For many of us, data and compute are abstractions, hidden away in the neatly packed ‘cloud’ and ‘credits’ of the hyper-scalers. But where does the cloud live? (am I the only one who read this as the live that rhymes with hive as opposed to the live that rhymes with give)
Cloud lives on servers, which are storage and compute devices mounted on racks, in data centers. This podcast series by Google gives you a sense of what a data center feels like.
While the idea of the cloud was floated in the 1990s, it began capturing the market in the post .com-bubble era, where servers put together in the warehouses of Amazon and Google began to first power the internal operations of these companies, and then in 2006 emerged as a service.
Just like how the steam engines ushered in Industrial Revolution 1.0 producing a shift from manual agriculture to mechanized crop production; the widespread access to electricity brought in Industrial Revolution 2.0 of mass manufacturing and consumption; and access to the internet and personal computers brought in the Industrial Revolution 3.0 of democratized access to compute; the emergence and widespread use of AI enabled by the shift of enterprises from in-house server infrastructure towards cloud is the biggest industrial transformation of the 21st century so far.
Apart from AI applications, the cloud has enabled the development of streaming services (Netflix etc), self-driving cars (Waymo etc.) and accelerated drug discovery for pharma companies.
What the cloud has effectively done is to take the operational expenses out of the hands of individual companies and concentrate them in data centers located anywhere in the world. Over the last few years, investments in data centers have skyrocketed, with $420 billion being invested in capex alone by the largest data center providers in 2025 alone, the number expected to reach $600 billion by end of 2026 ($80 billion already invested by Q1) and $7 Trillion by 2030.
Capex is only one part of the story. Opex is the other. While all the initial cloud data centers were built in the United States, companies slowly started branching out to Southeast Asia, India, Middle East and now Africa. While this is in part enabled by the governments in these regions offering generous special economic zone provisions to cloud companies and their co-location providers, this shift is primarily led by space, energy and water constraints in the United States.
Let’s look at each of these components one by one.
i. Space
How much space does a data center take up?
It ranges between 100,000 - 10,000,000 sq. ft depending on the scale of the data center. For context, a football field (not the American football xD) is 81,000 sq.ft. So a super large data center would be worth slightly more than 123 football fields.
The total land area on earth is 1.6 * 10^15 sq. ft. That feels like a lot and capable of accommodating all the data centers we would ever need, but not yet.
10% of this is taken up by icebergs (it’s obvious why data centers cannot be built on icebergs); 36% of this is agricultural land which is important for food security; 26% is forest land; 19% deserts; 8% grassland; 1% fresh water bodies and 1% human settlement.
Data centers started out by being built on grassland, but has now expanded to encroach on agricultural land, in addition to human settlements. Grassland and forests are great carbon sinks, and the loss of these implies rising global warming. Data centers are being built extensively on desert land (in Saudi Arabia, UAE etc) but there are real cooling challenges in deserts.
ii. Energy & Water
The total global data center energy consumption (as of 2025) is ~485 TWh, which can power, equivalent to what 139 million homes would consume in a year. When we break it down, the average data center consumes energy equivalent to that of 100,000 homes. Note that the biggest source of energy consumption and heat generation in data centers is the IT infrastructure (~60% of total energy), referring to the compute and storage devices, the server racks they are placed in and how they connect with each other.
The energy efficiency of a data center is typically measured in Power Usage Effectiveness (PUE), which is the ratio of total facility power by IT equipment power. PUE of 1 is ideal.
While conventional data centers (enterprise, smaller colocation) have very low power density (~5–15 kW/rack) and mediocre PUE (1.54–1.92), while AI- first data centers have soaring density (130–600 kW/rack) and better PUE (1.10–1.15 for hyperscale). This means that as you go up in size and AI-focus, energy consumption per rack explodes even as efficiency per unit of compute improves.
As a result of the PUE not being a perfect 1, waste heat is generated. This leads to the production of local heat islands (raising temperatures between 2-9 °C) in regions up to a 10 km radius from the data center. There are efforts to utilize waste heat from data centers and return them to the grid in some form, but it is yet to be implemented at scale, partly because of the infrastructure required to do that.
Data centers are typically cooled using water. For context, a Meta data center that opened in 2018 using 500,000 gallons of water per day (i.e.10% of the entire county’s water consumption) and all of Google’s data centers combined using 5 billion gallons of water across all its data centers in 2023, with 31 percent of its freshwater withdrawals coming from watersheds with medium or high water scarcity.
Out of the total water derived from different sources for data center cooling, at least 70% is consumed by evaporation and only the rest used in a closed loop. Nearly 57% of water used for cooling is derived from fresh water sources, with more recent efforts to leverage waste water. However, waste water typically has dissolved solids, salts and nutrients that can corrode data center equipment, necessitating water treatment before use, which again consumes energy in a different form. That is why fresh water is still the go-to source for cooling. Water Usage Effectiveness (WUE) is the analogue to PUE for water efficiency of data centers, and the average WUE achieved is 1.8 L/kWh.
