The Cloud World, Part One: The Big Three and the Resistance

Opening

Last time we walked through Scaleway top to bottom. The full catalog, every service from storage to quantum, with Pär's existing setup as the anchor. That was about going deep on one provider Pär already uses. This three-part series is the opposite shape. We zoom out to the entire cloud world, looking for interesting things to know about rather than necessarily things to use. Pär already has Modal for serverless graphical processing unit work, has fal dot eye eye for image and video model endpoints, has RunPod for persistent pods with network volumes, and has confirmed that ThinkDiffusion is the wrong architecture for the kind of training and inference he actually runs. So the goal here is to map everything else. What does the rest of the territory look like, who plays in it, and what is worth pointing at even if Pär never spins up an account.

Part one is the general-purpose cloud world. Hyperscalers, tier-two alternatives, deployment platforms, and storage specialists. Part two is the new generation of pure-play artificial intelligence infrastructure providers, the so-called neoclouds, and the specialty silicon companies that were supposed to break NVIDIA's grip but have instead been partially absorbed by it. Part three is where things get strange. Edge databases, the full Cloudflare developer surface, quantum hardware across multiple vendors, decentralized compute marketplaces, and the European sovereign cloud build-out that is actually happening right now.

The shape of the territory

The general-purpose cloud market in twenty twenty six is shaped like a pyramid with three layers that look completely different from each other. At the top sit the hyperscalers. Amazon Web Services, Microsoft Azure, and Google Cloud Platform together control somewhere between sixty and seventy percent of all global cloud spending depending on which analyst you ask. They sell everything to everyone. Compute, storage, networking, databases, machine learning, security, observability, more than two hundred services each in catalogs so deep that learning any one of them properly is a multi-year career commitment. They are the safe choice for any enterprise that wants a single vendor relationship, and they are also the most expensive choice for almost every individual workload they offer.

Below the hyperscalers sits the tier-two layer. These are general-purpose cloud providers that exist because the hyperscalers are too expensive, too complex, too American, or all three. Oracle Cloud. International Business Machines Cloud. DigitalOcean. Linode under Akamai. Vultr. Hetzner. OVHcloud. Scaleway. Cloudflare, depending on how you classify it. Some of these are big general purpose stacks in their own right. Oracle Cloud Infrastructure has the deepest enterprise database integration in the world and the most generous free tier of any provider, including ten terabytes of monthly egress at zero cost. Others are deliberately minimal. Hetzner sells you a virtual server, a load balancer, an object bucket, and a small handful of other things. That is it. The pitch is the price, not the catalog.

Below the tier-two sits the developer platform layer. These are not really clouds in the traditional sense. They sit on top of someone else's cloud and sell a much smaller, opinionated product. Push code, get a deployed application. Heroku invented this category and then killed its own free tier in twenty twenty two, scattering hundreds of thousands of indie hackers across a new generation of platforms. Railway, Render, Fly dot eye oh, Vercel, Netlify, and a long tail of newer entrants are the survivors. They each have a clear philosophy about what deployment should feel like, and they each have a niche of users who would not migrate to anything else.

That is the territory. Let us walk through it.

The hyperscalers, briefly

We are not going to spend much time on Amazon Web Services, Azure, and Google Cloud, because Pär already knows what they are. The interesting question is which slices of them are worth knowing about even if you have no intention of using the rest. Amazon's standout differentiators in twenty twenty six are mostly in the corners. Trainium and Inferentia, Amazon's own machine learning accelerators, are quietly becoming cost-competitive for inference workloads. The current generation of Inferentia claims roughly seventy percent lower cost than equivalent NVIDIA H one hundred capacity for workloads that fit its architecture. The catch is that you have to be inside the Amazon stack already, and you have to use the Neuron compiler toolchain, which is its own learning curve. Most teams that benefit from this are already running on Amazon for other reasons.

