The World Is Being Rewired: A Data Center Tour of Every Region on Earth

I grew up in Kathmandu when load-shedding was just part of the vocabulary. There were winters when we had scheduled power cuts of twelve, fourteen, sometimes sixteen hours a day. You learned to charge everything the moment electricity came back, to do your homework by the blue light of an inverter, to treat a fully charged laptop like a small fortune. So when I read that a single data center in Texas is being built to draw 1.2 gigawatts — enough electricity for roughly a million American homes — I feel something that is hard to put into one word. Awe, partly. Unease, mostly.

Because here is the strange irony I keep turning over: Nepal exports hydropower to India, sitting on some of the richest untapped hydroelectric potential on the planet, and yet we have almost no serious data center footprint to speak of. Meanwhile the rest of the world is pouring hundreds of billions of dollars into windowless warehouses full of humming chips, and fighting over every spare megawatt to feed them. I wanted to understand that whole picture — not just the headline about ChatGPT, but where the machines physically live, region by region. So I spent a good while reading through reports from the IEA, Synergy Research, CBRE, Goldman Sachs, and a stack of company earnings calls. This is what I found.

First, the scale — because the numbers are genuinely hard to believe

Let me set the stage with a few figures that reframed how I think about the internet.

There are somewhere north of 11,800 data centers operating worldwide as of late 2025, though the count depends entirely on how you define one. The United States alone hosts the largest share — depending on the database, somewhere between roughly 3,900 and 5,400 facilities — which is more than the next ten countries combined. The global data center market was valued at around $527 billion in 2025 and is projected to reach roughly $739 billion by 2030.

But the market value almost undersells it. The real story is capital expenditure. In 2025, the four biggest American hyperscalers — Amazon, Microsoft, Google, and Meta — spent a combined figure in the range of $388-410 billion on capex, most of it digital infrastructure. For 2026, their combined guidance climbed to roughly $725 billion, up around 77% year-over-year. Add Oracle and the "big five" total approaches $700 billion for 2026 alone. Amazon alone is guiding to about $200 billion in 2026. In its June 6, 2026 report Tracking Trillions, the Goldman Sachs Global Institute modeled $7.6 trillion in cumulative AI infrastructure capital expenditure between 2026 and 2031 — split into roughly $5.1 trillion for compute, $2.1 trillion for data centers, and $358 billion for power — with annual spend rising from $765 billion in 2026 to $1.6 trillion by 2031. These are numbers that used to belong to national governments, not individual companies.

Hyperscaler capex by company, 2025 vs. 2026 (USD billions):

Data from company earnings calls, Q4 2025 / early 2026.

What changed? AI. Specifically, the shift from training big models occasionally to running inference — answering billions of queries a day — continuously, at scale. That is the engine behind everything that follows.

The AI mega-projects: gigawatts as the new unit of ambition

A few years ago, a large data center was measured in tens of megawatts. Now the frontier players talk in gigawatts, and they say it casually.

The flagship is Stargate, the OpenAI–Oracle–SoftBank venture announced at the White House in January 2025 with a headline commitment of $500 billion and 10 gigawatts of capacity. Its first site in Abilene, Texas is already running, with the campus targeting 1.2 GW and eventually more than 450,000 Nvidia GB200 GPUs, according to Oracle's Larry Ellison. By late 2025, OpenAI said the program had expanded to nearly 7 GW of planned capacity across sites in Texas, New Mexico, Ohio, Wisconsin, and Michigan. Not everything has gone smoothly — a planned 600 MW expansion near Abilene was scrapped in early 2026 after financing negotiations stumbled and winter weather disrupted the liquid-cooling infrastructure, which tells you these are physical projects subject to physical problems, not just press releases.

Then there is xAI's Colossus in Memphis, Tennessee. Elon Musk's team built the first 100,000-GPU cluster in 122 days, which insiders called impossibly fast. By late 2025, Musk announced the purchase of a third building to push the site toward nearly 2 GW and a reported 555,000 Nvidia GPUs, at a cost of around $18 billion for the chips alone. The Memphis story is also a cautionary one: xAI ran dozens of gas turbines on-site — aerial imagery in April 2025 showed 35 turbines at a combined 422 MW — partly by exploiting a rule that only regulates generators kept in one place for more than 364 days. Local groups pushed back hard on the air quality and grid impacts.

The pattern repeats globally. HSBC estimated more than $2 trillion in AI infrastructure has been planned around the world. Whether all of it gets built is, frankly, an open question I'll come back to.