While air cooling (i.e. using fans and Computer Room Air Conditioners to circulate cold air throughout a facility, removing heat generated by IT equipment) is an alternative, it is typically less energy efficient than water cooling, leading to it not being widely adopted.
iii. Hardware replacement
While typical data center hardware has reasonably long lifespans (GPUs up to 8 years, SSDs up to 10 years, HDDs up to 7 years and tapes for up to 30 years), they are replaced within an 18-24 month timeframe. This is partially due to reader obsolescence, improved speeds and pre-emption of degradation. While some of the hyperscalers are talking about extending the lifespan of their hardware, this then results in more water required for cooling.
E-waste associated with the use of LLMs alone is predicted to reach 5 million tons by 2030 (1.2 million if LLMs are used sparingly).
Note: The extended climate impact of space, water and energy consumption coupled with e-waste boils down to 180 million tons of CO2 emissions annually, roughly double that of the emissions of the aviation industry. While CO2 capture and carbon credits is being used a way to mitigate these emissions, these initiatives are only at 50 million tons capacity per year as of 2026.
iv. Wild Card Entry - Security
Given the large space footprint of data centers, they are quite conspicuous. While data centers have a six to seven layer security mechanism for entry into their premises and to prevent fire incidents, they are quite vulnerable to enemy attacks, as indicated by the the recent attack on AWS data centers in the Middle East.
Blast resistant architecture, surveillance and defense facilities similar to those used in military bases and decentralization are emerging as ways to counter this threat. Decentralization is quite hard to achieve with the current infrastructure capabilities, as there is clear operational advantages to concentrating IT infrastructure in specific locations (this is the premise for the widespread adoption of cloud as we discussed earlier).
First Principles Approaches to Addressing These Challenges
i. Build more water and energy efficient data centers/data center networks
The starting point is to measure how much energy and water is being consumed and for what purpose. This then enables data centers to cut down on wasteful consumption. This has led to the rise of many IoT sensor companies and their products being adopted by hyperscalers.
Once this is established, data centers can then go on to retro fit different technology assets, such as heat pumps to turn waste heat into usable energy, immersion cooling instead of evaporative cooling and direct to chip (D2C) cooling instead of rack level cooling. Startups like Uravu Labs are helping data centers get fresh water from air to reduce reliance on local fresh water sources.
In terms of data center construction, water circularity (in-house treatment plants), material sustainability, picking location that enable providers to spend lesser on cooling due to natural factors (such as the Khazna data center in Mazdar City, Abu Dhabi) are being adopted as pathways to greater energy and water efficiency.
An extension of these approaches is building flexible data center-power grid networks. This means that power is directed towards specific data centers based on their compute loads at a given point in time, preventing idle servers from eating up energy by operating at their fullest capacity. Companies like Emerald AI, led by Varun Sivaram who previously worked with the US Department of State for Energy and Innovation, are enabling this systemic transformation.
ii. Cheaper and infinitely scalable sources of energy
There is a certain physical limit to which retrofitting, grid networks and better construction can enable efficiency. So we then start thinking about sources of energy that are not as limiting as fossil fuels in terms of availability. Enter renewable energy sources.
Solar, wind and hydroelectric power have already been widely adopted by data center providers,. Geothermal and nuclear energy are being looked at as the next frontiers powering data centers,. Both the public and private sectors are investing heavily in nuclear fusion.
Batteries for long term energy storage are being looked at as an important infrastructure asset to enable the benefits of renewable energy to accrue for data center use.
iii. Build data centers in a place where space and energy is abundant
Even if energy is infinitely scalable and cheap, what do we do about cooling and the large tracts of terrestrial space that data centers take up?
Two things you could - build data centers under the ocean or in space.
Microsoft attempted the former between 2018 and 2020 but data centers on the ocean floor are challenging due to the impact that it can have on marine life, the coastal communities that rely on them for livelihoods and how it would change the ocean’s ability to act as a carbon sink. In addition, we barely understand our oceans in comparison to outer space, and hence a lot of the infrastructure needed to enable ocean floor data centers needs to be built from scratch.
However, the space narrative is different. The (nearly) unlimited access to solar energy, radiation for cooling (instead of convection) and almost infinite physical space for expansion, coupled with recent advancements in rocket technology and space manufacturing makes orbital data centers a compelling alternative. As this also directly contributes to inter-planetary civilizations, like the moon/mars settlements that SpaceX attempts to build, orbital data centers are being heavily invested in.
There are definitely challenges to implementation there. Cooling by radiation sounds compelling but is not as simple, it requires massive infrastructure to capture the energy and radiate it back. While SpaceForge has shown that plasma, that is super important for semiconductor manufacturing in space, can be generated without human intervention, there is a long way to go towards building fabs in space given their operational complexity. Data centers require regular maintenance (unless a very new kind of space data center hardware gets developed) and we need to figure out how to do it economically. In addition, data center hardware is not built for space environments due to poor radiation tolerance. While NVIDIA is attempting to build more radiation resistant GPUs for space environments, Jenson himself admits that the cost curves will take time to fall exponentially.
