For neoclouds and AI platform teams

Flash throughput. Hard-drive economics. One cluster.

Octodec is the software layer between your applications and your drives. It combines flash and high-density SMR hard drives into a single cluster that presents as ordinary object and file storage — native S3 and standard NFS, no application changes — delivering more throughput and more capacity for a given budget than flash alone or disk alone.

Bigger SSDs don't mean more throughput.

An SSD's throughput is capped by its interface. A 60 TB drive delivers roughly the same throughput as a small one, at many times the price — so throughput per terabyte falls as capacity grows. The same money buys enough hard drives to match that throughput and far exceed the capacity.

Hard drives alone can't hold the throughput. Flash alone can't reach the capacity. Octodec uses flash where throughput and latency are won and disk where capacity is won, in one cluster behind one interface — and moves the working set between them at line rate.

Throughput

Competitive with flash on dollars per GB/s, on hardware priced like disk.

Built for streaming workloads — training, retraining and sustained ingest that move large datasets continuously.
Read the whole dataset in about a day where raw hard drives take three or four.

Capacity

More usable terabytes per dollar, without giving up throughput.

Higher per-drive, per-JBOD and per-rack density behind a standard interface.
Low compute and memory overhead — each management node orchestrates more JBODs than conventional software, so CPU, RAM and server count per petabyte all fall.
Erasure coding you configure — from 6+2 for durability (25% overhead) to 60+4 for density (6.25%), and anything between.

Operations

A large disk estate without a storage team standing behind it.

No centralised state. All cluster configuration and drive metadata lives in the distributed database — nothing to manage on the server itself.
Add, move and remove drives across the cluster without pausing operations.

Tiering

Move data between the capacity and hot tiers at line rate.

Slurm-style jobs stage data in a prepare step; Octodec moves the working set from disk to flash in that window, to a guaranteed speed and SLA.
Your flash tier only holds what the job needs now — not the whole dataset.

Nothing above it has to change.

S3 NFS 9P NBD

No application changes Native S3 and standard NFS, with 9P and NBD available. Anything written against S3 runs untouched.

Real POSIX, not close enough Filesystem semantics are verified rather than approximated — no recompilation, driver work or custom integration layer.

Hardware you already own Commodity servers and standard JBOD enclosures, with SMR and CMR supported in the same cluster. Deploy as software on your own hardware, or as a pre-configured rack with published per-rack performance.

Adopt incrementally Put Octodec in front of the drives you have and add high-density SMR as supply allows. With no centralised state to migrate, a cluster stands up alongside an existing system and takes over drives as you move them.

Built for hundreds of petabytes that get read, not archived.

Neoclouds and training infrastructure

You run the compute that trains foundation models and are moving off all-flash for the capacity tier. It has to keep the GPUs fed, at economics your customers accept — and, increasingly, inside a jurisdiction they can name.

Enterprise AI and ML platform teams

You train and fine-tune in house. On-premises capacity removes cloud egress charges and gives you dedicated, predictable throughput instead of a shared multi-tenant tier.

Physical AI and sensor-driven operations

Fleets of vehicles, robots or production lines generating continuous video and telemetry that has to stay accessible and be re-read for training.

Genomics and life sciences

Sequencing pipelines producing petabytes a year, re-read for every reanalysis. Erasure coding carries multi-year retention cheaply and no egress charge lands on each pull-back.

Media and post-production

Petabytes of 4K, 8K and HDR footage that editors stream concurrently rather than pull from cold storage. Repatriation is making on-premises active archives the default again.

Teams running Ceph, or a capacity tier on flash

You are paying flash prices for bulk capacity, or carrying the compute, memory and operational overhead of a general-purpose distributed store. This is the conversation we most want to have.

First access

Not generally available yet.

Leave an email and we'll tell you when it is. If you'd rather talk before then, please share what you're running and someone here will connect with you ASAP.