Newsletter · · Ashutosh Agarwal

Power Not Demand Is the Whole Data Center Story Now - Powering AI Infrastructure - Week of July 16, 2026

Powering AI infrastructure newsletter for the week of July 16, 2026. Electricity, not compute demand, is now the binding constraint on data centers, landlords who control power and build fast are winning, and the bull-bear argument over whether the buildout is a debt-fueled bubble got louder and more specific.

Powering AI Infrastructure

Week of July 16, 2026: Power Not Demand Is the Whole Data Center Story Now

The infrastructure story this week wasn't about how much computing the world wants, everyone agrees that's enormous. It was about whether the electricity, the land, and the buildings can arrive fast enough, and whether the mountain of debt paying for it all is as safe as the balance sheets make it look. Landlords who control power and can build fast are winning; the argument over whether the whole thing is a bubble got louder and more specific.

TL;DR

  • Power is now the binding constraint on data centers, full stop. One market tracker says committed billion-watt projects have gone from a handful to 17 in about two years, and "behind-the-meter" on-site power plants have shifted from clever workaround to necessity.

  • The bull-bear fight sharpened. Bulls point to real usage and rising returns ("there are no dark GPUs"); bears point to leases that are really 30-year debt, a spending-vs-revenue gap of roughly $755 billion against under $50 billion, and a memory-chip cycle that always turns.

  • The read-throughs are flashing green on the ground: industrial landlords are re-pricing space around electricity (Waymo paid up for a warehouse mostly because of its amperage), intermodal rail just had its second-best year ever, and cooling/electrical contractors are quoting year-plus lead times.

What's new

Power availability, not tenant demand, is now the thing that decides who builds. On The Data Center Frontier Show, Colby Cox of DC Byte, a firm that tracks every known data-center project, laid out the clearest numbers of the week. Going back to early 2023, there were three "committed" projects of 900 megawatts or larger and seven in early planning. Today there are 17 committed projects above 900 MW and 49 in early stage, and 15 of those early-stage projects are above two gigawatts "with a couple getting into the 10 range." His blunt summary of why so many projects are stuck: it's power. "Behind the meter" (building your own on-site power plant instead of waiting for the utility grid) has gone from a workaround to "a necessity for a little while." (The Data Center Frontier Show) That matters to landlords because whoever can deliver electricity first wins the lease.

The people actually driving power demand aren't the cloud giants, they're the AI labs behind them. Catalyst with Shayle Kann had the most useful power discussion of the week. The argument: when you see Amazon or Microsoft building gigawatts, a big share is really for OpenAI and Anthropic. And the grid simply can't keep up, the buildout is running at "tens of gigawatts per year… let's say 50% per year," while the grid is adding maybe 5-6 gigawatts of gas and 20-25 (nameplate) of solar and batteries a year. As the guest put it, when a lab like Anthropic wants to go from 1.5 gigawatts at the end of 2025 to more than 10 gigawatts by 2027, "you're building a Google in 2 years… That's not going to happen with the grid." So they bring their own generation. One striking detail: a gigawatt of capacity is roughly $50 billion of capital (more like $60-70 billion if someone else owns the power). (Catalyst with Shayle Kann)

A data-center landlord's-eye view: the winner is whoever starts sooner. On QTS's own podcast, The QTS Experience (Ep. 254, featuring Rob Hopewell), a senior QTS executive, a genuine operator/insider, since QTS is one of the largest data-center developers, owned by Blackstone, was refreshingly plain about the game. Everyone uses the same construction crews and the same equipment, he said, so "we're all building at the same timeline. It's just a matter of who starts sooner." QTS's edge is a "rapid design" that lets it put up a building in 8 to 10 months, and hyperscalers increasingly want the whole thing turnkey, "one throat to choke." His view on the bubble talk: it bothers him, because unlike the dot-com era "there's real revenue involved… and people are getting real value." A fun tell on how fast this snuck up on everyone: when Blackstone bought QTS in September 2021, he said, the deal paperwork "didn't even have the letters AI in it." (The QTS Experience)

The cost to build a data center has roughly doubled per gigawatt, and that's a feature for the pick-and-shovel names. On Motley Fool Hidden Gems Investing, the hosts walked through Meta's giant Louisiana project: the original estimate was about $10 billion for 2 gigawatts (~$5 billion per gigawatt); it's now $50 billion for 5 gigawatts, so ~$10 billion per gigawatt, cost per gigawatt has doubled, driven by pricier power, cooling and components. To power one site, "Meta's paying for 10 new natural gas plants." The read-through: contractors who do the specialized electrical and (now mostly liquid) cooling work are booming. They flagged Comfort Systems (FIX), revenue up 56% year over year, earnings per share more than doubled, backlog up 80%, "sold-out" language and lead times "well over a year", and Celestica (CLS), which assembles the custom liquid-cooled AI server racks. (Motley Fool Hidden Gems Investing)

Demand is still, in the operators' words, "insatiable." On All-In, Cerebras CEO Andrew Feldman, an operator selling AI chips into these buildings, said the company sits on a $25 billion backlog, with OpenAI, Anthropic, SpaceX's xAI, Google, Microsoft and AWS all wanting more capacity than can be built. He described individual buildings "the size of football fields that have more power coming into them than midsize cities," going up across the US, Canada, the Nordics, France, the Middle East, even Kazakhstan. His line that sums up the moment: "There is so much demand right now that there is no silicon that will go unused." (All-In)

The debate

This is the fight that decides whether infrastructure landlords are entering a multi-year growth cycle or standing next to a credit accident. Both sides showed up this week with specifics.

