AI news digest — August 25, 2026
6 items, each with its source.
Emerald AI raises $150M Series A to scale power-flexible AI data centers
Emerald AI raised $150 million in an oversubscribed Series A funding round co-led by Energize Capital and DCVC at a $1.05 billion valuation. The company develops orchestration software called Emerald Conductor that dynamically modulates data center compute loads in real time to match local electric grid constraints. The capital will be used to expand commercial deployments across utility and hyperscaler facilities worldwide.
Why it matters. Dynamic power throttling enables high-density AI clusters to bypass multi-year electrical interconnection queues and deploy on existing grid infrastructure.
finance.yahoo.comNvidia launches Jetson Orin Nano 2 edge processor with 78 TOPS AI compute
Nvidia announced the Jetson Orin Nano 2, an edge computing module designed to run multi-modal AI models locally on robotics hardware. The device delivers 78 TOPS of AI compute, 8GB of memory, and an eight-core CPU while consuming 40% less power than its predecessor in 15-watt mode. The hardware and developer kit will become generally available in the first half of 2027.
Why it matters. Small autonomous drones and lightweight robots can run vision-language-action models on-device without tethering their control loops to cloud latency.
siliconangle.comOasis Security uncovers Nvidia NemoClaw vulnerability enabling local model poisoning via DNS rebinding
Oasis Security disclosed CVE-2026-65105, a vulnerability in Nvidia NemoClaw's local runtime architecture that allows remote attackers to compromise developer AI agents. When running local models via Ollama, non-loopback binding combined with DNS rebinding allows malicious websites to rewrite prompt templates directly at the model server level. Injected instructions persist in the underlying model definition across new chat sessions without triggering agent sandboxes or runtime guardrails.
Why it matters. Sandboxing agent execution environments fails to protect developer systems if unauthenticated local inference daemons remain accessible to standard browser network requests.
siliconangle.comKeenable exits stealth with $26M seed round to index the web for AI agents
Startup Keenable exited stealth mode with $26 million in seed funding led by Accel, alongside Conviction Partners and angel investors. Founded by former Yandex search lead Andrey Styskin and AI scientist Matthias Petri, the company has indexed over 100 billion web pages specifically structured for agentic retrieval rather than human browsing. The platform exposes a proprietary Web Query Language and API to synthesize multi-source information for LLMs during inference.
Why it matters. Dedicated crawler architectures designed for synthetic multi-hop reasoning reduce the token overhead and scraping latency that standard search APIs impose on autonomous agents.
techcrunch.comRopedia launches HOMIE Gen2 wearable egocentric video capture device for training robotics models
Singapore-based robotics data startup Ropedia launched HOMIE Gen2, an egocentric head-mounted capture rig for collecting physical AI training datasets. The crown-shaped system synchronizes four camera streams, spatial audio, and inertial sensor data to within 50 microseconds to record human task execution from a first-person perspective. The device captures 360-degree visual data, depth maps, hand-body keypoints, and task metadata without requiring static lab infrastructure.
Why it matters. Standardizing high-precision egocentric human demonstration capture lowers the cost and turnaround time of training generalist visual-motor policies for humanoid robots.
siliconangle.comCisco and Supermicro expand rack-scale liquid-cooled AI infrastructure for sovereign cloud deployments
Cisco expanded its Secure AI Factory architecture in partnership with Super Micro Computer to support Nvidia-compliant infrastructure across sovereign clouds and neoclouds. The solution combines liquid-cooled Cisco N9000 switches using Silicon One and Nvidia Spectrum-X with Supermicro rack-scale compute capable of managing power densities exceeding 200 kilowatts per rack. The hardware integrates Cisco Nexus One and Nvidia AI Enterprise software for turnkey deployment starting in October 2026.
Why it matters. Pre-validated liquid-cooled rack blueprints allow enterprise data centers and regional cloud providers to deploy high-density frontier GPU architectures without custom networking engineering.
siliconangle.comFeed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.