AI Wants More Data. More Chips. More Real Estate. More Power. More Water. More Everything

It looks easy enough: Ask ChatGPT something, and it responds. But pull back the curtain, and you’ll find that every ChatGPT prompt and Microsoft Copilot task consumes vast resources. Millions of human beings engineering, correcting and training models. Enough terawatt-hours of electricity to power countries. Data center mega­campuses around the world. Power line networks and internet cables. Water, land, metals and minerals. Artificial intelligence needs it all, and it will need more.

Researchers have estimated that a single ChatGPT query requires almost 10 times as much electricity to process as a traditional Google search. Your typical search engine crawls the web for content that’s filed away in a massive index. But the latest AI products rely on what are known as large language models, or LLMs, which are fed billions of words of text—from the collected works of William Shakespeare to the latest forecasts of the Federal Reserve. The models detect patterns and associations and develop billions and billions of so-called parameters that help them mimic human behavior. Using these models, ChatGPT and the like create new content—hence the term generative AI.

The resource-intensive nature of AI will create winners and losers. Those with the most resources will have the most advanced AI systems. It’s leading to clashes over increasingly scarce commodities, as well as access to chips. It’s motivating tech com­panies to seek more efficient means of developing AI. They’re throwing ­billions of dollars into alternative energy solutions such as nuclear fusion that have spent years if not decades sputtering along without heavy spending and technological breakthroughs. At the same time, AI’s demands are adding to the ­pressure to keep burning fossil fuels to feed the power grid, even as the world is on track to blow past crucial emissions targets in the fight against climate change.

Although the AI buildout represents a tremendous opportunity for investors, businesses and societies, there is peril. Many have raised the specter of the potential harm and bias of such systems. Wall Street, meanwhile, is getting tired of waiting for the technology to translate into meaningful profits. Even the focus on efficiency could turn into a dark cloud for anyone who over­invests in infrastructure. Here’s a closer look at everything the AI industry needs to keep its models running.

Another 1,000 Terawatt-Hours of Power

AI largely lives and runs in data centers humming with ­motherboards, chips and storage devices. The electricity demand from these centers is now outstripping the available supply in many parts of the world. In the US, data centers are projected to use 8% of total power by 2030, almost triple the share in 2022 when the AI frenzy took off, according to Goldman Sachs Group Inc., which has described it as “the kind of electricity growth that hasn’t been seen in a generation.” Similar surges in demand have been forecast for Sweden and the UK. By 2034 annual global energy consumption by data centers is expected to top 1,580 ­terawatt-hours—about as much as is used by all of India—from about 500 today.

The data centers operated by Alphabet Inc.’s Google used more than 24 terawatt-hours of electricity in the 2023 fiscal year, up more than 31% from 2021. Microsoft Corp.’s overall use was roughly the same, representing a 70% surge from two years earlier. The world’s largest tech companies have grown acutely aware that power could be the most disruptive kink in the AI supply chain, and they’re racing to lock in long-term supplies. In May, Microsoft and Brookfield Asset Management Ltd.’s green energy arm signed the biggest corporate clean energy purchase agreement.

In October, the world’s largest solar and wind power generator, NextEra Energy Inc., said it had struck deals for the potential development of a remarkable 10.5 gigawatts of renewable energy and storage by 2030 for only two Fortune 50 companies. In a sign of the clashes to come, they aren’t even tech companies. The boom has created “even more of a premium on other industries outside of data centers to try to lock up low-cost renewable generation,” NextEra Chief Executive Officer John Ketchum told investors. “All ships are rising with the tide here.”

All the Fossil Fuels It Can Get—and More

Coal, among the world’s most ­carbon-­intensive sources of energy, is still burned to generate roughly a third of electricity supplies. Natural gas, which also creates global-warming emissions, fuels 20% of power. Wind and solar farms have gained ground in recent years, but absent giant batteries that can even out the supply for data centers, the intermittent nature of renewable energy has been problematic for these centers because they depend on a constant flow of power.

