Scaleway, Part One: The Floor Beneath the Floor

The Bucket You Already Pay For

Somewhere in a Paris data center sits a Scaleway account that already has work on it. Object storage holding training datasets for FLUX LoRA experiments, mirrored across to RunPod when training kicks off. A virtual private server at fifty-one dot fifteen dot one-ninety-four dot sixty-five carrying the public face of a handful of projects. And in the corner of that account, the faint footprint of an old experiment with quantum computing, ten minutes spent there a while ago, never returned to.

This three-part series is a guided tour of what's in the rest of that account. Not the marketing version. The version where every service gets the question, does this fit a one-person consulting and publishing operation in Kall, or is this a service waiting for a different kind of customer.

The brief covers everything Scaleway sells. Storage, compute, networking, containers, databases, serverless, data warehousing, artificial intelligence services, security tooling, monitoring, the quantum platform that has grown considerably since it was last tried, and the genuinely strange bits at the edge of the catalog. The catalog is large. The relevant slice is smaller. The point is to know what's there, because every cloud bill that runs away does so because someone forgot a managed service existed and rebuilt it on a virtual machine.

This episode is the foundation layer. Storage and compute. Boring on the surface, but the layer everything else sits on, and the layer where a clear picture saves the most money.

A Quick Word On What Scaleway Actually Is

Scaleway is the cloud arm of Iliad, the French telecom group, with three primary data center regions in Paris, Amsterdam, and Warsaw. As of this year there's a new presence in Italy as well. The company has been quietly insistent about its independence story while everyone else got loud about it. The whole stack runs from European hardware in European buildings under European law, which matters less when your project is a hobby and more when your project handles real customer data under the General Data Protection Regulation.

The other thing worth noticing is that Scaleway publishes pricing on every page. The default of European cloud providers is to make pricing readable. The default of American hyperscalers is to make pricing negotiable. Anyone who has spent an afternoon trying to predict an Amazon Web Services bill knows which approach reduces anxiety. Scaleway prints the per-hour cost on the same page where you click to provision the thing.

That sets up the rest of this tour. The services described here all have visible prices. The cost of being wrong about your initial choice is small. The cost of trying things is small. Most of what makes Scaleway feel different from the bigger clouds is not technical, it's behavioral. They expect you to experiment.

The heritage is worth a beat too. Iliad has been in the European telecom and hosting business since the nineteen nineties. The bare metal product line that became Scaleway's Dedibox brand was running dedicated servers under the Online dot net name long before cloud was a category. That heritage shows in how the company thinks about hardware. The bare metal options are not an afterthought. They are part of the inheritance.

The Storage Tier Pär Hasn't Reached For Yet

Object storage is the workhorse on the existing Scaleway account. S three compatible, multi-zone redundant, holding training datasets for image generation work. That's the standard piece.

The one that isn't being used is Glacier. Scaleway runs a cold storage class at a small fraction of the cost of standard object storage. The trade is retrieval time. Files come back in minutes to hours rather than instantly. For hot working data this is useless. For everything else, this is probably the most underused service in any cloud catalog.

Think about LifeLab. Photographs dating back to two thousand and two. Face recognition runs through InsightFace. Optical character recognition pipelines over scanned papers. Spatial enrichment from DuckDB pulling geographic coordinates onto images. The active working set is maybe a few hundred gigabytes. The full historical archive is several terabytes. The bit that gets touched this month is a sliver. The bit that gets touched this year is bigger. The bit that might be revisited five years from now is almost everything.

That is exactly the shape Glacier is built for. The original raw training photos behind the FLUX LoRA experiments, archived once you have the captioned dataset that actually matters. The unedited recordings from Årebladet interviews that already ran in print, kept because you might want them later but not because you'll touch them this month. The source files from the McDonald's commercial shoot and the Klarna spots and the ABB modeling work. The version of your archive from last year, sitting at maybe ten percent of the cost of keeping it warm. Glacier is where past selves live cheaply.

Block Storage is the next layer up. This is the disk attached to a virtual machine. When you spin up a Scaleway instance and it has a hundred gigabytes of working storage, that storage is block storage. You can grow it without rebuilding the machine. You can take snapshots, attach extra volumes for specific workloads, detach a volume from one machine and attach it to a different one. The boring part of cloud infrastructure that nobody talks about until they run out of it. Worth knowing how it works.

