The AI buildout has produced a genuine standoff between serious investors, and both sides are arguing about one thing. On the bull side, people like Gavin Baker of have made the case that the hyperscalers’ return on capital is actually higher today than before they started spending on AI, that much of their installed base is contracted at prices well below current spot rates so operating cash flow accelerates as those deals roll off, and that the demand is real. On the bear side, Burry has argued the returns are an accounting mirage: the hyperscalers depreciate Nvidia gear over five or six years when its economic life is closer to two or three, which he estimates understates depreciation and overstates industry profit by roughly $176B between 2026 and 2028. He singled out Oracle and Meta as the most overstated, by about 27% and 21% by 2028, and called extending useful lives to flatter earnings one of the more common frauds of the modern era. Strip the rhetoric and the two are arguing about the same quantity, the return on the capital being poured into data centers, net of what it costs to replace it. Baker says that return is rising while Burry says it is a fiction created by under-depreciation. That is not a matter of opinion and the metric that settles it is ROIIC, what each new dollar of the buildout earns, and whether it clears the cost of that capital. Static margins and total-capex headlines cannot answer it and only the incremental return can. This piece attempts to compute it for the seven largest builders, on the capacity already running and on the record spending still to come.Here is the short version, and the rest is the arithmetic behind it. The capital already in the ground earns well above its cost everywhere except the neoclouds. On the money still going in, the hyperscalers clear their hurdle comfortably and Oracle clears; only Meta is a genuine knife-edge, at or just above the line on the standard accounting and below it on a stricter one; and the neoclouds fail either way. Notice that the two names sitting closest to the line, Meta and Oracle, are the exact two Burry named. I get there in four steps, what the earning capital returns today, how the return has moved over a decade, what consensus implies for the capex still to come, and how much profit each company must add to justify it. On Burry’s specific charge, a note up front: I use the depreciation the companies actually report, so if he is right that the economic life is shorter, every return here is too generous and the truth is worse than I show. But first the two mistakes, because getting them wrong is what produces the scary headline versions of this number.Mistake one: not lagging the capital. A dollar of data-center capex spent this year does not earn this year. It has to be built, energized, and filled, which takes roughly one to two years. So the incremental return is this year’s profit growth divided by last year’s capital, not this year’s. Divide by the same year and you get a number that looks like collapse precisely when capex is accelerating, because you are putting this year’s ballooning denominator against profit produced by a much smaller prior-year base. Lag it properly and the cliff disappears.Mistake two: crediting the buildout with the monopoly. Company-wide operating profit includes the ad auction, Office, Windows, and the retail machine, none of which the AI capex created. Divide whole-company profit growth by capex that is almost entirely AI and cloud, and you flatter the buildout by handing it the monopoly’s profit growth. This enterprise number is real and useful, but it is not what the AI dollar earns. We report the enterprise number as the spine because it is what the filings let you compute cleanly, and label it as the ceiling it is. The math then is incremental NOPAT (operating profit after a normalized 21% tax) in year t, divided by capital deployed in year t minus 1. We lag it by a year, and we size the capital two ways, on gross capex and on net invested capital, because for gear that depreciates as fast as GPUs the denominator is a real choice that moves the answer. What the earning capital is actually returningStart with the realized number, on the capital that has already come online. This is the latest fiscal year, incremental profit over prior-year capital, against each company’s own cost of capital.Two numbers, because there are two defensible ways to size the capital, and the gap between them is Burry’s argument in miniature. The cash-on-cash figure divides incremental profit by the year’s capex, the money actually out the door. The textbook figure divides by the change in net invested capital, capex minus depreciation, which is what a valuation desk means by return on incremental capital. For gear that depreciates as fast as GPUs the two pull apart hard, and the honest read sits between them. The cash measure treats every dollar of GPU replacement as brand-new investment, which is too harsh and the net measure treats a five-year-old GPU as nearly free capital, which is too kind, and it is too kind if Burry is right that the real life is shorter than five years. At