Bringing the Cloud Back to Earth: The Internet Is Buildings, Electricity, and 0s and 1s

I used to work at a company that built on-premises software in Rust. “On-premises” means installing and running software in a self-managed server environment that the customer controls directly, rather than in a cloud provider’s environment. It was at that company that I first saw a dedicated server room inside an office. Until then, a server had been an abstract thing to me, something that existed only on my monitor as an IP address or a login. I already knew, in my head, that beneath all software, however abstract, there is a physical computer. But it was in that server room, I think, that the fact finally became real to me, seen with my own eyes.

These days I’m working as an on-site freelancer, moving a service that was delivered from the cloud onto a local, on-premises setup. Things that used to take a few clicks in the cloud console now have to be put together one by one, and the physical layer I had always waved off with a “surely it’s there somewhere” is rising back into view.

Beneath the Cloud, There Is a Building

In 2006, Amazon launched S3 for storage and EC2 for virtual servers, and the “cloud” services we use so casually today began to spread in earnest. Hosting services that let you rent servers existed before that, but the model of provisioning servers and storage through code, when you need them, and paying only for what you use took off from this point. You no longer had to buy hardware, find a place to put it, and run a connection to it. With an account and a payment method, getting a server became far easier. The origin of the name “cloud” is also interesting. From what I’ve read, it comes from an old convention in network diagrams, where a part of the network whose inner workings didn’t need to be drawn out, such as the internet, was sketched as a cloud. The cloud was never a sign that nothing was there. It meant: you don’t need to look inside here.

But just because something has been erased from the diagram doesn’t mean the physical things that make it possible have disappeared. The cloud is, in the end, renting servers that sit in someone else’s data center. Whether you’re on-premises or in the cloud, a server physically sits somewhere. What differs is whose building it is, who pays the electric bill and deals with the failed disk, and how you secure more resources and scale up when you need them.

What Became Visible Again When I Moved From the Cloud to On-Premises

What struck me most about this work was the difference in what you are directly responsible for. Cloud providers describe this relationship with something called the shared responsibility model: the provider manages the building, the power, and the hardware, while the operating system, applications, data, and configuration on top are split between provider and customer depending on which service you use. Moving on-premises, a large share of that responsibility comes back inside the organization. Someone has to replace the disks, take the backups, manage the network equipment, and apply the security patches.

The part that hits hardest in a migration is managed services. In the cloud, it’s common to consume databases, object storage, message queues, load balancers, and authentication as services. When you move on-premises, you have to decide anew how to keep providing what those services used to do. You might replace them with open-source software you install yourself or with commercial products, or you might change the architecture to merge some functions or drop them altogether. You only see how much was sitting behind a single checkbox once the checkbox is gone. The more deeply a system leans on one provider’s proprietary services, the harder it is to move, and that is the point where the phrase “vendor lock-in” stops being abstract.

Whichever way you go, you need space for the servers, electricity, cooling, and a network. The cloud isn’t a technology that removes that burden so much as a service that hands it to someone else and makes it invisible. So where does that responsibility end up in the cloud?

What Does a Data Center Actually Do?

The time I heard the words “data center” most loudly in the news, as I remember it, was in October 2022, when a fire at a data center in Pangyo, near Seoul, took Kakao’s services offline for a long time. It wasn’t only KakaoTalk, the messaging app that much of South Korea uses every day. Taxi hailing and payments went down at the same moment, and it made plain how many seemingly separate services had been depending on one building all along. Looking into it again, I found that the fire reportedly started in the batteries, and that afterward the question of whether the company had enough redundancy and recovery systems in place became an issue too, not just the fire itself.

A data center is a building that gathers servers, meaning computers, in shelf-like frames called racks, stacked layer upon layer, and provides everything needed to keep those machines running around the clock. First, there is electricity. Layers of backup keep the servers up if the grid fails: uninterruptible power supplies (UPS) and batteries, then backup generators to take over after them. Much of the power a server consumes ends up as heat, so the cooling equipment that carries that heat away needs a great deal of power as well. Telecom lines come in along multiple routes, and there are fire suppression systems, access controls, and security guards. To keep things running reliably, power, cooling, and network connections are often made redundant, designed so that if one piece of equipment fails, another picks up its role immediately.

