The Cloud World, Part Three: Where Things Get Strange

Opening

Part one mapped the general purpose cloud world. Hyperscalers, tier-two providers, storage specialists, developer platforms. Part two mapped the neoclouds and the specialty silicon reshuffle that has redrawn the artificial intelligence infrastructure picture over the last six months. This final part is the strange end of the catalog. The pieces of the cloud world that are interesting to know about even when there is no immediate practical use case, because they represent either where things are heading or genuinely weird bets on alternative architectures.

We are going to cover four territories. First, the database specialist layer that has emerged around modern serverless workloads, including the recent Databricks acquisition of Neon and the rest of the territory after PlanetScale pivoted away from indie developers. Second, the quantum computing cloud market across the four major platforms and the standalone vendors, which has matured significantly in twenty twenty six. Third, the decentralized compute marketplaces and the broader category that calls itself decentralized physical infrastructure networks. Fourth, the European sovereign cloud build-out, which is genuinely accelerating in twenty twenty six in ways that matter for anyone operating from European Union jurisdictions.

The database specialist reshuffle

The database market has consolidated around Postgres as the default choice for almost every new web application, and the interesting differentiation now happens at the platform layer rather than the engine layer. Five companies dominate the conversation. Supabase, Neon, PlanetScale, Turso, and Convex. Each occupies a clearly different position philosophically.

Supabase is the integrated backend platform, often described as the open source Firebase alternative. The pitch is a full backend in one product. Managed Postgres, built-in authentication, file storage, real-time subscriptions, edge functions, all behind a single dashboard with a generous free tier. Supabase has become the default choice for indie hackers building anything that needs more than a database. The catch is that the integration is opinionated. You use Supabase Authentication, Supabase Storage, Supabase Edge Functions, or you do not get the benefits of the integration. For a certain shape of project this is exactly what you want. For projects where each piece needs to evolve independently, Supabase can feel like a constraint.

Neon is the pure Postgres play. The pitch is serverless Postgres that genuinely scales to zero, which traditional managed databases including Supabase do not. When nobody is querying your Neon database, you pay nothing for compute, only for the underlying storage. Database branches per pull request let your continuous integration pipeline spin up isolated database copies for testing in seconds. The architectural innovation Neon shipped is decoupling compute from storage, similar to how Snowflake decoupled compute from storage for warehouses a decade earlier. Databricks acquired Neon in early twenty twenty six, which the press framing called a strategic partnership but the financial reality was probably closer to a lifeline. The good news for existing Neon customers is that Databricks reduced compute costs by fifteen to twenty five percent post-acquisition, and the product continues to operate independently under the Neon name.

PlanetScale is the cautionary tale of the modern database market. The product is genuinely excellent. MySQL built on Vitess, the same technology that powers YouTube. Schema branching that mirrors git workflows. Online Data Definition Language changes without downtime. Zero-downtime schema migrations. The technology is best-in-class. The business decision was less popular. PlanetScale removed its free tier in April twenty twenty four and went enterprise-only, with entry-level pricing now starting at thirty nine dollars per month. The free tier removal scattered thousands of hobby projects to Neon, Supabase, and Turso, and PlanetScale lost its position as the indie hacker default. The product still serves customers like Cursor, Intercom, and Block at scale where reliability and operational maturity dominate the cost calculation, and PlanetScale recently added managed Postgres alongside their core MySQL offering. But the indie market mindshare is gone.

Turso is the edge database play. The underlying engine is libSQL, which is an open source fork of SQLite extended to support multi-region replication. The pitch is SQLite-compatible databases that run at the edge, close to your users, with sub-millisecond read latency from cached replicas. The free tier offers a billion reads per month, which is wildly generous compared to anything in the Postgres market. The constraint is that it is SQLite. If you need heavy Postgres extensions like PostGIS for spatial queries, time series extensions, or advanced full-text search, Turso is the wrong choice. For applications where read latency dominates user experience, especially content directories, e-commerce catalogs, and programmatic search engine optimization sites, Turso has no real competition.