File Storage is in beta. This is the option for shared filesystems across multiple machines. Think network attached storage, fully managed. If Pärkit ever grew from one Postgres on one virtual machine to multiple application servers all needing the same file paths, this would be the bridge. Right now it's not needed, but beta services often come with generous early-adopter pricing, and beta is sometimes the right time to plant a flag.

The fourth piece in the storage corner is Container Registry. Every container image built for Pärception tooling, for the Baren project, for the article pipeline, has to live somewhere. Right now those images probably live on Docker Hub or get rebuilt from scratch every deploy. A Scaleway container registry costs almost nothing per month, lives in the same physical building as the machines that pull from it, and avoids the rate limits that Docker Hub has started enforcing on anonymous pulls. Small change. Real annoyance reduction.

The Compute Family

Now compute. The thing called a virtual private server in casual speech is actually a whole family of instances at Scaleway, and the existing setup uses one particular shape of it.

General Purpose instances are the default. Balanced central processing unit and memory, designed for the workload that doesn't have a strong opinion about which resource it needs most. The current virtual machine running Pärception services is almost certainly one of these. If you don't know what shape your workload wants, you start here and migrate later.

Development Instances are smaller and cheaper, intended for staging environments and testing. Same underlying architecture, less raw compute. If Pärception ever needs a separate environment for trying changes without risking the public face of Årebladet or PärCel or any of the shop projects, this is the affordable way to do it. A few euros a month buys a place where things can break without breaking production.

Compute Optimized Instances are for workloads that hit the central processing unit hard. Video encoding. Batch processing. The kind of work where the bottleneck is making the processor cycle faster rather than fitting more data into memory or reading from disk quickly. If the Whisper transcription pipeline currently running on local Mac ever needed to be hosted in the cloud, this is the shape it would request.

Memory Optimized Instances are for the opposite problem. Workloads where the bottleneck is keeping a large dataset in working memory. Large in-memory databases. Certain analytical queries that hate spilling to disk. Pärkit running heavy joins across BMW telemetry, Apple Health data, and zone transitions might eventually want this kind of machine. Not today. But worth keeping in the back of the mind for the day Pärkit grows to the point where it cares about query times.

The thing that's quietly important about Scaleway's instance family is that the underlying virtualization is the same across all of them. You can resize a machine from general purpose to memory optimized without rebuilding from scratch. You start small. You watch the actual usage. You move to the right shape based on what you observe rather than what you guess. For a one-person operation, this is exactly the workflow that fits. The cost of being wrong about your first guess is a few minutes of downtime, not a weekend of rebuilding from images.

Snapshots are worth a mention here. Every Scaleway instance can be snapshotted, which is to say the current state of the disk is frozen and stored as an image. From a snapshot you can create a new instance, identical to the one you snapshotted, in a few minutes. For someone running an rsync deploy workflow into a Scaleway virtual machine, this means the disaster recovery path is fast. The machine is gone. You spin up a new one from yesterday's snapshot. You rsync the most recent changes. You're back. The whole process is under fifteen minutes.

The Bare Metal Branch

Now things get more interesting.

Below the virtualization layer sits the actual hardware. Most cloud providers hide this completely. You get virtual machines, and that's it. Scaleway, partly because of the Iliad heritage, gives you the option to rent the physical machine directly. No virtualization tax. No noisy neighbor risk. Just a server, given to you, with full access. The whole physical machine, yours for the rental period.

There are two product lines that do this. Elastic Metal and Dedibox.

Elastic Metal is the modern bare metal product. You provision a physical server like you provision a virtual one, with application programming interface calls, automated installation, and integration with the rest of the Scaleway control plane. The naming convention is periodic table elements. Aluminium is the entry tier, dedicated hardware at the cheapest possible price point for the brand. Beryllium is the balanced workhorse for production-grade work. Iridium is the heavy hitter for serious applications. Lithium is the storage-dense option for keeping a lot of local data fast. Titanium is the bare metal with dedicated graphics processing units for serious machine learning.

