The Excess

The Paranoidist | Issue #22 By Paul Morin | July 6, 2026

Last week a company told its investors it had built more artificial intelligence computing power than it can use, and the stock went up almost nine percent. That is the sentence I keep turning over. Meta said it is forming a business, reportedly called Meta Compute, to sell its surplus AI capacity to outside customers, a move that would put it in competition with Amazon, Microsoft, and Google in the cloud. Wall Street read the news as a maturing bet: a company that has poured many tens of billions of dollars into data centers has found a way to turn that iron into a second revenue line, the way Amazon once turned its own spare servers into the most profitable business it owns. The commentary said Meta had finally figured out how to make its infrastructure pay.

Read generously, it is a good story, and I want to concede that before I doubt it. Cloud is a real, enormous, high-margin market, and Meta has hyperscale economics, custom silicon, and one of the largest computing footprints on the planet. If it has capacity to spare, selling it beats letting it sit idle. The problem is the premise buried inside the good story. To sell spare capacity, you first have to have spare capacity, which means you built ahead of your own need. For three years the entire case for the AI build-out has rested on one claim: that demand for compute is effectively infinite, that the scarce thing is the hardware and never the buyers. The thing that keeps me up is not that Meta is building a cloud business. It is that the most reassuring word in the announcement was "excess," and a builder with excess is a builder telling you, in the politest possible terms, that the demand ran out before the capacity did.

Issue #21, "The Ratepayer's Subsidy," ended on a question I said was sitting underneath all the others: how much of AI's apparent demand is genuine end demand, and how much is the manufactured demand of a small group of firms financing one another's build-outs. There I worried the electrical grid is being built and priced against forecast data center load that counts as firm when it sets the bill and stays optional for the tenant who ordered it. This issue turns that worry one notch upstream, from the grid to the builders themselves. The tenants are starting to tell us something, and the first thing one of the largest of them said is that it has more than it needs. Call the two readings of that admission what they are. Monetized excess is spare capacity turned into product: the build was right, demand is real, and the overage is a bonus that becomes a business. Admitted excess is a confession dressed as a strategy: the build ran ahead of the demand it was sized for, and the overage is the first crack in the infinite-demand thesis, booked as revenue in the one place a write-down might otherwise appear. The same idle server is a new profit center or a stranded asset, and the entire question is which one you are looking at.

Start with the comfort story at full strength, because on its own terms it is coherent. The most profitable idea in modern technology, Amazon Web Services, was born exactly this way: a retailer noticed it had built more computing capacity than its own operations required and rented the surplus to everyone else. Two decades later that surplus is the engine of the company. When Meta says it will sell spare AI compute, the bull case is that history is rhyming. Meta spends almost all of its money serving one internal customer, its advertising machine, and still gets roughly ninety-eight percent of its revenue from digital ads. A cloud business would diversify that, monetize a multi-hundred-billion-dollar asset base, and give the market a reason to value Meta as something more than an ad company that got into an arms race.

There is real demand on the other side of the trade. Smaller AI firms and enterprises cannot get enough computing power, and Meta selling into a supply-starved market is not absurd. The timing looks disciplined rather than desperate: a company committed to spend as much as one hundred forty-five billion dollars this year, having faced open investor anxiety about that number, is showing it can turn the spending into cash. On the comfort story's assumptions, this is a strong company converting a cost center into a profit center, and the nine percent pop was the market rewarding capital discipline.

Now read the same announcement from the other side, and start with the word. You do not sell what you need. You sell what you built and did not use. When a hyperscaler that has spent three years insisting it cannot get enough compute announces it will rent out the surplus, the surplus is the news, not the rental. For most of this cycle the operating assumption, stated on earnings call after earnings call, was that being short on compute was the one unforgivable mistake, that any amount of building was justified because demand would always outrun it. "Excess" is the first public admission from inside that consensus that, for at least one of the largest builders, it did not. The build reached the demand and passed it.

The resemblance to the grid problem I flagged in Issue #21 is exact, and it has the same cruel feature. There, the more speculative the demand, the stronger the scarcity signal it threw off, because a larger forecast lifted the price whether or not the load ever energized. Here the mechanism runs through the income statement instead of the auction, but the shape is identical: the more a company overbuilds, the more "spare capacity" it has to sell, and the more it can relabel an overbuild as a growth initiative. A cloud business born of genuine surplus and one born to find revenue for capacity with no internal use look identical from the outside. Both are announced as strategy. Both send the stock up. The market cannot tell, from the press release alone, whether it is cheering foresight or applauding a company for finding a buyer for its mistake.