The bull case: real usage, real returns, and no idle capacity. The clearest bull was investor Gavin Baker of Atreides Management on The a16z Show. His core point is a comparison to the year-2000 telecom bust, which was defined by "dark fiber" (cable laid in the ground and never switched on); at the peak, 97% of it was dark. Today, he argues, the opposite is true:

"Contrast that with today. There are no dark GPUs."

He backs it with numbers: the big spenders are all public companies, and their returns on invested capital have risen about 10 points since they ramped up spending; Google says it processed 150 times more AI "tokens" in the last 17 months; and Nvidia trades around 40 times earnings versus Cisco's 150-180 times at the 2000 peak. The buyers, he notes, generate roughly $300 billion of free cash flow a year and sit on ~$500 billion of cash, against roughly $40-50 billion to light up a gigawatt. (The a16z Show) The QTS insider above is the operator version of the same case: real revenue, real value, keep building. (The QTS Experience)

The bear case: the leases are debt, and the cycle always turns. On RiskReversal Pod, Louis-Vincent Gave and Peter Boockvar went after the accounting. The big three cloud companies are spending roughly $700 billion this year (McKinsey sees ~$6.5 trillion through 2030), and Oracle is spending essentially 100% of its revenue on capital projects. The part they say the market is missing:

"The leases that they're signing is essentially debt… a lot of them are guaranteeing out 30 years. Having to upgrade the technology and the new Nvidia chip, that's hugely expensive… there's no rainbow in year four here."

They also picked apart the "earnings are great" story: quarterly earnings up ~28% fall to ~17% once you strip out investment gains, and to mid-single-digits once you strip out semiconductors. (RiskReversal Pod) Boockvar sharpened it two days later on Wealthion: "The bubble is clearly on the CapEx side, not… the technology." His warning aims squarely at the chipmakers the market has been rewarding: Micron at ~85% gross margins and ~10x sales (about where it peaked in 1999-2000), with new supply arriving in 2027-2028 and Chinese competitors "coming for your gross margin." His rule of thumb worth remembering: on a highly cyclical stock, a low P/E usually means peak earnings, not cheapness. (Wealthion)

The bear case, escalated: the debt is hiding off the balance sheet. On the Providence Financial Retirement Show, host Anthony Saccaro made the most alarming version. His figures: AI data-center debt issuance jumped from $166 billion in 2023 to $625 billion in 2025 (nearly 4x); data-center asset-backed securities rose 19-fold between 2022 and 2025; and much of it sits in off-balance-sheet vehicles (he cites a $30 billion Meta data center and SoftBank's $10 billion into the $500 billion Stargate project). His single "clearest, most measurable" warning sign: hyperscalers are set to spend ~$755 billion on AI in 2026 against total AI industry revenue estimated under $50 billion. His analogy is 2008, not that data centers are subprime mortgages, but that "hidden leverage amplifying a downturn is exactly what turned a real estate correction into a global financial crisis." (Providence Financial Retirement Show)

Where the week landed. The bulls are mostly talking about usage and today's returns; the bears are mostly talking about how it's financed and what a normal cycle does to it. Notably, nobody this week was arguing the classic REIT-specific bear points, data-center oversupply or falling lease rates. Every operator voice said the opposite: demand outstrips supply and power is the only thing holding them back. The risk on the table isn't "not enough tenants." It's a financing/credit shock. Worth keeping the two straight.

Names in play

  • Meta. The clearest single-stock story. On The Rundown, Public.com's Zaid Admani framed Meta's roughly $145 billion capex plan for the year against the fact that it still makes 98% of revenue from ads, but last quarter grew revenue 33% to $56.3 billion, its fastest since 2021. The new twist: on July 1, Meta unveiled "Meta Compute," renting out spare AI capacity, and the stock jumped 9% on the day while neocloud names CoreWeave and Nebius fell. The playbook copies xAI, which Admani says now rents capacity to Anthropic and Google for more than $2 billion a month combined. Next catalyst: whether "Meta Compute" turns capex into visible rental revenue. (The Rundown)

  • QTS (private, Blackstone). Not investable directly, but the most useful landlord tell of the week: 8-10 month builds, turnkey triple-net leases, and "start sooner" as the whole competitive edge, a good proxy for how the listed data-center names (Equinix, Digital Realty) are competing. (The QTS Experience)

  • Prologis (PLD). Called out by name on The Industrial Real Estate Podcast as "the bellwether for the industry." A Kidder Matthews broker (an operator on the ground) said Prologis is pushing occupancy to 97% by year-end in his Seattle market (95% now) and, tellingly, is spending ~$4 million to bring extra power into a building so it can land a premium tenant (more below). (The Industrial Real Estate Podcast)

(No podcast this week moved a specific thesis on the tower REITs (American Tower, Crown Castle, SBA) or on Equinix or Digital Realty directly. When a real operator or analyst voice shows up on those, they'll go here.)

