Newsletter · 2026-08-06 · Ashutosh Agarwal

Power Becomes the Real Constraint and Oracle Blinks - Powering AI Infrastructure - Week of August 6, 2026

Powering AI Infrastructure synthesis for the week of August 6, 2026, as earnings season put power, not chips or money, at the center of the data-center trade: Alphabet signed roughly $700 billion of mostly-power contracts and AWS sold out into 2028, while Oracle admitted it is worried a major customer, widely assumed to be OpenAI, could break its data-center lease.

Powering AI Infrastructure

Week of August 6, 2026: Power Becomes the Real Constraint and Oracle Blinks

This was earnings season for the giant tenants, and two things came through loud enough to reorganize the whole thesis. The first is that power, not money or chips, is now the thing everyone is fighting over, and one company put a number on it that makes your eyes water. The second is that the bear case finally got a real fact to stand on, instead of a forecast: a landlord admitted, out loud, that it is worried a tenant will walk away from a lease. Both happened this week, and the space between them is where the money gets made or lost.

TL;DR

• Alphabet quietly signed roughly $700 billion of long-term contracts, mostly to lock up electricity for data centers. That is not a typo, and it tells you what the real shortage is. Meanwhile AWS says it is "largely sold out" of AI capacity through the end of 2027 and into 2028. Demand is not the problem. Power is.

• The bear case got a body. Oracle has said it is worried one of its big customers, everyone assumes OpenAI, will break its data-center lease. That is the first crack in the "long-term contracts protect the landlords" story, and it landed the same week analysts pointed out that roughly three-quarters of the big clouds' AI revenue traces back to just two loss-making customers.

• The split between the two poles widened. AWS and Microsoft can show you the invoice; Meta got billed and couldn't. Meta now has $279 billion of future lease commitments off its balance sheet, up 53% in a single quarter, and the market punished the stock for it.

What's new

Alphabet locked up $700 billion of long-term contracts, and it's mostly for power. The single most striking number of the week came from a plain-spoken read of Google's own filing. On The Investing for Beginners Podcast, "The AI CAPEX Illusion with Thomas Chua", Thomas Chua of Steady Compounding, an analyst, not a company insider, pointed to a footnote (note 10) most people skip:

"Usually it's a couple of billion dollars. This quarter is spike up to 700 over billion dollars. And these are contracts that are... deliverable over the next few decades... And these are largely for electricity for data centers."

Why it matters: the capital-spending headline everyone fixates on changes every quarter, but a multi-decade contract you can't escape "unless they pay a heavy penalty" does not. And it says the binding constraint is not chips or money. As Chua put it plainly: "the main bottleneck right now for a lot of these huge companies is actually power. They're not getting enough power to fire up a lot of these data centers." He also flagged the accounting reason the strain hasn't shown up in margins yet: half-built data centers sit in a line called "construction in progress," and that line isn't depreciated until the buildings switch on, so reported margins keep rising even as the cash pours out. That is worth remembering when the bulls point at fat margins.

AWS is "sold out" into 2028, and the boss says demand still outstrips supply. This is the cleanest demand signal a landlord could ask for, and it came straight from the biggest tenant of them all. On Bloomberg Talks, "Amazon Web Services CEO Matt Garman Talks Capex Investment", Matt Garman, CEO of AWS, an operator, was unusually specific:

"Much of our capacity is already spoken through the end of 27 and even through much of 28. And so as we're investing, we're getting five-year commitments from customers... today demand still significantly outstrips supply."

Two things travel from that. First, the demand is broad, not a house of cards: Garman went out of his way to say AWS is "not like some others maybe just concentrated on just one or two large customers," growth is coming "across the board," from financial services to healthcare to startups. Second, the money keeps going up: Amazon's 2026 capital budget is now $220 billion, and Garman expects to "keep investing in CapEx next year as well." For anyone who owns space these tenants might rent, five-year commitments and a sold-out book are exactly the words you want to hear.

Oracle blinked, and that's the first real crack in the bear case. For weeks the bear argument has been a hypothetical: what happens if a tenant walks? This week it stopped being hypothetical. On The Exchange, "The Spending Shift, Model Mayhem, and Wrong on Warsh?", Matt Maley of Miller Tabak, an analyst, said it out loud:

"We already know where Oracle's admitted that they're worried that one of their customers, which everybody assumes is OpenAI, is going to break their lease on some of these big data centers."

That matters because the entire bull case for data-center landlords rests on long, signed, take-or-pay contracts. If the tenant behind those contracts can wobble, so can the backlog. On the same program, Steve Whiting of CIO Group gave the arithmetic that makes people nervous: hyperscaler spending is up about 85% this year, and its three-year compound growth rate of roughly 65% "if we were to extend this for years ahead, it'd be bigger than the whole economy." It won't be, which is his point.

