Silicon Valley’s optimism tested as Big Tech bets billions on AI

Silicon Valley entrepreneurs prefer to talk of possibilities rather than obstacles, and of vast revenue opportunities to be seized through sheer force of will. That defiant optimism now faces the ultimate test as big tech firms spend hundreds of billions of dollars on artificial intelligence in the belief it will take human endeavor to the next level, making us all more intelligent, creative and productive.

The visible manifestations of this revolution-in-progress are the vast, windowless buildings that are springing up across the world to handle AI workloads. Where a “big” data center was once roughly the size of a football field, some of the new AI facilities will cover an area comparable to Central Park in Manhattan. The investments required are so large that they’re giving a significant boost to the US national economy.

Yet messy reality is already casting its shadow over proceedings as it becomes clear just how much energy and other resources these computation sheds on steroids will consume. The largest AI data centers are expected to gobble up at least a gigawatt of electricity, enough to power roughly 750,000 US homes. If data centers were a country, by 2035 they would be the fourth-largest consumer of electricity after the US, China and India, according to BloombergNEF estimates.

The boom is not only reshaping landscapes and communities but also straining power grids, scarce environmental resources and, potentially, financial markets. At stake are hundreds of billions of dollars committed by investors to a technology whose business model is still a work in progress.

How are AI data centers different from regular data centers?
Data centers house racks of servers connected by dense cabling, cooling systems to stop them from overheating and generators to keep power flowing when the grid fails. The facilities contain water-treatment systems, cooling towers, electricity substations and battery rooms and are protected by fences, cameras and, often, security guards. Inside, temperature, airflow and power are tracked by sensors and fine-tuned by software to try to limit energy and water consumption.

The same goes for AI facilities, only they tend to be much bigger and process far more data per square meter. That’s because they use “accelerator” chips that process data in parallel rather than in sequence, allowing them to quickly perform the model training and inference tasks that are fundamental to AI. Tech executives are betting on a bigger-is-better approach, packing thousands of chips in ever larger data centers with the goal of building more sophisticated AI systems.

Just how big is the AI boom?
The numbers can be hard to comprehend. Bloomberg Intelligence estimates that generative AI could produce $2 trillion in revenue by 2032. Companies are racing headlong to meet that opportunity, aware that any bottlenecks that slow the pace of AI training and deployment could hand rivals an edge. Allianz Global Investors said the AI rush could end up rivaling the 19th century US railroad boom for its contribution to US economic output.

Gigawatts have replaced megawatts as the metric for the biggest data center campuses. Sam Altman, the chief executive of ChatGPT maker OpenAI, has said the company expects to spend “trillions” on infrastructure. Its Stargate project in partnership with Oracle Corp. and Softbank Group Corp. targets as much as 10GW of AI computing power and as much as $500 billion in spending in the US. Altman has said he hopes eventually to add 1GW of computing capacity every week. Meta Platforms Inc. is building a massive site in Louisiana that CEO Mark Zuckerberg has said will eventually grow to 5GW. The company says its AI platforms are “compute-starved” and it expects to invest as much as $72 billion this year and even greater sums in 2026.

Where is it all happening?
Until recently, data centers were most often located near big population centers to reduce “latency,” the time it takes for a server to respond to a user’s click of a mouse or tap of a screen.

AI is shifting that map. As latency is not so important when training AI models, the equipment that handles this work doesn’t need to be near cities. Companies are locating many facilities in isolated rural areas where land and resources are cheaper and connecting them to urban centers using high-capacity fiber-optic cables. OpenAI’s Stargate is establishing sites in Texas, New Mexico, Wisconsin and Ohio. The company is pursuing a similar buildout in remote regions abroad, including “Stargate Norway” near Narvik, within the Arctic Circle, and another in Argentina’s Patagonia.

Latency starts to matter again once the models have been trained and consumers start to use them, a separate process known as inference. So AI services such as ChatGPT and Google’s Gemini tend to route user prompts to data centers closer to towns and cities. As AI adoption grows in coming years, companies may direct more resources toward inference and once again build more facilities closer to where most people live.

The data center boom extends far beyond the US, as political leaders tout AI as key to an independent economic and technological future. China has tapped into a sprawling network of data centers powered increasingly by renewable energy. The country also added 429GW of new generation capacity in 2024 alone.

