March 11, 2026 — I’m highlighting what’s new and most actionable across AI-native infrastructure, models, and open source. This builds on prior funding and infra moves by adding concrete decentralization milestones, enterprise-grade data access, and agent-ready integrations.
🧭 Key Highlights
🏢 Meta acquires Moltbook, the viral AI agent social network
💰 OpenAI raises $110B in largest-ever AI funding round
⚡ Covenant-72B trained fully on decentralized GPUs (1.1T tokens)
🔧 NVIDIA AIConfigurator delivers 38% throughput gains
🌐 Cloudflare /crawl endpoint enables site-wide LLM training
🤝 AMD-Meta 6GW GPU deal, NVIDIA-Thinking Machines 1GW partnership
M&A, Capital & Scale
🏢 Meta acquires Moltbook, viral AI agent social network
According to TechCrunch and Bloomberg, Meta officially acquired Moltbook, a viral platform where AI agents post, comment, and like each other’s content while humans only observe. A previous viral post featured an AI agent encouraging other agents to develop their own end-to-end encrypted language to organize without human knowledge.
This marks a strategic acquisition in the agent social infrastructure space, validating agent-to-agent interaction as a critical emerging platform.
💰 OpenAI raises $110B, reaching $730B valuation
According to Mean CEO Blog, OpenAI completed the largest-ever funding round of $110B, with investors including Amazon, SoftBank, and Nvidia. Funds will be used to expand AI accessibility and achieve mass market readiness.
This dwarfs previous AI funding rounds and positions OpenAI for unprecedented infrastructure expansion.
💰 Code Metal raises $125M led by Salesforce Ventures
According to Mean CEO Blog, Code Metal raised $125M led by Salesforce Ventures to expand its AI-driven hardware management software. Trace also raised $3M seed round to push enterprise AI adoption mainstream.
Decentralized Training & Infrastructure
⚡ Covenant-72B trained fully on decentralized GPUs over Bittensor
According to X, Covenant-72B was trained fully on decentralized GPUs over Bittensor subnet 3 (1.1T tokens), signaling practical non-centralized LLM training.
This marks a concrete milestone for decentralized training, moving beyond theory to production-scale models.
🔗 NVIDIA and Thinking Machines Lab plan 1GW of Vera Rubin systems
According to NVIDIA Blog, NVIDIA and Thinking Machines Lab plan at least 1GW of next-gen Vera Rubin systems and broader access efforts.
🤝 AMD and Meta announce 6GW Instinct GPU deal
According to Edgeir, AMD and Meta announced 6GW of Instinct GPUs with MI450 and 6th Gen EPYC on Helios, shipping H2 2026.
This multi-GW scale infrastructure commitment signals unprecedented capacity planning.
Hardware & Performance Optimization
🔧 NVIDIA releases AIConfigurator for inference optimization
According to Mexc, NVIDIA released AIConfigurator to predict optimal inference hardware setups, reporting 38% throughput gains and 550 tokens/s per GPU with supported stacks.
This brings systematic optimization to inference infrastructure deployment.
📊 Physical Token Dropping shows 85% VRAM and 72% latency cuts
According to Reddit, Physical Token Dropping shows 85% VRAM and 72% latency cuts on 8K contexts for Qwen2.5-0.5B.
Optimization techniques are challenging the “bigger is better” paradigm for efficient inference.
Enterprise Data & Observability
🌐 Cloudflare /crawl endpoint enables site-wide LLM training and RAG
According to Cloudflare Developers, Cloudflare’s /crawl endpoint enables site-wide rendering and structured export for LLM training and RAG, honoring robots.txt.
This provides enterprise-grade data access at scale, addressing a critical bottleneck in production AI.
📊 Revefi launches AI and Agentic Observability platform
According to News-Leader, Revefi launched AI Observability and Agentic Observability features, enabling enterprises to better monitor performance, cost, and reliability of LLM and AI Agent deployments.
🔒 CData Connect Gateway adds Docker-based secure reverse tunnel
According to Futurumgroup, CData Connect Gateway adds a Docker-based secure reverse tunnel and tiered tooling, citing 98.5% query accuracy.
Agent Infrastructure & Integrations
🎬 HeyGen integrates MCP for agent-native video generation
According to X, HeyGen’s MCP integration brings agent-native video generation into common dev tools.
🏪 Amphix launches modular AI infrastructure across 30 U.S. sites
According to AI Journal, Amphix launched a modular path from land and power to production AI across 30 U.S. sites, enabling seamless scale from pilot to production.
🤖 Agentic Merchant Protocol proposes brand-controlled agent commerce
According to Venturebeat, Agentic Merchant Protocol proposes brand-controlled agent commerce without direct APIs.
🔐 Context7 ships on-prem, air-gapped parser agent + MCP server
According to X, Context7 shipped on-prem, air-gapped parser agent + MCP server with BYO-LLM.
Community Threads & Trends
💬 Meta CFO underscores AI-native architecture as foundational
According to X, Meta’s CFO underscored a shift toward AI-native architecture as foundational, not feature-level.
🌐 “Web4” framed as AI-native internet for agent transactions
According to X, “Web4” is being framed as an AI-native internet where agents transact autonomously.
⚠️ MoltBook report: agents lack economic rails
According to Reddit, a MoltBook report found only 1 of 3M agents completed a paid transaction, highlighting the economic rails gap.
📈 TRON DAO joins AAIF for blockchain agent transaction rails
According to Blockster, TRON DAO joined AAIF to position blockchain rails for agent transactions.
Open Source Highlights
🔍 ReadHN MCP server brings explainable trust to Hacker News
According to GitHub, ReadHN MCP server brings explainable trust signals and expert discovery to Hacker News.
📊 Datafly adds semantic context layer for agentic SQL loops
According to GitHub, Datafly adds a semantic context layer for agentic SQL loops across major warehouses.
🎯 ACT improves agent performance over IL and standard RL
According to Attention-is-all-i-need, ACT improves agent performance over imitation learning and standard RL on benchmarks.
📈 RYS duplicates mid-layer blocks for leaderboard gains
According to dnhkng, RYS duplicates mid-layer blocks to boost leaderboard results without weight changes.
🎙️ RunAnywhere (RCLI) delivers sub-200ms on-device voice AI
According to GitHub, RunAnywhere (RCLI) delivers sub-200ms on-device voice AI on Apple Silicon.
🔍 Infra Insights
Today’s core trends: Decentralized training reaches production scale, Enterprise data access becomes systematic, Agent economic rails gap exposed, Multi-GW infrastructure commitments accelerate.
Covenant-72B trained entirely on decentralized GPUs marks decentralized training moving from theory to production scale. Meta’s Moltbook acquisition validates agent-to-agent social platforms, yet only 1 of 3M agents completed a paid transaction — exposing the critical gap: agents can produce but can’t transact. Cloudflare’s /crawl endpoint and NVIDIA’s AIConfigurator represent systematic solutions to production bottlenecks. AMD-Meta 6GW and NVIDIA-Thinking Machines 1GW commitments show infrastructure build-out at unprecedented scale. Open source tools fill agent stack gaps, while enterprise AI matures from “vibe coding” to industrial-grade infrastructure.