OpenAI and Broadcom Unveil 'Jalapeño,' a Custom AI Inference Chip Designed in Record Nine Months to Halve LLM Serving Costs

For years, the artificial intelligence boom has been constrained by a single, expensive bottleneck: the hardware required to run the models. Now, the industry's most prominent player is taking matters into its own hands. OpenAI, in direct collaboration with semiconductor giant Broadcom, has unveiled "Jalapeño," its first proprietary AI inference processor. The custom-built chip represents a massive strategic pivot for the ChatGPT maker, marking its transition from a pure software and research company into a vertically integrated hardware architect. Designed from a blank slate to serve large language models (LLMs), Jalapeño aims to drastically reduce the energy and financial costs associated with generating AI responses.[1]

The announcement underscores a fundamental shift in the AI supply chain. Historically, companies like OpenAI have relied heavily on general-purpose graphics processing units (GPUs), primarily from Nvidia, to both train their models and serve them to users. While GPUs are unmatched for the mathematically intensive process of training AI, using them for "inference"—the act of generating live responses for millions of users—is highly inefficient and astronomically expensive. By designing an Application-Specific Integrated Circuit (ASIC) exclusively for inference, OpenAI is targeting the exact phase of the AI lifecycle that drives its daily operational costs.[2]

What makes Jalapeño particularly notable is the unprecedented speed of its creation. In the semiconductor industry, designing an advanced custom chip typically takes two to three years. OpenAI and Broadcom transitioned Jalapeño from an initial concept to "tape-out"—the final design phase before physical manufacturing—in just nine months. This rapid development cycle was achieved through a unique feedback loop: OpenAI used its own advanced AI models to simulate, design, and optimize portions of the hardware architecture. As OpenAI noted, the same models served to users are now helping to engineer the infrastructure required to run future models.[1]

From an architectural standpoint, Jalapeño is built to solve the primary bottleneck in modern interactive AI: data movement. When an LLM generates a response, it must constantly shuffle massive amounts of data between the compute cores and the memory banks. To address this, Jalapeño utilizes a contemporary multi-chip module design. It features a large, centrally located logic tile flanked by eight stacks of HBM3E (High Bandwidth Memory), all connected via a high-speed interposer. This configuration prioritizes high throughput and low latency, ensuring that the chip can feed data to the processor as fast as the processor can calculate it.[2]

The chip's design is heavily informed by OpenAI's deep understanding of its own software stack. Rather than building a generic AI accelerator, OpenAI optimized Jalapeño's architecture around the specific kernels, memory movement patterns, and serving systems that power ChatGPT, Codex, and future agentic AI products. This tight integration between software and hardware is conceptually similar to Google's Tensor Processing Units (TPUs), allowing every layer of the technology stack to be tuned for maximum efficiency. By controlling the hardware, OpenAI can ensure that its future models are not constrained by the limitations of off-the-shelf silicon.[1][2]

Early laboratory testing suggests the gamble is paying off. Engineering samples of the Jalapeño chip are already running machine learning workloads, including an unreleased model dubbed "GPT-5.3-Codex-Spark," at target production frequencies and power levels. While final performance metrics are still being measured, OpenAI claims that the first-generation accelerator is delivering a performance-per-watt ratio that is "substantially better than current state-of-the-art" hardware. If these claims hold true in real-world deployments, Jalapeño could operate close to its theoretical maximum efficiency, drastically lowering the electricity required to power advanced AI.[1]

Bringing a chip of this complexity to market requires a massive coalition of industry heavyweights. Broadcom provided the crucial silicon implementation expertise and integrated its Tomahawk networking technology, which is essential for linking thousands of chips together in a data center. Taiwan Semiconductor Manufacturing Company (TSMC) has been tasked with the physical fabrication of the silicon, while Celestica is managing the engineering of the server boards and rack systems. This consortium approach allows OpenAI to leverage the manufacturing scale of established hardware giants without having to build its own fabrication plants.[1]

The scale of the planned deployment is staggering. Jalapeño is not an experimental prototype; it is the foundation of a multi-generation compute platform. Broadcom executives confirmed that initial deployments in enterprise data centers will begin in late 2026. Furthermore, Broadcom has secured a contractual commitment to deploy 1.3 gigawatts of compute infrastructure for OpenAI in 2027, as part of a broader 10-gigawatt agreement extending through 2029. To put that in perspective, a single gigawatt is roughly equivalent to the power consumption of a mid-sized city, highlighting the sheer physical footprint of the next generation of AI.[1][2]

For enterprise customers, the introduction of Jalapeño could fundamentally alter the economics of artificial intelligence. Currently, businesses face explosive costs associated with "tokenomics"—the pricing model based on the volume of data processed by an AI. As companies integrate AI deeper into their operations, their token consumption and associated costs have skyrocketed. By changing the economics of converting electricity into tokens, Jalapeño gives OpenAI the financial bandwidth to lower prices. Analysts suggest this could trigger a price war in the model ecosystem, making advanced AI significantly more affordable for end-users.

Beyond cost savings, Jalapeño represents a strategic declaration of independence. By diversifying its infrastructure away from a strict reliance on Nvidia, OpenAI gains immense leverage in the hardware market. While the company will continue to use third-party GPUs for the computationally grueling task of training new models, handling inference internally protects OpenAI from supply chain bottlenecks and the steep profit margins commanded by incumbent chipmakers. It ensures that as AI transitions from a research novelty to a ubiquitous utility, the underlying engine remains firmly under OpenAI's control.[1][2]

The ripple effects of this launch will be felt across the semiconductor industry. For Broadcom, the Jalapeño partnership validates its custom AI accelerator strategy, providing tangible evidence that its silicon roadmap is progressing and strengthening its revenue visibility for years to come. Broadcom expects its AI semiconductor revenues to exceed $100 billion in 2027, driven largely by sustained demand from hyperscale customers like OpenAI. Meanwhile, incumbent GPU manufacturers will face increasing pressure to prove that their general-purpose hardware can compete with the efficiency of application-specific designs.

Ultimately, the unveiling of Jalapeño marks the maturation of the artificial intelligence industry. Just as the smartphone era eventually required Apple to design its own silicon to achieve optimal performance, the AI era is now demanding bespoke hardware. By co-developing a chip that perfectly matches its software, OpenAI is not just cutting costs; it is paving the way for faster, more reliable, and more accessible AI. As these custom processors light up data centers around the world, they will quietly power the next wave of agentic AI, turning the theoretical promise of artificial intelligence into a scalable, everyday reality.[1]