For ten years I have been sold the same future: the cloud will be cheaper, simpler, and need fewer ops. I am still waiting.
TL;DR: the PaaS promise does not hold. Ops did not disappear, it changed names: FinOps, platform engineering, vendor management. The 2025 and 2026 outages dented the reliability argument. Real prices bury the cost argument. And AI just flipped the last one, simplicity: running your own cluster, assisted by an agent, is a reasonable choice again.
This article is for teams paying for a PaaS "to avoid doing ops". And for teams that do not dare to self-host.
Where I speak from
A PaaS, a platform as a service, runs your code without you ever seeing servers. Heroku, App Service, Railway, Render.
By day, I work on a healthcare platform with 25+ million users. Managed cloud, Kubernetes, proprietary services.
By night, I self-host everything. A full platform on Docker Swarm, a k3s cluster carrying my mail server, a mesh VPN deployed with Terraform and Ansible.
So I see both invoices, both on-call rotations and both kinds of outages. That is where I speak from.
"Fewer ops", really?
Start with the central promise: less operations. The 2026 numbers tell the opposite story.
Per the Flexera 2026 report, 63% of organizations run a FinOps team. FinOps is the discipline invented to understand and control the cloud bill. We created a job to read an invoice.
71% have a "Cloud Center of Excellence". Gartner predicted 80% of large engineering organizations would run a platform engineering team by 2026. And 85% name cost control as their top challenge.
Fewer ops? We have never paid this many people to manage what was supposed to manage itself.
My daily work confirms it. On managed Kubernetes, you no longer administer servers. You administer a vendor: forced version upgrades with deadlines, deprecated APIs, quotas, IAM. The skill did not disappear. It became vendor-specific, and perishable.
PaaS does not remove a dependency, it stacks one
Your PaaS runs on someone else's hyperscaler. You did not remove the infrastructure, you added a middleman.
On 19 May 2026, Google Cloud automatically suspended Railway's production account. A compliance robot, zero warning. About eight hours of outage, and the applications of three million users offline.
Those customers had picked a PaaS precisely to stop thinking about infrastructure. They went down because of a vendor they never chose, for a reason that had nothing to do with them.
That is the true face of this complexity: invisible on even days, uncontrollable on odd ones.
"It is more reliable", let's talk
The remaining hammer argument: a hyperscaler will always be more reliable than you. Look at the past year.
On 20 October 2025, a DNS race condition brought AWS us-east-1 down for about fifteen hours. Snapchat, Signal, payment apps, the UK tax website: all down.
Nine days later, Azure Front Door fell for about eight and a half hours, taking Microsoft 365 along. Three weeks after that, Cloudflare broke for six hours on a configuration file that doubled in size.
In May 2026, chillers failed in an AWS data hall. Coinbase stayed offline for about seven hours. Analysts now forecast major cloud outages as a yearly fact, not an accident.
I do not claim to beat their 99.99%. I claim something else. When my cluster goes down, my outage is actionable: I diagnose, I fix, I learn. When us-east-1 goes down, you refresh a status page along with half the internet.
"It is cheaper", public prices say no
Managed Kubernetes control plane (before your 1st container)
AKS, EKS, GKE: ~$0.10/hour, about $72/month
Small production app on Heroku
2 Standard dynos + Postgres + Redis: $130 to $160/month
My full self-hosted platform
Traefik, Postgres, Keycloak, monitoring, mail: ~90 EUR/month
The control plane is the part of Kubernetes the vendor manages for you. On all three hyperscalers, you pay for it before launching a single container. Free tiers exist, without an SLA.
One scale up, the movement has a name: repatriation. 37signals left the cloud and deleted its AWS account in the summer of 2025. Projected savings: over 10 million dollars across five years. And 86% of CIOs say they plan to repatriate some workloads, a record for the Barclays survey.
Honesty: only about 8% do a complete exit. Repatriation is selective, not ideological. But Flexera measures 29% of cloud spend as waste. That waste is on your bill.
What AI just changed
One serious argument remained for PaaS: knowledge. Standing up a cluster, writing the ingress, debugging DNS, keeping runbooks. That knowledge was expensive, in salary or in years of learning.
That is the argument AI just broke. My platforms are built and maintained with an AI agent as a pair. Terraform, Ansible, manifests, network diagnosis, documentation: the operational knowledge lives in the repo, and it stays current.
One important caveat: an agent under a harness. Required proofs, written rules, never live production. A free-running agent on your infrastructure is an outage in the making.
The result fits in one sentence. What used to require a full-time senior DevOps now requires a curious developer, an agent, and discipline.
Where managed still wins
Managed keeps legitimate wins. Imposed compliance, HDS or SOC 2, with certifications you do not want to carry yourself. A team where nobody can take the on-call. Huge, unpredictable traffic bursts.
The healthcare platform I work on has good reasons to run on managed cloud. At that scale, under regulatory audits, the math changes.
And self-hosting has a real cost: your time. Count it honestly. My point is not that this cost is zero. It is that its order of magnitude just changed.
The checklist before signing, or re-signing
Before renewing your PaaS or creating your cluster, answer these questions.
- [ ] Count the people who "manage the cloud" at your company, FinOps included
- [ ] List what you can fix yourself during a vendor outage
- [ ] Price the same stack on plain VMs at a simple hosting provider
- [ ] Check who holds your backups, and where they live
- [ ] Reread your vendor's last outage and note what you could have done
- [ ] Stand up a throwaway cluster with an AI agent, one weekend is enough
- [ ] Keep managed where it truly wins: compliance, on-call, burst
What to remember
The cloud is neither moral progress nor fate. It is an invoice and a dependency contract. Both deserve a reread.
The "fewer ops, cheaper, more reliable" promise does not survive 2026. FinOps teams, chained outages and public prices contradict it.
And the last argument, simplicity, just switched sides. Running your own cluster, assisted by an agent, has never been this accessible.
Want to price a repatriation, or build a clean self-hosted platform? Let's talk.
Sources: ThousandEyes, AWS outage of 20 October 2025 · Cloudflare, 18 November 2025 post-mortem · Railway, 19 May 2026 incident report · Coinbase, 7 May 2026 postmortem · Flexera, State of the Cloud 2026 · 37signals, Leaving the Cloud · GKE official pricing · Gartner, platform engineering
Top comments (0)