Morning Digest, August 26, 2026

14 newsletters, 10 overlapping stories

Top Stories

OpenAI’s first in-house chip beats Nvidia in its own benchmarks

(2 newsletters)

OpenAI published the first results for Jalapeño, the inference chip it built with Broadcom, claiming its 700-watt part answers up to 3.6x faster and does up to 1.9x more work per watt than Nvidia’s 1,200-watt flagship systems. The company says it used its own Astra model and Codex to help design the chip, going from first design to manufacturing-ready in nine months. Jalapeño will not be sold externally, is aimed at running models rather than training them, and reaches OpenAI data centers later this year with two more generations already in progress.

Apple’s M6 and M5 Ultra land in a Mac mini and Mac Studio pitched at local AI

(4 newsletters)

Apple refreshed the Mac mini and Mac Studio around the M6, its first 2nm chip, and the M5 Ultra, a quad-die part it calls its most powerful ever. The framing is the real story: Apple now describes the Mac mini as its leading desktop for always-on agentic computing, after developers spent the past year buying them to run models locally on unified memory instead of renting cloud inference. The mini starts at $899 with preorders open and deliveries from September 22, and a refreshed M6 iMac is reportedly queued for later this year.

Nvidia puts the Groq 3 LPX inference accelerator into full production

(2 newsletters)

Groq 3 LPX racks are shipping as an extension of Nvidia’s Vera Rubin platform, tuned specifically for fast token generation on latency-sensitive agentic workloads. Nvidia claims a 4x improvement in response times over the nearest alternative platform, turning agent tasks that took hours into minutes. Nebius is the first cloud provider publicly committed to deploying it.

SpaceXAI adopts Nvidia’s Vera CPU and puts AI compute in orbit

(2 newsletters)

SpaceXAI will run Grok’s agent infrastructure on Nvidia’s Vera CPUs and scale its Vera Rubin footprint toward gigawatts of capacity. The first Starmind AI satellite will carry a space-optimized Vera Rubin NVL72 system, with the orbital compute launch targeted for Q4 2027. Nvidia is billing Vera as its first CPU designed specifically for agentic workloads, and the announcement lands days before its earnings report.

Anthropic is telling IPO investors it sees a $30 trillion market

(2 newsletters)

Anthropic is reportedly preparing to pitch IPO investors on a total addressable market north of $30 trillion, edging past the $28.5 trillion figure SpaceX floated in May. Reports suggest the company could raise as much as $100 billion at a valuation near $2 trillion. Separately, Anthropic hired Amir Salek, the engineer who founded Google’s custom-chip program and ran its TPU business, for its own silicon effort.

Oura eyes a September IPO at more than $16 billion

(2 newsletters)

The smart-ring maker is reportedly seeking up to $3 billion in a September listing at a valuation above $16 billion, up from roughly $11 billion after last year’s raise. Oura says paid membership has quadrupled in two years and was on pace to pass 5 million this quarter, with more than 80% of members renewing after year one. The pitch is wearables as a health-data platform rather than a fitness gadget, following Whoop’s $10 billion mark in March.

Alibaba raises $10.2 billion earmarked entirely for AI and opens up Wan 3.0

(2 newsletters)

Alibaba is issuing $10.2 billion in new shares with all proceeds directed at AI investment, a move that pushed its stock down on the news. Alongside it, the company widened access to Wan 3.0, a video model that generates clips up to 30 seconds from text, video, documents, spreadsheets, slides, or web pages.

The argument that AI coding tools are blocking the path to expertise

(2 newsletters)

The “expert novice” problem: developers are expected to steer coding agents effectively before they have built the judgment needed to evaluate what those agents produce, and leaning on generated code skips the struggle that builds that judgment. The suggested alternative is using AI as a Socratic tutor rather than a code generator, preserving the planning, debugging, and trial-and-error that turn experience into expertise. It is worth reading next to the separate argument that verification, not authoring, is now the real bottleneck.

The anonymous Ox Alpha model is quietly setting records

(2 newsletters)

OpenCode users pushed 26 trillion tokens through the unattributed Ox Alpha model in its first four days, across 327,000 unique users and more than 8.3 million completed sessions. No maker, release date, knowledge cutoff, or output limit is listed, and it is free through an OpenAI-compatible endpoint. In a head-to-head on a real bug, Cline found Ox Alpha and Claude Fable 5 both fixed it, with Ox Alpha using 3x fewer tokens.

Fitbit’s founders are back with an LTE health and safety band

(2 newsletters)

James Park and Eric Friedman launched Luffu Link, an LTE band that pairs the usual sleep, step, and location tracking with elder-care features like summoning a family member and flagging missed medications. The startup is positioning itself as an intelligent family care system for monitoring multiple relatives from one place rather than as another fitness tracker.

Also Worth Knowing

Quick Hits

Shower Thoughts

Disposable chopsticks ought to ship with a couple of detachable toothpicks built in, so you have one ready when the meal is over.