OpenAI unveils custom AI inference chip Jalapeño

OpenAI’s chief financial officer, Sarah Friar, used a recent blog post to explain the company’s holistic approach to scaling artificial intelligence. At the centre of that strategy is a new piece of hardware – the Jalapeño chip – which the firm says marks its first foray into custom inference silicon. The announcement highlighted measured results from a public benchmark called InferenceX, where Jalapeño ran a 120‑billion‑parameter GPT‑OSS model and achieved higher peak throughput per kilowatt and lower token latency than the commercial accelerators it was compared against. The chip also posted strong numbers on other model families such as DeepSeek R1 and Kimi K2, suggesting the gains are not limited to a single architecture.

The Jalapeño chip is presented as a lever for OpenAI to gain tighter control over both performance and cost. By co‑designing the model, serving software, chip, memory, and networking layers, the company claims it can improve throughput, latency, energy efficiency, and overall economics as a single system rather than as isolated components. Friar noted that this first‑generation silicon is already in production, and that future generations are in development, indicating a longer‑term roadmap for in‑house hardware.

OpenAI’s compute strategy, according to the post, rests on an integrated stack that spans data‑center facilities, custom chips, frontier models, developer platforms, consumer and enterprise products, and AI‑native devices. Each layer reinforces the others: better software extracts more performance from hardware, while more capable models enable richer products that generate additional usage signals, which in turn feed back into system improvements. The company emphasizes that different workloads – such as frontier training, high‑volume inference, and always‑on agents – have distinct requirements across chips, software, networks, power, and latency, and that its portfolio is designed to meet those varied demands.

OpenAI’s portfolio of partners includes long‑standing collaborators like Microsoft’s compute infrastructure and NVIDIA’s GPUs, as well as newer contributors such as AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank. By maintaining a mix of premium systems for capability‑heavy tasks and optimized setups for cost‑sensitive scaling, OpenAI aims to stay on what it calls the Pareto frontier – the optimal balance of capability, speed, reliability, efficiency, and cost for each workload. The post argues that preserving choice across providers and hardware types lets the company direct demand toward the most cost‑effective solutions while retaining pricing discipline as market conditions evolve.

Beyond the chip itself, OpenAI highlighted a data‑center project named Camellia in Georgia. The initiative is described as a purpose‑built facility that aligns its design with customer workloads, creates local jobs, supports nearby businesses, and incorporates a closed‑loop water system. An independent public audit is said to verify the project’s commitments each year, underscoring OpenAI’s focus on sustainable and accountable infrastructure.

The post concludes by linking hardware efficiency to broader economic effects. Using an internal metric called the Artificial Analysis Coding Agent Index, OpenAI reported that its GPT‑5.6 Sol model achieved a new high on reasoning tasks while using 54 % fewer output tokens than a leading competitor. The company frames this as evidence that more useful intelligence per dollar enables new use cases – from contract review to live financial scenario planning – and triggers a Jevons‑paradox effect where greater efficiency expands overall consumption. OpenAI asserts that the compounding advantage of better technology, lower costs, and reinvested growth fuels a virtuous cycle of continued progress in AI capabilities and accessibility.