Then there's RISC five. This one deserves a pause.

Scaleway runs what they describe as the first commercial RISC five servers in any cloud. RISC five is an open instruction set architecture. Not Intel, not AMD, not ARM. A new processor family designed from the ground up to be free of licensing fees and to allow anyone to build their own variant. For most workloads in two thousand and twenty-six this is currently a curiosity rather than a production answer. The software environment is thinner. The performance per dollar is not where mature architectures sit.

But for a podcast producer who likes documenting unusual paths, this is interesting. The first PärPod episode rendered on a RISC five server is presumably still available for whoever wants to claim it. The first character LoRA trained on RISC five hardware would be a footnote with some claim to permanence. The cost of doing one weekend experiment on a RISC five box is small. The bragging rights are non-zero. And the documentation that comes out of doing something like that has its own life. Tutorials on running unusual workloads on unusual hardware tend to age well, because the next person who wants to do the same thing finds them through search and is grateful.

Dedibox is the older bare metal product. It traces back to before Scaleway was called Scaleway, when the parent company Iliad was running dedicated servers under the Online dot net brand. The naming there is functional rather than periodic. Start, Pro, Core, Store, GPU, and Dedirack for housing your own physical hardware in their data centers. There's also Dedibox VPS, which is a small virtualized presence available in sixty locations worldwide starting at under five euros a month.

The Dedibox brand reaches further geographically than the rest of Scaleway. If Pärception ever wanted edge nodes outside the central European triangle, Dedibox VPS is the cheapest way to plant a flag in São Paulo or Singapore or wherever. Not relevant for Årebladet or PärCel today. Quite relevant the day Napkincast launches and starts catering to students who are not all in Sweden.

If Pärception ever needed an actual physical box dedicated to one customer's workload, with no virtualization layer between the work and the iron, this is the family to look at. The pricing is competitive with Hetzner, which is the European bare metal benchmark, and the integration with the rest of Scaleway is tighter. Hetzner gives you a server. Scaleway gives you a server that already speaks the same application programming interface as the rest of your infrastructure. Small difference. Big in the day-to-day.

Mac Minis In A Data Center

The third branch of bare metal is, in a sense, the strangest. Scaleway rents Mac minis. By the hour. With full native macOS, full Apple Silicon, accessible over the network from anywhere with an internet connection.

The lineup spans the current Apple Silicon generation. The M one mini sits at the cheap end for simple experiments. The M two and M two Pro are general-purpose Mac development boxes. The M four is current-generation performance. The M four Pro is sold explicitly as built for artificial intelligence workflows and complex media work.

Why does this product exist. There is exactly one reason. Apple's software licensing requires Apple hardware. If you need to build iOS applications, run continuous integration for Mac software, render Final Cut projects, or do anything that requires macOS as a host, you historically had to own a Mac. The cloud version of that need was unmet for years. Now several providers, Scaleway among them, rent Apple Silicon by the hour.

For Pär specifically, this opens up something specific. The PärImage pipeline currently runs FLUX LoRA training on RunPod with NVIDIA hardware. That's correct. Training is NVIDIA's territory. But Mac Silicon is genuinely excellent for inference. The MLX framework experimented with locally, running Gemma four variants, is built specifically for Apple Silicon. An M four Pro mini in a data center is essentially an MLX inference server with proper bandwidth and proper uptime. You can run model captioning on a queue of LifeLab photos at meaningful throughput without keeping the local MacBook spinning all night.

The PärPod render pipeline is another candidate. Text-to-speech rendering through Inworld is mostly bound by network and orchestration, but post-processing the audio, splicing segments, generating effects, can be parallelized across multiple machines. If a session ever needed to render fifty podcast episodes overnight, a Mac mini cluster spun up for that purpose and torn down in the morning would do the job. Hourly billing. No commitment. The bill arrives the next morning along with a coffee, and the work is done.

The economics are interesting. The cheapest Mac mini in the fleet is meaningfully more expensive per hour than the cheapest Linux instance, but it's also a complete Mac with full local solid-state storage and Apple Silicon performance. The workload pattern that fits is the one you batch up rather than the one you run continuously. Not daily. Quarterly. The day you've accumulated three months of audio production tasks. Spin up four Mac minis for an evening. Run the queue. Tear it down.

