A Data Center Is a Factory: The Honest Case For and Against

I spent years in manufacturing before I moved into AI, IT, and teaching. The habit that stuck with me is this: if you want to understand a plant, read its utility bill before you read its brochure.

That habit is why the data center boom reads differently to me than it does in most headlines. Strip away the word "cloud" and a data center is a factory. Its product is compute. Its raw materials are electricity, water, and land. It runs three shifts forever, and its entire business case depends on never stopping. Once you see it that way, the arguments for and against building one get much easier to weigh honestly.

The Philippines is now building these factories at pace, including a large one about twenty minutes from where I am sitting. So it is worth being precise about what they give a country, what they cost it, and what separates an operator doing this well from one doing it badly.

Start with the scale, because it explains everything else

The International Energy Agency put global data center electricity use at roughly 415 terawatt hours in 2024, about 1.5 percent of world electricity. Its base case has that roughly doubling to around 945 TWh by 2030, and reaching about 1,200 TWh by 2035. That is growth of around 15 percent a year, more than four times faster than electricity demand from everything else (IEA, Energy and AI, 2025).

Two things in that paragraph deserve attention, and the second one matters more.

The first is obvious: that is a lot of power. The second is that only one of those three numbers is a measurement. The 2030 and 2035 figures are projections from a base case, carrying assumptions about chip efficiency, model demand, and how quickly grids can connect new load. I built the chart at the top of this article to keep that visible: solid bar for measured, faded and dashed for projected. That distinction gets flattened constantly in coverage of this industry, and flattening it is exactly how people end up arguing about forecasts as if they were facts.

The factory frame, and the one metric worth learning

In manufacturing we lived by OEE, overall equipment effectiveness: how much of your theoretical output you actually captured. Data centers have their own version, and it is called PUE, power usage effectiveness. It is total facility power divided by the power that actually reaches the computers.

A PUE of 1.0 would mean every watt entering the building did useful computing work. Real facilities land above that, because cooling, power conversion, and lighting all consume energy without processing a single query. So a PUE of 1.5 means half again as much power as the servers themselves need. A PUE of 1.2 means the overhead is down to a fifth.

That single ratio is the closest thing this industry has to an honest efficiency score, and it is where the engineering competition actually happens. When you hear about liquid cooling instead of air, that is a PUE argument. It also tells you something useful as a reader: an operator who publishes its PUE is inviting comparison, and an operator who talks only in adjectives is not.

What a data center genuinely gives a host country

Latency and data residency. Physics does not negotiate. If your application, your bank, or your government service runs on servers in Singapore, every request pays for that round trip. Hosting compute locally makes local software measurably faster, and it lets regulated industries keep data inside the country where the rules apply.

Serious capital, and the infrastructure it drags along. ST Telemedia's STT Fairview facility is reported at around 1 billion dollars for 124 megawatts, which would make it the country's largest (w.media). Money at that scale does not arrive alone. It pulls submarine cable landings, substation upgrades, and fiber routes with it, and those outlast any single tenant.

A real construction phase. Build work runs roughly 0.7 to 2.0 workers per megawatt, and a hyperscale campus can employ one to three thousand people across an 18 to 36 month build (iRecruit staffing analysis). For construction trades, that is genuine work.

A floor under a domestic tech industry. You cannot build a serious local cloud or AI sector while renting every rack abroad. Compute on home soil is a precondition, not a guarantee, and that distinction comes back later in this piece.

What it actually costs

Power, and in the Philippines this is the binding constraint. The Department of Energy reported that the country's average electricity price reached 12.43 pesos per kilowatt hour in June, edging past Singapore by 9 centavos to become the most expensive in Southeast Asia (BusinessWorld, July 2026). Set that beside the IEA growth curve and the tension is plain: the most power-hungry industry of this decade is arriving in a market with the region's highest power prices and a grid that has already been strained by plant outages. That is not an argument against building. It is an argument for being clear-eyed about who absorbs the cost.

Very few permanent jobs. This is the number most often left out of announcements. Once running, data centers employ roughly 0.15 to 0.35 full-time staff per megawatt. Facilities above 100 MW can operate with 20 to 40 permanent people. Google's 500-megawatt campus in Kansas City is expected to create about 1,000 construction jobs and 200 permanent ones (Data Center News). A 100 MW facility is an enormous piece of infrastructure and a small employer at the same time. Both halves of that sentence are true, and any pitch that mentions only one of them is selling.

Water, land, and neighbours. Cooling consumes water, siting consumes land, and generators and chillers make noise. These are ordinary industrial externalities, familiar to anyone who has worked in a plant. They are manageable. They are not zero.

What doing it right looks like: Schneider Electric

Schneider Electric was named the world's most sustainable company by TIME and Statista in 2024, scoring 88.86 out of 100 (The Manufacturer), and it has now held the top position for three consecutive years, through 2026 (Business Wire). It has also been on CDP's Climate A List for 15 straight years.

Before treating that as settled proof, it is worth knowing how the ranking is built, because reading a ranking properly is itself an analyst skill. TIME and Statista start from a pool of more than 5,000 large companies, apply a four-stage screen that excludes certain industries outright, weigh external sustainability ratings, reporting practices, and environmental and social indicators, and publish 500 names. So it is a rigorous, methodology-driven ranking of disclosure and performance among large listed companies. It is not a measurement of total planetary impact, and no ranking is.

