+May 13, 20260Hyperscalers are reporting backlogs for cloud service, how much is this backlog and who has booked this backlog?The hyperscale cloud backlog is driven by two very distinct groups: the AI labs themselves (OpenAI and Anthropic), and a wave of traditional enterprises modernizing their core operations.Here is exactly how these two groups are using this capacity:1. The AI Labs: The Dominant ForceOpenAI and Anthropic are securing massive contracts to outsource their infrastructure, accounting for about half of the $2 trillion total backlog .· Anthropic:· Google Cloud: A 5-year deal worth $200 billion for massive "5GW computing power" (over 40% of Google's backlog) .· AWS: A $100 billion, 10-year contract .· Purpose: Training next-gen models and running inference at massive scale .· OpenAI:· Oracle: A massive $300 billion contract .· AWS: Over $100 billion in commitments (80%+ of AWS's recent backlog growth) .· Microsoft: Responsible for a huge portion of Microsoft’s $627 billion backlog .2. Traditional Enterprises: The Production ShiftBeyond the labs, thousands of conventional companies are moving beyond "experiments" into full deployment.· The "Big-Ticket" Buyers:· Financial Services: A top-tier global bank just signed Snowflake's largest deal ever ($400M+) to run production AI .· Retail & Commerce: Walmart and Target are partnering with Google to deploy "agentic commerce" (AI that actually completes purchases) .· Software Giants: ServiceNow signed a $1.2 billion deal with Google Cloud to embed AI agents .· The ERP Shift (SAP):· Backlog: €77 billion (up 30%) .· Purpose: AI is now in two-thirds of SAP's new orders as companies modernize finance and supply chains .🎯 The Specific Use CasesThe "purpose" of this spending reveals three distinct phases of AI adoption:1. Training & Infrastructure (The Labs): This is the "raw compute" spending. OpenAI is expected to spend $45B on servers just this year to build foundation models .2. Agentic AI (The Enterprise): Instead of just chatbots, companies are buying infrastructure for "agents" that perform tasks autonomously (e.g., writing code, solving customer tickets) .3. AI-Ready Data (The Analytics): Snowflake reported that 50% of new bookings involve AI features, proving firms are building the governed data systems needed to support LLMs .
+May 12, 20260aren't this datacenter commodity? do hyperscalers have pricing power?If you look at a data center purely as a concrete box with electricity, cooling racks, and servers, then yes, it is absolutely a commodity.But hyperscalers do not sell concrete boxes. They sell cloud ecosystems, and because of how those ecosystems are engineered, hyperscalers possess some of the strongest pricing power of any industry on earth.When analyzing the fundamental unit economics of the cloud, the distinction between a "data center" and a "hyperscaler platform" is where the economic moat is built. Here is why they are not a commodity and how they maintain immense pricing leverage.1. The Moat of Switching Costs (Enterprise Lock-In)If cloud computing were a true commodity, an enterprise could seamlessly switch from Amazon Web Services (AWS) to Microsoft Azure overnight to save 5% on their compute bill. In reality, doing this is an operational nightmare.Once a company builds its infrastructure on a specific hyperscaler, it adopts that platform's proprietary databases, security protocols, identity management, and serverless architectures. Unraveling that code and migrating massive data lakes to a competitor costs more in engineering hours, downtime, and risk than the company would ever save on the server bill. Because enterprise clients are effectively trapped by this operational friction, hyperscalers have the leverage to steadily increase prices without suffering massive churn.2. The Egress Fee TrapHyperscalers historically engineered their pricing models to act like a roach motel for data: it is free to bring your data in, but incredibly expensive to take it out.While regulatory pressure (particularly in Europe) has recently forced hyperscalers to start waiving some "egress fees" for customers who are completely shutting down their accounts, the underlying friction remains. It is structurally cost-prohibitive for companies to run multi-cloud environments where data constantly travels back and forth between AWS and Google Cloud. This forces companies to consolidate their spending within a single provider, solidifying the hyperscaler's pricing power.3. Moving Up the Stack (PaaS over IaaS)If hyperscalers only rented out bare-metal servers (Infrastructure as a Service, or IaaS), margin compression would eventually turn them into a commodity.To prevent this, they aggressively push customers up the software stack into Platform as a Service (PaaS). They don't just want to rent you a server; they want to rent you the proprietary machine learning tools, the managed database software, and the AI agent frameworks that run on the server. These higher-level software services carry massive profit margins and are highly differentiated, insulating the hyperscaler from race-to-the-bottom commodity pricing.4. AI Differentiation and Custom SiliconThe AI boom has actually fractured the idea of compute