Jalapeño
An internally developed custom AI chip project by OpenAI, codenamed Jalapeño. It is regarded as an application-specific integrated circuit (ASIC) designed for AI workloads. The design process extensively employs large language models to assist with RTL code generation and verification.
Timeline (last 90 days)
In the InferenceX public benchmark, Jalapeño achieved 1.5 to 1.9 times the tasks per watt at peak throughput compared to NVIDIA's reference system, with end-to-end latency reduced by 1.7 to 3.6 times.
OpenAI unveiled its first in-house AI inference chip, Jalapeño, with a rated power consumption of 700W and a measured sustained power consumption of 550W or below.
OpenAI plans to deploy Jalapeño into its own infrastructure by the end of this year, but will continue deploying NVIDIA and other accelerators.
With the assistance of its own models, OpenAI's team completed the entire process from design to tape-out in 9 months, and adapted 3 open-source models in 2 months.
Richard Ho stated that Jalapeño simultaneously achieves high throughput and low latency on the same chip, which is a first in the industry.
OpenAI uses an open-source benchmarking tool called InferenceX to test Jalapeño's performance.
Jalapeño is an inference chip, not a training chip.
Jalapeño's performance per watt is 1.8 to 4 times higher than existing chips.
Once Jalapeño enters full production, token prices for customers are expected to decrease.
OpenAI has not yet disclosed Jalapeño's foundry partner, process node, or mass production timeline.
38 more timeline events
All Facts (20)
In the gpt-oss 120B test, Jalapeño achieved a peak throughput of approximately 85,448 per watt compared to GB200's 44,960, roughly 1.9×; end-to-end latency was reduced from 1.80 seconds to 1.03 seconds; and minimum TBT dropped from 1.87 ms to 0.69 ms.
60%UnverifiedIn the DeepSeek R1 670B (MXFP4 precision) test, Jalapeño achieved approximately 1.7× throughput per watt (19,641 vs. GB300's 11,781), and end-to-end latency was reduced from 5.99 seconds to 1.65 seconds.
60%Unverified"Jalapeño" is most likely a dedicated accelerator (ASIC) targeting AI workloads rather than a general-purpose computing chip.
60%UnverifiedWith the assistance of its own models, OpenAI's team completed the entire process from design to tape-out in 9 months, and adapted 3 open-source models in 2 months.
50%UnverifiedIn the InferenceX public benchmark, Jalapeño achieved 1.5 to 1.9 times the tasks per watt at peak throughput compared to NVIDIA's reference system, with end-to-end latency reduced by 1.7 to 3.6 times.
50%UnverifiedJalapeño will not replace existing suppliers, but will instead serve as part of OpenAI's chip fleet, sharing different workloads alongside chips from NVIDIA, Cerebras, and other vendors.
50%UnverifiedOpenAI uses an open-source benchmarking tool called InferenceX to test Jalapeño's performance.
50%UnverifiedRichard Ho stated that Jalapeño simultaneously achieves high throughput and low latency on the same chip, which is a first in the industry.
50%UnverifiedJalapeño is an inference chip, not a training chip.
50%UnverifiedJalapeño's performance per watt is 1.8 to 4 times higher than existing chips.
50%UnverifiedJalapeño is OpenAI's in-house chip dedicated to inference and is not responsible for training next-generation GPT models.
50%UnverifiedOpenAI states that Jalapeño went from initial design to tape-out in just 9 months, with AI models participating in chip development and arithmetic circuit optimization.
50%UnverifiedAt peak throughput, Jalapeño completes 1.5 to 1.9 times more AI work per watt than the comparison systems.
50%UnverifiedJalapeño has a rated power consumption of 700W with measured sustained power consumption not exceeding 550W; by comparison, the GB200 is rated at 1200W and the GB300 at 1400W.
50%UnverifiedJalapeño was designed from the ground up with co-design of chip, memory, network, software, and rack, enabling model state such as KVCache to be explicitly placed and retained as much as possible.
50%UnverifiedThe division of labor for the Jalapeño chip is as follows: OpenAI is responsible for architecture definition, Broadcom handles chip implementation and network connectivity, and Celestica is responsible for board and system integration.
50%UnverifiedThe core design solution of Jalapeño is to keep state such as the KV Cache locked in local storage as much as possible, trading larger local storage for lower data-movement latency—a system-level tradeoff of space for time.
50%UnverifiedJalapeño testing used Semi Analysis's InferenceMAX framework with data provided by OpenAI; the full test suite was not completed, raw logs were not made public, and test scenarios were limited to 8K input and 1K output with speculative decoding disabled.
50%UnverifiedOpenAI claims it took only 9 months from initial design to tape-out, while Semi Analysis estimates approximately 16 months when accounting for team formation time.
50%UnverifiedJalapeño integrates networking directly into the chip architecture, keeping requests within a single interconnect domain as much as possible for end-to-end execution.
50%