MSFT, AMZN, GOOGL, ORCL, CRWV, NBIS, IREN, APLD, and DOCN are all buying the same inputs right now: Nvidia and custom silicon, power, land, cooling, memory and networking. What each of them earns on those inputs is anything but the same. Across the group, the returns range from the best incremental business in large-cap tech to something closer to a levered construction project. The market has repriced part of that divergence, with Oracle down about 63% from its peak, and left the rest of it alone.The yardstickThe economic unit here is a capacity increment: a datacenter hall filled with GPUs, placed into service, rented out. The right questions for that unit are how much revenue a dollar of installed capacity produces, how much of that revenue depreciation immediately claws back, and whether the company generates enough cash internally to keep building without borrowing.The table below uses each company’s most recent completed fiscal year (calendar 2025 for most, May 2026 for Oracle, June 2025 for Microsoft and IREN, May 2025 for Applied Digital). PP&E is net and includes leased right-of-use assets, so companies that lease their datacenters don’t get flattered against those that own them. Asset turns use average PP&E across the year, since several of these capital bases doubled or tripled during the period and a year-end snapshot would understate everyone who is ramping.Two caveats before anyone quotes this table at me, the hyperscaler turns are total-company figures, so Amazon’s retail assets and Google’s search business inflate them relative to a pure cloud comparison. AWS on its own, at $128.7B of FY25 revenue, runs much closer to Oracle’s capital intensity than Amazon’s consolidated number suggests. And the capex line is cash capex from the cash flow statement, which materially understates the levered names. CoreWeave’s net PP&E grew about $24B in 2025 against only $10.3B of cash capex, because finance leases and vendor financing did the rest of the work. If you want to know how fast someone is really building, read the balance sheet, since the cash flow statement has become the polite version of the story and even with those caveats, the tiers are unmistakable.Three different businesses wearing the same hardware The hyperscalers are running the only version of this that is self-evidently a good business today. Microsoft generated $98B of EBITDA above its capex in FY25 while spending $64.6B. Alphabet cleared $62B of surplus while spending $91.4B. Amazon’s surplus is thinner at $18.6B, but that reflects a company that chose to spend $131.8B, more than any other entity on the planet, and still covered it internally. These three fund the largest capital program in corporate history out of operating cash flow, at a scale where reported 2026 plans run to roughly $110-120B for Microsoft, $175-185B for Alphabet, and around $200B for Amazon. They can be wrong about AI demand and the consequence is a lower return on capital for a few years.The middle tier is Oracle and the big neoclouds, and this is where the same hardware becomes a different business entirely. Oracle spent 83% of its revenue on capex last fiscal year, produced negative $23.7B of free cash flow, and carries $167B of debt and lease obligations. CoreWeave spent twice its revenue. Nebius spent nearly eight times its revenue, which is what happens when a $530M revenue company decides to build $4B of datacenters in a single year. None of the three covers its build internally. All three are, in effect, borrowing today to install capacity that a customer has contractually promised to rent tomorrow. The entire investment case is a race between backlog conversion on one side and depreciation plus interest on the other.The contracted demand is genuinely enormous, which is why these companies can raise the money at all. Oracle exited FY26 with $638B of remaining performance obligations, roughly half tied to OpenAI on a five-year contract that starts in 2027. CoreWeave’s backlog hit $99.4B in Q1 2026, with Meta in for $21B and the counterparty mix improving to the point where less than 30% of it comes from non-investment-grade AI natives. These are real contracts with real (mostly) creditworthy names. What they are not, yet, is revenue, and depreciation does not wait for revenue recognition.Then there is the third tier, the power-and-shell developers. IREN and Applied Digital came to AI infrastructure from bitcoin mining, which means their real asset was never compute expertise. It was secured megawatts, grid interconnects, and cheap sites at a moment when energized capacity is the scarcest input in the industry. Applied Digital’s revenue actually fell in its fiscal 2025 (May year-end) to $144M while it spent $682M building, because the company is converting from crypto hosting to AI leases and the new capacity had not yet turned on. IREN at least covers its operating costs, with $132M of EBITDA on $321M of revenue. I would not call either of these a cloud business. They are speculative real estate development with a GPU thesis, and they will produce either spectacular or catastrophic outcomes depending almost entirely on the price at which their capacity meets the shortage.DigitalOcean fits none of the tiers, which is exactly what makes it interesting. On 2025 numbers it looks like a miniature hyperscaler, $1.22 of revenue per dollar of installed capacity, capex at 16% of revenue, EBITDA