as a commodity even further.Proprietary Access: If a company wants API access to the best version of OpenAI's models, they have to use Microsoft Azure. If they want deeply integrated Gemini agents, they use Google Cloud.Custom Chips: Hyperscalers are increasingly forcing adoption of their own custom-designed silicon (like Google's TPUs or AWS's Inferentia/Trainium chips). Because these chips are proprietary and often cheaper/faster for specific workloads than standard Nvidia GPUs, workloads compiled for them cannot be easily ported anywhere else.The Real Commodity: The Data Center REITsIt is important to separate the hyperscalers from the companies that actually own the physical buildings.Companies like Equinix or Digital Realty (Data Center REITs) lease the physical floor space, power, and cooling to the hyperscalers. Their business is much closer to a commodity. They are highly dependent on the cost of capital, local real estate markets, and power availability.The hyperscalers, however, take that commodity physical space, layer an incredibly sticky, proprietary software ecosystem on top of it, and extract high-margin, recurring revenue from Fortune 500 companies that cannot easily leave. That is the definition of pricing power.
+May 12, 20260how are the hyperscalers funding capex?Historically, Big Tech companies possessed such massive, money-printing core businesses (like search, software licenses, and e-commerce) that they could easily fund all their infrastructure projects using the cash they generated every quarter.In 2026, that era is officially over.The projected $700 billion-plus capital expenditure bill for the "Big Five" (Amazon, Microsoft, Alphabet, Meta, and Oracle) has finally eclipsed their ability to pay for it entirely out of pocket. To fund this unprecedented AI buildout, hyperscalers are being forced to fundamentally change how they manage their balance sheets.Here is exactly where the money is coming from today.1. Squeezing Operating Cash FlowsThe first line of funding is still their core business revenue, but the AI bill is eating almost all of it.Historically, hyperscalers had massive margins of Free Cash Flow (FCF)—the cash left over after paying operating expenses and CapEx.Today, CapEx is growing much faster than revenue. According to recent Wall Street models, hyperscaler CapEx is now consuming over 90% of their operating cash flows. The result is a historic cash burn. Projections for 2026 indicate that Amazon will likely spend more cash than it brings in for the year. Meta and Microsoft are also projected to see negative cash flow in at least one quarter this year. As a group, their free cash flow is dropping to levels not seen in over a decade. 2. Tapping Global Debt Markets (The Big Shift)Because their internal cash flows are no longer enough to cover the massive checks they are writing for GPUs and power infrastructure, hyperscalers are aggressively pivoting to external financing. They are loading up on corporate debt. The Bond Binge: In 2025 alone, the Big Five raised over $100 billion in the corporate bond markets. Analysts project that tech companies could issue up to $1.5 trillion in debt over the coming years to sustain this buildout. Going Global for Cheaper Capital: Because US interest rates make domestic borrowing expensive, hyperscalers are getting creative. In early 2026, Alphabet began setting up its first-ever Yen-denominated bond offerings, while Amazon initiated massive multi-billion-dollar bond sales in Swiss Francs, Euros, and Canadian Dollars. They are scouring the globe for the cheapest possible debt to fund their data centers.3. Capital Allocation Trade-OffsWhen you have to find $200 billion in a single year (which is Amazon's stated CapEx target for 2026), you have to cut back elsewhere. Hyperscalers are funding the AI boom by pulling levers across their entire corporate structure:Slowing Shareholder Returns: While companies like Meta and Alphabet recently initiated dividends to appease investors, the pace of their massive share buyback programs is facing pressure. Every dollar spent buying back stock is a dollar that cannot be spent on an Nvidia server rack.Operational Austerity: The "Year of Efficiency" wasn't a one-off. Hyperscalers are continuing to quietly trim headcount, slash middle management, and eliminate speculative, non-AI "moonshot" projects. They are gutting the bloated parts of their legacy businesses to redirect that payroll money straight into data center real estate.Oracle's Equity Strategy: Oracle, which is trying to compete with giants ten times its size, is stretching its balance sheet the furthest. To fund its $50 billion 2026 CapEx goal without losing its investment-grade credit rating, it is turning to equity and equity-like instruments (essentially issuing new stock) to raise capital. The Bottom Line The hyperscalers are no longer self-funding fortresses. To pay for the AI revolution, they are spending almost every spare dollar of their own cash, initiating severe internal cost-cutting, and borrowing hundreds of billions of dollars from global bond markets. They are officially acting less like software companies and more like heavily leveraged utility, energy, and telecom giants.