The concentration problem got quantified, and it's ugly. The reason an Oracle lease wobble is scary is what sits underneath all these leases. On Better Offline, "The AI Demand Bubble with Ed Elson", Ed Elson walked through sell-side work (Wells Fargo, Barclays, UBS) showing that roughly 74% of Microsoft's AI revenue, and the majority of Amazon's and Google's cloud growth, traces back to just two customers, OpenAI and Anthropic, "two of the most unprofitable companies in the history of companies." Worse, a chunk of the money is circular: the clouds invested in those two labs, then book the labs' spending as their own revenue. The kicker Elson cited: Amazon recently handed OpenAI another $35 billion, "more than Amazon is estimated to make on AI this year in totality." Every landlord in this complex is ultimately leaning on a very short list of counterparties.

Microsoft and Amazon can show the invoice; Meta got billed for it. The market's verdict this quarter was brutal and specific. On The Real Eisman Playbook, "The AI Debate Gets More Complicated: Microsoft Has a Win, Meta Stumbles", Steve Eisman laid out the numbers. Microsoft's Azure accelerated to 43% growth with $19.6 billion of free cash flow (down 23% but still enormous). Meta, by contrast, missed, with earnings per share down 13%, research spending jumping from $13 billion to $22 billion, and free cash flow of a "mere $784 million... which is basically nothing." Then the line that should make every data-center bear sit up:

"Meta said that it has $279 billion in future lease agreements, mostly related to AI, that are not yet reflected on its balance sheet. That is up... 53% in just three months."

On Prof G Markets, "Microsoft Soars, Meta Sinks: Has The AI Narrative Flipped?", Gil Luria of DA Davidson summed up the mood by calling Microsoft "the adult in the room." The read-through for landlords: the tenants who can prove a return will keep leasing aggressively; the one who can't is now trying to become a landlord itself to justify the spend, which brings us to the debate.

The debate

The bull case: the shortage is real, it's physical, and it's getting worse.

The whole bull thesis reduces to one idea: whoever already owns powered, leased space owns something that cannot be reproduced quickly at any price. This week that got more support, not less. Start with the tenant demand: AWS sold out into 2028 on five-year commitments, and Alphabet signing $700 billion of mostly-power contracts is not the behavior of a company that thinks it has enough capacity. Then look at the grid. On Energy Evolution, "Grid operator PJM juggles reforms as power auction results hit price cap again", the PJM capacity auction, the market that keeps the lights on across the mid-Atlantic and Midwest, including the data-center capital of Northern Virginia, hit its price ceiling again. Without the cap, prices would have cleared at $554.72 per megawatt-day. Paul Siegel, CEO of LS Power, an operator that builds power plants, described a genuinely strange signal:

"Irrespective of the fact that natural gas prices have gone down over the course of this year... because supply and demand are getting tighter, we're actually seeing higher energy prices. And that's counterintuitive to the way that these markets typically behave."

When power gets more expensive even as fuel gets cheaper, you are watching scarcity, not cost-push. And the S&P Global analyst on the same program, Tanya Peavy, made the timing point that protects incumbents: even with a screaming price signal, "there isn't the time for the generation to come online" for years. New competing supply physically cannot show up fast. A landlord with power today is holding the scarce thing.

The most telling detail was structural. Siegel said the hyperscalers are increasingly "procuring new resources in a manner that falls on their balance sheet," and that going forward, matching AI demand to new supply is "much more likely to be bilateral than through a mandated auction." Translation: the biggest tenants are now signing private, long-dated power deals directly, because the normal market can't move fast enough. You saw exactly that on Telltales, "Weekend Update W2631", which flagged that Chevron just signed a 20-year agreement to supply Microsoft 2.67 gigawatts of behind-the-meter power, power generated on-site, bypassing the grid entirely, for a West Texas data center. As the host put it: "a hyperscaler deciding it would rather buy the gas than wait on the grid." Morgan Stanley, on Thoughts on the Market, "When a Data Center Comes to Town", framed the gap as a potential ~38-gigawatt shortfall through 2028. Scarcity, everywhere you look.

The bear case: it's a credit-fueled build, and the tenants are shakier than the leases suggest.

The bears didn't need to argue demand this week. They pointed at the counterparties and the balance sheets. Three facts, stacked:

• The lease backlog leans on two loss-making labs. Roughly 74% of Microsoft's AI revenue and half of Oracle's 600-billion-dollar backlog trace to OpenAI and Anthropic (per Eisman and the analyst work Elson cited). Concentration like that is the opposite of the "6,000 diversified customers" story the landlords prefer to tell.