The region-by-region tour

North America — the epicenter, and the friction

The United States is the undisputed heart of this. Northern Virginia's "Data Center Alley" — centered on Loudoun County — handles something like 70% of the world's internet traffic and hosts several hundred facilities. Texas is the new gigawatt frontier, thanks to cheap land, its own grid (ERCOT), and a permissive attitude. Together, the US consumed about 183 TWh of electricity for data centers in 2024, roughly 4.4% of the national total.

But the friction is real and growing. In Virginia, data-center demand drove an 833% jump in the PJM grid's capacity auction price for 2025-2026, and residents are seeing it on their bills. The state approved a new $600 million electricity-use tax on data centers in 2025 and created a special "GS-5" rate class for users over 25 MW, effective 2027, to shield households. Nationwide, community resistance had blocked or delayed around $64 billion in projects as of March 2025 according to Data Center Watch. This is no longer a fringe concern; it's mainstream politics in several states.

Latin America — the renewable-powered challenger

Latin America is having a genuine moment. Brazil leads, with roughly 75% of its electricity coming from clean sources (mostly hydro), which makes it magnetic for sustainability-conscious hyperscalers. São Paulo is the regional hub. Chile has attracted a $4 billion AWS commitment for a new cloud region, leaning on its stable grid and submarine-cable connectivity. Mexico is riding the "nearshoring" wave — AWS announced a $5 billion investment in Querétaro, which has become a real cluster. The Latin America construction market was valued at around $1.58 billion in 2024 and is projected to reach $4.40 billion by 2030.

But the same 2025 research that celebrates the boom flags a warning: studies of data centers in Chile and Mexico found local environmental stress, with water consumed faster than communities could track or contest. In Chile, the government quietly let some developers bypass environmental assessments.

Europe — the mature FLAP-D core and Ireland's cautionary tale

Europe's traditional heart is the "FLAP-D" markets: Frankfurt, London, Amsterdam, Paris, and Dublin. Germany leads the continent by facility count. France is growing fast — its operational capacity jumped 32% in 2025.

But the story everyone should study is Ireland. According to Ireland's Central Statistics Office (in its Data Centres Metered Electricity Consumption 2024 release of June 10, 2025), the share of metered electricity used by data centres rose to 22% in 2024, up from just 5% in 2015 — with quarterly consumption climbing from 290 GWh in early 2015 to 1,829 GWh by the end of 2024, an increase of 531%. (The CSO's July 2026 release put the 2025 figure at 23%.) In the greater Dublin area, where about 97% of Irish data centers cluster, they account for roughly half of regional electricity demand. EirGrid forecasts the national share could hit 31% by the early 2030s. Ireland effectively imposed a moratorium on new Dublin grid connections, then in late 2025 rolled out a new framework requiring large data centers to source at least 80% of their power from new renewables. It's a live experiment in whether a small grid can host the world's data without wrecking its own climate targets or household bills.

The brighter European story is the Nordics, which I'll cover under technology below, because their waste-heat reuse is one of the few genuinely hopeful things in this whole industry.

The Middle East — sovereign wealth meets desert compute

The Gulf states are spending oil money to become AI powers. Saudi Arabia launched Humain in May 2025, a Public Investment Fund company chaired by the Crown Prince, targeting something like 6 GW of capacity over the next decade and aiming to be "the third-largest AI provider in the world" after the US and China. It has signed partnerships with Nvidia (up to 500 MW), AMD, Qualcomm, AWS ($5 billion), Google Cloud ($10 billion via PIF), and xAI. The UAE unveiled a 5 GW AI campus in Abu Dhabi — described as the largest outside the US — and hosts the region's own Stargate node through G42.

The logic is straightforward: cheap electricity (around $0.05-0.06/kWh versus $0.09-0.15 in the US) and vast land. The obvious tension is cooling data centers in desert heat, and the fact that AI engineering talent is genuinely hard to recruit to Riyadh.

Africa — the "final digital frontier"

Africa remains strikingly under-served: as of mid-2025, the continent hosted roughly 223 data centers across 38 countries — well under 1% of the global total. South Africa dominates with about 40% of continental capacity, anchored by Johannesburg and Cape Town, with cloud regions from AWS, Azure, and Google. Kenya stands out for a renewable-heavy grid (over 60% geothermal, wind, solar, hydro) and a Microsoft–G42 geothermal-powered campus. Nigeria has the population and fintech demand but a punishing power problem — its national grid has rarely exceeded 6 GW for 230 million people. The African market was valued at around $3.49 billion in 2024 and is projected to reach $6.81 billion by 2030. The recurring theme across every African market: the constraint is not demand, it's reliable electricity.

