#103 - European Cloud Sovereignty with Mark Shine, Pawel Piwosz and Filipe Berti
Lower cloud bills depend on knowing which managed services the team can take back.
18 min read
Mark Shine, Pawel Piwosz, and Filipe Berti discuss why the default choice of AWS, Azure, or GCP is no longer automatic for every team. The conversation covers cost, managed services, open source, AI workloads, and what European cloud providers can offer instead.
Picking a cloud stopped being a default
You might not have picked your cloud provider. Take the ordinary case: an account already existed, someone chose AWS, Azure or GCP in the first week of a project, and nothing since has forced a second look. Three things are putting the question back on the table: the bill, the blast radius, and a compliance date in August that asks what role you play in every AI feature you ship.
None of that is a reason to move. It’s a reason to know what you’d be able to move. The useful version of the sovereignty question isn’t about flags, it’s operational. Which parts of your stack could you run somewhere else inside a week? What would your team have to learn to do it? When the provider has a bad day, what do you still hold outside it? If nobody has written those answers down, a $25,000 bill and a single-region dependency can both sit there for years without anyone deciding they should.
Three people who argue about this for a living joined us. Mark Shine is co-founder of Ankra, a Stockholm platform that manages Kubernetes clusters and installs versioned add-on stacks on them, with integrations for UpCloud, Hetzner, OVH and the hyperscalers. Pawel Piwosz is a developer advocate at UpCloud, the Finnish provider, and has worked with clouds since around 2014. Filipe Berti is UpCloud’s business intelligence manager and has spent five years there, close enough to the product to know which plan customers actually ask for.
Traditional cloud and AI cloud are two different products
Start with what the word “cloud” covers now, because the answer differs for each half of it.
Why does the AI stack change where you can run?
Because it’s new for everyone, and a long head start in managed databases doesn’t carry over to it.
Traditional cloud computing is largely figured out in Europe. Virtual machines, storage, load balancers, a managed database or two: the pieces are generic, and a VM is a VM. What European providers have less of is the depth of managed services, and pretending otherwise helps nobody. The AI side is a different shape. Its tools are custom to AI and useful for nothing else: inference servers, KV caches, embedding stores, routers that know what a prompt costs to serve.
In May 2025 Red Hat opened llm-d, a Kubernetes-native stack for distributed inference that runs vLLM and other engines across a cluster with cache-aware routing. None of the techniques were new; teams had been wiring embeddings, routing and batching together by hand for a year. The packaging is the product, and because the packaging is open source, it lands on whatever infrastructure you point it at, including a GPU you rent in Helsinki.
The models moved the same way. Open weight families like Mistral, Qwen and Kimi closed enough of the quality gap that cost per token sits beside quality as a thing you optimise. DeepSeek is the case people reach for. Its technical report for DeepSeek-V3 puts the training run at 2.788 million H800 GPU hours, about $5.576 million at an assumed $2 per GPU hour, and says the figure excludes the costs of prior research and ablation experiments on architectures, algorithms or data. DeepSeek-V3 was also trained partly on other models’ output. Neither caveat touches the part you can use: you don’t need the largest model for most of the work you were going to send to an API.
What you’re paying for when you pay for managed
That lands you in front of a price list, which is where the European argument gets concrete.
Is the managed premium worth it?
Sometimes, but check what it’s a premium over.
A fair number of managed services are an open source tool with a console in front and an SLA behind. Work out the multiple over running the same thing yourself before you accept it. One comparison we’ve heard put a service at roughly five times, close to the cost of an engineer who would know how it works. Treat that as one person’s pricing exercise, not a market rate. The only multiple that decides anything is the one you run on your own bill.
European providers tend to price nearer the fair-business end of that range, and slice capacity more finely because their customers ask them to. UpCloud revamped its server plans after users kept asking for the shape they needed instead of the shape above it, and the same revamp is coming to its managed databases. That follows from who these providers design for: platform and senior engineers who want control, think about architecture, and can tell excess from capacity.
The economics changed too. Ten years ago a land-grab price was acceptable because getting to market first paid for it. Now runway is what pays.
The bill went from $25,000 to roughly $800. One case, and a gap that size says as much about what the original setup was doing as it does about either provider.
Managed doesn’t mean handled
The bill is the easy half of the argument. The hard half is what you have to know either way.
What goes wrong when nobody knows what’s under the managed service?
You expect a ticked box to do something it was never built to do.
