Vitalik Buterin is running his own AI stack on a laptop, and he thinks you should consider doing the same. In an April 2 blog post, the Ethereum co-founder laid out a detailed blueprint for operating large language models entirely locally, without any reliance on cloud providers, their terms of service, or their data collection pipelines.
The setup isn’t just a thought experiment. Buterin is actively using open-weight models like Qwen3.5:35B on an NVIDIA 5090 laptop, hitting roughly 90 tokens per second. That’s fast enough for real-time conversational use, which makes the “local AI” pitch more than aspirational.
The hardware and the stack
Buterin tested his configuration across multiple machines. The NVIDIA 5090 laptop was the clear winner on speed, but he also ran the setup on an AMD Ryzen AI Max Pro with 128 GB of unified memory, which clocked in at 51 tokens per second. He additionally tested a DGX Spark.
The software stack is equally deliberate. Buterin runs NixOS for reproducible system configurations, meaning the entire environment can be rebuilt identically from a specification file. He uses llama-server to host the models locally and bubblewrap sandboxes to isolate processes from each other and from the broader system.
Buterin’s setup includes a custom messaging daemon that requires human confirmation before executing any sensitive commands. The AI can suggest actions, but a human has to approve them before anything touches the real world.
Why self-sovereignty matters now
Buterin has been building toward this vision since early 2024, when he began publicly emphasizing the need for trustworthy, private AI tools.
One data point from the blog post stands out: approximately 15% of AI “skills,” the modular capabilities that agents use to perform tasks, may contain malicious code. That’s not a hypothetical risk assessment. It’s a warning that the current agent ecosystem, where users download and run third-party plugins with minimal auditing, is structurally vulnerable.
The Ethereum AI play
Beyond the personal setup, Buterin made a broader suggestion that could have lasting implications for the Ethereum ecosystem. He called for the development of AI models specifically fine-tuned for Ethereum use cases.
Fine-tuned models are trained on domain-specific data, meaning they understand Ethereum’s particular quirks: its EVM architecture, its token standards, its governance patterns, the way its DeFi protocols compose with each other. A general-purpose model can answer questions about Ethereum. A fine-tuned model can reason about it.
Buterin’s vision positions Ethereum not just as a blockchain but as an economic and coordination layer for decentralized AI agents. If agents need to transact, verify identities, or coordinate with each other in trustless environments, Ethereum’s existing infrastructure, smart contracts, decentralized identity, programmable money, becomes the natural substrate.
No new tokens or protocols were launched alongside the blog post. This is infrastructure-level thinking, not a product announcement.