the top and bottom of the table the choice does not matter: the profitable builders clear their cost of capital on the money already working, comfortably for the hyperscalers, and the neoclouds destroy value on either measure. The name the range pins is Oracle, which sits right at its hurdle on both, and only because its assets are longer-lived than a GPU fleet, so its two measures nearly coincide near 11%. One note on Microsoft, because its number carries an adjustment the others do not. It builds roughly $20B a year of cloud capacity through finance leases rather than cash purchases, so cash capex alone understates what it is deploying and would overstate the return. We use capex including finance leases, the figure Microsoft itself guides to: on cash capex alone Microsoft would screen at 33%; on the true capital deployed it is 25%.Now three honest cautions sit under this table. First, the numbers at the top are flattered by mistake two. Meta’s 29 to 44% and Microsoft’s 25 to 30% are mostly the ad auction and Office, not the AI unit, which is closer to the low-to-mid teens. Google is the same story with a smaller wedge: its 25 to 36% is the highest-quality of the three because its own AI unit is the strongest in the group, around 20% once you strip Search and ads out, but even Google’s headline is lifted by a still-growing ad business the capex did not build. The rule is that the faster a name’s legacy business is still growing, the more its headline overstates the AI economics, and Google, Meta, and Microsoft all carry that lift in descending order of how much. Second, Oracle is the truest read, not because it lacks a legacy business but because its large database and license-support franchise is mature and barely growing, so almost all of its incremental profit comes from its cloud unit rather than from a still-expanding monopoly. It lands right at its hurdle, which is why it, not the hyperscalers, is the one to watch. Lastly, Amazon sits at the other extreme, and its 11 to 19% is the lowest of the majors and deserves an explanation. It is not that Amazon is a worse operator, it is the most mid-build of the group. Its incremental profit in 2025 was the smallest of the three while its capex was among the largest, and the drag is AWS itself, not retail. Isolate AWS and its own incremental return is the lowest in the group on the cash measure and low double digits on the net one. AWS deliberately let its margin dip, from 37% to 35%, to reinvest into capacity; hold the margin and it clears higher. And on the capital AWS deployed two years ago, before the ramp, it still earns in the mid-teens. AWS is spending the hardest and harvesting the least right now, which is why Amazon’s forward number recovers as that capacity earns and Trainium scales into the numbers. Google Cloud, by contrast, saw its operating income double in 2025, and Microsoft monetizes through Office and Azure immediately, so their incremental profit shows up alongside the capex rather than a year or two behind it. Even a properly lagged number flatters during a capex ramp, because it credits this year’s profit to a smaller prior-year capex. The return on the actual record vintage, the $90B Google and $130B Amazon spent last year, cannot be measured until 2026 and 2027 revenue arrives. What the market is underwritingThe backward number cannot see the record vintage, so turn it forward and build it from both sides, the way you would underwrite any project is what it costs to put in the ground, and what it throws off. Start with the cost and this is consensus capex, the investment side, and the scale is the story:Now the other side, the revenue and profit the Street expects that capacity to generate. This is consensus operating profit, and the contrast with the capex table is the whole tension:Put the two together, lagging the capital, and you get the incremental return the market is implicitly underwriting on the money being spent now.The hyperscalers clear, and so does Oracle, which faces much higher financing risk. Google’s incremental return declines across the window, from 33% to 12%, because even $126B of added profit cannot keep pace with capex tripling to $338B, but it stays above its ~10% hurdle the whole way, and Microsoft the same. Amazon and Oracle are thinner on the cash measure, brushing their hurdles in the middle year, but on the net measure both clear comfortably. Either way they stay on the right side of the line.Meta is the marginal major, the one the denominator decides. On the conservative cash measure its forward return runs 6 to 9% against a ~10% hurdle, below it in all three years, on the textbook net measure it runs about 8 to 12%, at or just above the line, and its required-profit coverage crosses from 0.8x to just over 1.0x. So Meta is the single name whose buildout is not clearly value-creating either way, clearing on the standard measure and missing on the harsh one, the only major on the knife-edge. The reason is in the two tables above: its capex nearly triples while consensus adds only $40B of profit, the thinnest revenue response to the biggest relative spend of any major, so