Cloud SQL’s high-availability option is the example we keep coming back to. Teams run it, pay for it, and send every query to the primary, which is exactly how the service works. Google’s own documentation describes the standby as a failover target: writes are replicated to disks in both zones before a transaction commits, and if the instance becomes unresponsive Cloud SQL switches to serving data from the standby. High availability shortens a zone failure. It doesn’t spread read traffic, because the standby serves nothing until a failover happens. Read scaling is a read replica, a separate resource you create and pay for, and the tick in the console gives you no hint of the difference.
The same gap shows up elsewhere. A managed queue has a limit on how many messages a second it will take, and finding that limit during an incident is finding it too late. Your Postgres parameters and your slow queries are still yours. Check Nginx configuration and resource limits before buying a bigger instance.
Teams that already have those skills are the ones who can move. One of them had run its logging on GCP’s managed Log Explorer long enough to know which logs were worth keeping. That team moved infrastructure and other pieces to European providers while its APIs stayed on GKE, a partial migration rather than an exit, and operating costs came down by about 60%, roughly $100,000 a year.
A second case points the same way at a larger scale. Filipe Berti described one of Finland’s largest companies moving a warehouse platform from AWS to GCP with fewer managed services, cheaper and more efficient at the end, because the team had already mastered many of the pieces it was taking back. Pawel Piwosz challenges that framing: you can cut the same bill without changing provider, because the understanding that lets you move is the understanding that lets you shrink what you already run. A five-minute proof of concept can sit in production for five years, eating money the whole time.
Hybrid is now the resilience argument
Cost is only part of why teams are reopening this. Last year supplied the rest.
Does a second provider actually buy you resilience?
Only for the parts you built so they can move.
On 20 October 2025 a latent race condition in DynamoDB’s DNS automation left the regional endpoint for dynamodb.us-east-1.amazonaws.com pointing at an empty record, and AWS’s own summary traces how that cascaded into EC2 launches, network load balancer health checks and Lambda polling. Four weeks later a database permissions change at Cloudflare roughly doubled a Bot Management feature file, pushed it past a 200-feature limit in the proxy, and the network returned 5xx until 17:06 UTC. Neither was an attack. Both exposed how much rides on a provider dependency nobody had priced: one regional endpoint, one configuration file on the proxy path.
Standardisation is the honest counter-question. If your product is Lambda functions, you aren’t moving it on the morning of an outage, and the same goes for ECS task definitions. Containers and Kubernetes are the closest thing to a portable floor, and the price of that floor is that you run it.
Backups are where this gets testable. A native RDS snapshot can be shared with other AWS accounts, something AWS has supported since 2015. It still doesn’t get your data out of AWS, because the snapshot stays an RDS artifact that restores into RDS. A scheduled pg_dump is old-fashioned, and it produces a file another vendor’s Postgres can load.
Three versions govern that file, not two, and the one people forget is the utility’s. PostgreSQL’s notes on pg_dump peg compatibility to the pg_dump binary rather than to the server it read. Output “can be expected to load into PostgreSQL server versions newer than pg_dump’s version”. A load into an older major version isn’t guaranteed, even when the dump came from a server of that version. So run a binary at least as new as the source, restore onto a server at least as new as the binary, and pass --quote-all-identifiers, which the docs recommend whenever versions differ. A dump also covers one database: roles and tablespaces are cluster-wide objects for pg_dumpall, and your schema’s extensions have to exist on the far side. Rehearse the restore on the destination you would really use.
CI/CD was built for professionals, not for agents
Recovery time is the other worry we keep hearing, and it has nothing to do with which cloud you’re on.
What belongs between an agent and production?
A human who reads the bill of materials before it applies, every time.
We’ve all heard the stories: databases wiped, Terraform applying while someone was asking a question about it. Those aren’t surprises, they’re missing safeguards. CI/CD has existed for twenty years and was built for a professional with a change process, not for something that types faster than you can read and is confident either way. An agent wired to your cluster and your repository with the full lifecycle in its hands feels like magic and is reckless, which is why the review gate is the part worth building first.
The numbers give you the measured half. Google’s DORA research puts AI use at 90% of the technology professionals it surveyed, not of developers generally, and reports higher adoption associated with an increase in both software delivery throughput and software delivery instability. Those are associations in survey data, not a finding that AI caused either one. Faros AI’s 2026 telemetry across 22,000 developers and 4,000 teams is blunter: bugs per pull request up 28%, median review time five times longer, incidents per merged pull request roughly tripled. That’s vendor telemetry and correlational too, so treat the multiples as direction rather than physics.
The worry about recovery has a shape of its own. Teams used to have someone who knew what had broken from the length of one log line, and that person isn’t always the one carrying the pager now. The cited studies don’t establish whether losing experienced responders explains longer recovery times. It’s an explanation people offer for the incident numbers, so test it against your own incident history before you accept it.