+May 12, 20260so where is the capex currently going? if they are constrained by power to actually power the data centers?Even if a hyperscaler cannot immediately hook up a new 500-megawatt facility to the grid, the capital expenditure meter is still running at maximum speed. In 2026, the four major US hyperscalers are projected to spend close to $700 billion. If they can't plug new data centers in, where is all that cash actually going today? It is flowing into a mix of hardware upgrades, supply-chain hoarding, and massive infrastructure waitlists.Here is exactly where the CapEx checks are being written right now:1. Silicon and Short-Lived Assets (The Lion's Share)Historically, you might assume "CapEx" means buying land and laying concrete. In the AI era, the vast majority of capital expenditure goes straight into fast-depreciating silicon. Roughly two-thirds of Microsoft’s recent CapEx, for example, went toward "short-lived assets" like GPUs and CPUs. The Upgrade Cycle: Because power is capped at existing data centers, hyperscalers are actively ripping out older, traditional servers and replacing them with the newest, highest-density AI hardware. They are spending hundreds of billions on Nvidia GPUs, custom in-house chips (like Google's TPUs or Amazon's Trainium), and the High-Bandwidth Memory (HBM) required to run them. This allows them to get drastically more compute out of the exact same power envelope. Hoarding: Hyperscalers are also buying silicon and racking it up in warehouses, letting it sit idle so that it is ready the precise moment a new facility finally gets power.2. Heavy Electrical Equipment (Buying a Spot in Line)To build a data center, you need massive power transformers, switchgear, backup generators, and electrical distribution units. As mentioned earlier, this equipment has a 2-to-4 year lead time. Hyperscalers are not waiting for a grid connection to order this gear. They are spending billions today to secure their place in the manufacturing backlog of industrial power companies like Eaton, Schneider Electric, and nVent. The cash has to be deployed now to ensure the physical electrical components exist three years from now.3. Advanced Cooling and NetworkingAI racks are incredibly dense and run exponentially hotter than standard cloud servers. A massive chunk of current CapEx is flowing into companies that build liquid cooling infrastructure (like Vertiv) to keep the chips from melting. Additionally, training massive AI models requires thousands of GPUs to "talk" to each other flawlessly. Hyperscalers are spending billions on optical networking equipment, fiber, and advanced switches to stitch these chips into massive supercomputers. 4. Real Estate and Colocation LeasesHyperscalers do not build every data center from scratch. They are pouring CapEx into massive, long-term leases with specialized Data Center Real Estate Investment Trusts (REITs) like Equinix and Digital Realty. Because these REITs specialize in securing land and power permits years in advance, hyperscalers are paying them premium, double-digit rate hikes just to absorb whatever available, powered real estate exists on the market right now.5. Energy Development (The Long Game)While it's a smaller piece of the pie compared to silicon, billions are being allocated toward the actual generation of power. This includes funding the engineering studies for Small Modular Reactors, signing massive Power Purchase Agreements (PPAs) with utility companies at premium rates, and investing in localized natural gas and battery storage solutions.The Financial RiskThis spending distribution is exactly what makes Wall Street nervous. The majority of the $700 billion being spent in 2026 is going toward chips that will be economically obsolete in three to five years.If a hyperscaler buys a GPU today, but it takes two years to get the power to turn it on, that chip will have lost half of its useful competitive lifespan before it ever processes a single token of data. They are racing against a clock where their most expensive assets are depreciating faster than the concrete around them can dry.