Read-throughs

  • Power / utilities & on-site generation. The behind-the-meter shift is the read-through with the longest legs. Catalyst detailed who's doing what: Meta running gas with almost no backup (10 MW for a 100 MW site); OpenAI and Oracle deploying 2.3 gigawatts of reciprocating engines in Shackleford County, Texas; Google and Amazon instead locking up grid power (together roughly 150 gigawatts of contracted utility loads as of end-2025) with Amazon doing large nuclear and gas deals (Talen, Comanche Peak). Everyone is scrambling for gas turbines from GE, Mitsubishi and Siemens; fuel-cell maker Bloom Energy is a beneficiary of scarcity. (Catalyst with Shayle Kann)

  • Electrical & cooling contractors. As above, Comfort Systems (FIX) with an 80% backlog jump and year-plus lead times, and Celestica (CLS) building custom liquid-cooled racks, are the clearest "picks and shovels" of the AI buildout. Data centers are now generally liquid-cooled, not air-cooled, a structural tailwind for that work. (Motley Fool Hidden Gems Investing)

  • Industrial real estate is starting to price power like a premium amenity. The most vivid example came from Seattle on The Industrial Real Estate Podcast: Waymo is signing with Prologis for a ~145,000-square-foot space carrying 4,000 amps of power versus ~1,500 standard, at roughly $1.70 per foot (well above market), largely because of the electricity. The broker's prediction: rents may soon get quoted with a power "override," because for these tenants "the location is almost secondary… if you don't have the power you can't actually do your operations." Same episode: Meta's $13 billion, ~1-gigawatt data center announced locally drew heavy community pushback, a reminder that NIMBY opposition and moratoriums are now a real gating factor. Local fundamentals are still soft, though: Seattle industrial vacancy is just over 10% (worse than the financial crisis locally), rents are down 20-35% from peak, and ~90-95% of tenants are just renewing rather than expanding. (The Industrial Real Estate Podcast)

  • Freight & inventories, a genuinely constructive demand signal. On FreightCasts, the intermodal association's director said 2025 was the second-best year for intermodal rail on record (18.5 million loads, just shy of 2018's 18.9 million), with domestic volumes up 10% over the last four weeks. Sonar analysts flagged that spot trucking rates are running 27 cents a mile above contract rates (a spread matched only during COVID extremes) implying contract rates could climb ~50 cents/mile (~20%). Rail carload volumes are up double-digits across most commodity groups (ex-coal), and retailers are restocking (imports up ~8% year over year in June). That broadening industrial and retail demand is the leading edge of warehouse absorption. (FreightCasts)

  • 3PLs & e-commerce fulfillment. On The New Warehouse Podcast, the growth lead at Basic 3PL said the firm is "doubling as we speak," now above 25 US warehouses, building vertically with 44-foot ceilings and heavy robotics to handle unpredictable social-commerce spikes. It's a small operator, but a useful ground-level signal that fulfillment demand, the bread and butter of the logistics REITs, is still expanding. (The New Warehouse Podcast)

  • Public REIT / capital-markets backdrop. CRE Exchange's midyear recap tied it together: data-center construction was the one bright spot in an otherwise soft construction picture (up double-digits, over 30% at a point) before power and grid limits bit; hyperscalers issued over $100 billion of new debt in the first half, meaning "the public unsecured market is now directly financing the data center buildout" with "very deep, very cheap capital pools." Listed real estate had a strong half (the NAREIT All Equity index up ~14%, with data-center, lodging and specialty REITs all up 20%+) helped by REITs entering the year at their largest valuation discount to the S&P in 25 years. The hosts flagged the key divide to watch: operators creating real value versus those just riding cap-rate compression. (CRE Exchange)

What changed

One tone shift is worth marking. For most of 2024 the story was hyperscalers self-building their own data centers. Multiple voices this week described the pendulum swinging back toward leasing: on Catalyst, the point that "the leasing markets really go up tremendously" as Oracle, Meta and Microsoft increasingly scale through third-party operators (who often bring the behind-the-meter power); and the QTS insider describing hyperscalers wanting turnkey, soup-to-nuts buildings. If that holds, it's a direct positive for the landlords who own the buildings and control the power, and the single most important trend for the infrastructure-REIT thesis right now. (Catalyst with Shayle Kann, The QTS Experience)