The easiest mistake in the AI capex debate is assuming the outcome has to be clean. and that it’s either the spending is a bubble or the spending will lead to massive ROI. That framing misses the stranger possibility sitting in the middle, AI can work beautifully for customers and still produce mediocre returns for the investors funding the wrong layer of the stack. The profit will not necessarily go to whoever spends the most but to whoever keeps control of scarcity, distribution, or workflow after the cost of intelligence falls.The spending numbers are now too large to dismiss AI as a science project. The five large AI spenders are expected to reach roughly $730 billion of capital expenditures in 2026. Alphabet, Amazon, Meta, and Microsoft alone could invest about $650 billion in AI infrastructure in 2026, up from roughly $410 billion in 2025. Amazon has said AWS AI services are running above a $15 billion annualized revenue pace, while AWS overall is around a $142 billion revenue run-rate, and Amazon has projected roughly $200 billion of 2026 capex. Meta raised its 2026 capex forecast to $125-$145 billion and sold $25 billion of bonds. The software side looks just as real. OpenAI reportedly generated $5.7 billion of revenue in Q1 2026 while burning $3.7 billion, and Anthropic has been reported around the $50 billion annualized revenue run-rate area, helped by Claude Code and enterprise adoption. Run-rate is not audited annual revenue, especially for companies growing this quickly, but the scale of the demand is no longer theoretical.The hard part is that the word AI capex hides several businesses that barely resemble each other. A powered shell leased on long-term take-or-pay terms is one kind of investment. A levered GPU cloud buying fast-depreciating NVIDIA systems is another. A hyperscaler can justify the same cluster through cloud pull-through, software attach, customer retention, internal workloads, and custom silicon. A model lab converts compute into intelligence, but only if training spend stops outrunning revenue. An application company may barely touch the physical infrastructure and still capture a huge share of the customer’s willingness to pay. So the question is not whether AI demand exists, because it clearly does, but who gets paid after the stack is done eating.CRWV is the cleanest public stress test for that question. The company finished 2025 with more than 850 MW of active power, 3.1 GW of contracted capacity, and $66.8 billion of backlog. In Q1 2026, revenue reached $2.08 billion, backlog rose to $99.4 billion, contracted power moved above 3.5 GW, and the company guided to $31-$35 billion of capex for 2026. Those numbers are extraordinary but they also sit next to a GAAP net loss of about $740 million in Q1 and operating expenses that exceeded revenue. CRWV proves the central tension in one company, the demand is real but the economics below EBITDA can be heavy.The 1 GW sanity checkTo keep the discussion from drifting into vague AI infrastructure language it helps to reduce the problem to a single unit: 1 GW of AI capacity. Public estimates for a 1 GW AI data center have moved into the $40-$50 billion range, depending on the hardware generation, rack density, memory, networking, cooling, power, land, and construction assumptions. Bernstein has reportedly estimated that a 1 GW Rubin-based data center could cost around $47 billion, which is broadly consistent with how expensive rack-scale NVIDIA systems have become before the owner even pays for the building and power. A 1 GW campus is therefore not just a large data center, it is a very large capital project with a short-lived compute core sitting inside a long-lived physical plant. The shell may last a long time but the GPUs may not.For a merchant GPU cloud, I would frame the annual burden in three separate buckets. First, there is the D&A or replacement burden, which is roughly $5.5-$7.0 billion per GW depending on the capital base and how harshly you treat GPU obsolescence. Second, there is data-center opex, roughly $0.9-$1.2 billion per GW, including power, cooling, staff, maintenance, and overhead. Third, there is explicit cash interest, which can easily run around $1.0-$1.5 billion depending on leverage and funding cost. The categories need to stay separate. Depreciation allocates the cost of the asset and Interest is the cost of financing it. Mixing those two is where things can gets slippery.Power deserves more than a footnote inside opex. At 1 GW, every $10/MWh change in electricity cost moves annual power expense by roughly $88 million before utilization adjustments. Availability matters even more than the price. If power or permitting delays a site, revenue gets pushed out while capital remains committed. Chevron’s Project Kilby, a dedicated 2.67 GW power project for Microsoft’s West Texas data center campus, is a good signpost. The scarce asset is no longer just GPUs. It is GPUs plus power, permits, cooling, and delivery certainty.The neocloud problemNow apply the 1 GW lens to a merchant GPU cloud. Using $7 billion of revenue per GW as a base case which is a reasonable middle ground from public data. CoreWeave’s 2025 revenue near $5.1 billion on more than 850 MW of active power points to roughly $6 billion per GW on an end-of-year active-power base, while Q1 2026 annualized revenue of roughly $8.3 billion suggests the run-rate moved higher as more capacity came online. A $7 billion revenue-per-GW assumption sits between those two readings.The loss range is wider than a single-point model suggests, but the conclusion is hard to escape. At a real-world EBITDA margin around 60%, which is closer to CoreWeave’s adjusted EBITDA experience, $7 billion of revenue produces about $4.2 billion of EBITDA. Subtract a $5.5-$7.0 billion