Meta locks in $27B with Nebius. Nvidia's networking division is quietly outpacing Cisco.
Mistral released Small 4 with a 256K context window under Apache 2.0, and launched Forge for training custom models on proprietary data. Google shipped Stitch, a design tool that turns text prompts into production UI. Meta also signed a $27 billion compute deal with Nebius, its largest AI infrastructure bet to date.
Meta signed a five-year agreement worth up to $27 billion with Nebius for dedicated AI compute, built around Nvidia's Vera Rubin chips. It's the largest single infrastructure commitment Meta has made for AI. Nebius stock jumped 14% on the announcement. The deal signals how far hyperscalers are willing to go to lock in next-generation compute capacity before competitors do.
Read full story→Mistral released Small 4, merging its previous flagship models for reasoning, multimodal input, and coding into one. The model has a 256K context window with native text and image support, ships under Apache 2.0, and handles 3x more requests per second than its predecessor. Completion times dropped 40%. Mistral is positioning it as a direct competitor to mid-tier offerings from OpenAI and Google.
Mistral launched Forge, a platform that lets enterprises train frontier-grade AI models on their own proprietary data. It supports pre-training, post-training, and reinforcement learning across both dense and mixture-of-experts architectures. Companies keep full IP ownership of the resulting models. The pitch is that enterprises can build context-aware agents for internal operations without handing their data to a third party.
Google Labs launched Stitch, a design tool that generates production-quality UI from text descriptions. You describe what you want in plain language, focusing on user goals and business objectives rather than pixel specs. It includes an infinite canvas, a design agent that works across projects, and DESIGN.md integration for maintaining design system consistency. Google is calling the approach 'vibe design.'
Nvidia's networking division, built on the 2020 Mellanox acquisition, has grown into the company's second-largest revenue source. The NVLink and Spectrum-X infrastructure that connects GPUs inside AI factories is now pulling in billions on its own. TechCrunch reports it's outpacing established networking companies like Cisco, though it gets far less attention than the GPU side of the business.
ByteByteGo reverse-engineers how OpenAI built Codex, and the punchline is that the model was the easy part. The team tried using MCP to connect the agent to VS Code but it couldn't handle streaming, mid-task approvals, or code diffs, so they built a custom protocol from scratch. The piece covers the agent loop, how prompts are assembled from five sources, and what happens when context windows fill up.
Uber rebuilt its ML infrastructure from a monolithic Kubernetes stack to a cloud-native mesh with custom resources and a federation pattern. The original Michelangelo platform couldn't scale to their needs, so they moved to a system that handles millions of predictions per second across distributed clusters. The migration also sets up multi-cloud batch orchestration for training workloads.
Engineering leaders say AI-driven code generation is forcing development teams toward test-first workflows, because the tests are now the specification that matters. The piece introduces a 'recursive improvement thesis' where better models make better tools, which make better models. One practical consequence: teams are building internal replacements for commercial dev tools faster than SaaS vendors can ship updates.
a16z makes the case that the internet degraded customer service by forcing businesses to choose between quality and scale. AI removes that tradeoff by making personalised, proactive support affordable even at massive scale. The piece walks through how AI-powered support agents can handle the kind of individual attention that used to require expensive human staff, turning support from a cost line into a revenue driver.
The Atlantic draws a parallel between ChatGPT and early social networks like Myspace. The thesis is that ChatGPT's first-mover advantage is eroding as competing models approach parity and switching costs stay low. Unlike smartphones where one winner took all, AI may end up more like streaming, where multiple services coexist because none can lock users in. The question is whether OpenAI can build enough stickiness before the window closes.
Euclid VC profiles Hanover Park, which scaled AI-powered fund administration from zero to $15 billion in assets under management in 12 months. Their model is 'human-in-the-loop Vertical AI,' where AI handles repetitive data parsing and reconciliation while humans manage exceptions. The company grew by targeting a back-office process that incumbents had neglected for decades, then automating 80% of the manual work.
An open-source physics simulation engine built on NVIDIA Warp that runs entirely on the GPU. Designed for robotics researchers doing sim-to-real transfer, it supports rigid body dynamics, articulated systems, and contact interactions. The GPU acceleration means simulations that used to take hours on CPU can run in minutes, which matters when you need thousands of training episodes for reinforcement learning.
An open-source tunnel server written in Rust that works like ngrok but is fully self-hostable. Exposes local services through encrypted WebSockets with HTTP and TCP proxying, includes a web dashboard for managing tunnels, and supports custom domains. Also ships with an MCP server and Claude Code skill for integrating tunnel management into AI agent workflows.
An open-source web UI for training, running, and exporting open models entirely on your own machine. Handles text, vision, audio, and embedding models, exporting in formats like GGUF and safetensors. Claims 2x faster training with 60% less VRAM than standard approaches. Also includes dataset creation from your own documents and fully offline operation, so nothing leaves your hardware.
A terminal IDE that creates pre-configured tmux layouts for running multiple Claude Code agents in a single terminal. One command sets up panes for a lead agent, teammate Claude sessions, and dev tools, all defined in live-editable YAML files. Includes a Claude Code skill that auto-detects your project type and configures the environment. Aimed at developers who run several agents in parallel across projects.