Newsletter Archive

Anthropic committed $15 billion annually through May 2029 for compute access to SpaceX's Colossus data centers in Memphis, per SpaceX's IPO filing—a deal worth more than SpaceX's entire 2025 revenue that signals major capital deployment by a leading AI lab to secure GPU capacity.

The Verge AI · 3 min read Industry

Groq researchers published details on SHIP, their SRAM-based inference pipeline now serving hundreds of billions of tokens daily on public cloud. The system uses synchronous low-latency interconnects across thousands of chips and specialized optimizations to achieve state-of-the-art decode latency while sidestepping GPU HBM bandwidth bottlenecks.

Semiconductor Eng · 5 min read Hardware

Datasette Agent, a new conversational AI assistant for the Datasette database platform, lets users query SQL data via natural language and generate charts. The tool runs on Gemini 3.5 Flash-Lite and can automatically write complex SQL queries from English questions—demonstrated querying a bird sighting database.

Simon Willison · 4 min read Tools

Nvidia reported an 85% revenue increase driven by surging demand for AI infrastructure, signaling sustained market momentum for GPU-based AI workloads and influencing which hardware practitioners choose for model training and deployment.

Ciodive · 4 min read Market

Daytona's CEO discusses how agents' need for reliable, API-accessible compute is driving the company's 74% month-over-month growth and 850K daily sandbox runs. The company is capitalizing on agents becoming the default development paradigm, providing bare-metal sandboxes that solve what developers couldn't do locally—giving AI agents a real computer to work with.

Latent Space · 69 min read Industry

Cursor's annualized revenue surged to $3 billion by late April—doubling from $2 billion in February—with 3,000+ enterprise customers paying $100K+ each, ahead of an expected SpaceX acquisition by Elon Musk.

Bloomberg Tech · 1 min read Market

Trump pulled an anticipated executive order on AI and cybersecurity hours before signing after David Sacks, Mark Zuckerberg, and Elon Musk pushed back against what they saw as unnecessary regulation. Trump cited concerns that rules would hamper U.S. competitiveness against China.

Axios · 3 min read Policy

AgentAtlas proposes a unified evaluation framework for LLM agents across diverse tasks (code, browsers, OS), using taxonomies for control decisions and failure modes. Testing eight models reveals that removing explicit supervision drops accuracy 14-40 points and collapses performance to a 0.54-0.62floor, suggesting most gains come from prompt engineering rather than genuine capability.

arXiv AI · 3 min (abstract) Research