March 5, 2026 — AI compute capacity continues its rapid expansion, decentralized Agent infrastructure achieves breakthrough, and developer tools accelerate alongside robotics integration.
🧭 Key Highlights
🏢 IREN expands to 150K GPUs, projecting $3.7B annualized revenue
⛓️ 0G launches dAIOS with on-chain Agent ownership
📚 Andrew Ng releases JAX-based LLM course
💾 Residuum introduces Observational Memory for sessionless Agents
🤖 ROSClaw bridges OpenClaw to ROS2 robotics
Compute & Cloud Infrastructure
🏢 IREN to 150,000 GPUs by H2 2026
According to Globenewswire, IREN Limited signed purchase agreements for over 50,000 NVIDIA B300 GPUs, taking its total AI cloud fleet to 150,000 units. The company projects this capacity to support an AI Cloud annualized run-rate revenue of over $3.7B by the end of 2026.
The B300 is NVIDIA’s latest data center GPU, and this procurement marks a major scale milestone for AI-native compute providers.
Open Source Ecosystem
⛓️ 0G launches dAIOS with on-chain Agent ownership
According to Aiverse, 0G introduced dAIOS, a decentralized AI infrastructure protocol and OS enabling AI agents to be owned, traded, and evolved on-chain as iNFTs using the new ERC-7857 standard.
Features include agentic ID, a no-code agent builder, integrated compute/storage/data availability, and intelligent routing for cost-efficient inference.
📚 JAX-focused LLM course from Andrew Ng
According to X, a new course “Build and Train an LLM with JAX,” developed with Google, walks through implementing a ~20M-parameter MiniGPT-style model using JAX, Flax/NNX, and related tooling, with training, inference, and a chat interface.
The course targets developers looking to master next-gen LLM development tools, as JAX’s advantages in automatic differentiation and compilation optimization make it a popular choice for research and new model development.
💾 Residuum adds Observational Memory for agents
According to GitHub, the Rust-based personal AI agent framework introduced Observational Memory to compress conversation history into a dense, persistent log that removes session boundaries, alongside multi-channel support and Pulse Scheduling to reduce token waste.
Residuum positions itself as a lightweight local Agent runtime.
🤖 ROSClaw bridges agents to ROS2 robotics
According to X, an open-source integration connecting OpenClaw with ROS2 via rosbridge, translating natural language commands into robotic actions, was recognized at the SF OpenClaw Hackathon.
The project demonstrates a path for digital Agents to extend into physical robotics.
🔍 Infra Insights
Today’s core trends: Accelerating compute scale expansion, Agents move toward decentralized ownership, Developer tools and robotics integration advance together.
IREN’s 150K GPU plan and $3.7B revenue projection signals AI-native compute’s transition from experimental to commercial scale. 0G’s dAIOS and ERC-7857 standard establish Agents as ownable, tradable on-chain assets — critical infrastructure for the Agent economy.
Andrew Ng’s JAX course, Residuum’s Observational Memory, and ROSClaw’s robotics integration drive Agent ecosystem maturity across three dimensions: developer education, Agent state management, and physical embodiment.