Pawel Piwosz holds that the more you know, the more you get out of AI. Mark Shine adds a caveat: willingness counts as much as prior knowledge, because someone who keeps asking “and how do I make that highly available” gets there fast, while someone who skims the answer ships it. Same tool, opposite outcomes, and we did this with Stack Overflow snippets for a decade.
Your secrets are what you’re moving, not your servers
This gets personal at the moment you ask a hosted model to debug something real.
What actually leaves your network when you do that?
Whatever the agent had to read to answer: environment variables, configuration, the shape of your infrastructure, sometimes the credentials themselves.
One correction worth making, because the fear is often aimed at the wrong part. Which terms apply depends on the service, not on whether you pay. Anthropic’s commercial terms govern the API, Claude for Work, Claude for Government and Claude for Education, and commit that customer content from those isn’t used to train models. Claude Free, Pro and Max sit under the consumer terms instead, and Anthropic’s August 2025 update says new models are trained on data from those accounts when the setting is on, including Claude Code used from them, with retention extended to five years in that case. A paid subscription is not a commercial agreement. Check which contract your own plan sits under before you decide what an agent may read. Training was never the whole exposure anyway: the content leaves your network, it’s retained for some period, and access to it is governed by somebody’s policy rather than by your network boundary.
The legal reach is the part you can’t configure away. US law lets US authorities compel a US provider to produce data it controls regardless of where the servers sit, which AWS documents itself. The question isn’t whether your bytes are in Frankfurt. It’s who holds the keys and who answers the order.
Self-hosting is the other end of the trade, and it got cheaper. A base model plus a fine-tune for the specific job runs on hardware you can rent by the hour, including on-demand L40S instances from a provider in Helsinki. Not free. Bounded, and the data stays where you put it. There’s a product argument under the security one. Generic model output can look repetitive: Lovable projects keep coming out with the same square grids. Your own data and a fine-tune move the output away from that, and they’re the same material you’d be pasting into someone else’s context window.
Article 50 splits the duties by role
Under all of this sits a date that doesn’t care which argument you found convincing.
Does calling someone else’s model move the duty to them?
No, and it doesn’t move theirs onto you either, because the EU AI Act attaches Article 50 duties to providers and to deployers separately.
From 2 August 2026 those duties divide like this. Providers have to build systems intended to interact directly with people so that people are informed they’re dealing with an AI system, and have to ensure systems generating synthetic audio, image, video or text mark their outputs in a machine-readable, detectable format. Deployers have to tell people when an emotion recognition or biometric categorisation system is operating on them, and have to disclose deep fakes and AI-generated text published to inform the public on matters of public interest. Ship a product that wraps a vendor’s model and you can hold both roles at once, on different surfaces.
The exceptions carry as much weight as the duties. The interaction notice doesn’t apply where it’s already obvious to a reasonably well-informed, observant and circumspect person. Machine-readable marking doesn’t apply where the system performs an assistive function for standard editing or doesn’t substantially alter the input data or its semantics. For evidently artistic, satirical or fictional work, deepfake disclosure is limited to a form that doesn’t hamper the display or enjoyment of the work. Published text is exempt where it has had human review or editorial control and a person or company holds editorial responsibility for it. Law enforcement carve-outs run through all four paragraphs.
Fines for breaching Article 50 sit in Article 99(4), up to 15 million euro or 3% of total worldwide annual turnover, whichever is higher. Article 99(6) inverts that for SMEs, start-ups included: the cap is whichever of the two is lower. The headline number is the wrong one to quote at a ten-person company.
Where the market pressure comes from
The compliance argument has a commercial twin.
The complaint we hear from European providers is that commercial terms rather than missing features keep them out of deals: gatekeeping and unfair-competition allegations, not a feature matrix. Regulators now have that in scope. In November 2025 the Commission opened three market investigations into cloud computing under the Digital Markets Act, two of them asking whether AWS and Azure should be designated gatekeepers despite falling below the usual quantitative thresholds. A designation wouldn’t change a price list by itself. It would change who sets the terms on which you leave, which is the thing you were trying to price all along.
What this means for teams
The gap that’s left is developer experience. Connecting a GitHub repository to Vercel is still easier than a VPS, an SSH key and installing your own database. European providers are closing that from behind, which means early workloads there are rougher and also that they shape what gets built next. Data processing, batch work and GPU jobs are where the trade is best, not your marketing site.