comfortably above capex, an actual profitable business renting compute to hundreds of thousands of small customers nobody else bothers to serve well. It is the only sub-$1B name in the group that self-funds.Inside the cloud P&L, line by lineThe consolidated numbers above blend cloud with retail, search, and legacy software, so strip all of that away and look only at the cloud businesses themselves. A cloud provider’s economic unit is a rack-scale capacity increment: accelerators and servers that depreciate over five to six years, sitting in a shell that depreciates over decades, drawing power that scales with utilization. Revenue from that unit is utilization times price per GPU-hour or vCPU-hour, or a committed contract that locks both in advance. The cost stack per dollar of revenue runs, in order of size: depreciation, then power and facility operations, then network and people. Interest lives below the operating line but belongs to the same unit whenever the capacity was built with borrowed money. Every cloud P&L is the same machine. What differs is utilization maturity, fleet vintage, pricing power, and the cost of the capital underneath.AWS is the existence proof as the mature, fully disclosed cloud P&L with $128.7B of FY25 revenue, $45.6B of operating income, a 35.4% margin. That margin absorbs depreciation on the industry’s largest installed base and still held at 35.1% in Q4 while revenue grew 24%. The street has FY26 at about $168B, and the backlog supports it: AWS’s contracted book grew from $189B in Q1 2025 to $244B by Q4. Whenever anyone asks what a cloud business earns at scale with utilization full, the answer is written down in Amazon’s segment note: mid-30s operating margins, after depreciation, at $129B of revenue. Every other row on the scoreboard is a claim about converging to that line. The backlog column is also worth a second look now that it is complete: Oracle carries 2.6 times AWS’s contracted book on one-seventh of AWS’s revenue, which is either the greatest land grab in enterprise software history or the industry’s most concentrated collection of promises. Google Cloud shows the convergence happening in real time, and it is the most instructive dataset in the public markets. The Google Cloud margin walkRun the incremental math on the Q4-to-Q4 move: revenue grew $5.7B and operating income grew $3.2B, so the marginal dollar of Google Cloud revenue carried roughly a 56% operating margin. That is what filling an existing capacity base looks like. Demand caught up to installed capital, utilization rose, and depreciation stayed relatively fixed while revenue climbed over it. The backlog says the filling continues, and consensus agrees: the street’s FY26 Google Cloud estimate moved from about $83B in February to nearly $97B by July, a 16% upward revision in five months. The caveat is that this margin walk was earned on the pre-2024 capex vintage. Alphabet’s capex ran 4.3 times its D&A in FY25 and the FY26 guide of $175-185B doubles it again, so a fresh depreciation wave should claw back a few hundred basis points even if revenue executes. Azure sits somewhere between the two, mostly undisclosed. Microsoft reports Azure revenue growth (39-40% in constant currency as of the March quarter) but not Azure operating income, and the street puts Azure at about $105B for the fiscal year that just ended in June. What Microsoft does say: demand still exceeds supply, its AI business crossed a $37B annual run rate, and Microsoft Cloud gross margin runs near 64% on a guided basis. One disclosure deserves more attention than it gets. Commercial bookings grew 7% excluding OpenAI but shrank 4% including it, and commercial RPO grew 26% excluding OpenAI but 99% including it. A single partner relationship is now large enough to bend Microsoft’s headline demand metrics in both directions, which is a version of the counterparty concentration problem usually reserved for the neoclouds, hiding inside the bluest chip in software.OCI is what the same P&L looks like when the capital arrives before the revenue. Oracle does not disclose OCI operating income, so this comes from my residual reconstruction (subtracting estimated license, SaaS, and hardware profitability from consolidated non-GAAP operating income): a 31-38% operating margin on $18.1B of IaaS revenue, which superficially reads AWS-like. The capital underneath tells a different story, and it deserves its own anatomy:What sits under $1.00 of OCI revenueThe margin looks fine but the return on the capital producing it is the open question, and it turns on fill rate: OCI today resembles Google Cloud circa early 2024, real margins on a huge, young capital base, with a $638B backlog that has to convert before depreciation (on a six-year server life Oracle extended from five) catches up. The pure-plays strip away all the blending, and the stacked P&L is stark. Normalize each to $100 of revenue and read top to bottom:Per $100 of revenue, latest fiscal yearCoreWeave has the best gross capture in the entire group, better than AWS at the EBITDA line, and depreciation consumes every cent of it before interest even enters the room. The current neocloud P&L is a pass-through from customer to depreciation schedule to lender, with equity holders owning whatever the utilization curve eventually leaves behind. Nebius is the same structure earlier in life, cash costs alone exceed revenue, and D&A nearly