HZ//AI

Live GPU exchange

Rent GPUs
by the second

A marketplace of provider GPUs you rent by the second. Run an SSH box, a batch job, or an inference endpoint, each one hard-isolated on the host. Or buy the hardware outright.

[R] Rent a GPU[P] Become a provider[B] Buy hardware
Two ways in

Rent the compute, or own the hardware.

Rent computeper-second

Start a node in seconds and only pay for the seconds you use. SSH in, run a batch job, or serve an LLM inference endpoint. Prepaid credits, transparent provider pricing.

  • /SSH & dedicated reservations
  • /Batch & training runs
  • /Inference endpoints
Browse marketplace →
Buy hardwareyou own it

Prefer to own the machines? We source the parts, build and test the boxes, and ship them to your rack. Itemized quote before anything ships.

  • /GPUs & full servers
  • /RAM, storage, networking
  • /Complete builds to spec
Explore hardware →Talk to sales
Isolation

One request path. Tenants never touch.

Every inference request follows the same isolated path from the public edge down to a sandboxed workload on the provider host.

Request pathedge → host

Buyer

HTTPS

Caddy edge

auto-TLS

Gateway

auth · meter

tlsproxy

mTLS

Workload

sandboxed

Prompts and outputs stay on the provider host. Only metering counts cross to the control plane.

Provider IP never exposed

On the inference path, buyers hit a per-deployment gateway. The datacenter's host address stays behind mTLS. It isn't in any buyer-facing API response.

Hardened, enrolled agent

Providers run a memory-safe Rust agent enrolled over mTLS. Private (RFC1918) networks and cloud metadata are denied by default and workloads run under gVisor.

Keys you can trust

Per-deployment API keys, compared in constant time and revocable instantly. Rate limits and request caps are enforced at the gateway, not bolted on.

For GPU owners

Have GPUs? Put them to work.

Install the agent, set your price, and earn whenever a buyer runs on your hardware. Every job is sandboxed and metered. You never hand out shell access to your box.

  • /One-command install on Ubuntu / Debian
  • /Set your own price per GPU-hour
  • /Specs auto-detected and shown to buyers
  • /Receive payouts in fiat
provider · setupbash
$ curl -fsSL 'https://github.com/wombatlabs/hzai-agent/releases/download/v0.1.2/install.sh' | sudo env HZAI_VERSION='v0.1.0' HZAI_BOOTSTRAP_TOKEN='hzai_bs_…' HZAI_API_URL='https://api.hz-ai.io' HZAI_MTLS_API_URL='https://api.hz-ai.io:8443' HZAI_PUBLIC_IP='203.0.113.10' bash
✓ agent v0.1.0 installed
✓ detected 2× NVIDIA A100 80GB
✓ enrolled over mTLS · listing live

$ hzai-agent --price 1.89
waiting for jobs…
FAQ

Your questions answered.

Per second. You prepay with credits and a running job is metered per second at the provider's GPU-hour rate, minus our platform fee. Unused credits stay in your account.

Yes, on-demand instances and dedicated reservations give full SSH access with your own public key. You can also run batch jobs or deploy inference endpoints instead.

Each job runs in its own container with dropped capabilities and a read-only root filesystem, and gVisor syscall-sandboxing where the runtime is configured. On the inference path the provider's host IP is never exposed to buyers, traffic goes through a per-deployment gateway over mTLS.

Whatever providers currently list. That spans consumer cards (RTX 4090/3090) and datacenter GPUs (A100, H100). Availability tracks real inventory, so it changes through the day.

Run the one-command installer on an Ubuntu/Debian box, set your price, and your node lists once the agent enrolls. Payouts are made via Stripe.

We don't claim a uptime SLA, multi-region HA, or SOC 2 certification today. Capacity depends on live provider inventory, and a few things (private registries, confidential-compute GPUs) are on the roadmap.

Pick a node. Start renting.

Browse live inventory, or list your own GPUs and get paid for the idle time.

Browse marketplace →