OpenAI Says Its In-House Chip Handles 1.9 Times More Work Than NVIDIA's Chips [Global AI Briefing]

[Financial News] OpenAI officially announced that the first performance test of its self-designed AI inference chip, Jalapeño, showed far better performance than NVIDIA's chips.
OpenAI plans to begin pilot deployment of Jalapeño in real AI infrastructure later this year and move to full-scale production next year, expanding its ecosystem from AI model development to in-house chip production.
On the 25th local time, OpenAI said that it tested its first in-house AI chip, Jalapeño, which it co-developed with Broadcom Inc., at Hot Chips 2026 held at Stanford University. Using SemiAnalysis' AI inference chip benchmark platform, InferenceX, the company found that the chip processed about 1.5 to 1.9 times more AI work per 1 kW of power than NVIDIA GB200 and NVIDIA GB300 NVL72 systems, while total response latency was about 1.7 to 3.6 times lower.
That means existing AI infrastructure could handle more AI requests with the same amount of power while also improving response speed.
This is the first time OpenAI has presented concrete performance figures since unveiling Jalapeño with Broadcom in June.

OpenAI emphasized that Jalapeño is part of its full-stack strategy.
The company said it aims to deliver more "useful intelligence" with the same computing resources by optimizing every layer together, including chips, computing, software, and data centers, as well as AI models.
OpenAI also said, "Different AI tasks require different hardware," adding, "Rather than solving all computation with a single chip, we will choose the most efficient computing resources for each task."
Because Jalapeño is an inference chip rather than a training chip, it is designed to handle the computations needed when a trained AI model answers user questions or performs agent tasks.
This also reaffirms OpenAI's strategy of not replacing NVIDIA GPUs with Jalapeño all at once, but instead adding in-house chips optimized for specific tasks to reduce inference costs and power consumption as usage of ChatGPT and AI agents surges.
Meanwhile, Anthropic has also officially announced that it has created an internal team to design custom chips for running Claude and will begin developing its own chips in earnest.
As global AI companies increasingly pursue in-house chip development to improve AI model performance and cut inference costs, observers say competition in the global AI market is expanding into a race to build full-stack systems that connect AI models, agents, and hardware.


[email protected] Lee Gu-soon Reporter