And notice who is moving. The chip and memory makers, the sellers of the picks and shovels, spent the first half of this year soaring while the companies doing the building came under pressure. The market is increasingly willing to pay the sellers of capacity and increasingly skeptical of the buyers, which is a strange thing to believe if you also believe the buyers' demand is infinite.

This can run for a long time before it resolves, because the two most important gauges, the accounting and the financing, are both built to read excess as scarcity. Take the accounting first. A data center full of graphics chips is only profitable if it lasts long enough to earn back what it cost, so how long you assume it lasts is not a footnote; it is the number that decides whether the whole build makes money. The hyperscalers have been depreciating their AI hardware over five to six years. The investor Michael Burry, among others, has argued in public that the real economic life of a leading-edge chip is closer to two or three years, because a far more efficient generation arrives roughly every year, and that stretching the schedule understates depreciation across the industry by something on the order of one hundred seventy-six billion dollars between 2026 and 2028, flattering reported profits by more than twenty percent. By one Goldman Sachs sensitivity analysis, as reported, shortening assumed chip life from five years to three would swing cumulative depreciation across the four largest builders from roughly three trillion dollars to roughly four trillion dollars over 2026 to 2031, a difference produced entirely by a single assumption that appears on no product page.

Here is why that matters for the word "excess." If the useful-life assumption is too generous, then a share of what is booked as growth capital expenditure is actually a permanent, recurring cost, and a chip that turns out to be surplus does not depreciate gently into a cloud business; it becomes a stranded asset that gets written down. The tell is not the announcement. It is the schedule. One large builder has already shortened the assumed life on part of its fleet while another extended it, in the same industry, on the same technology, which tells you the number is a choice, not a fact. The next builder to quietly shorten its useful-life assumption, or to take an impairment on hardware it can no longer profitably run, is telling you the excess was real before any income statement says so.

Then there is the financing, the direct continuation of the Issue #21 thread. A meaningful share of this cycle's demand runs in a circle: the dominant chip maker has committed to invest as much as one hundred billion dollars in the largest AI lab, that lab has committed hundreds of billions to cloud providers, and those providers buy the chip maker's hardware to fulfill the contracts. The chip maker has also agreed to buy the excess capacity of at least one cloud partner outright. Analysts tally more than eight hundred billion dollars of these arrangements. Every leg books revenue or backlog from the same underlying dollars, which makes demand look durable when a portion of it is a small cohort financing one another. If even part of the demand is circular, then "excess capacity" is not a temporary surplus waiting for real buyers; it is the first place the circularity shows up as unsold inventory of compute. The ledger cannot see this, because a committed contract and a genuine customer look the same until one of them does not pay.

What lifts this above one company's cloud strategy is that three things are arriving in the same window and pointing the same way. The first is the tape. Over June, the seven largest technology companies shed roughly two point three trillion dollars in market value, their worst month in more than a year, as investors began to reprice how long the payback on the build-out actually takes. At the same time, the traditional-economy names carried the Dow to record highs, briefly past fifty-three thousand, on a rotation out of the very AI leaders that had driven the market for three years. This is not a crash; the AI names bounced as this issue went to publication, and the chip suppliers remain far above where they started the year. It is something subtler: the market trying, in real time, to tell scarcity from excess, and visibly unable to hold one view for two sessions running. A trade that cannot decide whether it is looking at infinite demand or an overbuild is a trade that has stopped believing the demand is self-evidently infinite.

The second is the same-week evidence that the number underneath everything is softening. The June employment report showed the economy adding just fifty-seven thousand jobs against expectations near one hundred ten thousand, with the prior two months revised lower. A build-out sized for a roaring economy is being poured into a labor market that is quietly cooling: the demand-shadow problem from Issue #20, "The Look-Through," arriving from a different direction. There the worry was income that leaves the economy and never shows up in the inflation gauge; here it is end demand for AI products that may never arrive at the scale the capacity assumes.

The third is that the builders and the grid are now telling the same story from opposite ends. In Issue #21 the grid was pricing forecast data center load as firm, obligated to build for demand that might never energize. Now one of the tenants that load was forecast for says it has more capacity than it needs. Data centers were roughly four percent of national electricity use in 2023 and are projected to reach somewhere between seven and twelve percent by 2028, and the grid is being built against the high end as if it were certain. If the builders themselves have excess, the ratepayer financing the wires is being asked to treat as firm the very demand the tenant is quietly marking down. One build-out, seen from the auction and from the income statement, and both windows now show the same thing: capacity arriving faster than the demand it was sized for.