the incremental revenue only modestly outruns the depreciation the new assets throw off. This is precisely where bears and the arithmetic lands in the same place from the other direction.The neoclouds do not clear on either measure. The Street already models CoreWeave’s profit inflecting hard, from near zero to $7B in three years, and still its incremental return lands at 7 to 8% on the cash measure and only about 9% on the net one, against a 12% cost of capital, because the capex is simply too heavy and the funding too expensive. Nebius stays below its 13% hurdle throughout, both ways. The neoclouds’ problem is not that analysts are bearish. It is that even the bullish forecast does not clear the cost of the money, however you count the capital.Note the bear column, which haircuts consensus profit growth by a quarter, the commoditization case where AI compute prices soften as everyone’s capacity lands at once. On the conservative cash measure it bites hard, pulling Google, Amazon, and Oracle to or below their hurdles by year three; on the textbook net measure those three hold above the line and only Meta and the neoclouds fall below, which is the more telling result. Either way the most robust name is Microsoft and the most exposed major is Meta.What separates the namesThe spread is not random, and the names that clear have at least one working in their favor, and the strongest have two.The chip. Google runs its own TPUs for roughly two-thirds of its AI compute, by estimate, and Amazon’s Trainium is scaling. Custom silicon attacks the largest cost in a data center, the compute itself, cutting it by roughly half against merchant GPUs bought at Nvidia’s margin. A cheaper asset carries less depreciation, so each incremental dollar of that capacity earns more, and it also blunts bears depreciation charge, because a cheaper chip is a smaller number to depreciate no matter what life you assign it. That is why Google’s return, though it compresses as capex balloons, never crosses below its hurdle, and why Amazon’s, after dipping in the middle year as its heaviest capex lands, recovers by the end of the window as Trainium comes online and AWS’s investment-phase margin normalizes. Meta and Oracle buy merchant silicon at full price, so their incremental capacity earns the thin merchant return with nothing to defend it as capacity floods in and pricing softens. The balance sheet. Google and Microsoft fund the build off net cash with low betas, which gives them the lowest hurdles in the group, near 9 to 10%; Oracle’s leverage and the neoclouds’ expensive debt push theirs to 11, 12, and 13%. A cost edge on the asset side and a low hurdle on the liability side compound, which is why Google clears most comfortably even while spending the most, and why the neoclouds fail even when their profits inflect.Microsoft runs little custom silicon, but it sells the same GPUs up the stack as Copilot seats and Azure, lifting revenue per dollar of capacity the way a cheaper chip lifts it by cutting cost. Different route, same destination, and it is why Microsoft clears without a silicon edge. Line the three up and Meta is the one major with none of them working at the margin: merchant silicon, so no chip advantage; the highest hurdle of the ad-and-software group, from a high beta with no cheap debt to lower it; and capex running ahead of the demand that would monetize it, which is precisely why it is now trying to rent out the excess. No cheaper chip, a higher hurdle, and demand that has not caught up to the spend. That is the combination that leaves it on the knife-edge.How much profit the capex requiresFor the whole capex trajectory to earn its cost of capital, how much profit does each company have to add, and does consensus have it adding that much? Multiply the cumulative capex by the cost of capital and you get the incremental run-rate profit the buildout must produce just to break even on the capital. Set it against what consensus models.The margin-of-safety column is how far consensus profit growth could fall short and the capex would still earn its cost of capital. Read it, because it is the whole point. These figures use cumulative gross capex; on net invested capital the required profit is about a quarter lower, so every coverage ratio rises. Google and Microsoft could miss their consensus profit growth by roughly 40% and still clear on either measure. Oracle and Amazon have a fifth to a quarter of cushion on the gross measure and comfortably more on the net one. Meta is the borderline: its coverage is 0.8x on the gross measure, a quarter short, but crosses just above 1.0x on the net one, the only name that flips between the two. The neoclouds are far out on both: to justify their capex, CoreWeave has to beat consensus by more than half and Nebius by more than fourfold, and consensus is itself already modeling a steep profit ramp for each.Put in the plainest terms, this is what the market is asking each company to prove. Oracle has to grow its operating profit by about 142% over three years on the gross measure, and consensus expects 191%, so the bar sits well below the forecast. Meta