These are the approaches that the industry primarily looks at today.
But there is a fourth approach that goes deep into the root of the space, energy, water and security problems, all at once. It is the hardware that goes into data centers. Why don’t we tackle the problem at its roots by redesigning storage and compute hardware?
More of that in part 2 of this series
References/Footnotes
1. Now, there is greater focus on data sovereignty, which means that cloud providers have to make sure that data of a particular country stays within the country
https://www.delloro.com/news/data-center-capex-surges-57-percent-in-2025-as-ai-deployments-accelerate/#:~:text=Data%20Center%20Capex%20Surges%2057,Oro%20Group%20%2D%20Dell’Oro%20Group
https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-7-trillion-dollar-data-center-build-out-how-industrials-can-capture-their-share
This is specifically with respect to cloud data centers. Sify built data centers in India as far back as 2000, initially as a way to promote consumer internet use but then to provide co-location services to companies like Wipro and Makemytrip. Now Sify is a co-location partner for some of the hyper-scalers.
https://programs.com/resources/data-center-statistics/
https://www.iea.org/reports/key-questions-on-energy-and-ai
https://www.congress.gov/crs-product/R48646#fn16
https://www.frontiersin.org/journals/sustainability/articles/10.3389/frsus.2022.1008583/full
https://www.researchgate.net/publication/403073048_The_data_heat_island_effect_quantifying_the_impact_of_AI_data_centers_in_a_warming_world
https://local.microsoft.com/wp-content/uploads/2022/06/Azure_HeatReUse_Infographic.pdf
https://www.lincolninst.edu/publications/land-lines-magazine/articles/land-water-impacts-data-centers/
https://arxiv.org/pdf/2304.03271
https://www.nature.com/articles/s41545-021-00101-w
https://blog.google/company-news/outreach-and-initiatives/sustainability/our-commitment-to-climate-conscious-data-center-cooling/
https://www.mckinsey.com/capabilities/operations/our-insights/operations-blog/keeping-cool-in-the-data-age
https://www.fwpcoa.org/content.aspx?page_id=5&club_id=859275&item_id=130961
https://www.delltechnologies.com/asset/en-in/products/servers/industry-market/idc-server-refresh-white-paper.pdf
https://www.nature.com/articles/s43588-024-00712-6
https://about.bnef.com/insights/finance/global-carbon-capture-capacity-due-to-rise-sixfold-by-2030/#:~:text=Today%2C%20most%20capture%20capacity%20is,CO2%20to%20meet%20their%20goals.
https://www.cnbc.com/2026/04/07/aws-iran-threats-us-tech-data-centers.html
https://silentinfotech.com/blog/data-center-7/role-of-iot-in-real-time-data-center-monitoring-238
https://international.eco.de/topics/datacenter/white-paper-utilization-of-waste-heat-in-the-data-center/
https://www.infosys.com/iki/perspectives/data-centers-look-immersion-cooling.html
https://attom.tech/direct-to-chip-liquid-cooling-how-it-works-and-application/#:~:text=Headroom%20for%20Future%20Density%20Growth%20As%20chip,higher%20compute%20density%20within%20existing%20physical%20space.
https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai
https://news.microsoft.com/source/features/sustainability/6-projects-that-helped-microsoft-meet-its-renewable-energy-goal/
https://www.weforum.org/stories/2026/04/nuclear-energy-crisis-strait-of-hormuz-wef-small-modular-reactor/
https://about.fb.com/news/2024/08/new-geothermal-energy-project-to-support-our-data-centers/
https://www.iaea.org/newscenter/news/fusion-energy-in-2025-six-global-trends-to-watch
https://energy.mit.edu/wp-content/uploads/2024/09/MITEI_FusionReport_091124_final_COMPLETE-REPORT_fordistribution.pdf
https://www.datacenterknowledge.com/energy-power-supply/potential-energy-is-bess-the-answer-to-data-centers-gridlocked-future-
https://news.microsoft.com/source/features/sustainability/project-natick-underwater-datacenter/#:~:text=Microsoft’s%20Project%20Natick%20deployed%20an%20underwater%20datacenter,was%20tested%20and%20monitored%20for%20two%20years.
https://www.scientificamerican.com/article/data-centers-in-space/
https://www.reuters.com/business/aerospace-defense/spacexs-orbital-data-centers-could-face-same-hurdles-microsofts-abandoned-2026-04-01/
https://www.scientificamerican.com/article/the-push-to-make-semiconductors-in-space-just-took-a-serious-leap-forward/#:~:text=Space%20manufacturing%20is%20still%20in%20its%20early,reaches%20temperatures%20of%20around%201%2C832%20degrees%20Fahrenheit.
https://www.pcmag.com/news/nvidia-debuts-an-ai-chip-for-space-based-data-centers-gtc-2026#:~:text=Select%20your%20favorite%20topics:,of%20Use%20and%20Privacy%20Policy.&text=In%20contrast%2C%20Nvidia’s%20CEO%20Jensen,been%20developing%20orbiting%20data%20centers.