OpenAI Unveils 'Jalapeño' Inference Chip With Broadcom to Control Full Stack and Cut Compute Costs

OpenAI has officially entered the semiconductor arena, partnering with Broadcom to unveil "Jalapeño," the artificial intelligence company’s first custom-designed silicon accelerator. Billed as an "Intelligence Processor," the new chip represents a major strategic shift for the ChatGPT creator, moving the company beyond software and model development into the physical infrastructure that powers its products. The announcement, which also includes manufacturing partner Celestica, outlines a multi-generation compute platform designed to make advanced AI faster, more reliable, and significantly cheaper to operate at a global scale.[1]

Unlike the versatile graphics processing units (GPUs) that currently dominate the AI hardware market, Jalapeño is an application-specific integrated circuit (ASIC) built from a blank slate. It is engineered for one specific task: large language model inference. While general-purpose chips must balance the complex mathematics required to train new models with the speed needed to serve them, Jalapeño strips away the training architecture entirely. Instead, it focuses on the precise bottlenecks that slow down live AI products, optimizing for rapid data movement, memory bandwidth, and low-latency networking.

To understand the significance of the move, one must untangle the two distinct phases of artificial intelligence: training and inference. Training is the highly publicized, computationally massive process of teaching a model how to think by feeding it trillions of words over several months. Inference, by contrast, is the ongoing, everyday process of running live user prompts through that finished model to generate a response. While training requires a massive upfront investment, inference is a perpetual operational expense that scales directly with user adoption.[3]

The economic reality of the AI boom is that inference has become the industry's heaviest financial anchor. According to recent infrastructure analyses, inference workloads are projected to consume 65% of all AI compute by 2029. More critically, because inference runs continuously to serve millions of daily queries, it now accounts for 80% to 90% of the lifetime cost of a production AI system. Every time a user generates a document, writes a line of code, or asks a chatbot a question, a meter is running in a data center.

For OpenAI, which serves hundreds of millions of active users and powers thousands of enterprise applications through its API, those running meters translate to exorbitant cloud bills. The broader industry is feeling the same squeeze; a recent developer survey found that 44% of organizations now allocate over three-quarters of their total AI budgets strictly to inference. By designing hardware specifically tuned to the quirks of its own models, OpenAI hopes to break this cost curve and make high-volume AI deployment financially sustainable.[3]

The hardware architecture of Jalapeño reflects a deep collaboration between OpenAI's software engineers and Broadcom's silicon designers. Because OpenAI intimately understands the specific serving patterns, memory movement, and software kernels of its own frontier models, it was able to design a chip that eliminates the inefficiencies found in off-the-shelf hardware. Broadcom provided the foundational silicon expertise, including integrating its high-performance Tomahawk networking chips to ensure that thousands of Jalapeño processors can communicate seamlessly across massive data center racks.

This level of vertical integration—controlling the entire technology stack from the user interface down to the physical silicon—has long been the holy grail for major technology companies. By operating across the full stack, OpenAI can co-design its future models and its future chips simultaneously. If a new AI architecture requires a specific type of memory access, the hardware team can build that capability directly into the next generation of Jalapeño, creating a feedback loop that off-the-shelf hardware buyers simply cannot match.[1]

While OpenAI and Broadcom have kept exact performance benchmarks under wraps pending a full technical paper, the financial targets are aggressive. Early reports indicate that Jalapeño is designed to cut OpenAI's inference costs by roughly 50%. Furthermore, the companies claim the chip will deliver significantly higher performance per watt than current leading-edge hardware, a crucial metric as data centers worldwide strain against the limits of local electrical grids.[1]

The development timeline for Jalapeño has set a blistering pace for the notoriously slow semiconductor industry. The project moved from early schematics to physical engineering samples in just nine months—a cycle that typically takes years. OpenAI attributed this speed to a software-hardware co-development process that actively utilized the company's own AI models to accelerate parts of the chip design, effectively using artificial intelligence to build the infrastructure for the next generation of artificial intelligence.[1]

The chips are already moving out of the theoretical phase. Engineering samples of Jalapeño are currently operating in OpenAI's labs at target clock speeds and power levels. The company confirmed it is actively running production-grade machine learning workloads on the new silicon, including its prior-generation GPT-5.3-Codex-Spark model, proving that the blank-slate architecture can successfully handle the complex reasoning tasks required by modern agentic AI.

Despite the breakthrough, the launch of Jalapeño does not signal a divorce between OpenAI and Nvidia, the undisputed king of AI hardware. Nvidia's GPUs remain the gold standard for the computationally brutal task of model training. Earlier this year, Nvidia finalized a massive direct investment into OpenAI, securing an agreement to deploy gigawatts of computing capacity using its next-generation Vera Rubin platform. Jalapeño is designed to complement, rather than replace, this infrastructure by taking over the workload once the models are fully trained.[1]

OpenAI's foray into custom silicon mirrors a broader trend among the tech industry's heaviest hitters. Google has spent a decade refining its Tensor Processing Units (TPUs), Microsoft recently escalated its bespoke silicon efforts with the Azure Maia 200 accelerator, and Meta continues to deploy its MTIA chips. As AI models become the core engine of the modern internet, relying entirely on third-party merchant silicon for daily operations is increasingly viewed as an unacceptable strategic and financial risk.[1][2]

The financial implications of Jalapeño extend far beyond the server rack. As OpenAI lays the groundwork for a highly anticipated public offering, the company faces intense pressure from private investors and public markets to demonstrate a viable path to profitability. By drastically reducing its cost of goods sold—in this case, the compute required to serve its models—OpenAI is attempting to prove that the staggering capital expenditures of the AI era will eventually yield a sustainable, high-margin business.[1]

With manufacturing partner Celestica handling the board, rack, and system-level integration, OpenAI plans to begin deploying Jalapeño across active data centers by the end of the year. If the chip performs as promised at scale, it could fundamentally alter the economics of artificial intelligence, allowing developers to deploy more complex, reasoning-heavy models to millions of users without bankrupting the companies that build them.[2]