Microsoft Azure's most interesting corners are the partnerships with O p e n A I and the quantum surface. The O p e n A I relationship gives Azure customers access to the full Generative Pre-trained Transformer family through Azure-hosted endpoints with enterprise data residency guarantees, which is the main reason any large company uses Azure for artificial intelligence work. Azure Quantum, which we will cover in part three, hosts the broadest catalog of quantum hardware modalities of any cloud, including IonQ trapped ions, Quantinuum H-series, Rigetti superconducting, Pasqal neutral atoms, Atom Computing, and Quantum Computing Incorporated photonic systems, all available through a single Q sharp programming surface.

Google Cloud Platform's interesting corners are its custom silicon and its data products. The Tensor Processing Units are now in their seventh generation with the Ironwood release benchmarking at four thousand six hundred fourteen tera floating point operations per second per chip, which puts them roughly in line with NVIDIA's Blackwell generation on training throughput at significantly lower published cost. BigQuery remains the standout serverless data warehouse for anyone working with analytics at terabyte and petabyte scale, and Vertex A I has become the most polished managed machine learning platform of the three hyperscalers, ahead of SageMaker on developer experience. None of this is news, but the point is that even if you never touch Google Cloud for general compute, those specific corners are worth knowing exist.

For Pär specifically, the hyperscalers are mostly negative space. They define the price ceiling that every other provider competes against. Amazon Simple Storage Service charges nine cents per gigabyte of egress. Google Cloud Storage charges twelve cents. Anyone undercutting that, which is essentially every alternative storage provider on earth, gets to use the gap as their entire marketing pitch.

Cloudflare, the company that became a cloud by accident

Cloudflare is the most interesting general-purpose provider to discuss because it is not really a cloud provider in the way the others are. It started as a content delivery network and DDoS protection service, ran fiber to three hundred thirty five cities around the world to serve that core business, and then realized that having compute capacity within fifty milliseconds of ninety five percent of the world's internet users was a much bigger asset than just caching web pages. So Cloudflare turned that infrastructure into a developer platform, and the catalog they have built on top of it is genuinely strange.

The compute layer is called Workers. Workers run JavaScript, TypeScript, Python, or Rust in a sandboxed isolate-based runtime that achieves sub-millisecond cold starts. That is not a typo. Cold starts on Workers are faster than warm function invocations on most competing serverless platforms. The trick is that Workers are not lambda functions, they are V eight isolates, the same primitive that runs separate browser tabs inside the same Chrome process. Cloudflare's runtime is wildly different from anyone else's, which is both why it is fast and why some applications cannot be ported to it without rewrites.

Around Workers sits a constellation of storage primitives that do not quite map to anything elsewhere. R two is object storage that is S three compatible but has zero egress fees, which has made it a quiet favorite for serving large media files, machine learning weights, and analytics datasets. K V is key-value storage with eventual consistency, fast reads anywhere, slower writes. Durable Objects are stateful actors with strong consistency and global uniqueness. You can have exactly one instance of a given Durable Object class with a given identifier running anywhere in the world at any time, which makes them perfect for things like multiplayer game lobbies, collaborative editing, or live chat rooms. D one is serverless Structured Query Language built on SQLite with read replication across regions. Hyperdrive is a connection pooler for existing Postgres or MySQL databases that makes external databases feel local to Workers, with sub five millisecond latency on cached queries.

Then there is the artificial intelligence layer Cloudflare has bolted on. Workers A I exposes a catalog of open weight models running on Cloudflare's own graphical processing unit fleet, including text generation, image generation, embeddings, and speech to text, all priced per request rather than per second of compute time. Vectorize is a vector database for semantic search and retrieval augmented generation workloads. A I Gateway sits in front of any large language model provider and adds caching, rate limiting, observability, and fallback routing. A I Search is Cloudflare's managed retrieval augmented generation pipeline, built on top of Vectorize and Workers. The Agents Software Development Kit lets you build stateful artificial intelligence agents on top of Durable Objects with built-in inference, all running on the same network.

Cloudflare also has Browser Run, formerly called the Browser Rendering A P I, which gives you programmatic serverless browser instances for scraping and automation. Pipelines is a streaming ingestion service that batches and writes records to R two. Analytics Engine is an internal time-series and metrics database that Cloudflare uses to power its own product dashboards and exposes to customers for high cardinality event tracking. Containers, which launched into general availability last year, lets you run actual Docker containers alongside Workers for workloads that do not fit the isolate model. Mesh, announced more recently, is a private network specifically for agents to communicate over.

