This article first appeared in The Edge Malaysia Weekly on August 31, 2026 - September 6, 2026

ON Nov 3, Americans will vote in the midterm elections. The top issue is not US President Donald Trump, inflation and affordability, the Iran war or tariffs on trading partners and allies, but data centres. A recent Gallup poll found that 71% of the US population wants the massive data centre build-out that began three years ago to stop. Americans are saying: not in my backyard.

From Virginia to Texas, Ohio to Georgia, the once-coveted campus of computing power is being blamed for rising electricity prices, water scarcity, and noise and air pollution. Protesters say data centre operators get tax breaks from state governments, keep all the profits and create few local jobs. Rightly or wrongly, data centres have become a lightning rod for broader anxieties about Big Tech and powerful mega-billionaires who benefit from cutting-edge innovation. The backlash has turned the US midterm elections into a referendum on an economy overrun by artificial intelligence (AI).

Goldman Sachs estimates that the capital expenditure of hyperscalers, which build and operate giant data centres to provide global cloud computing and data storage for companies such as Google, Amazon, Microsoft, Meta Platforms and Oracle, will grow from US$792 billion (RM3.2 trillion) this year to US$1.05 trillion next year. On Aug 26, chip behemoth Nvidia Corp reported that its revenues for the last quarter ended July grew to US$96 billion, nearly doubling over the past year. It is on target to grow revenues to US$425 billion in the current fiscal year and another 70% in its next fiscal year starting February. By the end of its next fiscal year, analysts expect Nvidia to have annual revenues of over US$720 billion, or on a par with Walmart, the world’s largest retailer.

The big worry now is whether AI giants, hyperscalers and top frontier labs like ­Anthropic and OpenAI can generate enough real profits to justify the massive spending on Nvidia’s graphics processing units or GPUs, which power all the data centres.

What actually goes inside a data centre and what does it all cost? Also, what can a completed data centre earn every year and how long does it take to get a decent return on the original investment?

Research institute Epoch AI estimates that it costs at least US$38 billion to build the benchmark 1GW data centre in the US and US$900 million per year to operate it. Unlike malls or offices, data centres are measured in gigawatts because their primary operational limit is electrical power capacity rather than floor space. Servers powered by Nvidia’s top-of-the-line Blackwell GPUs make up about 60% of the total cost of a standard data centre. If you want to get more granular, the GPU chips alone cost about US$15.58 billion. Just so you understand the size and scale of what is being built, Meta Platforms’ Hyperion campus in Louisiana with a 5GW capacity will use US$78 billion worth of AI chips.

Let me break all the costs down so you have a better sense of the underlying economics. A current standalone Nvidia top-of-the-line Blackwell Ultra B300 GPU costs about US$53,000. An entire system with the eight-GPU DGX B300 costs between US$400,000 and US$500,000, and a full 72-GPU GB300 NVL72 rack would set you back US$3 million to US$4 million. That is US$42,000 to US$56,000 per GPU and it comes pre-packaged with high bandwidth memory. Unlike a standalone B300, a GB300 is built for high-density, rack-scale compute complexes where dozens of processors and GPUs operate. Think of GB300 as a fused “superchip” that combines the Blackwell Ultra GPU with Nvidia’s proprietary ARM-based Grace CPU.

Then there is the actual building, or “shell”, with secured power and fibre connectivity. You need to include the cost of land, concrete, networking gear, switches, cables, transformers, generators and all the backup. Add all that up and you need about US$11 million per megawatt of electricity for the physical facility before you rack the chips. Multiply that by 1,000 because it is a 1GW data centre and it is over US$11 billion. Tech firms have increasingly turned to private power sources as public grids struggle to feed the massive energy demand of the frenzied AI build-out. If you build your own natural gas-powered plant for a 1GW data centre, it will cost you between US$2.5 billion and US$4 billion. To build a nuclear plant, or four smaller 300MW plants, to get 1.2GW capacity, the cost could soar to between US$9 billion and US$12 billion depending on the type of plant. A gas-powered plant can be ready in two to three years. A nuclear plant takes much longer because of regulatory approvals.

Another big cost is cooling, since AI chips tend to run really hot. Back in the old days, when massive air conditioning was the norm, cooling costs were US$1.8 million per MW. These days, liquid cooling costs US$5.5 million per MW, or US$5.5 billion just for the hardware alone. Total cooling infrastructure costs can often balloon to US$7.5 billion. Not surprisingly, running a large data centre in a cooler climate like Michigan or Wisconsin is sometimes preferable to one in Johor, Singapore or Thailand.

To put it plainly, a data centre is a warehouse of very expensive chips on racks. And the cost keeps going up, not down. Nvidia CEO Jensen Huang last week mentioned that his firm would raise the prices of GPUs by 15% to 30% because suppliers of high bandwidth memory like South Korea’s SK Hynix were charging more. That US$56,000 chip I was talking about could soon cost up to US$72,000 and the Hyperion data centre might need at least US$100 billion worth of AI chips.

Now that we know what it costs, let’s look at what a data centre operator can earn from all that AI capacity. New Street Research analyst Pierre Ferragu titled his latest piece on AI “Everybody wants to rent a datacenter”. AI capacity leases for US$150,000 to US$250,000 per MW per month in the US, or between US$1.8 million and US$3 million per MW per year. That is nearly three times what a traditional non-AI data centre earns. Clearly, there is robust demand for AI compute and data centre rents are high.

