OpenAI posts first Jalapeño measurements: more work per watt than GB200/GB300

At Hot Chips 2026 OpenAI shared first results from its custom inference ASIC with Broadcom. SemiAnalysis InferenceX on GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T showed 1.5–1.9× more AI work per watt at peak throughput.

Cover image from OpenAI’s first Jalapeño results post

On August 25 OpenAI published the first measured results for Jalapeño, its custom inference ASIC built with Broadcom, at Hot Chips 2026. SemiAnalysis InferenceX compared the chip with Nvidia GB200/GB300 on GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T.

What we know

Jalapeño is a custom inference ASIC built with Broadcom; first measured results were shown at Hot Chips 2026.
On SemiAnalysis InferenceX, using GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T, it delivered 1.5–1.9× more AI work per watt at peak throughput.
End-to-end latency was 1.7–3.6× lower, and interactive performance was 2.1–4.1× versus the Nvidia GB200/GB300 comparison.
Richard Ho said a small-volume deployment comes at the end of 2026, with higher volume in 2027.

Takeaways

These are first lab measurements, not a shipping volume SKU.
The comparison set is GB200/GB300 on three named large models.
Ho’s calendar is late-2026 for a small run and 2027 for higher volume.

Source: OpenAI

Related reading

Image published with The Register on Perplexity’s Portable Computer

AIAug 263 min read

Perplexity’s Portable Computer runs its agent stack on a local DGX Spark

Portable Computer is a local version of Perplexity’s Computer agent: harness, orchestrator, and open-weight models on an Nvidia DGX Spark. Local inference avoids per-token API fees.

Image published with The Register on Debian’s LLM-usage General Resolution

AIAug 263 min read

Debian developers vote on how far LLMs can go in the project

Debian is running a General Resolution on LLM use, with eight overlapping proposals from a total ban to conditional acceptance. Project Lead Sruthi Chandran extended voting by a week; The Register counted fewer than 350 ballots when it reported.

Image published with The Verge on ChatGPT Work signing in to websites

AIAug 263 min read

ChatGPT Work can now sign in to websites to finish tasks

ChatGPT Work can sign in to websites for multi-step jobs such as booking a DMV appointment, canceling a reservation, or submitting a job application. OpenAI says it never sees the user’s login credentials.

All of Pulse

AIAugust 25, 20263 min read

OpenAI posts first Jalapeño measurements: more work per watt than GB200/GB300

What we know

Jalapeño is a custom inference ASIC built with Broadcom; first measured results were shown at Hot Chips 2026.
On SemiAnalysis InferenceX, using GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T, it delivered 1.5–1.9× more AI work per watt at peak throughput.
End-to-end latency was 1.7–3.6× lower, and interactive performance was 2.1–4.1× versus the Nvidia GB200/GB300 comparison.
Richard Ho said a small-volume deployment comes at the end of 2026, with higher volume in 2027.

Takeaways

These are first lab measurements, not a shipping volume SKU.
The comparison set is GB200/GB300 on three named large models.
Ho’s calendar is late-2026 for a small run and 2027 for higher volume.

Source: OpenAI