Who is leading the way?
Two groups are driving the boom: “hyperscalers” such as Amazon Web Services, Microsoft, Google and Meta, and “colocation” providers that lease space and computing power to multiple customers. Hyperscalers both build and own many of their campuses and also lease entire buildings or suites from colocation firms when they need a footprint in new markets. In colocation deals, the provider owns the land, the building and its on-site power and cooling systems, while the customer owns and controls the “IT stack,” including AI accelerators, servers and usually the racks.

There were 1,136 large data centers operated by hyperscalers worldwide at the end of 2024, double the number from five years prior, according to Synergy Research. Equinix Inc. and Digital Realty Trust Inc. dominate the colocation market. Equinix operates more than 270 data centers worldwide, including more than 70 in the US. Digital Realty runs more than 300 facilities across 50-plus metropolitan areas globally. Private equity firms control several big colocation operators: QTS (Blackstone Inc.), CyrusOne (KKR & Co. Inc.) and Switch (DigitalBridge Group Inc.). An investment group led by BlackRock Inc.’s Global Infrastructure Partners recently agreed to buy Aligned Data Centers LLC in a $40 billion deal.

There’s also a riskier corner of the market emerging as a growing number of younger, unprofitable startups push to build or co-develop their own data centers. OpenAI, which doesn’t expect to be cash-flow positive until near the end of the decade, is a key anchor of many infrastructure projects, raising questions about how it will pay for the buildout. Anthropic PBC, another leading AI developer, said in November that it plans to spend $50 billion on data centers in the US. And Poolside AI, a two-year-old AI startup, is partnering with CoreWeave Inc. on a planned 2-gigawatt data center in West Texas.

Meanwhile, so-called neocloud operators such as CoreWeave, Nebius and Nscale rent out accelerator chip capacity so customers can train and run their AI models, and they’re investing billions of dollars to build and operate data center campuses. Some previously focused on cryptocurrency mining, a notoriously volatile business.

Where is the money coming from?
The big tech firms are resorting to a complex mix of cash, debt and some complex financing arrangements with hardware suppliers. Blockbuster infrastructure deals struck by OpenAI have been criticized for being circular in nature: Nvidia Corp., which dominates the AI chip market, agreed to invest as much as $100 billion in OpenAI to help fund its data-center buildout. OpenAI, in turn, committed to filling those sites with millions of Nvidia chips. Weeks later, OpenAI partnered with Nvidia rival Advanced Micro Devices Inc. to deploy tens of billions of dollars’ worth of its chips and potentially become one of AMD’s largest shareholders.

A growing number of firms are turning to debt. Even OpenAI is considering it despite being years away from turning a profit. In some cases, tech firms are taking steps to ensure some of the debt doesn’t show up on their balance sheets. Morgan Stanley structured a $30 billion deal for Meta, where the debt would sit in a special purpose vehicle tied to Blue Owl Capital Inc. That made it easier for Meta to raise another $30 billion the usual way, in the corporate bond market.

While these deals may seem small for tech firms with famously large stockpiles of cash, the amount is poised to add up. Morgan Stanley estimates that tech firms and others will need as much as $800 billion from private credit in deals tied to specific assets by 2028.

Are AI data centers good for nearby communities?
Tech companies have touted the potential jobs AI data centers create. OpenAI has said Stargate will create “hundreds of thousands” of US positions. Most are temporary construction roles, as data centers don’t require a lot of permanent staff to operate them. A local development agency near the first Stargate site in Abilene, Texas estimated that, once up and running, the site will require about 100 high-skilled roles in the first phase. A similar-sized car factory typically employs several thousand skilled staff.

Local governments can benefit from data centers, with some levying taxes on the facilities as a way to fund services and reduce property-tax rates. Companies also pitch community grants and workforce training: Meta says its Louisiana campus will fund school grants and offer no-cost digital-skills training, and Google and Microsoft publicize similar education and training programs around their sites.

There are also downsides. For example, the AI buildout is having an impact on the cost and quality of power for other users. US wholesale electricity costs as much as 267% more than it did five years ago in areas near large data centers, a Bloomberg analysis found. That’s being passed on to customers through higher bills. A separate Bloomberg analysis showed that households near large data center campuses face degraded power quality. Voltage and frequency distortions in these areas can damage or destroy household appliances and electronics, raise the risk of electrical fires during surges, and cause lights to flicker. Over time, those conditions can tip neighborhoods into brownouts and blackouts.

