Run AWS workloads on any compute you want. Zero re-write.
AI inference to better GPUs, storage to somewhere cheaper, move the stuff that makes sense to move, leave the rest.
The diff is one flag: --endpoint-url.
The SDK clients, the Terraform provider and the AMIs you already
built carry over untouched.
Start on a Neocloud partner's fleet, move into your own racks, then push the same build out to a single box at the edge. One platform, every deployment shape.
AGPL-licensed and auditable, down to the storage engine. Read the source, run it in production, send us a patch. No vendor can lock you out of the platform your business runs on.
An emulator returns a plausible response. Spinifex boots the machine. A call lands on the gateway, crosses NATS to the daemon that owns the resource, and comes back having launched a QEMU guest, attached a Viperblock volume, or written an erasure-coded object into Predastore. Your software cannot tell the difference, because there isn't one.
Every object is split into data and parity shards by Reed–Solomon erasure coding and spread across nodes. Lose a disk and the bucket still reads. Rebuilds run online, with no operator in the loop.
Volumes sit on a write-ahead log behind an NVMe write-back cache that replays crash-consistently. On our reference rugged hardware that cut p99 write latency by 4–6×.
Instances are real KVM guests, with PCIe devices passed straight through so GPUs, SDRs and capture cards run at near-native speed. Cloud-init, user-data and SSH key injection behave as they do in EC2.
381 AWS API actions, implemented against the real API rather than a mock. The AWS SDK, the AWS CLI and Terraform point at one endpoint, and your workloads run unchanged at the edge, on-premise, or on a partner Neocloud.
Roadmap items ship under the same AWS API surface. Code written for AWS today keeps working the moment they land. Track what's shipped in the release notes.
Each platform here has been through a Spinifex validation on the actual silicon, and the reference architecture and benchmark numbers are published in our docs. The list grows with every vendor we work through.
An AWS workload moving out to your own edge and on-premise sites lands here, on hardware your teams already rack and support. Three references are published end to end: the platform benchmark, a production vision pipeline, and vLLM serving measured on Intel AMX against NVIDIA L4.
Move GPU work off AWS without touching the stack that runs it, from training through inference and the general compute around them. Supermicro gives us the deepest accelerator coverage of any platform we support: NVIDIA H200 partitioned through MIG, RTX Pro 6000 serving under EKS, and AMD Instinct MI350X on the H14 chassis.
Carries the same AWS workload out to sites with no rack and nowhere to put one. A single fanless HX401 runs a complete Spinifex deployment on its own, with no machine room behind it.
Run it on a partner Neocloud, in your own data centre, or at the edge. The AWS-compatible surface on top is the same.
Move workloads off the hyperscalers and onto our Neocloud partner ecosystem, without rewriting them. GPU capacity, cheaper, available now.
Explore neocloudsBring your own hardware. Host Spinifex in your data centre with full control of stack, data, and jurisdiction.
Explore on-premiseAir-gapped sites, vehicles, vessels, factories, clinics. Compute next to the data, running through disconnection.
Explore edgeSpinifex sits underneath the same goal for everyone: run AWS software on infrastructure you control. How that helps depends on who you are.
Reclaim sovereignty over your data and your bill. Repoint workloads, don't rewrite them.
ViewBreak out of the hyperscalers. Train and serve on Neoclouds with the same AWS APIs.
ViewRun AWS-native software at the edge: denied, disconnected, contested. Where it has to work.
ViewOffer an AWS-compatible surface so enterprise customers can migrate onto your infrastructure.
ViewPoint your AWS CLI or Terraform at a hosted Spinifex environment and watch your own software run. Nothing to install, no card required, credentials in under 60 seconds.