And then there is the recently announced Artifacts product, which is git-native versioned storage. It is billed as storage but it works like a git remote with cloud features wrapped around it. Branches, pull requests, replication, all the things you would expect from GitHub or GitLab, except as a storage primitive rather than a code hosting platform. This is the kind of product that sounds boring until you realize it is the missing piece for a category of workflows like dataset versioning, machine learning experiment tracking, or distributed configuration management that have historically lived in awkward bolt-on tools.

The whole Cloudflare stack costs five United States dollars per month minimum, gives you generous free tier limits across most products, and integrates so tightly that the boundaries between services blur. The downside is that everything is Cloudflare-specific. Workers code does not run anywhere else. Durable Objects have no equivalent in any other stack. R two is S three compatible at the application programming interface layer but does not behave identically. Once you build deeply on Cloudflare, migrating away is a rewrite. That is the tradeoff, and Cloudflare is open about it.

For Pär, the interesting Cloudflare angles are R two for media storage and Workers A I for cheap inference on small models. The rest is too far from his existing stack to be worth committing to, but the catalog as a whole is the most coherent vision of what serverless infrastructure looks like in twenty twenty six, and watching it evolve is genuinely educational.

The European tier two

Hetzner is the German hosting company that built its entire business on being cheap and reliable for almost three decades. The cloud product, which they launched comparatively recently, prices a virtual server with two shared virtual central processing units and four gigabytes of random access memory at three euros ninety nine per month. The equivalent on Amazon Web Services costs roughly thirty United States dollars per month. Every Hetzner server includes twenty terabytes of outbound network traffic. The equivalent egress on Amazon would cost over eighteen hundred dollars. Object storage costs six euros forty nine per month for one terabyte plus one terabyte of egress, and additional storage runs roughly half a cent per gigabyte per month. It is S three compatible, runs in three data centers across Germany and Finland plus newer locations in the United States and Singapore, and works with all the standard tooling.

The catch with Hetzner is that the catalog is deliberately small. About ten core services. Virtual machines, load balancers, firewalls, volumes, object storage, snapshots, floating internet protocol addresses, private networks, placement groups, and Secure Shell key management. There are no managed databases, no serverless functions, no platform-as-a-service offerings, no managed Kubernetes control plane, and no content delivery network. You install and manage everything yourself. For teams comfortable with self-hosted infrastructure who want costs as low as possible, Hetzner is unbeatable on price. For anyone who wants managed services, Hetzner is missing too many pieces.

OVHcloud is the French equivalent and the closest direct competitor to Hetzner. French headquartered, European Union data residency by default, and considerably broader in geographic footprint. OVHcloud operates more than forty data centers worldwide compared to Hetzner's six. The catalog is bigger than Hetzner's, including managed Postgres, managed Kubernetes, dedicated graphical processing unit instances, and a real public cloud product. The pricing is slightly higher than Hetzner but still dramatically below the hyperscalers. OVHcloud is particularly relevant for French public sector procurement because they hold the SecNumCloud certification, which is France's most rigorous sovereign cloud standard. They also famously had a major data center fire in Strasbourg in March twenty twenty one that destroyed one of their facilities and damaged another, which remains the cautionary tale every sovereignty advocate uses when arguing for multi-region redundancy.

Then there are the storage specialists. Backblaze B two is the original cheap S three alternative, running at about six tenths of a cent per gigabyte per month for storage with a generous egress allowance. Backblaze publishes its drive failure statistics openly, which is rare in the storage industry, and that transparency has made it a favorite for backup and archival workloads. Wasabi runs at about seven tenths of a cent per gigabyte per month with no egress charges at all, as long as your monthly egress stays below a one to one ratio with your total stored data, which is a constraint that catches teams off guard during migrations. Wasabi requires ninety days minimum retention on stored objects, so if you delete data early, you still pay for the full retention period.

