OpenAI has designed and built its first AI chip, named Jalapeño, in collaboration with Broadcom. The chip is purpose-built for LLM workloads powering products like ChatGPT, Codex, and the API, aiming to enhance scaling, serving, and access to AI.
SynthePulse Insight · AI deep reading
Version 1 · 2 sources
The Jalapeño chip, developed by OpenAI in partnership with Broadcom, is purpose-built to accelerate LLM inference. With a rapid 9-month development cycle, early tests show significant efficiency gains, but the ASIC's flexibility risks and a 2028 mass production timeline keep the market cautious.
On June 24, 2026, OpenAI announced its first in-house AI chip, Jalapeño, an application-specific integrated circuit (ASIC) designed for large language model inference workloads. The chip was designed from scratch by OpenAI, with Broadcom handling silicon implementation and networking technology (including Tomahawk), and Celestica involved in board-level integration. Jalapeño will directly serve real-time inference for ChatGPT, Codex, and future agent products, aiming to reduce per-query costs and dependence on external chip suppliers.
Inference is the core of daily AI product interactions, and its costs escalate sharply with user growth. OpenAI President Greg Brockman emphasized that the in-house chip is part of a 'full-stack strategy,' achieving system-level optimization by controlling more infrastructure layers. This move also reflects an industry trend: Google, Amazon, and Microsoft have all deployed their own chips, and OpenAI now officially joins that club with Jalapeño.
Jalapeño's most striking feature is its extremely short development cycle: just 9 months from design to tape-out. OpenAI believes this may be one of the fastest advanced chip development processes on record, with some design work accelerated by OpenAI's own AI models, creating a closed loop of 'using AI to design AI hardware.' Broadcom provided critical silicon implementation support, while Celestica handled board-level system integration. This speed results from close partnerships and a deep internal understanding of model behavior—the chip design was specially optimized around OpenAI's model kernels, memory movement, and serving systems.
However, the efficiency of 9 months also raises questions about design depth. Independent analysts point out that rapid tape-out may imply fewer design iterations and verification stages, so actual stability after mass production remains to be seen. OpenAI has not yet released a full technical report on chip architecture details or design methodology.
OpenAI claims that early internal tests show Jalapeño significantly outperforms current state-of-the-art chips in performance per watt, primarily due to optimization of data movement and the balance among compute, memory, and networking. However, the company has not released official benchmarks or any specific numbers. Industry observers note that such claims require independent third-party verification, especially since an ASIC's superiority in specific workloads may not hold across broader scenarios.
Furthermore, Jalapeño only handles inference; training still relies on Nvidia GPUs. This means OpenAI has zero hardware autonomy in training and cannot break its dependence on Nvidia in the short term. In the long run, if AI model architectures undergo a fundamental shift (e.g., Transformer replaced by another architecture), Jalapeño's specialized design could rapidly depreciate—an inherent risk of ASICs.
Jalapeño's launch marks further fragmentation in the AI hardware market. Google's TPU, Amazon's Trainium/Inferentia, and Microsoft's Maia have all been deployed for years. Broadcom, as a behind-the-scenes winner, has provided custom chip design services for multiple clients. For OpenAI, an in-house chip not only reduces operational costs—Broadcom CEO Hock Tan mentioned that 'small-batch prototypes' will begin delivery by the end of 2026—but also increases leverage in pricing negotiations with Nvidia.
However, Jalapeño's economies of scale will take time. The production plan indicates that large-scale deployment will only reach 'full speed' in the first half of 2028. Meanwhile, OpenAI has set an ambitious goal of 10 GW of computing power by 2029, with most computing currently reliant on Nvidia. The in-house chip cannot change this landscape in the short term, but it provides a strategic buffer for the future.
Despite optimistic early indicators, Jalapeño faces multiple uncertainties. First, technical risk: AI inference workloads evolve rapidly; architectures considered optimal in 2019 may become obsolete due to model scale explosions or new algorithms. The specialization of ASICs means that once the design is fixed, modifications are extremely costly. Second, production risk: from prototypes at the end of 2026 to full-speed production in 2028, there is a two-year ramp-up period during which any issues with capacity, yield, or supply chain could delay plans.
Economically, Jalapeño's benefits depend on whether it truly reduces the total cost of ownership for inference. OpenAI has not disclosed chip costs or TCO comparisons with Nvidia. If efficiency gains do not offset the fixed costs of large-scale deployment, its strategic value will be greatly diminished. The market will watch for subsequent white papers or third-party evaluations from OpenAI to validate performance. Finally, geopolitical risks cannot be ignored: semiconductor export controls could affect Broadcom's manufacturing nodes or supply chains, especially for high-end chips.
This analysis synthesizes OpenAI's official announcement and TechRepublic reporting. The official statement is a first-hand source but contains promotional elements; TechRepublic's report is based on official information and industry context, providing additional details and independent perspectives. All performance claims are unverified independently; the production timeline comes from Broadcom executives' statements.
Jalapeño is a strategic piece for OpenAI to lower inference costs and reduce reliance on Nvidia, with its 9-month rapid development showcasing engineering execution. However, the ASIC flexibility risk, the uncertainty of mass production only by 2028, and the continued dependence on Nvidia for training mean it is unlikely to disrupt the existing hardware landscape in the short term, serving more as a foundation for a long-term game.
Primary report
Primary source
Also covered by 1 sources