A technique that Google pioneered has emerged as a solution: Use software to hunt for clean electricity in parts of the world with excess sun and wind on the grid, then ramp up data center operations there. Otherwise, arguably the only dependable, around-the-clock source of zero-emissions power at the moment is nuclear. This explains why Microsoft struck a deal in September that will reopen a reactor at the Three Mile Island nuclear power plant in Pennsylvania, site of a notorious partial meltdown in 1979. About a month later, Amazon.com Inc. signed three agreements to develop small-scale nuclear reactors, and Google invested in and committed to buying power from a company similarly developing modular reactors. “I mean, my God, nuclear reactors. Are you kidding me?” Oracle Corp. Chairman Larry Ellison said to analysts in a September meeting. “That sounds completely made up, but it’s not. … Has anything like this ever happened before?”

A Hundred Times More Grid Capacity

Power lines and substations are the most underrecognized links in the AI chain. All the new data centers will need to be connected by a grid that’s already old, under stress and vulnerable when weather goes bad. (Note Hurricane Helene.) At a Bloomberg Intelligence event in April, Brian Venturo, co-founder of cloud services provider CoreWeave Inc., said companies like his are developing giant data centers that will strain the grid. Imagine a substation in an industrial area supplies 30 megawatts, and a data center in that area needs maybe 5 megawatts.

The rest goes to other offices and factories. Today, companies such as CoreWeave are saying, “ ‘I want 500,’ ” Venturo said. “You have to build new transmission lines. You have to do new substation builds.” And you’ll need transformers at those sub­stations, which may need to be ordered years in advance.

And that’s just for 500 megawatts. OpenAI co-founder and CEO Sam Altman is talking about data centers that could need 5,000 megawatts. Building a power system that can support that much load in a single place from scratch in short order is “functionally ­impossible,” Constellation Energy Corp. CEO Joe Dominguez says. Constellation is the owner of the Three Mile Island nuclear plant that’s reviving a reactor to feed power to Microsoft. Dominguez says data center builders need to be thinking about co-locating around giant, already existing power resources—such as his nuclear power plants. Build a mega­campus next to a couple of nuclear reactors, surround them with renewable energy resources and batteries, connect them all together with new wires and load-shifting controls, and you can create a self-contained grid.

Billions of Liters of Water a Day

Every watt of electricity that’s fed into a server generates heat. Temperatures too high can destroy equipment and slow systems. Right now, some of the most energy- and cost-efficient ways to chill the air in centers rely on water. Bluefield Research has estimated data centers use more than a billion liters of water per day, including water used in energy generation. That’s enough to supply 3.3 million people for a day. One 2023 study estimated that a conversation with ChatGPT consisting of roughly 10 to 50 questions and answers, requires a standard 16.9-ounce bottle of water. Training only one earlier AI model behind ChatGPT was estimated to have consumed almost 200,000 gallons of water. Making matters worse: Much of the water is of drinking quality, to avoid environmental problems and equipment failure.

In West Des Moines, Iowa, a network of Microsoft data centers that OpenAI used has turned the tech giant into the area’s largest water user, consuming more than the city itself, according to water district data. (The district says the company was also investigating a leak that significantly raised its usage.) In Talavera de la Reina, a small city tucked among Spain’s barley and wheat fields, Meta Platforms Inc. has clashed with locals over a plan to build a center that will use about 665 million liters (176 million gallons) of water a year.

Twice as Much Internet Bandwidth

The large language models that ­underpin generative AI learn by digesting huge amounts of data over the internet, and users of AI tools in turn will only add to the demand. AT&T Inc. CEO John Stankey said in May that the network’s wireless demand was already up 30% a year and won’t slow with AI raising usage. “If you’re going to continue to see usage go up 30% to 35% a year, you’ve got to build bigger highways to take that,” he said.

Over the past five years, the network traffic growth at Verizon Communications Inc. has more than doubled thanks to people watching and streaming videos, Verizon Consumer Group CEO Sowmyanarayan Sampath said in an interview at about the same time. In the next five years, he predicted, growth will double yet again because of prompts and data fed into AI models. AI, he said, “is the next growth machine for us.” Tech com­panies are so hungry to lock in fiber ­networks right away that the telecommunications company Lumen Technologies Inc. in August announced that it had secured $5 billion (and was in discussions to land another $7 billion) in new business tied to AI-driven demand for connectivity.