South Asia — India's surge, and the Nepal question close to home

India is the region I watch most closely, for obvious reasons. Its data center capacity roughly doubled in 2025, with new supply more than doubling to 387 MW added in a single year, pushing operational stock past 1,500 MW. Mumbai (and Navi Mumbai) is the anchor, holding about 47% of national capacity, followed by Chennai and Hyderabad. The market is projected to more than triple to over 4 GW by 2030. AI is now the dominant driver — reportedly 78% of leasing activity in 2025, up from 23% a year earlier — helped along by data-localization rules under the Digital Personal Data Protection Act. Investment commitments are staggering: Google's $6 billion Andhra Pradesh project, AWS's $12.7 billion by 2030, AirTrunk's planned $30 billion.

And then there is Nepal. This is where it gets personal. Nepal's newly elected government — the Rastriya Swatantra Party, which won the March 2026 elections — has placed digital infrastructure at the center of its economic vision, dreaming of turning the country into a hub for data centers and AI powered by hydropower. There are early projects: Ncell built the country's first certified data center, and Bichuten Data Vault has planned Tier IV facilities in Kathmandu, Birgunj, and Chobhar, reportedly with a Google Cloud partnership. On paper the logic is beautiful — Nepal generates clean hydroelectricity, has cool highland air, and could theoretically host green compute.

But I'd be lying if I said I'm not skeptical. As a Kathmandu Post investigation put it bluntly in April 2026, "Nepal wants to become a data centre hub. It has no rules for how." The country's environmental laws don't even mention data centers. Our transmission grid is patchy, our dry-season power supply is tight, and we have no experience connecting hydropower to a hyperscaler. An infrastructure advisor quoted in that piece said plainly that Nepal doesn't yet have the evidence the economics work. I want the hydropower-to-computing dream to happen. I just want it done with rules written before the concrete is poured, not after — which, sadly, is the opposite of what happened in Chile and Mexico.

East Asia — China's parallel universe, Japan and Korea's grid scramble

China is essentially a second, parallel data center universe. It's the world's second-largest market, consuming around 100 TWh of electricity for data centers in 2024 (about 25% of the global total). Its signature policy is "Eastern Data, Western Computing" (东数西算) — a national scheme launched in 2022 that routes compute-heavy workloads to eight national hubs in the resource-rich west (Guizhou, Inner Mongolia, Gansu, Ningxia) while keeping low-latency work near eastern cities. China is also reportedly drafting a $295 billion plan to build a national AI data center grid running largely on homemade silicon, a direct response to US chip export controls.

Japan has the largest data center capacity in Asia-Pacific — around 3.3 GW of IT load in 2025 — centered on Tokyo and increasingly Osaka. AWS has committed roughly $15 billion through 2027, and SoftBank announced a sovereign-cloud joint venture with Oracle in October 2025. The constraint is familiar: a multi-year power-connection queue in the metros.

South Korea made one of the boldest announcements of all. In June 2025, President Lee Jae-myung unveiled a national AI push targeting 18.4 GW of AI data center capacity by 2035, with an initial ~KRW 550 trillion phase led by SK Group, GS, and Naver. The catch is a serious grid problem — the Yongin semiconductor cluster alone needs 15-16 GW, and moving power across the country has historically taken decades.

Singapore deserves special mention for discipline. After a de facto moratorium from 2019-2022, it now runs one of the strictest regimes in the world. Its Green Data Centre Roadmap, launched in 2024, offers "at least 300MW of additional capacity in the near term, or more with green energy deployments," paired with tough efficiency targets (PUE ≤ 1.3). It's the rare government treating data center capacity as something to be rationed and greened rather than maximized.

Southeast Asia — Johor's oil-palm-to-servers transformation

The most dramatic regional story might be Johor, Malaysia. When Singapore froze new builds, hyperscalers jumped across the causeway. Johor went from an afterthought to Southeast Asia's fastest-growing data center hub, with live capacity around 487 MW and a pipeline of several gigawatts. Malaysia approved 143 data center projects worth around RM 144 billion between 2021 and mid-2025. Fields that were oil-palm plantations a few years ago are now filled with servers.

But the backlash has already started. By late 2025, Malaysian civil-society groups were openly calling for a moratorium, pointing out that data centers had secured electricity supply agreements totaling 5.9 GW — roughly 43% of the national utility's contracted capacity — while ordinary Malaysians still face water rationing. Johor has reportedly rejected close to 30% of applications over resource concerns. It's the Ireland story playing out again, faster.