doubles the hole. DigitalOcean, running a seasoned CPU-heavy fleet with high software attach, is the only pure-play whose stack resolves to a real operating profit. Now the part the annual stacks hide. Direction of travel matters more than level, and the incremental economics point in opposite directions:Incremental margins: who improves with each new revenue dollar**(change year over year, or Q4 vs. Q4 where noted) *(change year over year, or Q4 vs. Q4 where noted)*CoreWeave’s EBITDA margin actually fell from 62% in 2024 to 47.5% in 2025, and its operating income went from +$324M to slightly negative, because deployment costs and the depreciation wave from $24B of new assets ran ahead of the revenue those assets will eventually produce. That is not necessarily failure, it is what mid-build looks like, and management’s own framing (the “escape velocity” argument, with roughly $919M of its $1B adjusted operating income target loaded into H2) concedes the point. But hold the two ends of the table in your head at once, businesses filling installed capacity print incremental margins in the mid-50s, whether they are a $70B-run-rate hyperscaler segment or a $1B developer cloud. Businesses still installing print negative incrementals regardless of how good the underlying demand is. The whole neocloud investment case is a bet on migrating from the bottom row to the top rows before the financing runs out of patience. Google Cloud made that migration, on internal cash flow, with no interest line ticking underneath and no covenant clock running. CoreWeave has to make it on borrowed money, which is why its deployment calendar matters more than its backlog.One last observation on DigitalOcean, because its row is about to get interesting. The 55% incremental margins and 15-point D&A share describe the old, CPU-heavy DOCN. Its AI ARR (up 221% to $170M) is pulling GPU intensity into the fleet, and its committed capacity ramp will push its D&A line toward the neocloud shape. Watching DOCN’s D&A-to-revenue ratio drift up over the next eight quarters is a small, clean experiment in what AI mix does to a healthy cloud P&L. If it can hold incremental margins near 50% while that happens, the software-attach argument wins. If incrementals compress toward CoreWeave territory, the whole “AI-native cloud” premium deserves to come out of the stock.Put the dissection together and the cloud business has exactly one hinge line, depreciation relative to utilization maturity. AWS defines the end state, mid-30s margins with depreciation fully absorbed. Google Cloud defines the transition, 9% to 30% in eight quarters at 56% incremental margins once demand catches capacity. OCI and CoreWeave are wagers, one funded by an investment-grade legacy business and one by structured debt, that the same curve exists for them. Nothing in their current numbers proves it yet, and nothing disproves it either. The proof will arrive on specific, datable lines: Google Cloud’s margin in late 2026 when the big capex vintage hits depreciation in force, Oracle’s FY27 depreciation against the $12B path, and CoreWeave’s second-half margin ramp.Depreciation is where the argument gets settledLook at the D&A column again, because it is the closest thing to ground truth in this whole debate. Depreciation is the delayed bill for capacity you already bought. At Alphabet it consumes 5% of revenue. At Microsoft, 12%. At Oracle, 14% and rising fast, on our estimates toward $12B next fiscal year from $7.6B of pure depreciation in FY26. At CoreWeave it already eats 48% of revenue, running at $1.15B a quarter and accelerating. At Nebius it consumes 88%, essentially all of the revenue the assets have generated so far.Everything about the middle tier’s future hinges on one accounting input, how long a GPU actually earns money. Oracle depreciates servers and networking over six years, a life it extended from five in fiscal 2025, and the big cloud players cluster in the same range. The bear case says the true economic life of an AI GPU is closer to three years, because each Nvidia generation resets the performance-per-watt math and makes old chips uneconomic to run. Rerun Oracle’s numbers on a three-year life and the OCI business goes from marginally profitable to outright negative before a dollar of interest expense. The same math applied to CoreWeave is worse.Which is why the most important datapoint of this earnings season came from an unexpected place. In Q1 2026, CoreWeave disclosed that average pricing on its A100s, H100s, H200s, and L40s all rose quarter over quarter. DigitalOcean, in the same quarter, independently said it is seeing price increases on H100s, H200s, and legacy gear. Two operators with completely different customer bases reported the same thing: inference demand is bidding up old chips rather than stranding them. If that persists, the six-year lives are defensible, the depreciation schedules are honest, and the middle tier’s margin math works. If it reverses in the next downcycle, half the companies in this piece are under-depreciating right now and reported margins across the levered tier are fiction. I know of no single indicator more worth tracking than secondhand and spot GPU pricing. It is the entire useful-life debate settling itself in public, quarter by quarter.The financing tells you who survives a pauseSince the levered tier cannot fund itself, the terms on which it borrows are as