D&A/replacement burden and $1.0-$1.5 billion of interest, and the business is losing roughly $2.3-$4.3 billion per GW before tax. Under a much cleaner engineering case, where EBITDA margins are closer to 83% and the replacement burden is at the low end, the loss narrows to around $1 billion per GW. It still does not look like a good equity return. This is why EBITDA can be so misleading. The machine can look full, revenue can grow fast, backlog can be huge, but adjusted EBITDA still fails to capture the two lines that matter most, depreciation/replacement and financing. A GPU cloud is racing against obsolescence with borrowed money.At $10 billion of revenue per GW, the model begins to work in the better cases. At $5 billion, it is ugly. That spread tells you most of what you need to know. The neocloud business needs realized monetization to rise toward $10 billion per GW at the same time the industry is trying to push unit prices lower through more capacity, custom silicon, open models, and routing to the cheapest acceptable provider. That tension is the heart of the fragility.This does not make every neocloud a bad business. Long-duration contracts, customer prepayments, high utilization, cheaper financing, and strong residual value can improve the outcome. Some customers also need dedicated clusters faster than hyperscalers can deliver them, which gives neoclouds a real window. The question is how much that window is worth once the market normalizes and older GPUs lose their premium.The infrastructure label hides an important split. A powered shell and a GPU fleet are not the same asset. A powered shell is land, interconnection, cooling, building infrastructure, and access to electricity. If it is leased on take-or-pay terms to a strong counterparty, it can look like an infrastructure yield asset after development risk is gone. Hut 8’s Texas deal is a useful example, a 15-year, $9.8 billion take-or-pay, triple-net lease for 352 MW, with operations expected to begin in 2027. That pencils to about $1.85 billion of revenue per GW-year before renewal options.The compute layer behaves differently. A GPU can physically operate for years, but its economic life can shorten the moment a new generation resets rental rates. Hyperscalers can push older GPUs into internal workloads, batch jobs, or less demanding inference. A merchant cloud has fewer places to hide. It needs outside customers to keep paying enough for yesterday’s hardware. That is why the data-center owner and the GPU renter deserve different multiples, different discount rates, and different credit analysis. The market often blends them together because both sit inside the same AI campus. The cash flows are not remotely the same.Why hyperscalers can underwrite more aggressivelyThe hyperscaler does not need the GPU hour to carry the entire return. That is the biggest structural difference between Microsoft, Amazon, Google, Meta, and a merchant GPU cloud. The same cluster can support cloud consumption, storage, security, data services, model hosting, developer tools, software attach, advertising, recommendations, and retention. Azure can sell OpenAI access, GitHub Copilot, Microsoft 365 attach, and the surrounding cloud platform. AWS can use Bedrock, Trainium, and its enterprise distribution. Google has Gemini, Vertex, TPUs, Workspace, Search, and Google Cloud. Meta can use the capacity internally for ads and recommendations, then potentially rent excess capacity if it has more compute than internal workloads need.That attach revenue should not become a blank check. The disciplined approach is to credit only cluster-specific incremental economics, the cloud, software, retention, or advertising profit that plausibly exists because the AI capacity exists. A 20%-40% credit on cluster-specific pull-through can make sense. Applying that same percentage to the hyperscaler’s entire cloud business would be nonsense. Still, the advantage is real. A neocloud sells compute and hopes the spread is enough. A hyperscaler can use compute to deepen the customer relationship, sell more software, lower churn, and improve its own products. That is why the same 1 GW campus can be strategically rational for Microsoft and structurally dangerous for a levered GPU renter.Silicon gets paid before the risk is resolvedNVIDIA sits in a different position from the recurring software layers. The customer dollar moves through the app, the model, the cloud, and the data center over time. NVIDIA gets paid from the capex dollar up front: cash, debt, leases, equity, and long-term commitments raised against expected future AI demand. That timing explains why NVIDIA has been the best business in the stack. It gets paid before the downstream monetization question has fully resolved. It also explains why circular financing deserves attention. NVIDIA’s $2 billion CoreWeave investment does not make CoreWeave’s demand fake, but it does complicate the quality of that demand when the supplier helps finance the buyer.The custom-silicon risk is also more nuanced than NVIDIA loses, hyperscalers win. Some margin shifts to ASIC partners like Broadcom and Marvell. Some of the silicon profit pool already sits with HBM suppliers like SK Hynix, Micron, and Samsung. The better description is fragmentation. NVIDIA still has CUDA, networking, systems, scale, and developer lock-in, but the largest buyers will keep trying to move predictable inference workloads to cheaper silicon. Training and frontier research are harder to displace. Repetitive inference is where optimization pressure shows up first.The labs have revenue. They still need operating leverage.The model labs are the part of the stack where the product-market fit is clearest and the cash-flow profile is still hardest to underwrite. Once a model exists, the