There is “infinite” money available now to build out data centers. This has translated into record spending on the stuff that goes into them: chips, servers, HVAC systems, gas turbines, transformers, power plants and lines. (But) there are physical limits to how all that can be delivered (because) there’s no spare manufacturing capacity to build it faster. There are other bottlenecks, such as securing construction permits and connecting turbines to natural gas. Some projects are already delayed. AI companies have to generate new revenue to justify the $5 trillion committed meaning AI products have to create an additional $650 billion a year, indefinitely, to give investors a 10% annual return - more than 150% of Apple’s yearly revenue, and far from OpenAI’s yearly revenue of $20 billion. The projections of AI companies and their partners don’t reflect these shortages of equipment.
Here’s a way to evaluate the riskiness of the world’s collective investment in artificial intelligence: Can we even build all the necessary physical infrastructure? And if so, will the resulting AI-powered products generate enough revenue to pay back that investment?
Let’s look at the data on where we are—and where we might be going.
The rate at which tech companies and startups are investing in AI shows no signs of slowing. In their latest financial reports, all of the big spenders revealed that their current investments had grown significantly, and projected that this trend would continue.
There is effectively “infinite” money available right now to build out new data centers, says Jim Schneider, a senior research analyst at Goldman Sachs. All this investment has translated into record spending on the stuff that goes into data centers—aka “AI supercomputers”—all those chips, servers, HVAC systems, transformers, gas turbines, power lines and power plants.
There are absolute, physical limits to how quickly all of that can be delivered. As a result, some projects are already being delayed.
There are huge numbers of AI data centers in the planning stage, which means developers are in the process of securing land and getting permits, says Steve Tusa, a managing director at JPMorgan Chase.
Today’s record number of data centers at the planning stage is indicative of what Raymond James managing director Frank Louthan calls “kind of a gold rush mentality.” Connecting these planned data centers to power and fiber, and finding a buyer or tenant who wants one, in that particular location is another matter.
“A lot of people who have some land think they’re going to get into this industry and make some money, and they didn’t understand the risks,” he adds.
U.S. data-center build-out status by year of construction or planning, in gigawatts of capacity
80
GW
Stalled
60
Planned
40
20
Underway
0
Built
2010
’11
’12
’13
’14
’15
’16
’17
’18
’19
2020
’21
’22
’23
’24
’25
Despite these challenges, the biggest tech companies are spending more than ever on AI infrastructure. It’s claiming an ever-larger share of their revenue.
Amazon
Microsoft
Alphabet
Meta
Totals, quarterly
As a percentage of sales, annually
$100
billion
40
%
75
30
50
20
25
10
0
0
2017
’18
’19
2020
’21
’22
’23
’24
’25
2017
’18
’19
2020
’21
’22
’23
’24
’25*
An increasing share of that investment is fueled by debt. OpenAI, Anthropic and other startups continue to lose money, and must fund most of this investment by selling off pieces of themselves to investors and by issuing this debt. A recent Goldman Sachs report projected that OpenAI alone could spend $75 billion in 2026. Even cash-flush Meta Platforms is signing complicated debt deals involving private-equity firms.
Any key component of an AI supercomputer that is in short supply will determine how much AI infrastructure can actually be built. For example, even if there were enough power available for every new data center envisioned, there is currently a shortage of transformers, those gray, house-size boxes that connect buildings to the power grid, says Schneider of Goldman Sachs.
As a result, his projection for how much new data-center capacity will actually be built across the world through the end of 2027 is relatively conservative compared with other Wall Street and real-estate industry estimates.
Global data-center capacity, by energy usage
One square = 1 GW
For comparison, this is the
average hourly electricity
production for the entire U.K.
Current
capacity
End of
2026
End of
2027
70.8 GW
93.3
109.2
Global data-center capacity, by data-center footprint
One square = 6 million sq. feet
For comparison, this is the total
footprint of Costco warehouses
Current
capacity
End of
2026
End of
2027
455.0 mil.
570.5
645.6
Scott Strazik is chief executive of GE Vernova, one of the biggest U.S. manufacturers of equipment to generate electricity, such as transformers and natural-gas turbines. Strazik recently said on The Wall Street Journal’s Bold Names podcast that nearly all of the company’s output is booked through 2028.
Translation: There’s no spare manufacturing capacity to build that equipment any faster.
Strazik says that meeting the projected demands for U.S. electricity production isn’t something his industry can solve in the next five years, but rather in 10 or 15. Meanwhile, there are other bottlenecks, such as securing construction permits and connecting turbines to natural gas.
All that spending by the big AI and cloud-computing firms—which is translating into real estate, construction, energy and other massive costs—will have to be recouped somehow. The expectation is that increased spending by consumers and businesses on AI-powered products will make that happen. Today’s chatbots and image generators will, companies hope, become pricier AI agents that act on our behalf, and even humanoid robots.
In such a future, AI’s cloud-service providers could see rapid revenue growth. Based on best-case scenarios, the financial-services firm Raymond James projects that AI cloud revenue will explode in the next half-decade, to almost nine times its current value.
$400
billion
PROJECTION
Nebius
Amazon
300
Alphabet
CoreWeave
200
Oracle
100
Microsoft
0
2024
2025
2026
2027
2028
2029
2030
Recently, analysts at JPMorgan Fundamental Research created a financial model that projects total investment in global AI infrastructure through 2030, taking physical limitations into account. That figure is $5 trillion.
Then they calculated how much new revenue these companies will have to generate to justify that $5 trillion investment in AI supercomputers. The result: AI products would have to create an additional $650 billion a year, indefinitely, to give investors a reasonable 10% annual return. That’s more than 150% of Apple’s yearly revenue, and a far cry from OpenAI’s current revenue of about $20 billion a year.
This equates to every iPhone owner in the world paying an extra $35 or so a month for products and subscriptions, according to the JPMorgan analysts.
While these per-user revenue figures sound unlikely, there will be many ways to make money from AI, including advertising and specialized AI for high-paying enterprise customers, says Louthan of Raymond James.
The takeaway: The projections of AI companies and their partners don’t reflect shortages of equipment. At the same time, these projections assume a gargantuan market for AI-powered products and services. Analysts can’t agree whether that market will materialize as quickly as promised.
AI is already dramatically transforming our lives and businesses. But the real world limits how quickly companies can scale up these next-level supercomputers, and it’s unclear who will pay for all the resulting services.
It would be wise to moderate our expectations—or at least adjust the timetable—for the AI revolution.
Online Social Media Marketing is a powerful strategy that helps businesses connect with their audience, boost brand awareness, and drive meaningful engagement across platforms like Facebook, Instagram, TikTok, and LinkedIn.
ReplyDelete