What makes Schneider a genuinely instructive case is narrower and more durable than the trophy. Its sustainability position is load-bearing for its business model rather than decorative. The company sells the efficiency layer of the data center itself: uninterruptible power supplies, cooling, and the EcoStruxure software that monitors it. Its Galaxy VXL UPS is built to occupy about 52 percent less space than the industry average, it publishes reference designs for racks up to 132 kW, and it acquired Motivair to move deeper into liquid cooling. Its data center segment grew around 10 percent organically in 2025 (Data Centre Magazine).

Read that as an engineer rather than a fan and the lesson is simple. When a company's revenue rises because its customers use less power per unit of output, efficiency stops being a communications exercise. That alignment, not the award, is the thing worth looking for when you assess any operator.

What doing it wrong looks like: the Memphis and Southaven case

The clearest cautionary case in this industry right now is xAI's Colossus build, and it needs to be described carefully because it is unresolved litigation, not a settled finding.

The documented sequence is this. Colossus 1 in Memphis began operating in June 2024 using as many as 35 gas turbines without an air permit. The Southern Environmental Law Center, acting for the NAACP, issued a 60-day notice of intent to sue, after which xAI removed the unpermitted turbines from that site (SELC). Then on 14 April 2026, the NAACP, represented by Earthjustice and SELC, filed a Clean Air Act suit in the Northern District of Mississippi over turbines running at the Colossus 2 site in Southaven. The complaint alleges 27 turbines operating without a permit and asks the court to halt them, require Best Available Control Technology, and impose civil penalties of roughly 124,000 dollars per day per violation (NAACP, Earthjustice). Reuters reporting put the turbine count higher than the company had acknowledged. The plaintiffs emphasise that the site sits near homes, schools, and churches.

Those are allegations in an active case. No court has ruled on them. I am not presenting a verdict, and you should be wary of anyone who does while a case is still live.

What is already instructive, though, is where the failure sits. Nothing here is a story about compute being inherently dirty. It is a story about permitting, siting, and who lives downwind. The turbines were a workaround for the real bottleneck, which was grid capacity, and the workaround outran the paperwork. For a country deciding how welcoming to be, the transferable lesson is uncomfortable and useful: the promises matter far less than the permitting regime and the capacity to enforce it.

So what does it actually mean to host one?

Here is where I will be careful, because this is where enthusiasm usually gets ahead of the evidence.

Hosting compute is not the same as owning compute. A data center on Philippine soil can be full of servers owned by foreign companies, serving foreign users, taxed under negotiated terms. That is not a criticism, it is a description. The question that determines whether the country captures real value is which layer it participates in: the concrete and cooling, or the analytics, software, and models running inside.

Even the local numbers disagree, and that is the most useful thing on this page. Look at what published sources say about Philippine capacity. The DICT has said capacity may reach 1.5 gigawatts by 2028 (BusinessWorld). The Data Center Operators of the Philippines has described going from about 150 MW to 473 MW over two years. Commercial research firms have published figures such as 560 MW in 2025 rising to 1.3 GW by 2030, and separately about 633 MW rising to 853 MW by 2030.

Those cannot all be describing the same quantity, and they are not. Some count IT load, others total facility capacity. Some count what is energised today, others what is contracted, permitted, or merely announced. None of them is necessarily lying. They are answering different questions with the same word.

This is the single most transferable skill in this entire article, and it is the one I teach first. Before you compare two numbers, find out what each one counts. In practice that means asking three questions of any figure you are handed: what exactly is being measured, as of when, and who benefits if I believe it. Anyone who can do that reliably is already more useful than most dashboards.

Power policy will decide this, not marketing. Given the electricity prices above, the realistic Philippine advantage is not cheap power. It is location, cable landings, a large English-proficient workforce, and proximity to demand. Those are real. They are also not infinite, and the power question does not resolve itself.

The bigger board this sits on

There is a term that has been gaining ground among strategic analysts for the world order now being shaped by control of semiconductors: Pax Silica. It deliberately echoes Pax Britannica and Pax Americana, but replaces naval and military dominance with the ability to design and fabricate advanced chips (Real Instituto Elcano, GIS Reports).

The part worth holding onto is how wide the stack is. It runs from critical minerals and energy through fabrication and packaging, to compute, frontier models, software platforms, fiber and subsea cables, data centers, and logistics. A data center is one visible tile in that mosaic, which is precisely why it is a bad place to stop thinking.

I am going to spend the next few articles walking that stack: where the Philippines actually sits in it, which layers are realistically open to us, and which are closed for reasons that have nothing to do with talent. This article was the foundation, because you cannot reason about the geopolitics of compute until you can read a facility's power bill.

What I would do with this if I were starting out

If you are a student or a working professional in the Philippines watching these buildings go up, the honest read is this. The number of people who will ever work inside one is small. The number who will work on what they enable is not.

Those 200 permanent roles at a 500 MW campus are real careers, and mostly specialised facilities and electrical engineering. The much larger opportunity is one layer up: the analytics, cloud, automation, and AI work that exists because the compute is now nearby and cheaper to reach. That is the layer where a self-taught person with a portfolio can compete, and it is the layer I would point almost anyone toward.

Which is also why I keep insisting that people learn to interrogate a number before they learn to make a chart. This whole article was an exercise in it: one measured figure and two projections, four capacity estimates that disagree, one ranking that means something specific and narrower than it sounds, and one case where the allegations are serious and the verdict is not in yet. Getting comfortable with that kind of uncertainty is not cynicism. It is the job.

If you want to practise exactly that, the free resources here are built around it, messy real datasets and scored judgment calls rather than tidy examples. Start with whichever one looks least comfortable.