2026-08-26 · 48h window · synthesized from 42 sources / 652 articles · ~2 min read

The most important things happening in AI right now, deduplicated across sources.

1. OpenAI's Jalapeño chip lands as the labs design their way off Nvidia. OpenAI published the first benchmarks for its custom inference chip, claiming more tokens per user and more throughput per kilowatt than current state of the art, with a 128-chip rack quoted at 1.7 exaFLOPS and 27 TB of HBM. IBM's new mainframe processor, which runs Arm and Z instructions on the same cores, and early Groq 3 LPU numbers point the same way: inference is increasingly built around purpose-made silicon rather than general GPUs. Sources: OpenAI Blog · TechCrunch · The Register

2. Apple and Perplexity push inference onto hardware you already own. Apple's refreshed Mac mini (M6) and Mac Studio (M5 Ultra) are pitched explicitly at local AI development, with the memory shortage showing up as a $200 price bump on the mini. Perplexity, working with Nvidia, launched Portable Computer, an agent that runs entirely on the user's own machine at zero token cost, while Nvidia's Jetson Orin Nano 2 doubles inference for robots at the edge. Sources: Ars Technica · Wired · VentureBeat · The Register

3. The money is flowing to physical AI, not chatbots. Generalist reached a $3B valuation on a $200M extension only months after hitting $2B, XPeng's humanoid unit Dogotix raised $900M at a $6.3B valuation, and self-driving truck startup Gatik raised $200M off the back of its PepsiCo deal. Waymo is taking robotaxis to Munich and NEC has invested in Dexmate: capital is concentrating on machines that act in the world. Sources: TechCrunch · The Robot Report · Robotics & Automation News

4. The data center's physical footprint is becoming the story. US data centers tripled their water use in a decade, proposed gas plants are multiplying to feed them, and the demand is pulling solid-state transformer technology to market ahead of schedule. At the same time the EPA plans to drop the requirement that the public be notified about polluting data centers, leaving that call to state and local regulators. Sources: The Register · Wired · Ars Technica

5. AI security is moving from scanning to runtime enforcement. Prompt injection has topped the OWASP LLM list for three straight years yet ranks only No. 12 across 6,639 labeled real-world incidents, a gap that measures visibility rather than danger, because the attack is invisible to a scan. Microsoft is answering by shifting AI governance from written policy to runtime enforcement, while attackers work the other side: sponsored search ads for a fake OpenAI Codex are delivering Mac malware to developers. Sources: VentureBeat · InfoQ · The Register

6. Anthropic gives Claude one memory across apps and a voice in Slack. Claude and Cowork now share a single memory, so users stop re-briefing the assistant on projects and preferences; Anthropic says there is no option to separate it between the two services, and Claude Code stays outside it for now. A Claude Tag update lets the Slack agent read a full conversation and jump in unprompted, part of a "multiplayer AI" push away from the single-user chat window. Sources: TechCrunch · The Register · VentureBeat

Read the living page: https://jacopocastellano.com/ai-digest · Archive