The honest forecast comes in three parts, and only the first is close to mechanical. Near term, excess reads as scarcity and the comfort story holds. The AI names recover, as they already have this week; the chip suppliers stay strong; Meta's cloud plan is received as shrewd, and any hyperscaler that follows it is praised for the same discipline. Second-quarter earnings, due through late July, most likely show revenue growing fast enough for the bulls to say the payback is on track, and the useful-life assumptions stay where they are, because no one shortens a depreciation schedule in a quarter they do not have to. Nothing here looks like a problem, which is the point: overbuild keeps being sold as strategy, because a genuine surplus and a strategic one are hard to tell apart in advance, and every incentive favors calling it strategy.

The medium term is where the uncertainty lives, and it turns on one question with no date on it: does the demand convert. Do enterprises and consumers actually rent the excess capacity at a price that earns back its cost before the hardware obsolesces, or does the surplus grow as more builders reach the end of their own demand at once. Watch, above all, the depreciation disclosures, because a builder shortening its assumed chip life, or taking a charge on stranded silicon, is the excess becoming visible in the audited numbers rather than the press release. Watch the capex-to-revenue gap and the group's free cash flow, which several tallies already show squeezed hard as spending outruns income. Watch whether the specialized cloud providers, whose largest customers are becoming their competitors, start losing contracts or cutting prices. Watch the pricing on any capacity Meta and its peers actually sell, since a high price says scarcity and a low one says they are clearing inventory. And watch the gap I flagged in Issue #21, between forecast large load and load energized, because that is the same excess measured at the meter.

The counter-move, when it comes, will not be a clean reversal, and this is the part to brace for. Expect the overbuild to be relabeled rather than written down. Capacity built for demand that did not materialize gets recast as a thriving cloud business; stranded silicon gets described as strategic optionality; the free cash flow that vanished into concrete gets framed as the cost of owning the future. The base case is not a bust. It is a surplus renamed as a product line, holding as long as the market prefers the renaming to the write-down, and as long as enough real buyers show up to make the new business look like foresight rather than salvage. The scoreboard reads as healthy right up until a depreciation schedule, or a canceled contract, says the capacity everyone paid for is looking for a buyer it was promised already existed.

Four areas of board exposure, mapped against the admitted-excess reading rather than the comfort story.

Investors holding hyperscaler equity and debt, and anyone whose portfolio leans on the AI premium. This is where the reading cuts first and hardest. If a portion of reported hyperscaler profit rests on depreciation schedules that assume the hardware lasts twice as long as the chip cycle, then the earnings that justify the valuations are, in part, an accounting choice that can be revised. A single builder shortening its useful-life assumption, or taking an impairment, does not just hit one company; it re-rates the assumption across the group, because they are all making the same bet with the same hardware at the same time. The question for the board or the investment committee is whether your models value these companies on reported earnings or on cash earnings after a realistic chip life, and whether you have stress-tested a world in which the "excess capacity" being sold is a write-down wearing a revenue label.

Specialized cloud providers, data center developers, and their lenders. The exposure here is immediate: your largest customers are becoming your direct competitors. A hyperscaler that sells its own spare compute competes for exactly the buyers the independent providers were built to serve, and it can subsidize price during a land-grab in a way a smaller balance sheet cannot survive. The take-or-pay contracts, the long-dated commitments, and the debt underwritten against them all assume the demand is scarce and the pricing holds; if the surplus is real and industry-wide, both assumptions weaken at once. The question for the board is whether your contracts and covenants survive a scenario in which your anchor tenant becomes your competitor and compute pricing falls, and who holds the hardware and the substation if a marquee customer walks.

Companies selling into the AI capital-expenditure chain: power, cooling, memory, electrical equipment, and real estate. Your order book is the capex line of a handful of buyers, and that line is the thing now under scrutiny. As long as the build runs, demand looks bottomless and pricing is yours to set. But a capex line sized for infinite demand is the first thing that gets cut when the builders decide they have enough, and "we have excess" is the leading edge of that decision. The question is whether your capacity additions and long-term supply commitments assume the current build rate is permanent, and whether you can survive the builders slowing from a sprint to a walk once the surplus is undeniable.

The boards of the hyperscalers themselves, and anyone underwriting or insuring the build. This is the accountable-judgment layer. The same unit of computing power is being booked, implicitly, as certain future revenue for planning and as a durable asset on the balance sheet, while the market has started pricing it as a contingent claim. Selling the excess is a reasonable response to a surplus; announcing it as a triumph while carrying the same capacity at a six-year life is having it both ways. The question for the board is whether your disclosures treat forecast AI demand as firm in the capital plan and contingent in the risk footnotes, or whether you have quietly booked the same server as certain in both places, the way the grid booked the same megawatt as firm and optional at once.

The excess may be trivial and the demand may simply be real. Compute has been genuinely scarce for three years, inference use keeps rising, and selling spare capacity may be nothing more than ordinary capacity management, the same instinct that produced the most profitable business in technology. On that reading, Meta has surplus the way any large operator has surplus at the margin, and I am reading a routine efficiency as a confession. If demand outruns even these forecasts, the builders who hedged by selling capacity will look prudent and the ones who did not will look short.

