TL;DR
· OpenAI and Broadcom have dropped the first self-developed AI inference chip, Jalapeño, aiming for initial deployment by the end of 2026.
· This chip is designed for LLM inference, not as a replacement for training GPUs, and early performance claims still come from the company's testing metrics.
· OpenAI takes its first step to reduce reliance on Nvidia, but the extent of cost savings, production scale, and real-world performance have yet to be disclosed.

On June 24, OpenAI and Broadcom announced Jalapeño, a self-developed AI accelerator aimed at large language model inference, marking OpenAI's first chip dubbed the 'Intelligence Processor.' According to both companies, Jalapeño is still in the testing and sampling phase, with the goal of starting initial deployment by the end of 2026, and plans for continued expansion in the following years. For OpenAI, the focus isn't on immediately swapping out all Nvidia GPUs, but rather on gradually transitioning the growing inference requests from ChatGPT, Codex, API, and future intelligent agent products to a software and hardware combo better suited to its own models.

Starting from inference, focusing on the cost of queries happening daily.

AI chip demand can generally be divided into training and inference. Training determines the upper limit of model capability, while inference determines whether the model can respond to massive user requests at acceptable costs after training is completed.

Jalapeño is targeting the latter. OpenAI states that this chip will be used for large language model inference workloads, serving ChatGPT, Codex, API, and future intelligent agent products. For regular users, it won't directly change the chat interface, but it may impact the cost, speed, and scalability of backend request processing.

This is also the common direction of AI companies developing their own chips. Google has TPU, and Amazon and Meta are also advancing custom accelerators. They don't necessarily aim to completely replace Nvidia but to move their most stable, large-scale internal workloads onto chips that are better suited to their own models and software stacks.

OpenAI had previously relied heavily on external GPU supplies, especially Nvidia's high-end accelerators. As the model call volume increased, relying solely on purchasing generic GPUs led to dual pressures: costs and supply constraints from external supply chains, and generic chips may not achieve maximum efficiency under specific models and service patterns. The direct change with Jalapeño is that OpenAI is starting to participate in defining computing power instead of just purchasing it.

Broadcom is responsible for silicon implementation and connectivity technology, while the wafer manufacturing partner remains undisclosed.

Jalapeño is designed from scratch by OpenAI, with Broadcom providing silicon implementation, networking, and connectivity technology, and Celestica participating in the landing of the board, rack, and system levels. The official announcement did not disclose the wafer manufacturing partner, so it cannot simply be written that Broadcom is responsible for manufacturing.

Broadcom has deep experience in custom ASICs and data center networks and has become one of the suppliers behind several AI giants developing their own chips in recent years. For OpenAI, choosing to collaborate with Broadcom is more like handing over the internal architectural needs of the model company to a mature semiconductor and systems supply chain for transformation into deployable products.

Performance aspects still need to leave some boundaries. Both companies state that Jalapeño aims to combine the throughput capabilities of today's leading AI accelerators and achieve significant improvements in performance/watt over current advanced solutions. The officials also emphasized that engineering samples are already running machine learning workloads in the lab at target frequencies and power consumption, with final performance still under measurement, and a more detailed technical report will be released in the coming months.

Currently, third-party benchmarks have not been publicly released, and key metrics like precise throughput, latency, power consumption, and the degree of reduction in per-query costs have not been disclosed. Some market interpretations will directly compare it alongside Nvidia's Blackwell or Google's TPU, but until real deployment data comes out, cost reduction and performance leadership can only be considered early testing metrics from the company.

9 months to tape-out, AI-assisted chip-making has become another clue.

Jalapeño has another detail that the officials highlighted: it only took 9 months from initial design to silicon tape-out. This timeframe doesn't equal the full production cycle but refers to the speed at which the chip design entered manufacturing prep.

In traditional chip development, the timeline from architecture design, verification, EDA processes to tape-out usually takes a long time. OpenAI claims that this process is accelerated by its model. The company is trying to prove that AI can not only write code and generate content but also enter the hardware engineering process to help shorten chip design and verification times.

If this approach can be reused in subsequent products, Jalapeño will not just be a single-generation chip but the starting point for OpenAI's multi-generational custom computing platform. OpenAI and Broadcom had previously announced a collaboration on a 10GW-level custom AI accelerator set to deploy in the second half of 2026 and complete by the end of 2029. Jalapeño is the first sample of this collaboration entering the public product phase.

However, fast design speed doesn't equate to fast deployment speed. Whether Jalapeño can enter OpenAI's core infrastructure depends on subsequent mass production, packaging, high-bandwidth memory supply, server system integration, and data center scheduling. Any bottleneck in these areas could impact the initial deployment pace before the end of the year.

Reducing reliance on Nvidia is the direction, not something that has already been completed.

Jalapeño is most easily interpreted by the market as 'OpenAI challenging Nvidia.' A more accurate statement is that OpenAI is starting to seek custom paths outside of Nvidia for certain inference workloads.

Nvidia's advantages are not just in single GPU performance but also in the CUDA ecosystem, training and inference toolchains, system-level interconnects, developer base, and supply scale. Even if OpenAI successfully deploys Jalapeño, it will not immediately break free from Nvidia, especially in cutting-edge large model training, general R&D, and diverse workloads where high-end GPUs remain critical infrastructure.

The inference phase is more suitable for custom chips. The request patterns are relatively stable, the model structures are more controllable, and the service targets are clearer. If OpenAI can migrate high-frequency, standardized inference tasks from ChatGPT, Codex, or API onto Jalapeño, it may achieve actual benefits in cost and supply flexibility.

For Broadcom, AI giants developing their own chips doesn't mean chip companies are being sidelined; rather, it brings more custom projects. However, the profit margins for such projects still depend on component costs, delivery scale, and system complexity, especially the prices and availability of key supplies like high-bandwidth memory.

What Jalapeño really needs to answer next is not whether it can match Blackwell or TPU at launch but rather how large it can scale up by the end of the year, what real workloads it can handle, and how much it can reduce inference costs. For OpenAI, this 'Mexican pepper' is just the first taste of getting rid of the GPU bottleneck; whether it can become the main course will depend on deployment data.