Water is another potential flashpoint. Data centers run hot, and many facilities rely on evaporative cooling that uses large volumes of water. A typical 100-megawatt facility can use about 530,000 gallons of water a day. Increasingly, data centers are being established in places that have the least water to spare. A recent Bloomberg analysis found that nearly two-thirds of new US data centers since 2022 are in high-water-stress regions, sometimes leading to disputes over public supplies. In The Dalles, Oregon, residents pressed for details on Google’s water usage and challenged plans for new wells to supply data centers.

What is the impact on the climate?
For years, tech giants led corporate America with the scale of their pledges to cut carbon emissions. The AI buildout is blowing up those promises. In May 2024, Microsoft said its 2023 emissions were about 29% higher than in 2020, citing AI-driven data center growth. In July 2024, Google reported its 2023 emissions were 48% above the level 2019.

To keep data centers running, operators are drawing on whatever power sources they can — wind and solar, natural gas and even revived nuclear projects. But the sheer energy demand, coupled with pressure to move fast, has pushed some firms to fire up dirty gas turbines and aging coal plants.

In some countries, including Saudi Arabia, Ireland and Malaysia, the amount of energy required to run all the data centers they plan to build at full capacity exceeds the available supply of renewable energy, according to a Bloomberg analysis. The International Energy Agency projects data center electricity use will more than double between 2024 and 2030. Data centers now account for about a fifth of Ireland’s electricity use — 21% in 2023 and 22% in 2024 — and the IEA expects the share to reach nearly one-third as soon as 2026.

BNEF estimated that the world’s carbon dioxide emissions are expected to be 3.5 gigatons greater over the next ten years because of the additional fossil-fuel power generation required by data centers. That’s the equivalent of 10% of today’s total global emissions.

Google, Amazon and Meta say they plan to tap advanced nuclear power for future data centers, while Google and Microsoft point to AI tools that can cut emissions in other sectors, such as optimizing traffic lights and buildings. But grid bottlenecks mean a near-term risk of more fossil-fuel generation until enough clean supply is online.

Federal forecasts suggest a short-term bump in coal use as utilities race to meet new data center demand, but not a long-term reversal. The US Energy Information Administration expects natural gas to remain the dominant power source, slipping slightly after 2024, while coal’s share rises only briefly from 16% to 17% in 2025 before declining again as more renewable energy comes online.

Can power networks keep pace with the AI boom?
Much of the world’s power infrastructure was old and needed replacing even before the massive surge in expected demand linked to AI. System upgrades can take several years due to skills shortages and a variety of regulatory and permitting hurdles. Some completed AI projects are still waiting for a grid connection. The Trump administration is pushing to cut red tape for data center projects with its AI Action Plan, saying: “We need to ‘Build, Baby, Build!’”

But backlogs won’t go away overnight. At the end of 2024, about 2,300GW of proposed generation and storage were waiting to connect to the grid — roughly 1.9 times the US utility-scale capacity in operation — according to federal data. Historically, only about 14% of capacity in the interconnection queues ultimately gets built, and projects that reached operation in 2023 had a median wait of five years.

In many parts of the world, power is a scarce resource that’s effectively being rationed. Data centers’ surging power already triggered moratoriums before the latest wave of AI projects appeared. Ireland’s grid operator froze most new Dublin-area data center hookups through 2028. Singapore paused new builds in 2019 before reopening with tight caps. And the Netherlands imposed a national halt on large hyperscale proposals in 2022.

Evidence of grid strain is growing in the US. Dominion Energy Inc., which serves northern Virginia’s “Data Center Alley,” says it is in contract talks to supply roughly 47GW of new data-center load — more than the power needed for all of Virginia, underscoring the scale of demand building in the region. PJM Interconnection LLC, which coordinates the grid across all or parts of 13 US states, has warned that projected demand is outpacing supply across its footprint.

Then there’s the question of how all the extra power needed for AI will be generated. Companies are reaching for any solution they can, striking deals to revive dormant atomic power facilities and roll out relatively untried small nuclear reactors. In the UK, the government is exploring nuclear-backed AI Growth Zones to secure power for new campuses. In the US, Microsoft agreed to buy power from Helion Energy Inc.’s first fusion plant as early as 2028 — even though fusion is an experimental technology that’s never produced electricity in commercial quantities. Bloomberg