Cloudflare R two we covered already. Tigris is a newer entrant in the same zero-egress object storage category, with versioning and object lock support that R two lacks. Impossible Cloud is another. IDrive E two prices storage at four tenths of a cent per gigabyte per month, which is the cheapest published rate in the category, though they operate under United States jurisdiction with CLOUD Act exposure. Storj is the interesting decentralized option, using erasure coded shards distributed across thousands of independent storage node operators worldwide. The pitch is geo-distributed redundancy that you cannot get from a single provider, with prices around four tenths of a cent per gigabyte. The catch is that egress is not free and the consistency model is different from S three.

For Pär, the storage specialist that genuinely competes with what he already uses on Scaleway is Hetzner Object Storage and Backblaze B two. Hetzner because the cost per terabyte per month is meaningfully lower than Scaleway Standard at scale and the data residency is European Union. Backblaze because the egress allowance is generous enough for podcast distribution and image hosting workloads where access patterns are unpredictable. Neither replaces Scaleway in his stack today, but if Pärkit ever needs to archive terabytes of historical telemetry beyond what fits comfortably on the current setup, Hetzner is probably where it goes.

Oracle and the rest

Oracle Cloud Infrastructure is the dark horse of the tier two. The catalog is genuinely competitive with Amazon Web Services on most enterprise workloads. The pricing is roughly fifty percent lower for equivalent compute. The free tier is the most generous of any cloud provider on earth, including two virtual central processing units and twenty four gigabytes of memory on always-free Arm instances, plus ten terabytes of monthly outbound egress at zero cost. That egress allowance alone is worth paying attention to. Anyone running bandwidth heavy workloads, video streaming, large file distribution, podcast hosting, can use Oracle Cloud as the cheapest possible delivery network simply by exploiting that free egress.

The downside of Oracle is the operator experience. The web console is slow and dated compared to Amazon or Google. The documentation is uneven. The brand is, to put it charitably, not aspirational for indie developers. And the deepest integrations are with Oracle's own database products, which most modern stacks do not use. So Oracle Cloud sits in a strange position. Technically excellent, financially compelling, culturally repellent to the audiences who would otherwise be its biggest evangelists.

International Business Machines Cloud is the other major tier two enterprise option. The pitch is hybrid cloud, deep integration with on-premise IBM Power and zee Series hardware, and the longest lineage of any computing vendor on the planet. IBM Cloud's most genuinely interesting product is IBM Quantum, the original cloud quantum computing service, which we will cover in part three. Outside of that, IBM Cloud serves a narrow set of enterprise customers who already use IBM products elsewhere and value the unified procurement.

DigitalOcean is the original indie hacker cloud. Simple pricing, clean interface, predictable bills. The catalog has grown over the years to include managed databases, managed Kubernetes, serverless functions, object storage with the Spaces product, and graphical processing unit compute through the Paperspace acquisition. DigitalOcean Spaces is S three compatible with built-in content delivery network integration at five dollars per month for two hundred fifty gigabytes. The pricing is higher than Backblaze or Hetzner but the developer experience is much smoother.

Linode, now Akamai Cloud after the twenty twenty two acquisition, occupies a similar position to DigitalOcean. Clean pricing, decent catalog, strong reputation among individual developers. Akamai's integration brings global edge presence through their existing content delivery network footprint, which is one of the largest on earth. Vultr is the third in this triumvirate, with thirty two data center locations globally and pricing that starts at two dollars fifty per month for the smallest virtual machine. None of these three are exciting in twenty twenty six. They are solid, reliable, predictable, and that is the entire pitch.

The developer platform layer

Heroku invented the modern platform-as-a-service category in two thousand seven. Push code via git, get a deployed application. No servers to manage, no operating system to patch, no networking to configure. The free tier was the on-ramp for an entire generation of developers. Then Heroku killed that free tier in twenty twenty two and scattered everyone across the new generation of platforms that had been quietly building in the background.

