Deploy GPU LLM
endpoints in minutes.
Provision on-demand GPU compute, expose OpenAI-compatible APIs, and route fallback traffic automatically — with crypto-native billing built into the protocol.
A control plane for the open GPU market
Astra Compute turns raw marketplace GPUs into production LLM endpoints. Pick a model and a GPU offer; the platform rents the instance, installs the inference serving stack, wires up health checks, and hands you an OpenAI-compatible URL with its own API key — usually in under a minute.
Billing is crypto-native by design. Instead of a credit-card form, runtime is funded up-front through an x402 payment challenge — the HTTP 402 flow that lets any API meter itself in USDC, per deployment today and per request tomorrow. If your machine ever can't serve, traffic fails over to the managed fallback network so callers never see downtime.
And compute doesn't have to sit idle: any machine can join the grid, Astra Compute's marketplace where staked operators sell inference by the token and earn USDC + SGL rewards.
- ComputeOn-demand GPU marketplace — RTX 4090 up to 8× H100 clusters
- ServingInference engine with OpenAI-compatible /v1 routes on every machine
- GatewayMachine-scoped keys, metering, health checks, failover
- Paymentsx402 challenges settled in USDC on Solana (simulated in demo)
- FallbackThe fallback network picks up traffic when a machine can't serve
- MarketplaceThe grid — staked operators selling inference per 1K tokens
Everything you need to run production inference
From spot-market GPUs to metered, monetizable endpoints — one deploy flow, no DevOps.
GPU marketplace, one click
Provision RTX 4090s to 8× H100 clusters from the on-demand GPU marketplace — filtered by price, VRAM, region, and reliability.
OpenAI-compatible endpoints
Every machine exposes /v1/chat/completions and /v1/models. Point any SDK at your base URL and ship.
x402-native payments
Runtime is settled up-front in USDC over the HTTP 402 flow — the primitive for per-request, crypto-native API monetization.
Automatic fallback
When your machine is cold, degraded, or over capacity, traffic reroutes to the managed fallback network automatically. No dropped requests.
Earn on the grid
List spare capacity on the Astra grid. Stake SGL, serve inference, earn USDC per 1K tokens.
Full observability
Live logs, GPU utilization, p95 latency, request metering, and health checks on every machine — no agents to install.
Pick a model
Llama, Mistral, Qwen, DeepSeek — or any custom model repository ID.
Choose compute
Live offers from the GPU spot market, ranked by $/hr and reliability.
Settle via x402
Sign a USDC payment challenge. Runtime is funded before the instance boots.
Serve & earn
Call your OpenAI-compatible endpoint — or list it on the grid and get paid.
If it speaks OpenAI,
it speaks Astra.
Swap one line — the base URL — and your existing stack runs on your own GPUs. LangChain, LlamaIndex, the Vercel AI SDK, or raw curl.
- Machine-scoped API keys with one-click rotation
- Streaming, usage accounting, and cost estimates
- Automatic fallback failover per request
from openai import OpenAI
client = OpenAI(
base_url="https://m-atlas.astra.run/v1",
api_key="sk-sing-...",
)
response = client.chat.completions.create(
model="llama-3.1-8b",
messages=[{"role": "user", "content": "Explain how GPU inference works in one paragraph"}],
)
print(response.choices[0].message.content)Questions, answered
Do I need API keys to try it?
No. Demo mode ships with realistic GPU offers, simulated payments, and simulated inference out of the box. Add your compute and fallback provider keys to enable live GPU offers and real fallback completions — everything stays server-side.
Is real money involved?
Not in this MVP. The x402 flow — challenge, signature, X-Payment retry, settlement — is simulated end-to-end, but the payloads match the spec so a real facilitator and wallet can drop in later without changing the product.
What happens when my runtime expires?
The machine flips to 'expired' and stops serving. Nothing is deleted — extend the runtime from the console or the machine page and it comes back. Destroy is the only irreversible action, and it always asks first.
Can I serve my own model?
Yes — pick 'Custom model' in the deploy wizard and give it any model repository ID. It's served through the same inference engine as the presets, with the same OpenAI-compatible endpoint.
How do node operators earn?
Stake 50,000 SGL, contribute a machine, and join the grid. Your endpoint gets listed on the marketplace and you earn USDC per 1K tokens served, plus SGL rewards per epoch. The flow is simulated on the Operators page.
Is my prompt data private?
Private endpoints only accept your machine's API key. Grid endpoints advertise a privacy mode — standard, TEE, or ZK-attested — so buyers can choose the guarantees they need. Attestation is a placeholder in the demo.