Archil wants every machine to see all of your data

Cloud computing has lived with an awkward mismatch for twenty years.

The place where data accumulates is object storage. Amazon S3 and its imitators are cheap, effectively infinite, and durable enough that nobody thinks about losing a byte. So that is where everything ends up: training data, logs, genomics reads, media archives, model weights, backups of backups.

The way software reads data is something else entirely. Programs open files. They expect directories, paths, seeks, and appends, the POSIX interface that has defined "a filesystem" since before the web existed. Databases assume it. ML frameworks assume it. Nearly every tool ever written assumes it.

Between the bucket and the file sits a gap that the industry has papered over with copies. Sync scripts that pull objects down before a job runs. Provisioned block volumes that hold a duplicate of data that already lives in S3. Fragile FUSE adapters that pretend a bucket is a disk and fall over the moment you ask for real filesystem behavior. Whole data engineering pipelines exist mainly to move bytes from where they are stored to where they can be read.

Archil closes that gap directly. Point it at an object storage bucket and the bucket shows up on your machines as a real local disk: a POSIX directory tree you can list, open, and write to, backed by a shared SSD cache that makes it fast. The company was founded by Hunter Leath, who spent eight years at AWS building Elastic File System, Amazon's own managed filesystem. In April 2026 it raised an $11 million Series A led by Standard Capital, bringing its total funding to $18 million, on the strength of a sharper thesis: AI agents want to work in files, and Archil intends to be the filesystem they work in.

The tax everyone pays without naming it

Look at how a typical AI or analytics team actually handles storage and you find a quiet, expensive workaround economy.

The canonical dataset lives in S3 because that is the only place it fits. But the GPU instances that train on it cannot read S3 like a disk, so before a job starts, someone copies a slice of the data onto attached block storage. That volume has to be provisioned at a fixed size, paid for whether it is full or not, resized when it runs out, and cleaned up when the job ends. Multiply by every instance in the fleet, then add the scripts that keep all those copies roughly in sync with the source of truth.

None of this is anyone's product. It is plumbing that every team rebuilds, and it fails in the boring ways plumbing fails: a job dies at hour six because a volume filled up, two instances see different versions of the same dataset, a forgotten disk bills for months.

Archil's pitch lands on exactly this pain. Its volumes are infinite, so there is nothing to provision or resize. They are shareable, so a hundred instances can mount the same data at once and see the same thing. And they connect straight to the bucket, so the copy step disappears. The company claims the result is about 30x faster than reading S3 directly and roughly 90 percent cheaper than managing your own fleet of block volumes, since you stop paying for idle provisioned capacity and start paying for what you actually use.

A filesystem, not a trick

The reason FUSE adapters earned their bad reputation is that mapping object semantics onto file semantics is genuinely hard. Objects are immutable blobs; files get appended to, renamed, and locked. A thin shim that translates one API into the other breaks precisely when applications exercise the parts of POSIX that make a filesystem useful.

Archil is built as an actual storage system in the middle. Reads pull from a high-performance SSD caching layer that sits between your instances and the bucket, so hot data behaves like local flash rather than a network round trip to S3. Writes land in the cache and flow back to the bucket, which remains the durable source of truth in your own account. The data stays yours, in standard formats, in your bucket; Archil makes it usable at machine speed.

The team also made an unusually opinionated protocol choice. Rather than serving volumes over NFS, the default answer for shared filesystems since the 1980s, Archil built its own protocol in the style of AFS, the Andrew File System, which was designed around aggressive client-side caching. Leath is one of a small number of people entitled to that opinion: after eight years building and operating EFS, which is an NFS service at enormous scale, he knows exactly where the protocol runs out of road.

And Archil is deliberately promiscuous about what it mounts. S3 is the headline, but it also fronts Google Cloud Storage, Cloudflare R2, Azure Blob, MinIO, DigitalOcean Spaces, Wasabi, and Backblaze B2. Whatever bucket your data is in, it can show up as a directory.

From launch to rename to thesis

The company entered Y Combinator's Fall 2024 batch as Regatta Storage and launched with a straightforward framing: turn S3 into an infinite local filesystem. A $6.7 million seed led by Felicis followed in June 2025. Somewhere along the way the company renamed itself Archil and made a subtler shift, from selling a faster filesystem to selling the data layer for AI.

You can see the progression in its own language. The early product line was about replacing EBS volumes with something infinite and shareable. The current homepage says "real data for every agent." Those are the same technology, but they are aimed at different decades.

Then AI made the gap urgent

For most of those twenty years, the bucket-to-file gap was an annoyance teams absorbed. AI turned it into a tax on the most expensive hardware in the building.

Training is the obvious case. A GPU cluster billing thousands of dollars an hour does nothing useful while data loads. Datasets outgrew local disks long ago, and the choreography of staging shards onto instances is a large share of what ML infrastructure engineers actually do all day. Inference has its own version: model weights are tens or hundreds of gigabytes, and every cold start means pulling them from storage before the first token comes back. Between those sit checkpoints, embeddings, and evaluation data, all shuttling between object storage and machines that need them as files. Archil's unified cache over S3 aims at exactly these workloads, along with adjacent heavy readers like genomics and analytics.

Agents are the newer and more interesting case, and they are the center of Archil's current thesis.

Files are the interface agents already speak

Ask what interface an AI agent should use to reach a company's data and the reflexive answer is an API. Build endpoints, write tool definitions, wire up a connector per source. Archil's answer is older and simpler: give the agent a filesystem.

The argument has real teeth. Language models learned to work from billions of lines of code and documentation that read files, walk directories, and pipe things into grep. A hierarchical tree of folders is a data interface the model already understands natively. Hand an agent a mounted directory and it can browse, search, and process what it finds without anyone writing a custom tool definition for each data source. Every API integration you did not have to build is schema you did not have to maintain.

Archil has pushed the idea one step further with serverless execution: a filesystem you can send bash commands to, which runs them next to the data and returns the results. Instead of an agent pulling gigabytes across the network to look at them, the compute goes to where the files already are. That reframes the filesystem as the surface agents work on, which is the promise inside the homepage line about real data for every agent.

If agents keep taking on more real work, some layer will end up being the standard way they touch enterprise data. Archil is betting that layer looks like a directory tree, because that is what the models grew up on.

The round, and what it is for

The Series A closed in April 2026: $11 million led by Standard Capital, with participation from Y Combinator, Felicis, Peak XV Partners, and Wayfinder Ventures, less than a year after the seed. Total funding stands at $18 million. The company is small, around eleven people in San Francisco, which is roughly the size of the team you would expect to be rebuilding a filesystem protocol rather than integrating one.

Adoption friction is low by design. Archil is pay as you go, with a free developer tier, and trying it amounts to mounting a volume on a machine you already have, against a bucket you already own. Storage infrastructure usually spreads exactly this way, one mounted volume at a time, until one day it is load-bearing.

Storage is the least glamorous layer of the stack, and the most durable. Whichever models win and whichever agent frameworks come and go, the data sits in buckets, and something has to make it readable the way software actually reads. Archil built that something, and the founder spent eight years at AWS learning precisely which parts of the old answer to throw away.