accelerator into existing model-serving ecosystems. Scoring Rationale OpenAI's first custom inference ASIC with a named production partner (Broadcom), nine-month development cycle, and gigawatt-scale deployment plan is a significant infrastructure milestone. Self-reported performance claims require independent verification, but the partnership structure and deployment timeline make this a near-term industry-altering development. More OpenAI news → Practice with real Ad Tech data 90 SQL & Python problems · 15 industry datasets Used by DS/ML engineers at top companies Active Search Campaigns by Budget Easy High CPC Clicks & Poor Landing Pages Medium Campaign ROAS by Attribution Model Hard 250 free problems · No credit card See all Ad Tech problems