• A landlord just admitted a tenant might walk. Oracle's worry about an OpenAI lease is the first time the "what if they default" question stopped being theoretical.

• The financing is getting expensive and creative. Eisman noted the AI data-center firm CoreWeave is raising $2.6 billion of debt at a yield "of more than, get this, 9%." On Eurodollar University, "AI Credit Spreads Are Suddenly Blowing Out... Just Like 2008?", Jeff Snider noted the four big spenders issued roughly $194 billion of AI-related debt by early July. And on Big Technology Podcast, "How The AI Bet Pays Off", David Cahn of Sequoia put the whole build at roughly $3 trillion of cumulative spend needing $4 trillion-plus of lifetime revenue to break even, with the majors "massively accelerating" because "people just don't care about the math anymore."

Thomas Chua drew the cleanest line between the two camps. The hyperscalers, Microsoft, Amazon, Alphabet, even Meta, can survive a demand air-pocket because they have core businesses (search, ads, retail, enterprise cloud) that can soak up spare compute. But the pure-play "neoclouds" and Oracle, he warned, are "levered up to their eyeballs" with no such cushion: "the thing that is going to stay constant is their interest payment." If demand growth so much as slows, that's where it breaks first.

Where I come out. This week the two sides were arguing about different clocks. The bull evidence is about now: AWS sold out into 2028, Alphabet locking up power for decades, the grid physically unable to add supply. The bear evidence is about what's underneath: two shaky labs behind most of the revenue, rising borrowing costs, and one landlord already nervous about a lease. Both are true. The honest way to hold them together is this: the demand is real and the scarcity is real, but the quality of the demand is thinner than the headline backlog numbers imply. Own the scarcity; watch the counterparties. And note, as always, that this week's bulls were mostly operators with capacity to sell, while the bears were mostly commentators. Weigh accordingly in both directions.

Names in play

Digital Realty (DLR), no US executive was out this week, but its Japan joint venture chief was, and he's worth hearing. On datacenterHawk, "Data Center Scale: When 1 Megawatt Was a Big Deal", Chris Hahn, COO of MC Digital Realty (Digital Realty's Mitsubishi JV, and he also runs DLR's Korea business), an operator, made the point that AI has broken the old playbook. Cloud demand used to reach Asia on a predictable 3-to-5-year lag behind the US; AI doesn't, because "no one knows where we're going to be in three to five years." The operational tell: liquid cooling is now "almost a requirement," and DLR is paying up front to build that flexibility in. His bottom line is the bull case in a sentence: "development times are getting longer. The power is getting scarcer... it doesn't matter whether you're building an AI data center or a traditional data center. You're still coming across the same constraints regarding power and construction and land." Next catalyst: whether development yields hold as future-proofing (liquid cooling, redundancy) keeps pushing build costs up.

Alphabet (GOOGL), Amazon (AMZN), Microsoft (MSFT), the tenants that matter most to the landlords, and this week they sorted into a clear order. Amazon and Microsoft are the "adults," AWS sold out into 2028, Azure at 43% growth with real cash flow. Alphabet is spending hardest of all ($700 billion of long-dated contracts, capex guided to $195–205 billion) and the market is uneasy about the free-cash-flow hit. For colocation landlords, the read-through is simple: these three are leasing and contracting aggressively and can absorb a slowdown. Watch whether Alphabet's power contracts translate into leased third-party space or purely self-build.

Meta (META), the cautionary tale. $279 billion of off-balance-sheet future leases (up 53% in a quarter), free cash flow down to almost nothing, and a plan to rent out spare compute that management still won't put numbers on. On The Investing for Beginners Podcast, Chua relayed that Meta's CFO said the company is fielding offers to buy its compute at "multiples of what they paid for," real demand, poorly explained. Next catalyst: whether Meta actually names a compute-rental customer. If it does, colocation landlords have a new, deep-pocketed competitor; if it doesn't, this was a distraction from a spending problem.

Oracle (ORCL), moved from "aggressive builder" to "the tenant everyone's watching." A $600 billion-plus backlog with roughly half tied to a single loss-making customer, an admitted worry about a lease breaking, and widening credit-default-swap spreads. This is the name that would sound the first alarm if the demand story cracks.