Cloud providers operate physically separate availability zones to limit the impact of failures. An availability zone consists of one or more data centers, and if you design a service to be spread across several availability zones, you are more likely to be able to keep it running even when one facility goes down.

The server room I first saw at that company was a very small version of a data center. It’s a rare experience at a startup, but plenty of places that need to keep their data inside their own buildings, such as banks, hospitals, factories, and public agencies, build and run their own server rooms, even if they are nowhere near data-center scale. At the other end, the large data centers run by cloud providers hold thousands or tens of thousands of servers, and I’ve read that the biggest campuses are planned to draw as much power as a city. The scale differs, but what’s needed is the same: machines, electricity coming in, heat going out and being cooled, and lines to connect to the outside. A cloud price list ultimately includes the cost of keeping all of this running day and night.

The Smallest Server Room Is on My Desk

If a data center is the biggest home a server can have, the smallest one is on my desk. Developers often talk about spinning up a server locally. To spin up a server here means to run a program that receives requests, processes them, and sends back responses (serving), and “local” means that program is running on the computer in front of me. The moment I run a server program, my computer’s CPU, memory, disk, and electricity are used just as they are. Opening localhost in a browser isn’t a trip out to somewhere on the internet. It’s my computer talking to itself, which is why that address is called the loopback address. If I shut down my laptop, the server stops with it. Unlike a data center, which is built to run around the clock, my laptop was never meant to play that role.

Seen this way, a development server running locally, an on-premises server run by an organization, and a server running in the cloud all work on the same principle: a program that receives and handles requests is running on a computer somewhere. What differs is where that computer is, who manages it, and how it connects to the outside world. Moving a service from the cloud to on-premises is, in the end, bringing a program that ran on a cloud provider’s infrastructure into a server environment the organization controls. The scale and the way it’s operated differ, but running a server program on a computer is the same principle as spinning up a local server on a laptop. And yet, even having come down this far, the word “server” still remains something of a black box.

Beneath the 0s and 1s Are Electricity and Light

Going one level deeper into the server, the CPU holds billions of transistors, and a transistor is a tiny switch operated by electricity. People often explain 0 and 1 as electricity being connected or cut off, but what a computer actually handles are two electrical states that a circuit can tell apart, such as a high voltage and a low voltage.

Memory works the same way. The main memory, DRAM, stores information as charge held in tiny capacitors. Because the charge leaks away little by little, it has to be refreshed periodically even while the power is on, and when the power is cut, the information is lost. SSDs rely on the electrical properties of charge stored in flash memory cells, and hard drives record data using the magnetic state of a magnetic material. Even information that looks formless on a screen, like a line of chat or a single photo, always exists inside a server as the state of some physical material.

Then there is “wireless internet.” The term is so widespread that it’s easy to assume the internet really is wireless, but Wi-Fi only connects your device to the router. Behind the router, in a typical home with a wired connection, an Ethernet cable runs to the building’s telecom room, and from there fiber-optic cable runs through the ground. In fiber-optic cable, data is carried not by electrical signals but by changes in light, and that light is converted back into an electrical signal at the destination equipment. In the end, the road out to the rest of the world is also a matter of sending electrical and optical signals along a wire.

Communication between continents is mostly carried by submarine cables laid on the seabed. You might expect satellites to link the continents, but from what I’ve read, the overwhelming majority of international data traffic passes through submarine fiber-optic cables, and there are landing stations on the Korean Peninsula where those cables come ashore. Whenever I remember that a single line of text has to travel along cables on the ocean floor to cross the sea, I think about how many physical facts the word “internet” covers up.

Closing Thoughts

Moving a service from the cloud to on-premises means facing, one by one, these layers that used to be out of sight. Setting up by hand what used to take a few clicks in a console shows me how much equipment and how many people it takes for even a single app to open. Whatever architecture I choose from here on, I plan to start by checking which building, which machines, and whose electricity a service is running on.

These days, a single message on my screen no longer feels entirely weightless. Behind that one line are the building where the server sits, cooling units that run through the night, charge inside chips, and optical fibers beneath the sea, along with the facilities and people who make and supply the electricity. Perhaps one of the greatest things the cloud has done is let us stop thinking about all of this every time. But what has become invisible has not gone away.