much a competitive moat as any technology. Here the details diverge in ways the market lumps together.CoreWeave got its DDTL 4.0 facility rated the equivalent of A-minus by three agencies and priced below 6%, by pledging contracted revenue from investment-grade customers. That is a repeatable template, and management thinks it becomes the industry norm for IG-backed contracts. If they are right, CoreWeave has manufactured a structurally lower cost of capital than its balance sheet would ever earn on its own. The offset is that interest expense ran $536M in Q1 alone, guided toward $690M in Q2, and on our math lands somewhere in the $2.5-2.8B range for the full year, a number that quietly sits outside the adjusted operating income the company guides on.Oracle borrows on its legacy software cash flows rather than on contract collateral, and that engine has limits. On a lease-inclusive basis its net debt already runs about 4.3 times its non-GAAP operating income, close to the mid-4s threshold that bond commentary flags for a downgrade to the BBB-/Baa3 boundary. A downgrade would raise the cost of the entire remaining build. Oracle is, in a real sense, spending its investment-grade rating to buy market share in a business where its fully-loaded return on capital, on our numbers, sits at 5-7% against a roughly 9% cost of capital. The in-service assets alone earn 8-12%, which is the honest version of the bull case: the productive capacity is fine, and the drag comes from the $40B of construction in progress that earns nothing yet. Whether that resolves bullishly depends on backlog converting before the depreciation wave crests.DigitalOcean chose equity. It raised $888M, paid off its term loan, and entered its expansion with the cleanest balance sheet of any levered builder, mostly by declining to be one. The subtlety is in the lease footnotes: $668M of datacenter obligations not yet commenced as of March, plus another $1.27B signed after quarter-end. DOCN is committing to roughly 135 MW of capacity to chase a 2027 revenue guide of 50%-plus growth.The cleanest conclusion is the least exciting one. The hyperscalers are the only names here where the AI buildout is an unambiguous unit-economics success already, funded internally, with turns above 1.4x and enormous cash surpluses even at record capex. Whatever your view on AI demand, they are the tier that gets to be wrong cheaply. The debate about them is a valuation debate, never a solvency one. The more actionable observations sit in the mispricings between tiers.DigitalOcean is being paid compounder prices at the precise moment it is adopting the capital model of the tier that trades at a fraction of its multiple. The stock re-rated from about 6x EV/revenue at the end of 2025 to roughly 14.5x FY26 guidance, on genuinely excellent numbers: 22% growth, AI ARR up 221% to $170M, 41% adjusted EBITDA margins. But those trailing economics describe the old asset-light DOCN. The forward plan, with its MW commitments and lease obligations, will push capex intensity up and asset turns down toward neocloud territory. Either the software layer (inference routing, managed agents, core-cloud attach, which is already more than 80% of its AI ARR) keeps revenue per megawatt well above commodity GPU rates, or the multiple is describing a company that is ceasing to exist. Revenue per MW is the number to demand from management.CoreWeave presents the mirror image. It carries the ugliest current-year optics in the group (0.19x turns, D&A at half of revenue, billions in interest) and trades near 3x its contracted exit-2027 ARR, with more than 75% of that already signed. The market prices it as the riskiest name in the complex while its revenue is arguably the most contractually secured, increasingly with investment-grade counterparties, at a financing cost below what its balance sheet deserves. The genuine risk is executional: management needs roughly $919M of its $1B adjusted operating income target in the back half, which requires the deployment calendar to run clean. I would rather own contracted revenue at 3x with execution risk than a business-model transition at 9.6x forward revenue, but reasonable people will land differently there.Oracle sits between tiers and is priced between tiers, and after the drawdown the setup is more balanced than the discourse suggests. At 5.6x FY27 EV/sales the market is not paying hyperscaler prices for OCI. The stock works from here if in-service returns hold while construction commissions, and it breaks if depreciation outruns the model. One number arbitrates: fiscal 2027 depreciation against the roughly $12B path. Materially above that and the shorter-life bears are right about the whole tier, Oracle included.Nebius and the developers are optionality, not businesses you underwrite on current economics. Spending seven or eight times your revenue on capex is a bet that demand meets you at the other end of construction. Sometimes it does. I would size accordingly. And across all nine, watch the same three signals: secondhand GPU pricing (the useful-life referee), the hyperscaler capex guides (the demand referee, since the big three are also the marginal buyers of everyone else’s capacity), and the lease footnotes, where this industry now hides its real commitments. The income statements in this group are lagging indicators, the balance sheets are where the future has already been ordered.