Real Estate for Thousands of Data Centers

Globally, there are more than 7,000 data centers built or in various stages of development, up from 3,600 in 2015. And that still probably won’t be enough. The demand for data center services had been growing dramatically even before ChatGPT, mostly because companies have been increasingly moving their data processing off-premises and turning to remote cloud services. And every major country wants its own homegrown AI hubs, touching off a global infrastructure race.

Data centers require land. For reference: The data-­center-focused real estate investment trust Equinix Inc. bought 200 acres for a multihundred-­megawatt campus. Another company recently signed a lease development agreement on 2,000 acres for a gigawatt-size one. Finding land that works just right for the power requirements of a data center is tough, leading to bidding wars. These complexes also need construction materials and crews to install all of it. Material is on backorder, and there’s a shortage of workers. Meanwhile, Venturo of cloud services provider CoreWeave says some of his clients want him to devote entire campuses just to their business. “The market is moving a lot faster than supply chains that have historically supported a very physical business have been set up to do,” Venturo says.

Chips, Chips, Chips

Graphics processing units, or GPUs, are the workhorses for training AI models. They’re designed to handle thousands of tasks simultaneously, a concept known as parallelism. A data center may use hundreds or even thousands of these processors, each one costing more than a family car. Virtually every major tech company came up short on this type of chip when the generative AI boom first took hold. Microsoft and Google were among those that cited low GPU inventory in past earnings calls as a challenge.

Nvidia Corp. has upped the stakes for everyone by moving to annual introductions of new technology. That’s brought further strains to an already stretched supply chain. The company said in November that its new Blackwell product is back on track and ahead of predictions for the amount it’s getting out the door. But, crucially, it will be many quarters before it has enough to meet all the demand.

Silicon, Steel, Quartz and Copper

Many of the items above require metals and minerals. Consider silicon, the foundation for chips, circuits and processors. China is the largest producer of raw silicon and refined silicon materials, which has raised concern as tensions between the Asian nation and the US and its allies rise. The most recent supply chain scare surfaced in North Carolina. In October, Hurricane Helene, in addition to killing dozens of people and stranding others across the eastern US, disrupted operations at two mines in the state that together produce about four-fifths of the highest-quality quartz. It’s used to create crucibles where silicon is heated, melted and reformed into the single-­crystal ­structure that makes an ideal base to manufacture semiconductors.

Semiconductors contain gold, silver, aluminum and tin. There’s enough of these metals to keep the factories humming. But two obscure chip metals have emerged as potential bottle­necks: gallium and germanium. In December, China announced a ban on exports of the metals to the US—part of an escalating tech war. Copper is in everything including chips, data centers, electrical equipment and cooling units, potentially setting the stage for a clash between the demands of AI, renewable energy and electric transportation. And then there’s steel, which is critical to building data centers and for infrastructure such as cables.

More People Than You Think

Much has been said about the jobs that AI may eliminate. In February, the Swedish fintech company Klarna Bank AB made waves after saying its AI assistant was doing the equivalent work of 700 full-time customer service agents. Global research and analysis companies have warned that employment in sectors such as finance, law and customer service will be hard-hit. The International Monetary Fund has estimated that AI could replace, or augment, almost 40% of employment globally.

But AI companies themselves directly employ millions of people today. Among AI workers are computer scientists, data architects, researchers, mathematicians, software engineers, chip designers, product and program managers, and compliance attorneys. That’s not to mention the armies of in-house analysts, marketers and salespeople. In early November, Salesforce Inc. announced plans to hire more than 1,000 workers to sell its new generative AI product.

Talent bottlenecks have emerged across much of these professions amid the rush to recruit for AI. Tech investors and AI startups have lamented the lack of properly educated and experienced candidates. The phrase “AI-vies”—a play on the Ivy League—has emerged in Silicon Valley to refer to a few companies (among them, Alphabet, Microsoft and OpenAI) that have trained the talent everyone else wants to poach. Even more have been recruited abroad, in countries such as India, to build and clean up the high-quality datasets necessary to train AI systems.

