Inside OpenAI’s Jalapeño: How Custom Silicon Is Reshaping the AI Infrastructure Landscape

The compute race has a new entrant. The custom AI accelerator chips market was valued at $45.8 billion in 2025 and is projected to reach $746.2 billion by 2035, growing at a CAGR of 32.18%. That growth trajectory reflects a structural shift already underway across Big Tech: general-purpose GPUs are giving way to purpose-built silicon, and every major player wants a chip with its name on it. On June 24, OpenAI joined that race officially, announcing Jalapeño, its first custom inference accelerator, co-developed with Broadcom and supported by Canadian manufacturer Celestica. 

The announcement is short on specs and long on ambition. What it signals about the broader AI infrastructure landscape, however, deserves a closer look.

What’s inside Jalapeño 

Jalapeño is an inference accelerator, purpose-built for running large language models rather than training them. OpenAI co-developed it with Broadcom, which has become the dominant custom AI chip architect across the industry. Broadcom alone carries a $73 billion AI backlog and is targeting $100 billion in annual AI chip revenue by 2027. 

Richard Ho, head of hardware at OpenAI, stated the team optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier models, with early testing showing performance close to the hardware’s theoretical limits. OpenAI also claims it brought Jalapeño from design to manufacturing tape-out in nine months, crediting its own models for accelerating parts of the design and optimization process.

Engineering samples are currently running ML workloads in OpenAI’s lab, including GPT-5.3-Codex-Spark. A detailed technical report is promised in the coming months. Public benchmarks have not been released.

Why every major player is going custom

OpenAI is following a well-worn path. Google built its first Tensor Processing Unit in 2016 for TensorFlow workloads. Amazon debuted AWS Inferentia in 2018, followed by Trainium in 2022. Microsoft launched its Azure Maia AI Accelerator in 2023. Meta’s MTIA custom accelerator has entered production deployment in its data centers. Reuters reported in April that Anthropic is also weighing its own chip program.

The motivation across the board follows the same logic. Custom ASIC shipments are projected to triple by 2027 compared to 2024 levels, and by 2028, ASIC shipments will surpass GPU shipments for the first time in history. Every AI workload running on a proprietary chip keeps more margin inside the company, reduces reliance on third-party suppliers, and lets teams optimize for their specific model architectures rather than adapting to general-purpose hardware. 

As Ben Bajarin, CEO and principal analyst at Creative Strategies, put it on X: “Those serious about platforms should be serious about silicon.”

For OpenAI specifically, the inference focus makes strategic sense. Inference workloads now account for two-thirds of all AI compute, surpassing training for the first time in 2026. Building a chip optimized for inference at scale targets exactly where OpenAI’s operational costs are concentrated.

The full-stack play

OpenAI frames Jalapeño as infrastructure in service of its users: better chips mean more efficient compute, which means better models, which means better products, which means more revenue to reinvest. The company states the chip will make models faster, more reliable, and more affordable.

The subtext is a straightforward vertical integration strategy. Jalapeño moves OpenAI from being a model company that rents compute to a platform company that controls the layer below its models. Slated for deployment at gigawatt scale in Microsoft’s and other partners’ data centers by the end of the year, the chip represents the first step in what OpenAI calls a multi-generation compute platform roadmap.

OpenAI states Jalapeño was designed to work with all LLMs, current and future. That claim positions the chip as broadly useful infrastructure rather than a closed proprietary system. Whether that positioning holds as the roadmap develops remains to be seen.

What developers should watch

The lack of technical details in the announcement is a real gap. OpenAI claims performance substantially better than the current state-of-the-art but offers no benchmarks to support that claim. Developers evaluating infrastructure decisions cannot act on unverified performance assertions.

The more substantive question is what full-stack ownership means for developer flexibility over time. When a single company controls the model, the API, and the underlying silicon, the levers for switching costs multiply. OpenAI’s current framing emphasizes openness and broad LLM support. How that framing survives competitive pressure from subsequent chip generations is worth tracking.

Companies deploying custom AI inference chips report an average cost saving of 25% compared to third-party GPU alternatives. If OpenAI passes a meaningful portion of those savings downstream through lower API pricing, the developer value proposition becomes concrete. If the savings stay inside OpenAI’s margin structure, the chip’s benefits remain largely invisible to external builders.

The multi-generation roadmap hint at the end of OpenAI’s announcement is the most telling detail. Jalapeño is a starting position, not a finished product. What OpenAI chooses to optimize, control, or restrict in subsequent generations will reveal far more about its infrastructure strategy than the launch announcement does.