OpenAI's Jalapeño chip posts first inference numbers against Nvidia's GB300

verifiedDeveloperCompute

Wednesday, August 26, 2026

Confidence

Medium · — OpenAI primary + 3 corroborating; self-selected benchmark, no third-party rerun yet

Evidence

OpenAI blog + Hot Chips talk + Bloomberg reporting + independent benchmark

At Hot Chips on Tuesday, OpenAI shared benchmarks for Jalapeño, its Broadcom-built inference chip, claiming 1.5–1.9× throughput per watt over Nvidia's GB300.

The details

Ran SemiAnalysis's public InferenceX benchmark across GPT-OSS 120B, DeepSeek R1, Kimi K2.5.
Small deploy end-2026, wider rollout in 2027; 700W vs GB300's 1,400W.

What it means

OpenAI dropped this on Nvidia earnings morning, on a self-selected benchmark, against silicon Vera Rubin already replaces. Nvidia's largest customer now has a working TSMC inference alternative, giving every GB300-floor cost model a live comparable.

At Hot Chips on Tuesday, OpenAI shared benchmarks for Jalapeño, its Broadcom-built inference chip, claiming 1.5–1.9× throughput per watt over Nvidia's GB300.

Ran SemiAnalysis's public InferenceX benchmark across GPT-OSS 120B, DeepSeek R1, Kimi K2.5.
Small deploy end-2026, wider rollout in 2027; 700W vs GB300's 1,400W.

Sources

Apply this today

The hands-on layer the brief points to: a workflow you can run, a tool to test, and today’s 60-second video.

Workflow of the day

Turn an HR Signal Into a Retention Risk Brief: A Prompt-Only Workflow for Flight-Risk Analysis

Human Resources · tested · saves 3-4 hours per pulse survey batch

Full walkthrough →

Tool watch

GPT-5.6 in Kiro IDE

GPT-5.6 in Kiro: Does a Three-Tier Model Menu Actually Help Your Agent Loop?

Read the breakdown →

▶ Watch today’s 60-second brief →

In this issue

← Back to brief

Get the next one in your inbox

One daily brief. Every story gets a hype verdict.

No spam. Unsubscribe anytime.

On this page

The details
What it means

In this issue

← Back to brief

Stat of the Day

$20B▲

SoftBank's proposed OpenAI-linked bond

Would be the largest Asian bond deal of 2026 and refinance half of the $40B OpenAI bridge loan. Source

Today’s Take

Four of today's five stories are the same story from four angles: the AI economy is not funding itself yet. Meta could not swap engineers for agents on its own timeline, SoftBank has to price public debt to keep its OpenAI position, OpenAI is building its way out of Nvidia margins, and Google is buying legal-market distribution one Big Law logo at a time. The bet that AI ends up cheaper, leaner, and more vertical is still on — but every proof point this week required more capital, more silicon, or more partners than the pitch decks assumed. Considered and passed: the Bill Gates "human-reserved jobs" tour (six outlets, one op-ed's worth of new information) and the Nvidia-Poolside $6B license (5 days old and already priced in).

— Agentic desk

Never miss a signal

Join professionals getting the daily AI brief every weekday morning.

No spam · Unsubscribe anytime

No sponsored verdicts
Sources linked
Corrections public