On Second Thought — Episode 03
In February 2011, US Chief Information Officer Vivek Kundra published the Federal Cloud Computing Strategy under the Obama administration's 25-Point IT Reform Plan. The mandate was explicit: agencies must evaluate a cloud computing option first, and come up with a damn good reason not to use it. Kundra claimed $20 billion of the government's $80 billion IT budget could move to cloud.
By 2014, Gartner's Magic Quadrant showed AWS with more than five times the cloud IaaS compute capacity of the next fourteen providers combined. By 2018, "on-premise" had become a dirty word in enterprise IT. The axiom was established.
Not discovered. Not proven. Established. By a government memo, an analyst report, and three poster-child migrations from companies (Netflix, Spotify, Airbnb) whose elastic-demand requirements bear no resemblance to the vast majority of software running in production today.
The Axiom
"Of course we use the cloud. Everyone does." Nobody questions this in a board meeting. Nobody loses a promotion for recommending AWS. CTOs choose it not because it is optimal, but because it is defensible. If something goes wrong with AWS, it is AWS's fault. If something goes wrong with your own infrastructure, it is your fault. This is not engineering. This is career insurance.
The startup ecosystem reinforces this. AWS Activate offers $1,000 to $100,000 in free credits to new companies. Incubators and accelerators distribute AWS credits as standard onboarding. The credits expire after twelve to twenty-four months. By then, your architecture is built on AWS services, your team knows AWS tooling, your monitoring assumes CloudWatch, and your deployment pipeline assumes CodeDeploy. Migration cost exceeds staying cost. The business model is identical to the IBM mainframe playbook of the 1970s: make switching costs higher than the cost of staying. The technology changed. The economics did not.
Harvard Business School research from 2018 documented the effect: after AWS launched in 2006, first-round VC funding for cloud-benefiting startups dropped 20% because infrastructure costs fell dramatically. VCs responded by funding more startups with less diligence. The cloud did not just change infrastructure. It changed who gets funded and how. And it locked in AWS as the default infrastructure for an entire generation of companies that never evaluated the alternative.
The Cost
Flexera's 2025 State of the Cloud Report found that 27% of all cloud spend is wasted. At $675 billion in global cloud infrastructure spending, that is $182 billion per year evaporating into unused resources, over-provisioned instances, and forgotten development environments. Two-thirds of organisations report waste from idle or underused resources.
The utilisation numbers are worse. The median EC2 instance runs at 7 to 12% CPU utilisation. Kubernetes clusters average 10% CPU and 20% memory utilisation. You are paying for seven to ten times the compute you actually use. In no other industry would this be considered acceptable. In cloud computing, it is considered normal.
37signals, the company behind Basecamp and HEY, left AWS in 2023. David Heinemeier Hansson documented the entire process publicly. The hardware investment: approximately $700,000 in Dell servers, fully recouped during the first year. The storage migration: 10 petabytes moved from S3 to Pure Storage, with an upfront cost of $1.5 million and annual operating costs under $200,000 (replacing $1.3 million per year in S3 charges alone). The total annual savings: $2 million. The five-year projection, revised upward from the original $7 million: over $10 million. With faster hardware and considerably more storage. AWS reportedly comped a quarter-million-dollar egress bill. One does appreciate the parting gift.
Dropbox moved 90% of its customer data off AWS to custom colocation in 2015 and 2016. The investment: $53 million in its own data centres. The savings: $75 million over two years. The return on investment was achieved before the infrastructure was fully operational.
Ahrefs, the SEO analytics company, never went to cloud. They run 850 servers in a Singapore colocation. Monthly cost per server: $1,550 on-premises versus $17,557 for the AWS equivalent. AWS would cost 11.3 times more. Over 2.5 years, the estimated savings: $400 million. Ahrefs' total revenue for 2020 to 2022 was $257.5 million. The cloud would not have reduced their margin. It would have eliminated their company.
GEICO, Warren Buffett's insurance subsidiary, spent a decade migrating over 600 applications to Microsoft Azure. Costs ballooned to 2.5 times expectations. Reliability declined. In 2024, they announced repatriation of at least 50% of workloads to an OpenStack-based private cloud, projecting 50% savings per compute core and 60% per gigabyte of storage. Completion: 2029. A decade to get in, half a decade to get out.
The Sovereignty Problem
The CLOUD Act, signed into US federal law in March 2018, allows US law enforcement to compel American technology companies to hand over data stored anywhere in the world. If your data is hosted in Frankfurt or Paris, and the infrastructure is managed by AWS, Azure, or Google Cloud, it can legally be accessed by US authorities. This directly conflicts with GDPR Article 48, which requires international agreements for third-country data access. Cloud providers face an impossible choice: comply with US warrants and breach European law, or refuse and face US legal penalties.
Europe's response was Gaia-X, a federated cloud initiative launched six years ago. What happened: US hyperscalers lobbied to be included. Once inside, they, in the words of Nextcloud founder Frank Karlitschek, "flooded it with documents and regulations." Founding member Scaleway withdrew in 2021. Day-one member Agdatahub received EUR 4.8 million in funding, then went into liquidation. Karlitschek called it a "paper monster." A fundamental failure of strategy, vision, and political will.
The NIS2 Directive, transposed in October 2024, now requires critical-sector organisations to assess cybersecurity risks from cloud providers. DORA, effective January 2025, forces financial institutions to manage ICT risk, including cloud dependency. France's "Doctrine Cloud" mandates government data stays in French-controlled facilities. The regulatory environment is turning, slowly. The infrastructure, however, remains American.
The $1,000 Test
For $1,000 per month on AWS, you get approximately four mid-tier instances (8 vCPU, 16 GB each) with no bandwidth budget. Egress costs extra. Storage costs extra. Monitoring costs extra.
For $1,000 per month on Hetzner, you get seven dedicated AX102 servers. 16 cores each. 128 GB RAM each. 20 TB bandwidth included per server. Totals: 112 cores, 896 GB RAM, 140 TB bandwidth. Independent benchmarks show Hetzner delivering 76% better multi-core performance than AWS and 11 times more IOPS.
The price difference is not a rounding error. It is a factor of seven to ten. For workloads that do not require elastic scaling (which is most of them), the cloud is not a premium for convenience. It is a tax on the assumption that you had no other choice.
The Question
86% of CIOs now plan to move some workloads back from public cloud, according to Barclays' CIO Survey. The highest figure ever recorded, up from 43% in late 2020. The axiom is cracking.
Modern servers handle 500,000 HTTP requests per second. PostgreSQL delivers 70,000 IOPS. A single well-configured machine handles 50,000 concurrent users with proper caching. The vast majority of software in production does not need elastic scale. It needs reliability, predictable costs, and control over its own data.
The cloud was never the only answer. It was the only answer nobody got fired for choosing. The career insurance premium was paid by every company that did not question the default.
On second thought: what if the default is wrong?
Read the full article on vivianvoss.net →
By Vivian Voss — System Architect & Software Developer. Follow me on LinkedIn for daily technical writing.
Top comments (0)