I would happily go back to the server in the cupboard, and let me explain why

Posted on: 23 May 2026

Let me say it upfront, before building any argument around it: I would happily return to keeping a server in the cupboard. The real one, physical, running day and night in a corner of the office, with its little light pulsing in the dark when you walk past at midnight to fetch a glass of water. Twenty years ago this was normal, then it became quaint, and now, looking at the numbers carefully, it is starting to become rational again. Not out of nostalgia. Out of calculation.

The story of the past two decades in digital can be told as the story of one great delegation. Back then I had a server under my desk and an entire rack hosted at COLT Telecom that I could visit whenever I needed to. It was normal, that was how you did things if you ran a serious online operation. Then the SaaS promise arrived: don't worry about the technical side, we will take care of it. Shopify, WordPress hosting, Salesforce, AWS, Google Cloud Run. The message was unanimous: you focus on the business, rent the infrastructure from us.

Twenty years on, we can start adding up the bill, and the bill shows that homogenisation is the hidden price of SaaS. When you sell through Shopify, your user experience is essentially identical to that of nearly five million other merchants. When your blog runs on WordPress.com, your editorial presence belongs to an architecture that decides what you can and cannot do. When your CRM lives on Salesforce, your sales data sits inside a perimeter you do not control. I don't think there is any need to continue, the point is clear enough. For years this was considered an acceptable trade-off, because in exchange you gained speed, automatic maintenance, support, but a trade-off makes sense only when differentiation matters little and operational efficiency is everything. When the wind shifts, the trade-off starts to cost.

Today the wind is shifting for three reasons converging in the same moment, and this is what makes the conversation interesting now rather than five years ago.

The first is economic: the total cost of ownership of SaaS over five-to-seven-year horizons has stopped being attractive for anyone with predictable workloads and a minimum of internal technical capability.

The second is infrastructural: consumer hardware has reached a point where it can do what twenty years ago required server rooms, and open source AI models run on that hardware without depending on any cloud.

The third is methodological, and it is the most surprising: agentic coding is lowering the technical barrier that until recently made self-hosted accessible only to those with a development team.

Let us start with the economics, because it is the least romantic and most verifiable part. Shopify Plus starts at $2,300 per month, roughly $27,600 per year, plus transaction fees, plus app fees, plus everything else. Over seven years that comes to more than $200,000 in rent paid to a vendor giving you a user experience largely identical to that of your direct competitors. Professional WordPress hosting costs less for smaller merchants, but as you grow you also pay similar rents across hosting, plugins, maintenance. The "we'll handle it" promise has turned into a perpetual rent for those who run the platforms, and a rising cost for those who use them.

The alternative is self-hosted headless commerce. Medusa, Vendure and Saleor are modular open-source backends that sit behind Next.js frontends or equivalents for the user experience. A production build of medium complexity costs between £50,000 and £170,000 upfront, amortised over five years, with hosting costs starting at around $35 per month for lean setups and reaching a few hundred for more serious architectures. The calculation is dry: if you have an in-house technical team or a stable partner, you own what you sell through instead of renting it for life. If you did not, until recently SaaS remained rational by default, but it is worth doing the maths now.

This is where the third element enters, the one that genuinely changes the geometry. Across various professional marketer and developer communities, an observation has been circulating for some months that would have sounded like a joke a year ago. Many websites that used to be built on WordPress are now built directly through agentic coding, meaning by conversing with systems such as Claude Code or equivalents, starting from templates or even mock-ups and pre-built skills, with a solid design foundation. The times reported are in the order of a few hours of work for a functioning store, not weeks. This is not hype. It is operational observation from people who build sites for a living, and who now face a productive equation different from the one of three years ago.

The numbers sit behind this observation. WordPress peaked at 43.6% of global market share in mid-2025 before falling to 42.2% by May 2026 according to W3Techs. The overall CMS share has gone from 65.2% to 60.2%, while SaaS website builders are growing at 32.6% year on year. It looks small, but it is the first meaningful decline since 2011, and across hundreds of millions of websites it represents a shift the technical community is already rationalising in two directions. On one side, smaller clients are moving to no-code SaaS builders; on the other, more technical clients are bypassing traditional CMSs altogether and having static or headless sites built through agentic coding. WordPress is not dying, but it is losing its monopoly on the middle segment.

It is worth pausing on WordPress for a moment, because it illustrates well the structural problem of the plugin model. Anyone who has managed WordPress sites of any complexity knows that the core is powerful, but the moment you start stacking plugins to obtain the functionalities you need, you lose control. Security becomes a problem, performance degrades materially, updates turn into a minefield of incompatibilities, and you find yourself constantly going back in because plugin auto-updates do not always behave as they should. In 2025, 91% of WordPress vulnerabilities came from plugins, not from the core. Those running professional high-traffic sites have over the years developed a careful process of assessment and review before adopting any plugin, however the structural paradox is that the very thing that made WordPress's fortune, its plugin ecosystem, is also what makes it heavy when you actually want to own control of your site.