Convex is the reactive backend play. The pitch is a fully reactive query system where the client subscribes to queries and the backend automatically pushes updates whenever underlying data changes. TypeScript-native, with a developer experience that most TypeScript teams find genuinely pleasant. The catch is total platform lock-in. Convex is not Postgres compatible. It is its own data model, its own query language, its own everything. Migration out of Convex means rebuilding your application. For TypeScript teams building real-time collaborative applications who are willing to accept the lock-in tradeoff, Convex is one of the fastest paths from idea to production.

Beyond these five, the rest of the database market includes Cockroach Database for globally distributed Structured Query Language with automatic multi-region failover, MongoDB Atlas for the document database segment, ClickHouse Cloud for analytics workloads at petabyte scale, Tiger Beetle for financial transaction workloads where double-entry bookkeeping correctness matters, and a long tail of vector database specialists including Pinecone, Weaviate, Qdrant, and Milvus Cloud through Zilliz that compete on retrieval augmented generation workloads. None of these are particularly relevant to Pär's existing setup, which runs Postgres directly on the Scaleway virtual private server through Pärkit, but the territory is worth knowing.

The truly strange angle here is what happens when the database, the storage, the compute, and the orchestration all converge into a single integrated product. Cloudflare's full stack, which we touched on in part one, is moving in this direction. Workers compute, R two storage, D one SQL database, K V key-value, Durable Objects for stateful actors, Vectorize for embeddings, Workers A I for hosted inference, Hyperdrive for connecting to external Postgres or MySQL databases, all integrated through a single binding system in the Workers runtime. The same pattern is showing up at Supabase, where the database, authentication, storage, real-time, edge functions, and now vector embeddings all live behind one dashboard. The bet is that integration is more valuable than choice for the audience that buys this kind of product.

The quantum cloud market in twenty twenty six

Last episode covered Scaleway's quantum offering in detail, including the four hardware modalities Scaleway aggregates. Superconducting from I Q M, trapped ion from A Q T, neutral atom from Pasqal, and photonic from Quandela. The broader quantum cloud market has four major platforms beyond Scaleway, each with a different posture.

Amazon Braket reached general availability in twenty twenty and is the longest-running multi-vendor quantum cloud service. The catalog hosts IonQ's Forte and Aria trapped ion machines, Rigetti's one hundred eight qubit Ankaa three superconducting hardware released to general availability this month, I Q M Garnet superconducting, QuEra Aquila with two hundred fifty six neutral atoms, and Oxford Quantum Circuits superconducting systems. The Braket Software Development Kit supports multiple quantum frameworks including Qiskit, PennyLane, and the recently added compute unified device architecture quantum integration. Amazon also exposes its own Center for Quantum Computing in Pasadena, which is developing superconducting processors that are not yet publicly available through Braket but will be eventually.

Microsoft Azure Quantum reached general availability in February twenty twenty two and matches Braket on hardware breadth while adding deeper software integration. The Azure catalog includes IonQ trapped ion, Quantinuum's H series trapped ion systems which produced the most widely cited logical qubit milestones of twenty twenty four and twenty twenty five, Rigetti superconducting, Pasqal neutral atom, Atom Computing with the twenty four entangled logical qubit record set in twenty twenty five, and Quantum Computing Incorporated photonic. Microsoft adds the Azure Quantum Elements layer specifically for chemistry, materials science, and density functional theory workflows, and exposes the open-sourced Resource Estimator for sizing fault-tolerant workloads. The Q sharp programming surface remains Microsoft's first-party quantum language. Microsoft also introduced Majorana one in February twenty twenty five, the first chip on what Microsoft calls a Topological Core, using topoconductor materials to store information in the topology of the quantum state itself, which is naturally resistant to local perturbations. The topological qubit approach is the longest-shot bet in quantum hardware, but if it works the path to millions of logical qubits shortens dramatically.

International Business Machines Quantum Platform launched in May twenty sixteen as the original quantum cloud service. International Business Machines runs only first-party superconducting hardware, no multi-vendor catalog, but reaches the largest user base of any quantum cloud thanks to a decade of head start. The Qiskit software framework is the industry standard quantum programming environment, used across most other platforms even when running non-International Business Machines hardware. International Business Machines has demonstrated quantum processors at the four hundred qubit scale and has published a roadmap targeting fault-tolerant quantum computing by twenty twenty nine.

