Vitalik Buterin wants to cut the cloud out of his AI workflow entirely. In an April 2, 2026 blog post, the Ethereum co-founder laid out a detailed update on his local large language model setup, one designed to keep every token of inference running on hardware he physically controls.

The bigger pitch buried in the technical walkthrough: Ethereum needs its own fine-tuned AI models, purpose-built for tasks like verifying transactions and auditing smart contracts.

The hardware and the stack

Buterin’s setup runs Qwen3.5:35B, an open-weight model, locally on an Nvidia 5090 laptop. The performance numbers are genuinely impressive for consumer-grade hardware: up to 90 tokens per second.

He also tested alternative hardware. An AMD Ryzen AI Max Pro with 128 GB of unified memory hit 51 tokens per second. A DGX Spark managed 60 tokens per second.

The software stack is equally deliberate. Buterin uses NixOS for reproducibility, meaning every aspect of the operating system configuration is declarative and version-controlled. llama-server handles actually serving the model. Bubblewrap provides sandboxing for tasks, isolating processes so a misbehaving AI agent can’t reach beyond its designated box.

Rounding out the setup: a custom messaging daemon and a local Wikipedia dump. By storing a copy of Wikipedia locally, the system minimizes the number of external web searches the model needs to make. Fewer outbound requests means fewer opportunities for data to leak or for third parties to observe what questions are being asked.

Why local matters more than convenience

Buterin’s post frames the move away from cloud AI as a genuine security concern. He cites security research suggesting that roughly 15% of AI agent skills may contain malicious instructions. In English: when you let an AI agent browse the web, use tools, or interact with third-party plugins, there’s a meaningful chance some of those integrations are compromised or designed to manipulate the model’s behavior.

Buterin has been explicit about the dangers of giving AI agents unrestricted wallet access. His setup reflects that caution: sandboxed processes, local-only inference, minimal internet connectivity. The architecture is designed to keep humans in the loop for high-risk actions, even as AI handles routine analysis.

Ethereum-specific models and the bigger vision

The more consequential idea is Buterin’s push for models fine-tuned specifically for Ethereum use cases. He identifies several target applications: local transaction proposal and verification, where an AI model reviews a transaction before it’s signed and flags anything suspicious, and smart contract auditing, where models are trained on Ethereum’s specific patterns, vulnerabilities, and Solidity idioms.

This builds on a framework Buterin introduced in February 2026, which outlined four pillars for Ethereum’s role in the AI era. First, trustless private AI tooling, meaning infrastructure that lets people use AI without trusting a centralized provider. Second, Ethereum’s economic structure as a natural fit for AI agents that need to transact autonomously. Third, self-sovereignty through local verification. Fourth, enhancing governance and markets with AI support.

What this means for investors

Buterin didn’t announce a protocol upgrade or a new funding initiative. This is a blog post about a laptop setup. The emphasis on zero-knowledge privacy technologies and local processing could drive renewed interest in privacy-focused projects within the Ethereum ecosystem.

Fine-tuning models for Ethereum-specific tasks requires significant investment in training data, evaluation frameworks, and community coordination. None of that exists at meaningful scale yet. And the 90-tokens-per-second benchmark on an Nvidia 5090, while impressive, still represents a fraction of the throughput available through cloud providers.