has to grow profit by about 61% on the gross measure, and consensus expects 48%, so the bar sits just above; on the net measure the bar drops to roughly the forecast, and the two meet. Meta is the one company priced for its buildout to break even, no more and no less, whichever way you count.Meta caveat Meta is the marginal major, its forward return at 6 to 9% against roughly 10% on the conservative measure and just clearing on the textbook one, its coverage straddling 1.0x. It is the only name whose verdict the denominator decides, and it, alone among the majors, has a specific rebuttal that deserves an answer rather than a footnote. On July 1, 2026, Meta announced Meta Compute, a business to rent its AI infrastructure to outside customers in two tiers: access to models hosted on its servers, and raw GPU capacity of the kind the neoclouds sell. The stock rose about 9%. If Meta can monetize the buildout externally, the argument runs, its return is no longer hostage to internal ad demand, and the sub-hurdle number is too harsh. Consensus figures here do not include Meta Compute and the 9% pop was the market paying for optionality. More telling Meta Compute exists to sell capacity Meta cannot use internally, and a company spending toward $200B a year that needs an outside buyer for the surplus is confirming the thin-return diagnosis, not refuting it: the internal roadmap does not absorb the capex. Still, selling excess capacity beats leaving it idle, so it should lift the return. The question is by how much, and the arithmetic is not kind. Renting a slice of the incremental capacity adds profit on capital that was earning little, but it earns cloud economics, not ad economics, and Meta is explicitly pitching a raw-GPU tier that sells at commodity neocloud rates. Take the forward incremental return, about 8% on the conservative measure, and add what external rental contributes, at two rental shares and two price points, a hosted-model unit return of about 11% and a raw-GPU-rental return of about 5%: So Meta Compute narrows the gap and, in its most favorable and least likely corner, just closes it, but only if Meta rents a large share at real cloud margins rather than commodity rates, and only if it executes a business it has never run against three incumbents that have run it for fifteen years. The honest verdict does not flip, it firms: Meta is the major whose buildout only pencils out on the friendlier assumptions, with the balance resting on an unproven cloud business monetizing capacity it admits it overbuilt. That is the optionality the stock repriced. It is real, and it is a mitigant, not a fix.Reconciling the two lensesStep back to the whole group, because there is one reconciliation left to make explicit. Why do the enterprise numbers here (Meta 29%, Microsoft 25%) sit above the pure-AI unit returns from our bottom-up model (Meta ~11%, Microsoft ~15%)? The gap is the legacy subsidy, and it is largest exactly where the monopoly is still growing fastest.Microsoft and Meta carry a giant asset-light franchise still growing in the numerator, so their headline says almost nothing about the AI unit underneath. Google’s gap is smaller because its cloud is a larger share of the incremental story and its AI unit is genuinely the best in the group. Oracle’s gap is near zero, not because it lacks a legacy business but because its database and support franchise is mature and barely growing, so its enterprise return is essentially its cloud unit’s. Amazon is the one name where the enterprise number is below the unit, because AWS is reinvesting its margin and carries the heaviest construction lag. The more useful version of this is forward, because it asks whether the consensus we leaned on is grounded in what the capacity can actually earn. Run the bottom-up unit model on each company’s forward capex, marrying a revenue yield on the new capacity to the cost of building it, and set it beside the consensus-implied forward return:For the vertically-integrated hyperscalers the two methods agree, consensus’s forward is essentially what a dollar of their own capacity earns, so their number is a cross-checked economic return, not a hope. For Meta, consensus sits below even the unit economics, because its capex is not all revenue-generating capacity, a large share is frontier-model research and the surplus it now wants to rent out. And for Oracle and the neoclouds, consensus runs five to seven points above what merchant data centers earn: the market is pricing Oracle’s cloud pricing power and a neocloud ramp that the raw unit economics, on someone else’s silicon bought at full price, do not support. That last row means the thin consensus returns we showed for those names are themselves generous. On the physics of the capacity, Oracle and the neoclouds are more exposed, not less, which is the second place our numbers meet Burry’s.The scoreboard for the next 6 to 12 monthsThe whole case rests on two things being true, that consensus profit growth is roughly right and that the record vintage earns like the prior ones, and both are testable over the next four-six prints. Recompute the realized