A.I.
Hardware · Chips
Jalapeño
Jalapeño is the first chip OpenAI designed itself. It is built for inference, which is running a trained model to answer requests. Broadcom turns the design into silicon and supplies the network chips, and Celestica builds the boards and racks (OpenAI and Broadcom). OpenAI says it plans to start deploying Jalapeño in its own data centers by the end of 2026 (OpenAI, first results).
The chip
OpenAI presented these figures at Hot Chips 2026, a yearly chip design conference (ServeTheHome, Hot Chips 2026, reporting OpenAI's slides).
| Figure | One Jalapeño chip |
|---|---|
| Compute | 13.4 PFLOPS MXFP4; 3.4 PFLOPS MXFP8 |
| Memory | 216 GiB HBM4 |
| Memory bandwidth | 15.4 TB/s |
| Power (TDP) | 700 W |
| Network to nearby chips | 600 GB/s, up to 128 chips |
| Network to the full system | 200 GB/s, up to 2,048 chips |
HBM4 is the newest generation of high-bandwidth memory, stacked next to the chip (see HBM memory). OpenAI says the chip measured at or below 550 W on the work it tested, under its 700 W rating (OpenAI, first results).
The system
| Figure | 2,048-chip Jalapeño system |
|---|---|
| Compute | 27 EFLOPS MXFP4 |
| Memory | 432 TiB HBM4 |
| Memory bandwidth | 32 PB/s |
Up to 128 chips share one fast local network, and a two-level network of Broadcom Tomahawk 6 Ethernet switches joins up to 2,048 chips into one system. Ethernet is the common networking standard most data centers already use.
What OpenAI claims
- 1.5 to 1.9 times more AI work per watt at peak throughput than the comparison systems, and 1.7 to 3.6 times lower end-to-end latency, across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 (OpenAI, first results). Latency is the wait before an answer arrives.
- The comparison systems were Nvidia GB200 and GB300, normalized to each chip's published power rating. These are OpenAI's own tests, not outside results.
Units on this page: a PFLOPS is a thousand trillion math operations per second; an EFLOPS is a thousand PFLOPS. FP4 and FP8 store each number in 4 or 8 bits, and MX in front (MXFP4, MXFP8) means a shared-scale version of that format. GB is gigabytes, GiB is about 7% more than a GB, and TB is a thousand GB. TB/s and PB/s are terabytes and petabytes per second, how fast data moves. W is watts and kW is a thousand watts. TDP is the most power a chip is built to draw.