How quickly does the investment pay back? A GPU cluster’s payback can be anywhere between two and five years, depending on the size and location of the data centre, utilisation rate, who it is leased to and how much rent users of chip clusters are willing to pay. Here is another way to look at it. If you can rent out the GPUs at premium rates of, say, around US$3.50 per hour and have a high utilisation rate, the payback could be just 2½ years. However, if you are renting out at discounted long-term contract rates of US$1.80 per hour, the payback period can stretch to eight to 10 years. The economics are great if the facility stays at near-full utilisation at high enough prices but can quickly turn mediocre or loss-making if utilisation slips or prices fall. There is very little margin for error.

Here is the thing: 60% of a data centre’s huge cost wears out quickly or goes obsolete fast because the lifespan of a top-end GPU chip is between 3½ and five years. High temperatures over long periods dry out the thermal paste and wear down the chips. So, AI infrastructure risks ageing before the initial capital in data centres and chips can be recovered. Even as the prices of AI chips soar, their finite lifespan remains the key focus. A four-year-old GPU begins to lose computational efficiency, although it might still be able to handle everyday apps, video editing and e-sports games. If kept cool and clean, a graphics card’s life might be stretched to seven or even 10 years, though the Anthropics and OpenAIs of the world would most likely pass them on to the second-hand market after their 3½ years in service because a new, better generation of AI chips lands every 18 months or so. In a fiercely competitive industry where US$1 trillion is being spent annually on infrastructure, nobody wants to use a four-year-old chip that is degrading fast.

However, Meta, Microsoft and Google now depreciate GPUs only over five to six years on their books. That is just creative accounting. In reality, Nvidia stops adding new features for older cards after five years, even though it might continue to provide security updates for up to 10 years.

Over the past five years, the data centre business model has moved away from traditional “cloud storage” to GPU-as-a-Service. Data centre operators buy tens of thousands of Nvidia chips for billions of dollars, cluster them together and then lease out the computing power. The real economics are dictated by two major constraints: power and cooling. How long does it take to build a 1GW data centre? Typically, 18 months to three years. Data centres and conventional power infrastructure have long construction cycles while grid connection timelines can exceed five years. Accelerated AI projects such as Elon Musk’s Colossus in Tennessee had a compressed 12-month window, reusing industrial shells and deploying temporary on-site power generation.

There are 5,400 data centres in the US, 3,360 in Europe, 500 in China and 308 in Southeast Asia. But numbers are misleading metrics because some of the data centres are small and some are huge. A better number is operating capacity. The US has 30GW, Europe 13GW, China 30GW and Southeast Asia just 2GW. You need to focus on the new capacity being built to understand the economics of the current AI boom. The US has 15.9GW of data centre capacity under construction, of the total 23.1GW being built globally. Another 35GW has been announced, though some of it is likely to be shelved as political opposition grows. China will build at least 8GW more capacity by 2030. In Southeast Asia, 9GW of new capacity is planned for Malaysia or is under construction, Thailand has 3.5GW under construction and Indonesia 2.87GW.

So, will protests in the US eventually lead to more data centres in Southeast Asia? Not necessarily. For one thing, despite all the protests, so far only about US$80 billion or so worth of data centres has been actually blocked in the US. Most of the new capacity being blocked there tends to relocate within America. When Virginia or Pennsylvania says no, the project usually moves to Tennessee. Data centre operators cannot serve American users well from Johor because of latency, or the time delay that occurs when data travels from one point to another across a network, and data sovereignty. Governments increasingly require citizen data to stay within the country, which keeps data centres local regardless of cost. Domestic backlash might reshuffle the American map, but it won’t lead to capacity being exported out. A recent McKinsey study noted that Asean has moved “beyond the spillover” from a Singapore-overflow market into a self-sustaining one, powered by 700 million people, growing cloud adoption, digitisation and data sovereignty rules. With 20% annual demand growth forecast for the next three years, the region would be booming even if Americans were not protesting to stop data centres.

Data centres are not a zero sum game. A facility not built in Ohio will not suddenly appear in Johor. Compute is regional. Data centres serve nearby users because of latency. A facility serving Jakarta will not be built in Chicago or vice versa. One exception is frontier training in giant, concentrated clusters that build the most advanced models like Anthropic’s Claude. They are not tied to local users and can easily migrate to wherever power is cheap. Gulf states such as Saudi Arabia, the United Arab Emirates and Qatar, with their money and cheap power, are attracting those kinds of data centres rather than Southeast Asia, whose own grids are power-constrained.

So, will data centres make money? Yes, those running at full capacity without resorting to huge discounts certainly can. The problem is how long can an industry that is spending a trillion dollars on infrastructure annually to grow capacity have a 100% utilisation rate and still charge premium prices? As long as demand keeps growing faster than the chips depreciate, the data centre business will be fine. The four-year depreciation clock on the chips makes things difficult unless AI becomes so useful that end-customers continue to pay for cutting-edge frontier models.

Assif Shameen is a technology and business writer based in North America 

Save by subscribing to us for your print and/or digital copy.

P/S: The Edge is also available on Apple's App Store and Android's Google Play.