Storage engineers who got tired of telling AI teams to just buy more SSDs.
Founded in Tallinn in 2023. Building the filesystem the AI training pipeline should have had from the start.
The wrong answer keeps winning because the right answer is hard to build.
Alexander and Raimo spent years working on storage infrastructure for compute-intensive workloads. Every time a new AI project started, the question was the same: where do we put 200 terabytes of training data? The answer was always "buy more NVMe" or "use object storage with a custom data loader". Neither answer was good.
NVMe is expensive and much of its cost pays for write characteristics that AI training never uses. Object storage requires a data loader rewrite and charges egress fees on every training epoch. Both answers were accepted as necessary costs of doing AI at scale.
SMR hard drives offer 2.5x the capacity per dollar of NVMe and 60% lower per-TB cost than managed object storage, but every attempt to use them for AI training ran into the same problem: the filesystem. Standard Linux filesystems are designed for random writes. SMR drives require sequential writes. When you combine the two, the drives run at 20% of their rated speed and everyone concludes SMR is too slow for training workloads.
It is not too slow. The filesystem is wrong.
Leil Storage is the answer: a distributed filesystem that routes every write through a band-aligned write scheduler, eliminating the read-modify-write penalty that makes SMR look unusable. The drives run at their rated sequential speed. The training code sees a standard POSIX mount. The cost per terabyte drops by 60%.
We started development in late 2023, ran our first internal benchmark in early 2025, and began working with pilot partners in mid-2025. We are registered in Estonia, operating in the EU, and angel-backed since 2025.
Two engineers who build it, and do the sales calls.
Years building high-throughput data infrastructure and distributed storage systems for compute-intensive workloads across research and industry contexts. Co-founded Leil to solve the SMR write alignment problem that kept teams from using cheap high-density drives at AI training scale.
Filesystem internals, kernel-level IO scheduling, and magnetic storage engineering. Years working at the block device and VFS layer in Linux. Wrote Leil's SMR write scheduler from the kernel allocation path down. The person who actually understands why generic filesystems kill SMR drive performance.
We are hiring for a third engineer with kernel storage experience. If you have worked on VFS, block layer code, or SMR drives in Linux, send a note to [email protected].
The things we will not compromise on.
No hidden costs
No egress fees. No per-request charges. No pricing tiers that penalize you for reading your own data. We sell storage capacity; using that capacity is the whole point.
Written benchmarks
We do not publish unattributed "up to X GB/s" numbers. All performance claims include the cluster configuration, drive model, workload parameters, and test date. If you cannot reproduce it, we want to know why.
POSIX or nothing
We will not build a proprietary SDK. If your training code cannot use a POSIX path, we are the wrong product. We would rather lose the customer than compromise the interface contract.
EU data residency
Leil Storage OU is incorporated in Estonia. Our pilot clusters are in EU data centers. Your data stays in the EU unless you explicitly configure cross-region replication to a non-EU location.
Tallinn, Estonia
We work from Tallinn, a compact city with deep engineering talent and a government that treats software companies as real businesses. Tallinn has reliable low-cost colocation facilities connected to both Western European and Nordic network backbones, which makes it a practical location for running pilot storage clusters without prohibitive power costs.
Leil Storage OU5 Tornimae
Tallinn 10145
Estonia
[email protected]