On my list for my next SF trip 👊🌭

It’ll stand right next to me existing rack 😆

The new place is starting to come together nicely. I might have gone a little far with the network, though 😆

Per request, at the gateway. Each outbound call is checked against the user's delegated grant, the sandbox's identity and egress policy. Reachability is scoped to specific endpoints, Workspace APIs or domains, and the delegation itself covers specific work, not everything.

Oh. TP=4 vLLM

I can see about 4K prefill and 105 count 100 / 83 code / 44 prose. 2m KV Cache. Didn’t get to play with it long, though.

It already runs Gemini, Antigravity and Google's own eval and RL pipelines. Preview on Google Cloud is targeted for the coming months. Deep dive: cloud.google.com/blog/topics…

The other detail I like is that most agents spend most of their time waiting, for the model or for you. So an idle agent gets checkpointed down to near-zero compute and wakes on the next tool call with its files and memory intact. For coding agents and RL, you snapshot an environment once and fork it into thousands of copies that share memory.

Google's agent sandboxes assume the attacker already controls everything inside. So there is nothing inside worth stealing. Agent Compute Engine gives every agent its own hardware-isolated microVM. No API keys, no OAuth tokens, no network card in the box. A gateway outside the sandbox adds credentials to each request, only after it checks identity and policy.

Cool. Might give that a spin!

For now it supports the official assistant models.

Landed support for @googlegemma in @NVIDIAAI’s new inference engine Flare - today is a great day! github.com/NVIDIA/flare-loca…

So.. how many hot dogs were ordered to find the most photogenic one for the video? Love the demo! 😅 – at Großbeeren, Deutschland

Announcement: aistudio.google.com/learn/in… Repo: github.com/google-gemini/gem… Install on macOS or Linux: curl -fsSL raw.githubusercontent.com/go… | bash

The part I like most is how it works with coding agents. Media commands write to disk and print only the file path, so no base64 ends up in the context window. Every command describes itself offline with --help, --schema and --dry-run.

One terminal command, one image. No SDK script, no base64, no docs tab open. gemini-api image "Create an isometric image for a post on X of a server rack with 8 Nvidia DGX Spark systems (GB10) neatly clustered together" --out . It nailed the vibe, not the Sparks. Those are much smaller boxes. The new Gemini API CLI is one zero-dependency binary for text, images, video, TTS, music and Deep Research. Open source under Apache-2.0.

Wouldn't be able to tell from the display. Could be a mix of both?

Loved hearing about Sovereign AI at GTC Berlin. Throughout the sessions, the keynote, and conversations at the @GoogleCloudTech booth it was crystal clear that this topic is top of mind for people across the industry

Another cool setup I’ve seen at @NVIDIAAI GTC Berlin. This one by ABA Technology was very much about deploying a cohesive unit powered by the DGX Spark FEs

I thought the @Dell one looked very sleek but I’d make due with any of them 😉