Oceania — Australia's investment magnet, New Zealand's clean grid

Australia has become a global top-five market, with deployable capacity around 1,350 MW in 2024 projected toward 3,100 MW by 2030. Sydney and Melbourne are the hubs. The signature deal was Blackstone's roughly $16 billion acquisition of AirTrunk in 2024 — the largest infrastructure transaction in Australian history. AWS committed AUD 20 billion, and OpenAI signed on with NEXTDC for a AUD 7 billion Sydney campus. New Zealand is smaller but interesting for one reason: clean power. Both AWS (a NZ$7.5 billion region) and Microsoft (launched December 2024) built regions explicitly running on New Zealand's ~80% renewable grid, with Microsoft's facilities designed to be 100% water-free using air cooling.

The frontier — underwater, Arctic, and (maybe) space

Then there are the genuinely strange ideas. Underwater data centers had their moment: Microsoft's Project Natick proved servers on the seafloor were eight times more reliable than on land, but the company quietly confirmed in 2024 it wasn't building any more. China picked up the baton — a $226 million commercial underwater project off Shanghai's Lingang district aims for 24 MW cooled by seawater and powered largely by offshore wind. The Arctic and Nordic cold-climate sites are the more practical version of this idea (below). And space-based data centers remain, for now, mostly proposals and press releases — SpaceX has floated a factory for orbital data centers, but I'd file this firmly under "watch, don't believe yet."

The part that keeps me up at night: electricity, water, and carbon

Here are the numbers that matter most, and where I'd point anyone who wants to understand the stakes.

According to the IEA's Energy and AI report (April 2025), data centres "accounted for around 1.5% of the world's electricity consumption in 2024, or 415 terawatt-hours (TWh)," and this is "set to more than double to around 945 TWh by 2030" — close to 3% of the world's power and roughly equal to Japan's entire electricity consumption today. The US (45%), China (25%), and Europe (15%) led 2024 consumption, and the US and China together account for nearly 80% of the growth to 2030. AI is the primary driver: electricity for AI-focused servers is projected to grow around 30% a year.

Global data center electricity consumption, 2020-2035:

Source: IEA, Energy and AI (2025). 2024 marks the inflection point where annual growth accelerates.

The water side gets less attention but is just as visceral. Depending on climate and cooling technology, data centers evaporate between about 1 and 9 liters of water per kWh of server energy. US data centers directly consumed around 17.4 billion gallons in 2023, projected to rise sharply. Google's facilities alone withdrew about 7.8 billion gallons in 2024. Training GPT-3 was estimated to evaporate around 700,000 liters of freshwater on-site. On the per-query question, researchers at UC Riverside and UT Arlington (Shaolei Ren, Pengfei Li et al., "Making AI Less 'Thirsty,'" reported by The Washington Post in September 2024) estimated that writing a 100-word email with GPT-4 consumes approximately 519 milliliters of water — about a full bottle — as a full-lifecycle figure that includes both on-site cooling and the water used to generate the electricity. Newer estimates for simple text queries are much lower, and the industry is fighting hard to bring this down. In Memphis, xAI is building an $80 million plant to reuse 13 million gallons of wastewater a day precisely because the water math is politically radioactive.

Top countries by number of data centers, 2025:

Sources: Statista, Cloudscene, Visual Capitalist (2025). The wide range for the US and China reflects differing counting methodologies.

On carbon, the IEA's central case actually offers some perspective: data-center emissions might reach about 1% of global CO2 by 2030. That's meaningful but smaller than transport or heavy industry. The catch is concentration — a gigawatt facility lands on one local grid all at once, unlike EVs which spread out. That's why the grid-strain stories — Ireland, Virginia, Texas, now Johor and Yongin — keep repeating. The IEA warns that "unless these risks are addressed, around 20% of planned data centre projects could be at risk of delays," pointing to grid connection queues and transformer and cable lead times that have doubled in the past three years.

The technology fighting back

It's not all grim. The engineering response is genuinely impressive.

Liquid and immersion cooling are becoming standard for AI racks, which now run at 50-135 kW versus 10-20 kW for traditional servers — far too hot for air alone. Nuclear power has become the hyperscalers' favorite new obsession: Microsoft signed a 20-year deal to restart Three Mile Island's Unit 1 (renamed Crane Clean Energy Center, ~835 MW, targeting 2028), Google committed to 500 MW of small modular reactors via Kairos Power, Amazon invested in X-energy, and Meta issued an RFP for up to 4 GW. Collectively, big tech has signed well over 10 GW of nuclear deals — though nearly all deliver between 2028 and 2035, which doesn't help the immediate crunch.