Google Quantum A I operates more like a research program than a commercial cloud service. Access is gated, primarily available to research collaborators rather than open commercial users. Google's quantum work has produced the most-cited quantum supremacy results, including the original twenty nineteen Sycamore demonstration and subsequent error correction milestones. Google has set the public goal of an error-corrected quantum computer capable of solving real-world problems by twenty twenty nine.

The standalone quantum vendors are interesting beyond their cloud presences. IonQ became the first quantum computing company in history to cross one hundred million dollars in annual generally accepted accounting principles revenue, the milestone every public quantum company is now measured against. IonQ's trapped ion systems are accessible across Azure Quantum, Amazon Braket, and Google's research collaborations. Rigetti reported a record four point four million dollars in first quarter revenue this year driven by on-premise system shipments and the general availability of the one hundred eight qubit system, and is investing up to one hundred million dollars in the United Kingdom to deploy a one thousand plus qubit system. Quantinuum raised six hundred million dollars at a ten billion dollar valuation and has filed its S one for a public listing widely expected to value the company above twenty billion. PsiQuantum, the photonic quantum computing company, has raised over one point three billion dollars and is anticipated to pursue a public offering in twenty twenty six. I Q M became the first publicly listed European quantum company in February twenty twenty six. Infleqtion listed on February seventeenth this year. Six pure-play quantum companies now trade on United States exchanges. Quantum has transitioned from a laboratory curiosity to a measurable industry.

For Pär's Pärception consulting practice, the quantum angle remains the same as the Scaleway part one episode covered. A demo connecting through Qiskit to remote quantum hardware is a strong showcase for clients exploring whether quantum computing might fit their roadmap. The choice of provider matters less than the demonstration itself. Anyone interested in the topic finds the IonQ trapped ion modality more visually intuitive than superconducting, since you can show actual photographs of ytterbium ions levitating in electromagnetic traps. Amazon Braket and Azure Quantum both expose this through their standard interfaces.

The decentralized cloud marketplaces

This is the section where the cloud world genuinely gets weird.

Akash Network is the most mature of the decentralized compute marketplaces. The model is a reverse auction over a Cosmos blockchain. Tenants who need compute resources specify their requirements and the maximum price they are willing to pay. Providers who have spare capacity bid against each other to host the workload. The blockchain coordinates the matching, escrows the payment in the native A K T token, and settles the lease. The architecture is described as Airbnb for cloud computing. Decentralized matching, no single operator with control over the platform, prices set by market dynamics rather than published rates.

In practice Akash prices typically land sixty to eighty five percent below the hyperscalers for equivalent specifications. The catch is reliability. Workloads on Akash run on independently operated provider hardware with varying uptime guarantees, no unified support, and a self-service developer experience that has improved dramatically over the last two years but still lags polished commercial alternatives. Akash's roadmap for twenty twenty six includes Lease to Lease Private Networking that mimics virtual private cloud isolation, preemptible instance support for cost reduction on interruptible workloads, sovereign artificial intelligence agent infrastructure with confidential computing through Trusted Execution Environments, and a Managed Service Market that lets developers and service creators earn directly from hosting managed offerings on the network.

The most interesting twist in Akash's twenty twenty six story is the Project Twilight hard fork that activated a Burn Mint Equilibrium mechanism. The model burns A K T tokens when users purchase compute, minting non-transferable Akash Compute Tokens for actual spending. This anchors token scarcity to real network demand rather than speculative trading. The economics are deflationary by design, which is unusual in the cryptocurrency space and aligns the interests of token holders with actual platform usage.

Bittensor takes a different approach. Rather than matching tenants to compute capacity, Bittensor is a marketplace for machine learning model contributions. Participants run subnets that specialize in particular tasks like text generation, embedding production, or scientific computing. Models compete within each subnet, with the highest-scoring models earning the native T A O token. The economic flywheel is supposed to be that better models earn more tokens, which incentivizes participants to contribute better models. T A O trades at around two hundred eighty nine dollars per token with a market capitalization near two point eight billion dollars as of this month.

