OpenAI's first custom inference chip, Jalapeño, is now posting first results, and they're strong: 1.5-1.9x more AI work per watt at peak throughput and 1.7-3.6x lower end-to-end latency than leading commercial systems, with 2.1-4.1x gains on highly interactive agent workloads. The architecture delivers both throughput and latency at once, where rivals typically trade one for the other. It holds up across OpenAI- and third-party models alike, and marks a move toward owning the whole silicon-to-software stack.
Apple's M5 Ultra is its most powerful chip yet, a first-of-a-kind quad-die architecture pushing unified memory bandwidth to 1.2TB/s, 50% more than the M3 Ultra, with up to an 80-core GPU. It pairs with the M6, Apple's first 2nm processor, which adds a Dual 16-core Neural Engine for up to 2x peak compute. Together they power new Mac Studio and Mac mini models aimed squarely at running large AI workloads on a desktop.
On the outskirts of Shanghai, an R&D hub hosting 100-plus robotics firms threw a public robot carnival: kids buy mini robot dogs while humanoids do drunken boxing and fold shirts. It is entertainment with a strategy behind it. Embodied AI was a pillar of China's latest five-year plan, and nearly 90% of the 13,000-plus humanoids shipped globally last year were made in China. The softer side of an industrial push, and a reminder that China's bet on human form is running ahead of the skeptical West.
Anthropic's flagship Fable 5 has drawn sluggish corporate demand since its July 24 launch: Ramp billing data from 70,000 companies shows the pricey model pulling just 8% of Anthropic spend, behind the older Opus 4.8's 28%. The miss isn't hurting the business, though. Annualized revenue hit $65bn in July and Anthropic expects Q3 profit. The signal: customers increasingly refuse maximum capability at maximum cost.