The most genuinely inspiring model, to me, is Nordic waste-heat reuse. In Mäntsälä, Finland, a data center heats the equivalent of 2,500 homes — about two-thirds of the town's needs. Microsoft is building a data center region near Espoo designed to become the world's largest waste-heat recovery scheme, expected to heat much of Finland's second-largest city. In Ireland, a scheme in Tallaght uses waste heat from an Amazon facility. This is the circular-economy version of the industry: server exhaust becomes civic warmth. It only works where you have cold climates and district-heating pipes already in the ground, but it points at what "responsible" could look like.

So — is this a bubble?

I have to be honest about the thing hanging over all of this. In 2025 and into 2026, serious people started using the word "bubble" out loud. At Italian Tech Week in Turin in October 2025, Amazon's Jeff Bezos called it "a kind of industrial bubble," adding that industrial bubbles "are not nearly as bad" as financial ones because "when the dust settles and you see who are the winners, society benefits." OpenAI's Sam Altman warned "people will overinvest and lose money." Goldman Sachs' CEO said he expects "a lot of capital that was deployed" won't deliver returns. The MIT Media Lab's NANDA initiative, in its August 2025 report The GenAI Divide: State of AI in Business 2025, found that despite $30-40 billion in enterprise investment, 95% of organizations were getting zero return on their AI pilots — only 5% of integrated pilots drove measurable profit-and-loss impact. Hyperscalers have taken on over $120 billion in new debt, and there's a genuinely dizzying circularity — Nvidia investing in OpenAI, which buys Nvidia chips, while renting capacity from CoreWeave, which Nvidia also part-owns.

My honest read, from Kathmandu, is this: the demand for AI compute is real and probably durable, but the pace and financing of the buildout look frothy, and some of these gigawatt promises will quietly not materialize. The physical constraints — grid connections, transformers, gas turbines, water permits, skilled labor — are stubborn in a way that spreadsheets aren't. The Abilene expansion that got scrapped over winter weather and financing is, I think, a small preview of the friction ahead.

What I keep coming back to is the asymmetry. The world is spending the GDP of mid-sized nations to build machines that think, while I grew up rationing electricity and Nepal still can't figure out how to turn its rivers into computing power. There's something both thrilling and deeply unequal about this moment. The infrastructure of intelligence is being poured in concrete right now, mostly in a handful of rich countries and oil states — and the rules, the fairness, and the environmental accounting are all running to catch up.

I don't have a tidy conclusion. I just think it's worth watching closely, because whoever controls the data centers will, in a real sense, control a large part of the twenty-first century. I'd rather understand that from where I sit than have it happen to me.

Sources

• International Energy Agency (IEA), Energy and AI and Key Questions on Energy and AI (2025)
• Synergy Research Group; Dell'Oro Group (data center capex, hyperscale counts)
• Goldman Sachs Global Institute, Tracking Trillions (June 2026); Goldman Sachs Research
• CBRE, JLL, Cushman & Wakefield, Savills, Colliers, Mordor Intelligence, Arizton (regional market data)
• Statista, Cloudscene, Visual Capitalist, ABI Research (data center counts)
• Company earnings and announcements: Amazon, Microsoft, Google/Alphabet, Meta, Oracle, OpenAI, xAI, Nvidia
• OpenAI, SoftBank, Oracle (Stargate); SemiAnalysis, Epoch AI (Colossus, Stargate analysis)
• Central Statistics Office Ireland (Data Centres Metered Electricity Consumption 2024, June 2025); EirGrid; CRU; KPMG (Ireland grid data)
• White & Case; DCD; FTI Consulting (Malaysia/Johor)
• S&P Global; PIF; CNBC (Saudi Arabia/UAE, Humain)
• Mordor Intelligence; Arizton; Pulse Kenya; African Energy Chamber (Africa)
• Kathmandu Post; myRepublica (Nepal)
• UC Riverside / UT Arlington (Ren, Li et al., "Making AI Less 'Thirsty'"); Lawrence Berkeley National Laboratory; Google & Microsoft environmental reports (water)
• Bloomberg; World Economic Forum; atNorth (Nordic waste heat)
• MIT Media Lab NANDA (The GenAI Divide, Aug 2025); NPR; Yale Insights; Fortune; The Register; CNBC (AI bubble commentary)
• Data Center Watch; Virginia SCC; PJM (US community pushback, grid)