Render Network competes with Akash in the graphical processing unit segment but focuses specifically on rendering workloads, particularly for visual effects studios, three dimensional animation, and architectural visualization. The model is similar. Render owners list capacity, customers bid for time, the network coordinates settlement.

Io dot net operates as a more conventional graphical processing unit marketplace with decentralized aggregation. The token jumped more than fifty percent in early May this year, which placed it among the most-trended assets for that week. Internet Computer takes the most architecturally ambitious approach. Rather than coordinating compute through smart contracts on an external blockchain, Internet Computer runs its own validator nodes in independent data centers worldwide, and the platform itself runs as a decentralized network of these nodes. Smart contracts on Internet Computer can host entire applications including the front end, the back end, and the storage, with no centralized servers required at any layer. The pitch is true decentralization rather than partially decentralized coordination.

For most practical workloads decentralized clouds are not yet competitive with neoclouds. The reliability gap is real, the operational tooling is less mature, and the cost advantage often disappears once you factor in the additional engineering work required to run production workloads on this kind of infrastructure. The interesting case for paying attention is the long term thesis. If artificial intelligence inference becomes the dominant compute workload globally, and if the supply of graphical processing unit capacity remains constrained, decentralized marketplaces have a structural argument for being the most efficient allocators of that capacity. Whether that argument holds in practice is the question being tested over the next three to five years.

The European sovereign cloud build-out

The final territory worth covering is the European sovereign cloud push, which has accelerated dramatically in twenty twenty six in ways that matter for anyone operating from European Union jurisdictions like Pär does from Kall.

The structural context is uncomfortable. United States hyperscalers control roughly seventy to seventy two percent of the European cloud market as of recent measurements. Private artificial intelligence investment in the United States runs approximately twenty four times higher than in Europe in absolute dollar terms. European artificial intelligence startups have routinely trained models on United States clouds because no domestic alternative existed at meaningful scale, which means European data, European intellectual property, and European compute economics have flowed primarily into American infrastructure.

That is changing rapidly in twenty twenty six. The European Commission's InvestAI program is funding five artificial intelligence gigafactories across Europe with a total budget of two hundred billion euros over the program lifetime. Each gigafactory is targeting one hundred thousand specialized graphical processing units, with access open to large industrial companies, startups, and research institutes. The first phase of funding committed twenty billion euros specifically to four of these facilities. The locations and operators are being selected through twenty twenty six, with construction expected to overlap with the European Union artificial intelligence Act high-risk rules that take effect on August second twenty twenty six.

Below the gigafactory tier, EuroHPC's existing federation operates fourteen supercomputers and nineteen so-called artificial intelligence factories, smaller-scale facilities aimed at research and startup access, backed by roughly ten billion euros of combined Commission and member state funding from twenty twenty one through twenty twenty seven. The artificial intelligence factory tier is where European researchers can actually access compute today without United States dependencies.

Private capital is now flowing at unprecedented scale into European artificial intelligence infrastructure. French champion Mistral confirmed a one billion euro capital expenditure plan for twenty twenty six, including a one point two billion euro Swedish data center and a Paris-area facility. The Mistral Swedish build is interesting for Pär because it places hyperscale-class artificial intelligence infrastructure in the same country he operates from, with European data residency by default. Mistral also completed an eight hundred thirty million euro institutional debt raise in early twenty twenty six, which marked the first time a European artificial intelligence company financed a hyperscale data center without United States venture capital.

The federated sovereignty layer is also taking shape. EURO three C, launched in March twenty twenty six and backed by the European Commission with seventy five million euros of initial funding, brings together more than seventy organizations including Telefónica, multiple national telecom operators, technology companies, startups, and small to medium enterprises. The architecture is a federated network of national infrastructure nodes that operate across European Union borders rather than a single new European cloud platform built from scratch. The pitch is digital sovereignty without the construction timeline that a single new hyperscaler would require.