More (Good) Data Than the World May Have

Generative AI models need ­high-­quality data the way human beings need food. Large language models are “trained” by ingesting text that’s broken down into small units called tokens. From this text, LLMs identify patterns that help predict—in a process repeated over and over again—the text that should follow another set of text. The world’s foremost LLMs were trained off more than a trillion tokens each. To put that in context, consider that 2,048 tokens is roughly equivalent to 1,500 words. Estimates for exactly how many tokens of cumulative text data exist in the world are all over the place, ranging from a few trillion to thousands of trillions.

Amazingly, this abundance of data might not be enough to keep AI development moving forward as quickly as some hope. Some of the world’s most power­ful AI model developers such as OpenAI are already finding it increasingly difficult to locate new, untapped sources of high-quality, human-made training data to advance their models.

There’s limited data in non-­English languages and even less that isn’t focused on Western or White communities. This lack of diversity threatens to result in AI products that show bias against minorities, women and other underrepresented populations. A Bloomberg analysis this year, for example, found that the underlying AI model behind ChatGPT shows bias against certain racial groups based on names alone when ranking resumes. OpenAI says that the results may not reflect how its customers use its models and that it works to identify potential harms.

Producers of data and content, from media organizations to financial institutions, are waking up to the fact that their information is increasingly valuable to AI developers. Hollywood actors and writers went on strike in 2023 to protect their craft from the technology. The New York Times as well as major record labels are suing AI companies for training their data on copyrighted work. AI companies say that training on publicly available materials is a legally permitted fair use.

In a recent call with investors, S&P Global Inc. CEO Martina Cheung summed it up: “A large language model is only as good as the quality and quantity of data that it’s trained on, and we have lots of high-quality data.” Just in the past year, OpenAI has struck deals to use content from News Corp., Condé Nast, Hearst, Reddit, People magazine publisher Dotdash Meredith and Axel Springer.

Tech companies are experimenting with training models on “synthetic” datasets, content created by AI itself. In theory, this helps AI companies meet their bottom­less need for data while avoiding the legal, ethical and privacy-­related concerns surrounding scraping information from the web. But some researchers have warned that AI models may “collapse” if they’re trained on content generated by AI rather than humans. One 2023 paper on so-called model collapse showed that AI images of humans became increasingly distorted after the model retrained on “even small amounts of their own creation.”

Or Maybe Less of Everything Than Some Fear. Or Hope

Investors, data center operators, energy companies and other businesses are pouring hundreds of billions of dollars collectively into different parts of the supply chain that feeds AI. Every major bank and private financier is positioning itself for a piece of an estimated $1 trillion in spending on AI ­infrastructure. The capital ­expenditures of Alphabet, Amazon, Meta and Microsoft are set to collectively exceed $200 billion in 2024. An S&P 500 utility-­sector index has gained 22% over the past year, and data center-­focused REIT Equinix has almost doubled its market cap since late 2022. Nvidia shares have surged by nearly 700% over the past two years, turning the company into one of the most valuable ones on Earth.

And yet, in the end, nobody knows whether AI will keep booming. Some Wall Street analysts are starting to predict an end to the frenzy. Investors have begun questioning whether Big Tech’s heavy spending will ever result in the AI profit machine they’d envisioned. Arguably the biggest threat to the hundreds of billions of dollars being invested in AI is that the world’s most advanced model developers and their suppliers have grown obsessed with efficiency.

Gary Dickerson, CEO of chip equipment maker Applied Materials Inc., told investors in November that some AI companies are aiming for “100x improvements” in efficient computing within five years. Others are shooting for 10,000-fold gains in 15 years, he said. Efficiency, Dickerson said, “is emerging as a unifying driving force for the industry.” —With Dina Bass, Mark Burton, Mark Chediak, Katherine Chiglinsky, Jackie Davalos, Stephanie Davidson, Jennifer Duggan, Brian Eckhouse, Seth Fiegerman, Shirin Ghaffary, Evan Gorelick, Ian King, Christina Kyriasoglou, Jane Lanhee Lee, Yoolim Lee, Naureen Malik, Rachel Metz, Saritha Rai, Josh Saul, Olivia Solon and Will Wade

Doan is a managing editor at Bloomberg News in New York covering technology and cybersecurity.