There's a deeper engineering joy to it. The fact that you can ask a public application programming interface for a Mac, get one within minutes, push code to it, run a build, get the artifact back, and release the machine, all in a single afternoon, is a small daily miracle that the world has gotten used to. Worth not getting used to.

Multi-Zone Redundancy And Why It Matters

A short detour on something that runs through everything we've covered.

Most Scaleway services in a single region are spread across multiple availability zones. An availability zone is, in physical terms, a separate building with separate power, separate cooling, separate network connectivity. The idea is that one zone can fail completely, and the services that were spread across multiple zones keep working.

For object storage this is the default. The bucket that holds the FLUX training data is automatically replicated across the zones in Paris. If one zone goes dark, the data is still there in the others.

For managed databases, the option exists. You can run a primary in one zone with a synchronous replica in another. The day the primary's building loses power, the replica takes over within seconds.

For virtual machines, you choose. The simple version is one machine in one zone, which is what most one-person operations run. The redundant version is two machines in two different zones with a load balancer in front of them. Same application, different buildings. If one zone fails, the load balancer routes around it.

For Pärception today, the multi-zone story is mostly academic. One virtual machine in one zone is fine. The reason to mention it is that the day a particular service genuinely matters, the upgrade to multi-zone is straightforward. You add a second machine. You add a load balancer. The application code does not change. The cost roughly doubles, and the chance of an actual outage drops by an order of magnitude.

The Pricing Surface Behind All Of This

A note before moving on, because the pricing model shapes what kinds of experimentation are viable.

The virtual instances are billed by the hour. No setup fees. No termination fees. No minimum commitments. The dedicated servers are billed monthly. The object storage is billed by gigabyte stored plus egress. Egress, the cost of data leaving the cloud, is famously the place where Amazon Web Services and Google Cloud bills surprise people. Scaleway includes generous egress in many products, and where they charge for it, the rates are competitive with European norms rather than American hyperscaler norms.

What this means practically is this. Running an experiment on Scaleway for a weekend costs almost nothing. If you spin up a Beryllium bare metal server on Friday evening, run a training job through the weekend, and tear it down Sunday night, you have probably spent ten or fifteen euros. The same experiment on Amazon Web Services, depending on the egress involved, can be three to five times that.

For someone running a one-person consulting operation, this matters more than the hyperscaler benchmarks suggest. The cost of trying something new should be cheap enough that the decision is whether you have an idea worth testing, not whether the test is worth the cloud bill. Scaleway sits comfortably in that zone. The decision is yours. The bill is small.

There's also the internet protocol version six story. Every Scaleway resource gets an IPv six address by default, in addition to the IPv four address. This is a small thing that turns into a real thing when you realize that IPv four addresses are a scarce resource that costs extra on most clouds, and IPv six is approximately free. Building anything new on IPv six first is the correct approach, and Scaleway makes it easy.

What This Episode Was For

This was the floor. Storage and compute and the physical hardware below them. The unglamorous layer that everything else stands on.

Object storage with a cold tier called Glacier. Block storage for instance disks. File storage in beta for shared filesystems. Container registry for images. Four flavors of virtual machine plus snapshots and multi-zone redundancy. Three branches of bare metal, including the curiosity of RISC five, the geographic reach of Dedibox, and the Mac minis that exist because Apple's licensing made them necessary.

Next episode, the plumbing. Virtual private clouds and load balancers and edge services. Kubernetes. The three flavors of serverless. The managed databases, including the brand new Data Warehouse for ClickHouse that just went general availability this quarter. Apache Kafka, Apache Spark, the message queues, the topics. The stuff that turns a pile of servers into a working system.

After that, the wild end. The generative artificial intelligence services with models hosted entirely in Europe. The graphics processing unit instances and the custom-built clusters that train foundation models. The security primitives. The Internet of Things hub. Cost management. Environmental footprint. And a return visit to quantum, which has expanded considerably since the brief experiment ten minutes ago.

What's there is the Scaleway map. Two episodes left.