inference business can be attractive. Serving costs fall, utilization rises, and the lab can sell tokens into a growing base of applications and enterprise workflows. The frontier business is different because training the next model behaves like maintenance capex for the franchise. A lab that stops training may keep selling the current model for a while, but pricing power decays as open models, hyperscaler models, and cheaper alternatives close the gap.OpenAI’s reported Q1 numbers show the tension. Revenue was reportedly $5.7 billion, while burn was $3.7 billion. That is not evidence of weak demand. It is evidence that staying at the frontier is expensive. Anthropic’s reported move toward the $50 billion annualized run-rate area shows how quickly revenue can scale when the product hits, especially with Claude Code. The investor question is whether training intensity falls as a percentage of revenue before competition pushes pricing down. A lab can become a very good business if revenue scales faster than the training treadmill. If every frontier model requires a larger check just to defend the same relative position, the lab becomes a brilliant product attached to a difficult cash-flow machine.Apps can be the best business but only with a moatThe application layer has the cleanest upside when it owns a real workflow. If software saves a lawyer hours, helps an engineer ship code, processes insurance claims, resolves support tickets, or produces better financial research, the vendor can charge for the outcome rather than the token. Falling token prices then become a margin tailwind. The same falling token prices also lower the barrier to entry. A thin wrapper gets copied faster as models become cheaper and better. The winners need workflow depth, proprietary data, distribution, or system-of-record status. Without one of those, cheap intelligence destroys the moat as quickly as it improves gross margin. This is why the app layer will probably produce both spectacular winners and a lot of vapor. The best apps will price on customer value and the weak ones will discover they were just reselling model access with a nicer interface.The two ways this can go wrongThere are two bear cases worth separating. The first is a demand-side bear. Customers try the tools, like the demos, but do not get enough durable ROI. Seats churn, agentic usage hits a ceiling, token growth slows, and the productivity uplift is real but smaller than the stack needs. This version hits everyone because the customer dollar disappoints.The second is a supply-and-price bear. Usage keeps growing, but capacity arrives too quickly, open models get good enough, ASICs take more inference, and routing pushes workloads to the cheapest acceptable provider. Realized revenue per unit falls faster than volume grows. This version is bad for GPU renters and model APIs, mixed for hyperscalers, and often good for strong apps and end-users. Elasticity decides which version of cheap intelligence the market gets. If price falls 50% and volume rises 50%, revenue falls 25%. A 50% price cut needs a full doubling of volume just to keep revenue flat. Growth requires more than that. Users can win long before infrastructure owners do.Where the economics likely settleThe cleanest map looks like this. Contracted power and colocation can work after development risk is gone. Hyperscalers can justify aggressive spending because the cluster supports a broader platform. NVIDIA and parts of the silicon supply chain have captured the best current economics because they get paid early from the capex dollar, though ASICs, memory economics, buyer concentration, and circular financing complicate durability. Frontier labs have enormous revenue momentum, but they need training intensity to normalize. Workflow apps can be excellent if they own the customer problem.The fragile layer is merchant GPU. It needs high utilization, long useful life, disciplined financing, and realized revenue per GW moving higher while the rest of the industry works to push unit prices lower. At roughly $7 billion per GW, the business looks loss-making across a reasonable range of margin and replacement assumptions. At $10 billion, it starts to work only in the better cases.This makes AI capex look less like one bubble and more like a transfer mechanism. Capital markets fund the physical buildout. Silicon gets paid early, Hyperscalers turn the spend into platform control. Labs turn compute into intelligence and apps turn intelligence into workflows. Customers get cheaper cognition. The margin remains with the layers that keep scarcity, distribution, or workflow control after the cost curve falls.That is why the fiber analogy is only partly useful. Fiber changed the world and bankrupted plenty of investors. Railroads and shale did the same. Cloud is the other side of the story: the companies that owned both infrastructure and distribution became some of the best businesses in the market. AI will probably contain both outcomes. The neocloud piece may rhyme with fiber, the hyperscaler piece may rhyme with cloud, the application piece may rhyme with software. NVIDIA is harder to reduce to a Cisco-style equipment analogy because CUDA, networking, systems, and developer lock-in are real, but even great platforms face harder questions once their largest customers start optimizing around the bottleneck.The buildout can be real, useful, and transformative while still producing mediocre returns in the wrong layer. That is the uncomfortable part and the reason the debate keeps talking past itself. AI does not have to fail for some AI infrastructure equity to disappoint. Customers can win, Apps can win, Hyperscalers can win. The question is who gets left owning the depreciating asset when intelligence keeps getting cheaper.