Selling excess may be exactly the disciplined move, and a sign of strength rather than weakness. Monetizing an idle asset is good stewardship, not an admission of failure; the companies that refuse to sell spare capacity out of pride are the ones wasting shareholder money. On this view the market was right to cheer, because a hyperscaler willing to rent its surplus is managing the build like a business instead of a religion, which is precisely what the skeptics have been asking for.

The depreciation debate may be less than it looks. Free cash flow is identical whether you depreciate a chip over three years or six, because the cash left when the hardware was bought, and sophisticated investors see through non-cash charges to the underlying economics. If the market already values these companies on cash rather than reported earnings, the useful-life argument is an accounting curiosity, not a coming cliff, and the write-down I am bracing for is already in the price.

And the bounce may be the truth. The AI names recovered this week, the chip suppliers remain far above where they began the year, and the second-quarter numbers may show revenue growth that retires the whole worry. Markets that stare into an abyss and then climb are often telling you the abyss was not there. If the demand converts, "excess" was just a word, the cloud business becomes real, and this issue reads as pattern-matching a healthy company's smart move to a bubble that never popped.

The DeepStrategy.ai signature method requires sorting risk, which is quantifiable, from uncertainty, which is not, at every major analysis. The numbers here are risk. The combined 2026 capital expenditure of the largest builders, on the order of seven hundred twenty-five billion dollars and up roughly seventy-seven percent year over year, the capex-to-revenue gap, the group's forward free cash flow, the depreciation schedules and any change to the assumed useful life, any impairment or asset-retirement charge, the roughly two point three trillion dollars of market value the largest names lost in June, and the spread between the chip suppliers and the hyperscalers can all be tracked, charted, and put on a board dashboard tomorrow. They tell you the size and direction of the move.

What will not resolve to a figure is what the move means. Whether "excess" is a temporary surplus or a structural overbuild is an uncertainty, because the same idle server is a bonus or a stranded asset depending on a demand curve that has not happened yet. Whether the useful-life assumption is prudent or a choice that defers a write-down is an uncertainty about judgment, not arithmetic, since the cash is already spent either way. Whether AI's end demand is genuine or partly circular is the uncertainty sitting underneath all the others, the same one I could not resolve in Issue #21, and it is the one that decides whether selling spare compute is foresight or salvage. The precise figure is the announcement; a builder spending as much as one hundred forty-five billion dollars this year is a fact. Whether that spending built a business or an inventory of compute nobody ordered is the part that will not reduce to a probability, and the two answers point in opposite directions from the same word.

The institution that consumes the analytical process as preparation for multiple futures has what the forecast cannot provide: adaptability. When the next hyperscaler announces it too will sell its spare compute, the headlines will say the AI economy is maturing and the builders are finally making the iron pay, and the rising stock will be cited as proof. The architecture says something quieter. You do not rent out what you need. The title of this issue has two readings, and the market has chosen the flattering one: the excess as a product, a surplus turned into a business. The other reading is the excess as the whole problem, the overage that means the demand ran out before the capacity did, booked as revenue in the exact place a write-down would otherwise sit. The same idle server is a new profit center or a stranded asset, and the difference between those two readings is what three years of infinite-demand conviction is now, quietly, being asked to prove.

The boards best positioned for this are not the ones reading the cloud announcements as vindication. They are the ones asking, of every unit of AI capacity under their plans and their portfolios, whether it is need or surplus, and who holds it if it turns out to be surplus, and pricing that answer before a depreciation schedule tells them which one it was.

Capacity built for a demand that had to be infinite. A builder now selling what it swore it could never have enough of. And an excess being cheered as a business in the one place a reckoning was due.

The Paranoidist publishes weekly, with flash issues when events warrant.

If this changed how you think about one thing, consider subscribing. If it didn't, tell me what I'm missing.

Paul Morin is the founder of DeepStrategy.ai, author of Uncertainty: When Risk Is Not Enough (a guide to decision-making when probabilities fail), and publisher of The Paranoidist, BoardroomRadar, and ScenarioWatch. He has spent more than three decades in entrepreneurship, finance, risk management, and insurance, which is why he worries about the things that keep other people awake at night.

Researched, written, and edited in collaboration with Claude by Anthropic.

All information reflects publicly available reporting verified as of Monday, July 6, 2026. Market levels, market-value changes, capital-expenditure figures, and the June employment data are revised on a schedule; re-verify the Meta cloud reporting, the hyperscaler capex totals, the Magnificent Seven June drawdown, the depreciation figures, and the current index levels against the latest releases immediately before posting.