The AI revolution requires an unprecedented build-out of data centers. At least, that’s how the narrative goes. A quick peek at the data reveals a funny detail—most aren’t being built. An energy shortage? Nah, most don’t even apply for a connection to the power grid. This is about a little tax code change that sparked a stampede to become landlords of the internet. A multi-trillion dollar speculative bubble, complete with flippers. Seriously.
The Scramble To Build Data Centers
A scramble to build data centers is a bit of an understatement. In Canada, over 200 data centers have been announced or are in development. That includes 159 hyperscale facilities—those with at least 50 megawatts (MW) of capacity. Out of those, 116 are planning for over 100 MW of capacity. That’s a big jump from the 5 hyperscale facilities in Canada today.
The US is seeing a much bigger boom. Vancouver-based Aterio tracks 8,434 data centers, from announcement to operations. Their data shows 2,145 in operation, 888 under construction, and 4,442 announced. “The remaining 959 are delayed, land banked, not approved or cancelled, which we keep tracking because they often come back,” notes the predictive analytics firm.
If that seems unrealistic, wait until you crunch some numbers.
Data Center Energy Math Doesn’t Add Up
Big numbers are hard to appreciate, so let’s make them relatable. The average home in Canada and the US uses ~11,000 kilowatt hours (kWh) of electricity per year, about 1.25 kilowatts (kW). A megawatt (MW) is 1,000 kW, equal to the power draw of ~800 households. A gigawatt (GW) is 1,000,000 kW or 1,000 MW, equal to the power draw of ~800,000 households. Easy peasy, right?
Now let’s apply that. Canada’s 116 hyperscalers that need a minimum of 100 MW need 11.6 GW of power. That’s equal to the power of roughly 9.28 million households in a country with only ~16.85 million households. Canada’s Minister of Artificial Intelligence stated there are 20 GW in the pipeline.
Just 750 large projects in the US will be using 45 GW of additional power, equal to 36 million households. That’s a lot of automated Realtor emails for our spam boxes.
Does that sound a little unrealistic? Canada’s government seems to agree. When asked about the AI minister’s 20 GW claim, officials didn’t just concede it was an exaggeration. They outright stated, “most of this proposed capacity is not expected to proceed.”
Oh, cool. Wait… what? We’ll circle back, but first, let’s talk about the scramble to become landlords of the internet. That’s not me being cute. It’s in your tax code.
The Scramble To Become Landlords of The Internet
The internet is synonymous with giants like Google and Meta, but they’re mostly tenants. They’re now building data centers, but for the most part, they pay rent to “internet landlords,” a term coined by University of Maryland Associate Professor Daniel Greene.
“A tremendous amount of money and innovation may go into Google’s advertising networks, but those networks are largely housed in cages rented from a third party,” he explains.
He argues that internet landlords extract rent through three primary asset types: Internet Exchange Points (IXPs) for public network peering, co-location facilities where tenants rent space and server bandwidth, and managed storage. Internet companies may build products, but they have to rent infrastructure to get it to you.
One Weird Tax Trick That Ruined The Internet
This isn’t a happy accident but a long battle that only ended in 2016, which is why you’re hearing about it now. “To secure their status as [Real Estate Investment Trusts] REITs, internet landlords had to prove that the internet was a physical place you could own, manage, and lease, like an apartment building or a storage unit,” explains Greene.
IRS Private Letter Rulings (PLRs) from 2007 to 2014 helped to close gaps, including classifying payments as “rent.” The idea took off in 2015 after a PLR cleared the way for data center Equinix to become a Real Estate Investment Trust (REIT). Congress then codified the rules with the PATH Act (2015), and the IRS in 2016. Data centers were fully cleared to become tax-advantaged, globally traded, liquid financial assets.
What’s so important about being a REIT? Until the 1960s, real estate was incredibly illiquid, reducing its value. To stimulate investment, REITs became eligible for a pass-through structure. Corporate profits are usually subject to “double taxation.” They pay income taxes on net profits, and then they can pay out dividends from the profits. Investors then pay their own income taxes on that additional income. REITs distribute virtually the entire net profit as dividends, and investors pay income taxes on that income.
Roughly 10 years ago, the internet’s landlords secured the same treatment. It’s relatively new, which is why you’re only recently hearing about it. That and the recent rush to target retail investors with IPOs, such as Brookfield’s Csquare and Blackstone Digital Infrastructure Trust (BXDC), the latter being a REIT itself.
On that note, let’s circle back to those data centers that are very unlikely.
Nearly 5 Data Centers Announced For Each Built & Braggawatts
The announcements (and debt) to build data centers are booming, but the building—not so much. In Canada, only 1 of 8 hyperscale facilities slated for a 2027 launch has actually broken ground. JP Morgan recently warned that over 60% of US data center capacity for 2027 hasn’t broken ground. Nearly 5 data centers are announced for every one built. That’s excluding cancelled projects.
The media widely attributes the delays to NIMBYs, who don’t want to share their energy. However, most data centers haven’t even asked for energy. A recent study of 252 active US data center announcements found 129 had no utility filings, SEC disclosures, or issued permits. That’s like announcing you’re going to bake a cake, but don’t have flour, sugar, an oven, or a kitchen.
Unproven and exaggerated data centers are so common that the industry has a term—braggawatts. These are, um, inspirational announcements with virtually zero chance of being built. “The amount of power, and even the amount of financing required to meet these upper end projections, it doesn’t exist,” explained Chris Sharp, CTO at Digital Realty, the firm that pioneered data REITs in 2004.
Sharp warns that developers are increasingly entering the space and overstating their abilities. Raw land of that scale is expensive. The cost to build out a hyperscale is almost unimaginable to most. Lucky for the speculators, they don’t need to build it if they can flip it first.
Data Center Flippers
Is it really a real estate bubble without flippers? A whopping 72% of US AI data center power requests are phantom loads. Phantom loads are speculative requests that require utility companies to set aside dedicated power. The artificial shortage pushes energy prices higher, but improves the project value. The flipper may not have the funds to build it out, but institutions that need to allocate investor funds do.
The problem isn’t going unnoticed. In Canada, B.C. Hydro now requires C$25k per MW in bid security, or C$2.5 million per 100 MW facility. Fail to build by the deadline, the funds are forfeited. It’s a little light in contrast to what’s emerging in the US.
Earlier this year, Texas adopted a US$100k study fee plus a deposit of US$50k per MW. A deposit of US$5 million is a sign of a pretty strong commitment. Though Ohio has the toughest test we’ve come across. Applicants pay US$10k-US$100k per load study, and 85% of contracted capacity for 12 years, even if less is used.
The scale of debt in the pipeline suggests that it may be an underreaction. Canada’s financing is more opaque, but experts have noted that debt raises “start at several hundred million dollars and exceed well over a billion dollars.” Though that’s chump change compared to the mounting liabilities in the US.
US data centers and related infrastructure are tied to US$1.3 trillion in debt, according to Atrium. Morgan Stanley also estimates a US$1.5 trillion “gap” in financing for data centers through 2028. Only the equivalent of Canada’s GDP? I guess we all have to tighten our purse strings in this economy.
AI has hit an inflection point, and that means more infrastructure is needed. But technological revolution and a bubble aren’t mutually exclusive. The internet was revolutionary, but that didn’t stop the Dot Com bubble. AI isn’t different—it’s just wrapped around a real estate bubble and cushioned by more tax dollars.