Now to hardware, because this is the part that makes the conversation genuinely different from five years ago. An Intel N100 mini PC, a quad-core processor with a 6W TDP, costs less than £130 fully kitted out with 16GB of RAM and storage. It draws between eight and twelve watts under typical load, less than a light bulb, so a year of electricity costs between £10 and £20. On that machine you can run a serious homelab: a dozen Docker containers, Home Assistant, a media server, Nextcloud, a reverse proxy, possibly a small database. The community that has discovered these mini PCs over the past two years is called homelab, and it has grown considerably, because anyone capable of basic arithmetic can see that paying $300 a month for cloud services on predictable workloads makes no sense when a £350 machine does the same job for the following seven years.

For more demanding workloads, a base Mac mini M4 costs $599 and runs language models with eight billion parameters at 28-35 tokens per second. A Mac mini M4 Pro with 48GB of unified memory, under $2,000, comfortably runs 30B-parameter models. Five years ago, doing the same thing required a GPU instance on AWS costing several hundred dollars a month. Now that same power sits on a desk, draws thirty watts, and your data never leaves your network. Which, I would add, is no small thing.

The structural point is precisely this, and it is not a technical detail: for the past twenty years the narrative has been "you need cloud, because you need power you don't have", and that narrative was true and justified the entire cloud-first architecture of modern software, and today it is no longer true for many commerce use cases such as product recommendations, semantic search across catalogues, first-line customer service, automatic classification of requests, generation of product descriptions. All these workloads run today on reasonable consumer hardware, with open models such as Qwen, Llama, Mistral, without sending a single byte of your customers' data to OpenAI or Anthropic.

Anyone who lived through the personal computing cycle of the 1980s recognises this pattern. In the 1970s the centralised mainframe was the only viable answer for serious computation, and everyone paid rent to IBM, then the PC arrived and suddenly the power you needed sat on the desk. Between 2005 and 2020 the pendulum swung again: everything in the cloud, everything centralised, everything rented, and now it is starting to settle in the middle, because local hardware is once again powerful enough to handle a significant share of workloads without depending on anyone.

There is another factor that weighs, and it is data. In an era when privacy regulation has tightened, when customers increasingly ask where their data lives, when security breaches at major cloud providers have become routine, physically keeping your data on a machine you control is a positioning asset rather than a technical constraint. A small merchant in Switzerland or Italy who can truthfully state that their customers' data never leaves their office has a real commercial argument, particularly with institutional clients or discerning private ones.

None of this means that everyone should return to the server in the cupboard tomorrow. That would be both stupid and untrue. SaaS remains rational for anyone with unpredictable workloads requiring rapid scaling, for anyone who prefers paying rent in exchange for predictability, for anyone unwilling to spend an hour learning how a Docker container works. I am not waging a crusade against Shopify or WordPress, I am simply pointing out that the mathematics has changed, and that anyone thinking about their stack the way they did in 2018 is missing a significant part of the conversation.

The picture I see is a structural bifurcation of the market in which on one side a mass market based on standardised SaaS will remain, where smaller merchants will continue to pay rent to Shopify, BigCommerce, WordPress, because the cost of internal management exceeds the cost of subscription, and on the other side a sophisticated segment will emerge of those who reclaim control over the three fundamental layers: commerce backend, frontend experience, AI models. That segment will use control as a structural moat, because in a world where AI agents are starting to do the buying for users, where differentiation matters more than time to launch, where data sovereignty becomes a commercial argument, owning the stack is worth considerably more than it was in 2018.

The surprise in this bifurcation is that the sophisticated segment will not be composed only of large players with substantial budgets, because thanks to agentic coding, even small and medium merchants with a minimum of technical curiosity and a solid starting design can now build their own infrastructure at accessible costs. This widens the pool of potential "owners of their own stack" well below the mid-market threshold that until yesterday was the barrier to entry. For the first time in twenty years, owning has become a plausible choice even for those without a technical team.

The strategic question for the years ahead is not "self-hosted or SaaS?". It is more austere: "do you want to rent for life, or do you want to own?". For two decades the first answer has been obviously the more rational one for almost everyone, but it is no longer so obvious, and when the evidence changes, changing one's mind is not inconsistency but rather intelligence.

I would happily return to the server in the cupboard. And looking at the numbers without romanticism, I am beginning to think I would not be the only one.