The tower and industrial REITs, quiet, and we'll say so. No operator or analyst voice this week moved a specific thesis on American Tower, Crown Castle, SBA Communications, Prologis, Rexford, First Industrial or EastGroup. The closest to an industrial read was buy-side: on Commercial Real Estate Secrets, "The REIT Playbook", Peter Zebrek of Easterly Funds named industrial one of his top three sectors, "a silver medalist" transitioning into the top tier on reshoring, alongside data centers and senior housing. And on the valuation question, Dividend Stockpile, "The State of REITs Part 2", David Auerbach of Hoya Capital made a useful distinction: REITs broadly are cheap versus their own history and trade at a discount to private-market values, "but more REITs than not are trading at discounts," while "data centers... towers... public storage are, frankly, trading at a premium." In other words, the market has already paid up for the towers and data centers; the value, if it's anywhere, is in the mid-caps. We're not going to manufacture views on the towers from silence.

Read-throughs

Utilities and independent power producers: the market is going private and bilateral. The most important structural shift this week is that the biggest tenants are increasingly buying power through direct, long-dated deals rather than the public auction, Chevron's 20-year, 2.67-gigawatt deal with Microsoft being the marquee example (Telltales). LS Power's Siegel expects most new supply to be contracted this way, and had a pointed warning about regulated utilities trying to get back into building power plants: what they really want, he argued, is to "pass that risk... on to the families and the businesses that pay the utility bill" (Energy Evolution). For utility investors, the tension is whether data-center load flows through the regulated rate base (utility upside) or gets contracted away privately (IPP and merchant upside).

On-site power vendors: the "speed-to-power" trade is real, and it's reaching logistics. The clearest concrete example this week wasn't a data center at all, it was a warehouse. On Energy Capital Podcast, "How linear generators can fit into the Texas grid", Craig Spiegel, CEO of Mainspring, an operator, described a Prologis site at the Port of Long Beach where the tenant, Maersk, needed to charge 96 electric trucks, and the local utility said it was years away and would cost millions in upgrades. So Prologis built its own islanded microgrid on-site, about 3 megawatts of generators paired with 18 megawatt-hours of batteries, no grid connection at all. That is the same "buy the power, don't wait on the grid" logic now showing up in industrial real estate, not just data centers.

Electrical equipment and generators: the picks-and-shovels kept beating. In the earnings run-through on The Real Eisman Playbook, two power-infrastructure names put up eye-popping numbers: Bloom Energy (BE), which makes on-site fuel cells that turn natural gas into electricity for data centers, grew revenue past $1 billion for the first time (up 166%) with earnings up 680% and the stock up over 90% on the year; and Quanta Services (PWR), the firm utilities hire to build power plants and grid, grew earnings 71% and raised guidance. Both are direct plays on the same power scarcity the landlords are wrestling with.

Nuclear as a data-center power source: slow, but Google is committing. On Engineering Matters, "#372 Sustainability Champion", the discussion covered Google's partnership with Kairos Power on a small modular reactor (its Hermes design, ~20–30 megawatts) to feed data centers, a reminder that the hyperscalers are shopping every power option, but that nuclear is a next-decade answer to a this-decade problem.

Freight and warehouse demand: still mixed. On FTR State of Freight, "Trucking Market Update, Week Beginning August 3, 2026", dry-van spot rates slipped to their lowest since May and volumes fell to nearly the lowest of the year, soft on the surface, but the manufacturing side firmed, with the ISM production reading up 6.3 points to 58.5% and order backlogs up 4.5 points to 55%. A recovering factory pipeline with still-soft trucking is a "watch, don't act" signal for warehouse demand.

What changed

Power went from "a" constraint to "the" constraint, with a number attached. Last week the story was that grid connections had slipped to 2030–2031 and building a megawatt now costs $10–15 million. This week it got concrete and enormous: Alphabet's $700 billion of mostly-power contracts, the PJM auction hitting its cap again, and a 20-year gas deal replacing the grid outright. Power is no longer a footnote to the AI trade; it is the AI trade.

The bear case stopped being a forecast. Last week the bears' sharpest argument was analytical, that this is a credit-driven real-estate cycle that will break when demand growth slows against fixed supply. This week they got a fact: Oracle's admission that a tenant might break a lease, plus hard numbers on how concentrated the whole backlog is (74% of Microsoft's AI revenue, half of Oracle's backlog, from two loss-making labs). The "what if a tenant walks" question is no longer hypothetical.

Meta's landlord ambitions went from an org chart to a balance-sheet number. Last week the news was that Meta had hired Amazon's former cloud chief to lead a compute-rental effort. This week we got the scale of the commitment behind it: $279 billion of future leases off the balance sheet, up 53% in three months, and a CFO confirming buyers are offering "multiples" for the capacity. The strategy is now measured in a quarter-trillion dollars, and the market still isn't sure it's a business.