MinIO

At a data center somewhere outside Frankfurt, a rack of GPUs is busy fine-tuning a model nobody outside the building has seen. The training data - a few petabytes of it - is sitting one cable hop away on commodity drives, served by a piece of software you can install with one command. The software is MinIO. The company shipping it is also, somewhat confusingly, MinIO. And for the last decade it has been the answer to a question most enterprises did not realize they were asking: what if the storage layer of the cloud was not actually the cloud?

MinIO is a Redwood City company that builds high-performance, S3-compatible object storage. That sentence does a lot of work. Translated for humans: it makes the kind of bottomless bucket Amazon invented, except it runs on hardware you own, in a datacenter you control, on an operating system of your choosing. It is open source at the root, commercial at the top, and used today by more than half of the Fortune 500. The product has a 61,000-star GitHub repo and a $1 billion valuation. Most storage companies would settle for one of those.

A bill, a bucket, and a billion-dollar idea.

The 2010s sold a tidy story: move everything to the cloud and the storage problem goes away. It was a beautiful story right up until the invoice arrived. Egress fees. API surcharges. A pricing model that punished the data-heavy and rewarded the data-light. Meanwhile, the actual interface that defined cloud storage - the Amazon S3 API - had quietly become a standard. A grammar everyone understood, owned by exactly one vendor.

MinIO's founders looked at that situation and made a slightly heretical bet. What if the API was the moat, and not the building it sat in? What if you could keep the S3 grammar and replace the S3 landlord? It was, depending on your point of view, a very clever idea or a very obvious one. The problem with obvious ideas is that almost nobody actually ships them.

Three engineers, one Go binary, zero patience.

Anand Babu Periasamy - everyone calls him AB - had already done storage once. He co-founded Gluster, the distributed filesystem company Red Hat acquired in 2011. He knew what he did not want to build again. He did not want a kernel module. He did not want a sales motion that started with a six-month proof of concept. He did not want enterprise storage as the world had defined it for thirty years.

So in 2014 he started MinIO with Garima Kapoor and Harshavardhana. The product was written in Go, packaged as a single static binary, and licensed under AGPLv3. You could download it, run it, and have an S3 endpoint listening on port 9000 in under a minute. No SAN. No clustered filesystem with its own bestiary of acronyms. Just a process and some disks.

It was, in retrospect, the right product at the right moment. Kubernetes was arriving. CI/CD pipelines needed somewhere to dump artifacts. ML teams needed somewhere to stage training data. Every one of those workflows had been written against S3. MinIO showed up with a drop-in replacement and a permissive enough license that nobody had to ask for permission.

From a binary to a product line.

The MinIO of 2026 is not a single thing anymore. There is the commercial flagship, MinIO AIStor, purpose-built for AI workloads with native PyTorch, Iceberg and Hugging Face support. There is AIHub, which lets a regulated bank or a defense contractor run a Hugging Face-compatible model registry inside an air-gapped network. There is AIStor Pods, racked appliances that aim to deliver hyperscaler economics in your own datacenter. And underneath it all, still, is the open source server that started everything.

The arrival of S3 Express API support in AIStor is the kind of detail that sounds boring until you think about it. Amazon released Express to give its own customers low-latency buckets for AI training. MinIO became the first non-AWS vendor to implement that interface - meaning a developer can write to the same API surface and run the workload anywhere. The lock-in story gets a little less compelling every time someone ships a thing like that.

Numbers that do the talking.

The customer list, where MinIO has been willing to share it, is the other half of the argument. Intel runs on it. Nomura runs on it. So does a long parade of banks, telcos, manufacturers and the kind of three-letter agencies that prefer not to be named in marketing copy. Partnerships with NVIDIA, F5, Databricks, Snowflake and Starburst slot MinIO into more or less every reference architecture for modern AI infrastructure that ships with a logo wall.

Who picks up the phone

Bring the hyperscalers home.

If you want the mission in one sentence, it is this: software-defined object storage that runs anywhere, with the API the industry already agreed on. The pitch is unromantic and unusually durable. Most infrastructure startups die when the platform underneath them shifts. MinIO is, in a sense, the platform underneath - or at least the thin layer that turns commodity drives into something an AI team can use without filing a ticket.

That stance has costs. The community edition's recent move into maintenance mode rattled the open source crowd, and at least two forks have already appeared. License debates are part of MinIO's permanent weather. But the commercial proposition - that the same enterprises spending nine figures on GPUs need a storage layer that does not double their cloud bill - keeps growing, because the underlying math keeps getting worse for the hyperscaler-only path.

The data layer is now load-bearing.

A decade ago, storage was the part of the stack the buyer wanted to think about least. Today, training runs idle when the data does not arrive fast enough. Inference clusters get throttled by their object store, not their GPUs. The economics of every AI program at the Fortune 500 are now partly a storage question - and the storage question, increasingly, is which S3-compatible system you trust to sit between your drives and your models.

MinIO is one of maybe three or four serious answers. It is the answer with the largest developer footprint, the simplest install story, and a customer base that already runs the world's mission-critical data. That is not a moat in the classical sense. It is something stranger and more useful: a default.

Back at the datacenter outside Frankfurt, the training run finishes around dawn. The model gets versioned, the dataset gets a new tag, the operator nods at a dashboard and goes home. Nobody on the team thinks much about the storage layer. That is the point. MinIO has spent a decade making the boring layer reliable enough to ignore. In an industry that confuses noise with progress, that is a quietly subversive thing to do.

Where to find them

Watch & listen