In partnership with
Welcome, AI enthusiasts
Compute is rewriting who works with whom in AI. Anthropic just leased every GPU at Musk's Memphis facility, and, gave Claude agents a way to learn from their own mistakes overnight. Let's dive in!
In today’s insights:
Anthropic Strikes Deal with SpaceX for Compute and Unveils Claude Dreaming
Chrome Caught Silently Installing 4GB Gemini Nano
Nvidia Bolts AI Data Centers onto Homes
Read time: 5 minutes
LATEST DEVELOPMENTS
Evolving AI: Anthropic locked in SpaceX's full Colossus 1 data center and previewed a "dreaming" feature for Claude agents.
Key Points:
Anthropic gets all 220,000+ Nvidia GPUs at Colossus 1, on top of recent megadeals with Amazon, Google, Microsoft, and Fluidstack.
Users feel it right away, with Claude Code five-hour limits doubling for paid plans and Opus API limits rising sharply.
Anthropic also previewed "dreaming," which lets Claude Managed Agents reflect on past runs, catch recurring mistakes, and reshape memory between sessions.
Details:
Both announcements came out of Anthropic's Code with Claude event in San Francisco. The SpaceX deal brings 300 megawatts online within a month, and the two companies floated multi-gigawatt orbital data centers, a goal SpaceX has been openly chasing. Musk, who called Anthropic "misanthropic" back in February, reversed course after meeting the team and said Claude will "probably be good." On the product side, dreaming runs as a scheduled review between agent sessions, surfacing patterns a single agent can't see on its own, including recurring mistakes, workflows agents converge on, and preferences shared across a team. It goes further than the in-conversation compaction most chat models already do, and is in research preview for Managed Agents on the Claude Platform.
Why It Matters:
The compute side fixes an immediate problem. Anthropic admitted last month that demand was straining Claude's reliability, and doubling rate limits responds to that directly. Even Musk, one of Anthropic's loudest critics, swallowed his words to do business, a sign of how tight power and compute have become across the industry. Dreaming is the quieter half, but it's solving something compute can't. Agents running for hours lose the thread once context windows fill up, and giving them time to review their own work points at where the next round of progress comes from.
TOGETHER WITH DATADOG
🔐 Developer Toolkit for the AI Era
Evolving AI: The developer toolkit for shipping AI features with confidence
With release cycles speeding up in the era of AI, developers need to move fast without losing visibility in production. Get 4 resources covering everything from catching flaky tests and pipeline bottlenecks to instrumenting LLM calls and controlling rollouts before regressions reach users.
You'll learn how to:
Track every CI pipeline run and cut test suite instability slowing your AI delivery cycles
Catch LLM quality, latency, and cost issues before they surface in production
Measure and improve release confidence as AI drives higher commit volume across your team
Source: Digwatch
Evolving AI: Chrome's latest update is auto-downloading Google's on-device LLM without user consent.
Key Points:
Researcher Alexander Hanff found Chrome writes a 4GB weights.bin file for Gemini Nano without asking.
The model re-downloads itself if deleted, unless the user disables AI flags entirely.
The behavior may violate the EU's ePrivacy Directive and GDPR.
Details:
Hanff verified the install on Windows 11, macOS, and Ubuntu, with the 4GB weights.bin file landing in a folder called OptGuideOnDeviceModel. It powers Help Me Write, scam detection, and a Summarizer API, features most users never touch. The visible AI Mode in the omnibox still routes queries to Google's cloud. Hanff estimates a 500-million-device push produces 6,000 to 60,000 tonnes of CO2e. Google added a settings toggle in February but did not notify existing users.
Why It Matters:
A toggle added a year later is a strange definition of consent. On-device AI was meant to flip the script, less cloud, more user control. What it is quietly turning into is the opposite, vendors using your hardware as a place to stage their next feature. Chrome just made that pattern impossible to miss.
Your best prompts are the ones you'd never bother typing.
The detailed ones. The ones with examples and edge cases. Wispr Flow lets you speak them instead — clean, structured, ready to paste into any AI tool. Free on Mac, Windows, and iPhone.
Source:Shutterstock
Evolving AI: Nvidia, Span and PulteGroup are testing AI data center nodes mounted on residential homes.
Key Points:
Nvidia is teaming up with smart-panel startup Span and homebuilder PulteGroup to install fractional data centers on residential homes.
Each Span's XFRA unit packs 16 liquid-cooled Blackwell GPUs and runs silently with no fans.
The network taps unused household power, framed as a distributed alternative to centralized buildouts.
Details:
Span estimates the average home runs at just 40% of peak power capacity, leaving roughly 19 kilowatts unused. XFRA captures that. Each unit holds Dell servers with liquid-cooled Blackwell GPUs, paired with Span's smart panel and a backup battery. Homeowners get cheaper power and Wi-Fi while hyperscalers tap the network as if it were a regular cloud region. Span claims the model deploys six times faster and costs five times less than a 100-megawatt facility.
Why It Matters:
Big AI has been built on the assumption that compute scales by stacking it in one place. That assumption is getting expensive. Communities blocked or delayed $98 billion in centralized data center projects last year, and over 30 states are weighing bills to slow the buildout. Spreading the load across suburbia is less of a product launch and more of an admission that the centralized model is starting to crack.
👀 Our tip
Anthropic published a fascinating deep dive into how people are using Claude for life advice, relationships, career decisions, and emotional support. The report reveals where AI assistants can become surprisingly persuasive or overly validating, and why that matters if you increasingly rely on AI for decisions, feedback, or even difficult conversations.
Source: Anthropic
👀 Click on the image you think is real
QUICK HITS
🐳 DeepSeek eyes $45B valuation in first ever funding round.
🔍 Google AI Overviews tap Reddit threads for niche answers.
🏭 SpaceX may pour $119B into "Terafab" chip megafactory.
💸 Snap and Perplexity quietly call off $400M AI deal.
📈 Trending AI Tools
🐰 CodeRabbit - Ship higher quality code with AI-powered code reviews*
🎨 Beautiful AI - Create professional, client-ready slide decks in minutes
📄 Kickresume - AI resume builder
🤖 Vapi - Build, test, and deploy advanced voice AI agents in minutes
*partner link