The Important Project of Common European Interest on Next Generation Cloud and Services, branded I P C E I, channels three billion euros into European Union data and semiconductor projects. This is the public funding mechanism that supports the more sovereign-oriented private companies, including Scaleway through Iliad Group, OVHcloud, and various smaller players. The France-Germany Digital Sovereignty Summit in November twenty twenty five launched a joint task force, and the European Union Council signed a Declaration for European Digital Sovereignty in December that committed member states to coordinated procurement strategies favoring European providers.

The honest skeptical reading is that none of this fully closes the dependency. Europe operates fourteen supercomputers and nineteen artificial intelligence factories, but the graphical processing units in those facilities are designed by NVIDIA and manufactured by Taiwan Semiconductor Manufacturing Company. NVIDIA holds roughly eighty five percent of the artificial intelligence graphical processing unit segment as of twenty twenty six, with analysts projecting a drift toward seventy five percent over the next year as Advanced Micro Devices and custom silicon scale, but the dependency at the silicon layer is not going away on any reasonable timeline. European sovereignty at the application and data layer is genuinely achievable. European sovereignty at the silicon layer is not.

For Pär specifically, the European sovereign cloud build-out has both pragmatic and ideological angles. Pragmatically, the Mistral Swedish data center represents real hyperscale-class artificial intelligence capacity in his home jurisdiction, which means that if any Pärception client needs sovereign artificial intelligence infrastructure for compliance reasons, the answer is increasingly available within Sweden rather than requiring a German or French location. Ideologically, the broader story aligns with the values that show up across Pär's other operations. Local journalism through Årebladet. Pärkit running on his own virtual private server rather than third-party platforms. The bias toward self-hosted infrastructure that runs through every project. European sovereign cloud is the institutional version of the same instinct, applied at continental scale.

The far-fetched ideas

Closing thoughts on what is worth doing with all of this. The honest answer is that almost none of it maps directly to immediate Pär workloads. Pärkit Postgres on Scaleway is fine. Modal handles serverless graphical processing unit. RunPod handles persistent pods. Fal handles polished image and video generation. The summer edition of Årebladet ships first, and infrastructure changes are not on the critical path for that deadline.

What might be worth doing later, in approximate order of feasibility. A SambaNova Cloud free tier account, just to have access to Llama three point one four hundred five billion parameters at no cost, gives Pär the option to compare large model behavior against the smaller models that Scaleway Generative APIs hosts. Five minutes of setup, persistent access, no commitment.

A Cloudflare Workers experiment that uses R two for cheap storage, Workers A I for small model inference, and Durable Objects for stateful coordination, builds a working prototype of a fundamentally different serverless model than Scaleway exposes. The point would be education rather than production, but the integration is tight enough that something useful might emerge.

A quantum cloud demonstration through Amazon Braket or Azure Quantum that connects to IonQ trapped ion hardware. Five minutes of Qiskit code, runs against actual quantum hardware, produces a result that can be shown to Pärception clients during a presentation about emerging compute approaches.

An Akash Network deployment of some non-critical workload, either for the educational value of understanding how decentralized compute marketplaces work in practice, or as a price-comparison data point against Modal and RunPod. The setup friction is real but the architecture is genuinely different from anything else Pär uses.

A Hetzner Object Storage account configured for offsite archival of Pärkit historical telemetry, providing geographic redundancy and a sovereign European Union backup destination for data that currently lives entirely on Scaleway. This is the closest thing to a practical change in his actual stack that comes out of this entire three-part survey.

Beyond that, the value of all this is mostly cartographic. Knowing the territory exists, knowing roughly where each player sits, knowing what conversations are happening at the edges of the artificial intelligence infrastructure market in twenty twenty six, even if Pär never opens an account anywhere new. The summer edition is still the anchor deadline. After that ships, the optionality opens up.

That is the full cloud world map. Three episodes, somewhere north of two hours total. The whole thing is a snapshot of mid May twenty twenty six, which is to say it will be partially out of date within six months. But the shape of the territory tends to evolve more slowly than the headlines suggest, and the structural questions worth asking, where is value concentrated, where is competitive pressure highest, where are the genuinely novel architectures, are the same questions whether the answer is CoreWeave or some other company that does not exist yet.

That is all from this series.