What gets built next is partly your decision, because these providers build toward demand they can see, and three asks keep surfacing. Paying for inference by token or by request is the clearest: UpCloud is looking at it with care, without a product announcement attached. A European answer to Snowflake and Databricks is the second, still an open question rather than a roadmap item, because media and larger companies keep asking where they go if they want to leave and nobody has a settled answer. The third is completeness over breadth: if you work with tool A through tools B and C, developers want all three in one ecosystem rather than two plus a workaround.
Four things worth doing this week:
- Pick the managed service you depend on most and write down what you’d hold outside it tomorrow. If the answer is a snapshot in the provider’s own format, schedule a
pg_dumpor its equivalent into storage at a second provider, and test the restore. - Price one real workload, the data-heavy one, at a European provider. Compare total cost including the hours, not instance to instance.
- Put a review gate on anything an agent can apply. No merged pull request, no
terraform apply, and someone reads the bill of materials. - List every surface where model output or an AI conversation reaches someone in the EU, and mark each one with the role it puts you in, provider or deployer. That mapping is your Article 50 scope, and August is closer than it reads.
Common questions, answered
How do I move a Postgres database off a managed cloud service?
Start taking your own logical dumps first. A native RDS snapshot shares to other AWS accounts, but it restores into RDS, so it doesn’t get you off AWS. A scheduled pg_dump gives you a file another Postgres can load. The version rule hangs on the pg_dump binary, not the source server. Run a binary at least as new as the source, restore onto a server at least as new as the binary, and add --quote-all-identifiers across versions. Roles, tablespaces and extensions need handling separately. Rehearse the restore on the real destination. The parameters and tuned queries around it are the real migration work, and knowing them also cuts the bill without moving.
Do I have to label AI output in my product from August 2026?
Work out your role first, because Article 50 gives providers and deployers different duties from 2 August 2026. Build the system and you mark synthetic output machine-readably and tell people they’re interacting with an AI. Operate someone else’s and you disclose deep fakes and AI-generated text published to inform the public on matters of public interest. Exceptions apply, including assistive editing, obvious interactions, and text under human editorial responsibility. Fines reach 15 million euro or 3% of worldwide turnover, whichever is higher, and whichever is lower for SMEs and start-ups.
How do I estimate what leaving managed services would really cost?
Price the hours before the instances. Take one workload, write down instance and storage cost at the new provider, then add the operational work the managed service was doing: upgrades, failover, backups and restore tests, patching, and the on-call that now owns it. Don’t predict the answer from your team’s size or skills. Run it on one workload, measure what it actually took, and use that number to decide about the second.
Related episodes
- #90 K8s vs Managed Services: Cost, Lock-In, and Reality is the same argument at cluster level, with Mark Shine, and where the lock-in maths got worked through.
- #79 Going Local: What’s Driving the Move? opened our European infrastructure thread and asked what actually drives teams off hyperscalers.
- #87 EU Compliance 101: AI Act, DORA, NIS2 explained maps the regulations behind the August date, including how AI systems get classified.
- #89 Agents, Reviews, and Secrets: Real Talk on AI in Dev covers what happens when coding agents can read your environment variables.
Resources
- Regulation (EU) 2024/1689, the AI Act: Article 50 in the official text, where the provider and deployer duties and their exceptions are set out.
- DeepSeek-V3 technical report: the training-run GPU hours and cost assumption, and the sentence saying what the figure leaves out.
- llm-d: the inference stack, and what it packages that teams used to assemble by hand.
- Ankra docs: what a “stack” is in practice, and which clusters and providers the platform works with.
- UpCloud on-demand NVIDIA L40S GPUs: hourly GPU capacity in Helsinki, for pricing self-hosted inference against an API bill.
- Commission market investigations into cloud computing under the DMA: what the Commission is asking about AWS and Azure, and on what basis.
- DORA, balancing the tensions of AI-assisted development: AI use among surveyed technology professionals, and the association it reports with both throughput and instability.
- Anthropic, updates to our consumer terms: which plans fall under consumer terms rather than commercial ones, and what that means for training and retention.
Since we recorded
Article 50 took effect on schedule on 2 August 2026. By the end of July 2026, about 190 organisations had signed the Code of Practice on Transparency of AI-generated Content, which the Commission and the AI Board confirmed as an adequate voluntary route to compliance. On 25 June 2026 the Commission told Amazon and Microsoft it preliminarily considers AWS and Azure should be designated under the DMA despite not meeting the quantitative thresholds, and both can respond before any final decision. On 29 June 2026 we returned to the models themselves with the same three guests in #105 EU LLMs, on whether European models are good enough to keep company data at home.