Railway emerged as the platform with the best developer experience of the post-Heroku era. Push from GitHub, Railway detects your runtime automatically, builds it, runs it. One click templates for Postgres, Redis, MongoDB, dozens of common stacks. The dashboard is genuinely pleasant. The free tier got removed in twenty twenty three, and a thirty day five dollar trial credit replaced it. Entry pricing starts at five dollars per month with usage-based billing on top of that. Railway is the right starting point for indie developers who want zero infrastructure decisions in the first months of building.

Render took the Heroku replacement positioning more literally and built a platform-as-a-service that closely mirrors what Heroku used to offer, including a real permanent free tier with the limitation that web services sleep after fifteen minutes of inactivity. The first request after sleep takes about a minute to respond as the service spins back up. For prototypes that is acceptable. For anything customer-facing it is not. Paid tiers start at seven dollars per month per service and eliminate the sleep behavior. Render's value proposition is predictable flat-rate pricing without usage-based surprises.

Fly dot eye oh took a completely different philosophical approach. Instead of running your code on shared infrastructure in a single region, Fly runs your application as a Docker container in any of thirty plus regions worldwide, with the option to deploy across multiple regions simultaneously for global latency. Fly is the platform you pick when geography matters to your users. The catch is that Fly is significantly more complex than Railway or Render. You write Docker files, you manage volumes for persistent state, you think about region placement. The free tier was removed in twenty twenty four, and entry pricing is roughly five dollars per month per virtual machine. Fly Postgres at about thirty four dollars per month for a small instance is meaningfully cheaper than Railway's managed Postgres at around ninety dollars.

Vercel is the platform for front end deployment, specifically Next dot j s applications. The company that builds Next is also the company that runs Vercel, so the integration is unmatched. Push a Next application, get edge-deployed serverless functions, automatic preview deployments per pull request, and a content delivery network that handles static assets globally. Free tier for hobbyists, twenty dollars per month for the professional tier, real bills for production teams. Vercel had a real security incident in April this year where environment variables that were not explicitly marked as sensitive were exposed, which prompted a wave of secret rotation across customer accounts. The platform remains the default choice for Next dot j s teams despite that incident.

Netlify is Vercel's closest competitor, focused on static sites and Jamstack workflows. One hundred gigabytes of monthly bandwidth on the free tier, three hundred minutes of build time per month. Netlify is excellent for static sites generated by Hugo, Jekyll, or Astro. Beyond static, Netlify Edge Functions and serverless functions cover dynamic workloads but with stricter limits than Vercel.

For Pär, the developer platform layer is mostly irrelevant. He has Scaleway as his primary cloud, has his own systemd services on the virtual private server, and has built rsync deployment aliases that work well for his rapid prototype to deploy workflow. The closest thing to a Railway use case in his stack would be quickly deploying a side project that does not need to live on the main virtual private server. But for a single side project, spinning up another Scaleway instance is functionally equivalent and keeps everything in one place.

What is worth knowing

The shape of the general purpose cloud world in twenty twenty six is more fragmented than it was five years ago and more competitive than it has ever been. The hyperscalers still dominate revenue, but they no longer dominate developer mindshare. Cloudflare has built the most coherent vision of what serverless infrastructure looks like, even if you cannot easily port off it. Hetzner and OVHcloud have made European data residency genuinely cheap. Oracle has made egress essentially free for anyone willing to overlook the brand. The developer platforms have absorbed the indie hacker market that Heroku used to own.

For Pär, the practical conclusions are short. Scaleway remains the right primary cloud given his current setup. Hetzner is worth knowing about for European cold storage at extreme scale. Cloudflare R two is worth knowing about for any project where egress costs matter. Backblaze B two is worth knowing about for backup workflows. Oracle Cloud's free tier is worth knowing about for any workload where ten terabytes of monthly egress would be useful. Everything else is mostly negative space.

In part two we will cover the neoclouds. CoreWeave, Lambda Labs, Crusoe, Nebius, Nscale, Together A I, Fireworks, Replicate, Hyperbolic, and the specialty silicon story including the NVIDIA acquisition of Groq, the Cerebras initial public offering happening this very week, and the Meta CoreWeave thirty five billion dollar commitment that has reshaped the entire artificial intelligence infrastructure market. That episode is where things stop being incremental and start being weird.