AI Waves # 1: Anthropic sues the Pentagon, and why Circle just doubled
Anthropic sues the Pentagon, and why Circle just doubled
͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏
Welcome to AI Waves, a weekly note from Nazaré Ventures on what we’re watching in AI infrastructure, what’s happening in our portfolio, and where we’re going next.
Agents Are in the Wild. Everyone Is Scrambling
We are seeing the first AI agents deployed in the wild, and every major platform is racing to own the layer.
OpenAI acquired OpenClaw, the open-source agent framework that crossed 100,000 GitHub stars faster than React, TensorFlow, or Kubernetes. Meta bought Moltbook. Anthropic shipped Dispatch, letting you control desktop agents from your phone. In the space of a few weeks, agents went from demo to default. The scramble is on.
OpenClaw is the one that matters most right now. It’s persistent, always-on, running on your machine, connected to your tools. Not a chatbot you open and close. Closer to a digital coworker that sits in the background and acts on your behalf.
But there’s a problem nobody is talking about. 40,000 exposed OpenClaw instances have already been found on the public internet. Roughly 7% of the skills on Clawdhub are malicious. Users are giving root access to their computers to AI agents, and most of them have no idea what that means.
This is Linux in 1999. The open-source infrastructure is real, adoption is exploding, and nobody has solved the enterprise layer yet. Red Hat built a $34B business by doing exactly that for Linux: packaging, securing, certifying, and supporting open-source software for organizations that needed it to work and needed it to be safe.
The same opportunity exists now for agents. 86% of CIOs plan to move workloads back on-premise. A recent court ruling found that AI tool outputs lack attorney-client privilege, which means every law firm, bank, and healthcare company now has a compliance reason to keep agents on infrastructure they control (sovereign).
The company that configures, deploys, secures, and maintains AI agents on sovereign infrastructure will build one of the defining businesses of this cycle. That’s where we’re investing.
Read the full piece on Robot Wave
What We’re Watching
Anthropic sues the Pentagon. Dario Amodei drew a hard line: no AI for autonomous weapons, no mass domestic surveillance. The Pentagon declared Anthropic a supply-chain risk. Trump called them “radical left, woke.” Anthropic sued. The deeper question: who controls the AI infrastructure stack, under what conditions, and who decides? “Sovereign AI infrastructure”, systems that can’t be held hostage by any single government or platform, is becoming a core investment thesis, not just a talking point. Full story →
Circle is up 125%+ in a month. CRCL has doubled, outpacing Coinbase (+8.5%). The drivers: higher-for-longer rates boosting reserve income, tokenized assets ballooning to ~$26.5B, and USDC becoming the default settlement layer for prediction markets ($22B+ volume on Polymarket alone in 2025) and an emerging, but still small mechanism for agent payments. I was in Circle’s first round in 2013 as a partner at Pantera Capital. It’s gratifying to see stablecoin infrastructure finally getting repriced as the critical payments rail it always was. CoinDesk →
NVIDIA is building on OpenClaw. At GTC this week, NVIDIA launched NemoClaw, an open-source security and privacy layer built on top of OpenClaw, the emerging OS for personal AI agents. It runs open models locally via Nemotron, routes sensitive requests through a privacy router, and deploys with a single command across consumer GPUs through DGX. When NVIDIA builds its enterprise stack on top of an open-source agent framework, that’s the clearest validation signal in infrastructure. We’re actively using OpenClaw internally at Nazaré. The agentic layer (orchestration, memory, coordination, security) is a big focus for us now.
Portfolio Updates
Dimensional :1,800+ GitHub stars in first week of open-source launch. DimensionalOS shipped v1.0 of their framework for programming humanoid robots, launched on a Unitree G1, and went viral. Now supports a large number of OEM robotics hardware systems. We invested at seed in November 2025. Series A is coming. Reach out if you’d like an introduction. Watch the demo →
Prime Intellect : Official NVIDIA collaboration announced at GTC. Prime Intellect is now building the “open superintelligence stack” on NVIDIA Blackwell and upcoming Vera Rubin NVL72 systems. NVIDIA Dynamo powers their global inference orchestration; their RL sandbox infrastructure runs on NVIDIA Vera CPUs (176 VMs per socket, 30% greater throughput vs. AMD Zen 5). They’re also integrated with NeMo Gym, Nemotron models, and NemoClaw. When NVIDIA builds its GTC keynote demos on your platform, you’ve arrived. Our seed position is now our strongest unrealized markup in the fund. Read the announcement →
Vast.ai : Named fastest-growing vendor by Ramp and Brex. Both rankings based on actual transaction data across 85,000+ businesses. New Startup Program, Model Library, and SkyPilot enterprise integration all shipped in Q1. Revenue doubled in the last 12 weeks. Read more →
LayerLens : Agent-as-a-Judge. Launched Partner Evaluation Spaces with Fireworks AI and Nebius as inaugural partners, independently benchmarking 200+ models on accuracy and latency (88 live evaluations for Fireworks, 200 for Nebius). Also shipped dynamic, agent-driven evaluations replacing static scorecards. LayerLens is becoming the neutral benchmark layer for inference providers. Think Consumer Reports for AI inference. Blog →