lagged ROIIC print by print through the Q3 2026 reports (late October and November 2026) and the Q4 reports (late January and February 2027). If it inflects back up as the 2024 and 2025 capex energizes, Baker is winning and the current down-leg was the lag. If it keeps grinding lower while capex stays at these levels, the commoditization bear is winning and the forward table above is too generous. Watch the depreciation line and the useful-life disclosure in the next 10-Ks. If any hyperscaler extends useful lives again, it confirms the bears charge and mechanically lifts reported profit without lifting cash, if depreciation instead accelerates toward capex, the reported returns were real. Oracle’s and Meta’s depreciation-to-revenue ratios are the two to track, because they are the two he named and the two we flag. Meta, the cleanest single-name tell. Meta Compute’s first disclosed pricing, named customers, and utilization will settle which business it is. Hosted-model margins carry Meta over its hurdle; commodity GPU-rental rates do not. Alongside it, watch the Q3 and Q4 capex guide against the operating-profit trajectory: if capex intensity keeps climbing while incremental profit does not, the coverage slips further below 1.0x and the knife-edge tips the wrong way. Amazon, the recovery trade. The tell is the AWS operating margin. It dipped to 35% in 2025; a move back toward or above 37% over the next two prints is the signal that the invest-phase trough is passing and Amazon’s forward recovery is real. Watch too for any disclosed Trainium price-performance number, the first hard evidence of a silicon edge. Google has the widest cushion but is not immune. The two tells are the Google Cloud operating margin, which should keep inflecting up, and the TPU external-sales ramp scheduled to land in 2027, the revealed-confidence signal. As long as its serving-cost curve keeps falling and cloud margin keeps rising, the top of the table holds. Watch Azure’s growth rate and the capex-including-finance-leases guide; the finance-lease line is where its true spending hides, so a jump there without a matching Azure acceleration would be the first crack in the most robust name. Oracle, the highest-hurdle major and the RPO test. Everything hinges on converting the $638B remaining performance obligation into revenue at the margin consensus assumes. Watch OCI revenue growth, the cloud gross margin (compressing from 63% toward 59% as it builds), and free cash flow, which was negative $23.7B and deepening. A further leg down in free cash flow, or a credit-rating action on the $130B of debt and $260B lease book, turns a valuation question into a financing one.The neoclouds face the solvency clock. CoreWeave and Nebius both report within weeks. Watch the gap between adjusted EBITDA and free cash flow (the tell for how much of the “profit” is being consumed by capex and interest), the debt maturity walls (CoreWeave’s DDTL facilities from 2028), and customer concentration (CoreWeave was about two-thirds Microsoft). Because their forward returns are below the hurdle even on consensus, any profit-ramp slip converts the question from valuation to survival. The macro tell, above all of it is the GPU rental spot rates and utilization are the fastest read on whether AI compute is commoditizing. Softening rates would validate the bear column across the board and pull the marginal names, Meta and the neoclouds, decisively under. A hyperscaler capex pause or cut, which none has signaled, would be the loudest signal of all that the buildout has outrun the demand.Applying it both waysTwo things cut in the bulls’ favor, and we should say so plainly, because they are the bulls points and they are right. The lag correction is the first and the earning capital is returning above its cost across every profitable builder, the collapse narrative is an artifact of dividing this year’s profit by this year’s ballooning capex. The denominator is the second: on the textbook net-invested-capital measure, not the harsher cash-on-cash one, the returns are about a third higher, which lifts Amazon and Oracle from clearing-by-a-hair to clearing comfortably and lifts Meta from below its hurdle to just above it. But Burry is not entirely wrong either, and the forward math shows exactly where. Meta is the marginal major, sitting on its hurdle and clearing only on the more generous accounting; Oracle’s consensus runs well above what merchant data centers actually earn; and the neoclouds do not clear on any measure. Those are his two named companies and the segment most dependent on the accounting, and our arithmetic reaches them independently. And the deepest caveat is the one both men would accept: no lag, however careful, can yet measure the return on the largest, newest vintage of capex in corporate history. The honest position is that the capital already online is earning its keep, that Google and Microsoft carry the widest cushion on the money still going in, that Meta’s buildout is a coin toss on the accounting and Oracle’s is priced ahead of the physics, and that the neoclouds are the only names underwater however you count.