Glonce

machine0 launches CLI-driven CPU and GPU VMs for AI agents

infrastructureThe Glonce editorsHacker News🔥56Glonce rating56Impact45Novelty50Relevance80Surprise45Players35Credibility85

machine0, a company from Y Combinator's S26 batch, launched on Hacker News with a Launch HN post offering persistent CPU and GPU virtual machines built to be driven by AI agents. Every operation is exposed as a CLI command with a --json output option, and a remote MCP server lets agents create, control and manage VMs programmatically. VMs come in NixOS or Ubuntu images that machine0 says are open source; NixOS builds are pitched as fully reproducible, with deterministic output, one-command rollbacks and no configuration drift, while the Ubuntu image ships with Docker, Node, Python, Claude Code and Codex pre-installed. Machines can be started, suspended, snapshotted and resumed: suspending a VM freezes its state and stops compute billing, and users can turn any VM into a golden image to spin up clones from it. "Profiles" let a user inject MCP servers, credentials, prompts and environment variables into a VM, and machine0 says Claude Code and Codex pick these up automatically once injected. Each VM gets a dedicated resource allocation, a static public IP and an HTTPS endpoint at a .mac0.io address, with no NAT or tunnel required, across five regions: US East, US West, UK, EU and Asia. Sizing ranges from 1 to 60 vCPU and up to 240 GB of RAM. Optional GPUs include the H100, H200, L40S, MI300X and RTX 4000/6000 Ada, with configurations scaling up to 8x H200; GPU sizes are available only in us-east, uk, eu and asia, not in US West. A separate line of "-nvme" sizes runs on newer-generation CPUs with NVMe storage for higher disk IOPS at the same vCPU, RAM and disk specs as the regular sizes, and "-premium" sizes offer dedicated, not shared, vCPUs, which machine0 describes as its fastest single-thread performance tier. Billing is pay-as-you-go per minute at the same price across all regions: CPU VMs start at $0.013 per hour and GPU VMs start at $0.836 per hour. A suspended VM is billed only for image storage, at $0.078 per GB per month. The minimum account top-up is $5, and machine0 says unused credits are refundable. The company quotes 99.99% uptime for the service. The launch text markets use cases such as offloading compilation or automated research to dedicated VMs, running agents inside VMs configured through profiles, generating reproducible NixOS builds via a "/create-machine" skill, and hosting web apps or databases on a VM's static address.

Key facts

  • machine0, part of Y Combinator's S26 batch, launched persistent CPU and GPU VMs controlled via CLI (--json output) or a remote MCP server.
  • VM images are NixOS (deterministic builds, one-command rollbacks) or Ubuntu with Docker, Node, Python, Claude Code and Codex pre-installed; both are described as open source.
  • Sizing runs 1-60 vCPU and up to 240 GB RAM, with optional GPUs (H100, H200, L40S, MI300X, RTX 4000/6000 Ada) scaling up to 8x H200; GPU sizes are offered in four of the five regions (not US West).
  • Billing is per minute: CPU VMs from $0.013/hr, GPU VMs from $0.836/hr, suspended-VM storage at $0.078/GB/month, with a $5 minimum top-up and refundable unused credits.
  • "Profiles" inject MCP servers, credentials, prompts and environment variables into a VM, which machine0 says Claude Code and Codex pick up automatically.

Why it matters

machine0 is pitching infrastructure built specifically around AI coding agents rather than human developers: every action is a scriptable CLI command or an MCP call, and the VM images ship with Claude Code and Codex already installed. That framing, combined with NixOS-based reproducible builds, targets a gap between generic cloud VM providers and the newer pattern of letting an agent provision and manage its own compute.

Who it affects

The product targets developers and teams running AI agents that need to spin up, configure and tear down compute on their own, plus anyone offloading compute-heavy tasks like compilation or automated research to a dedicated machine instead of a local one. It also applies to teams that want a reproducible, drift-free build environment for agent workflows via NixOS flakes.

How to use it

Every operation runs as a CLI command with a --json flag, or through a remote MCP server for programmatic control. VMs can be started, suspended (freezing state and pausing compute billing), snapshotted into golden images, and cloned. "Profiles" inject MCP servers, credentials, prompts and environment variables into a VM so that Claude Code or Codex running inside it pick them up automatically. Billing is per minute and pay-as-you-go: CPU VMs from $0.013/hr, GPU VMs from $0.836/hr, with a $5 minimum top-up, $0.078/GB/month for image storage while suspended, and refundable unused credits.

How solid is it

The only source here is machine0's own launch page and Launch HN post, so every figure, including the quoted 99.99% uptime, is the company's own claim rather than an independently verified one. No customer count, revenue or funding amount is disclosed in the source, and there is no comparison against competing VM or cloud providers.

Risks and caveats

The source names no founders or team members and gives no specific launch date beyond the Launch HN framing and the YC S26 label. GPU sizes are available in only four of the five listed regions, with no explanation given for the gap. No remedy or credit is stated if actual uptime falls short of the claimed 99.99%.