+May 12, 20260so currently 100% of the power is coming from grids and build your own meter is pipe dream currently?You have hit the nail on the head. For all practical purposes today, virtually 100% of the power driving the AI boom is coming from the traditional, public utility grid. However, calling the "build your own" strategy a pipe dream isn't quite accurate. It is more of a heavily-funded pipeline with a massive latency problem.Here is the exact reality of the situation today and why "escaping the grid" is proving so difficult.The Reality of Today's PowerRight now, hyperscalers do not have fleets of their own private power plants running. When they say they are "powering with 100% renewable energy," what they actually mean is they are buying power from the local grid (which is a mix of coal, gas, nuclear, and renewables) and then buying "Renewable Energy Certificates" (RECs) from wind and solar farms elsewhere to offset their carbon footprint on paper.The actual electrons flowing into the data centers are coming from the exact same utility grid that powers the surrounding cities.The Amazon/Talen Reality Check (The Grid Strikes Back)The one major exception—Amazon’s $650 million purchase of a data center campus directly hooked up to the Susquehanna nuclear plant—recently became a perfect example of why building "behind-the-meter" is so hard.Amazon and Talen Energy tried to expand the amount of power they could pull directly from the nuclear plant without passing through the public grid. In late 2024, the Federal Energy Regulatory Commission (FERC) actually voted to block the expansion.Why? Because even if you build right next to a power plant, you are still technically connected to the broader grid for backup and stability. Utilities and regulators argued that Amazon would be "freeloading"—pulling massive amounts of power and destabilizing the local system without paying the transmission and maintenance fees that normal grid customers pay.This ruling sent shockwaves through the tech industry. It proved that even if you buy a nuclear plant, you cannot simply declare yourself immune from grid regulations.Not a Pipe Dream, Just a 10-Year WaitThe massive plans for Microsoft to restart Three Mile Island or Google to build Small Modular Reactors (SMRs) are not pipe dreams. The contracts are signed, the capital is committed, and the engineering firms are hired.But money cannot alter the laws of physics or the pace of federal bureaucracy.The SMRs: The technology for commercial Small Modular Reactors is still being proven. Google’s target of 2030 is highly ambitious.The Gas Plants: Building on-site natural gas requires environmental impact studies, securing gas pipeline rights-of-way, and passing emissions regulations.The Bottom LineToday, hyperscalers are entirely tethered to the public grid. The "bring your own power" era is real, but it is currently in the heavy-construction and permitting phase. You will not see a meaningful percentage of AI compute powered by private, off-grid hyperscaler power stations until at least 2028 to 2032. Until then, they are fighting for grid access just like heavy manufacturing and traditional industry.
+May 12, 20260how many of these power stations hyperscalers have successfully build so far? and how many are underconstruction ? and how many are in plan?If you are looking for a massive, gigawatt-scale power plant that a hyperscaler has designed, funded, and built from scratch that is currently operational, the number is zero.Software companies do not have the expertise to be heavy-industrial construction firms. Instead of building from scratch, hyperscalers are taking three different approaches: buying existing plants, paying to resurrect dead plants, or funding startups to build next-generation reactors.Here is the current scorecard of dedicated hyperscaler power projects as of mid-2026.1. Successfully Built & Operational (The "Bought" Approach)There is currently only one major, operational example of a massive data center co-located with a dedicated power station, and the hyperscaler didn't build it; they bought it. Amazon (AWS) & Susquehanna: In early 2024, Amazon paid $650 million to buy a data center campus directly adjacent to the Susquehanna nuclear power plant in Pennsylvania. They subsequently signed an agreement with Talen Energy to secure up to 1,920 megawatts of power from the plant. This is the only major, live example of a hyperscaler successfully tapping a dedicated, gigawatt-scale power source. (Note: Due to recent federal regulatory pushback, this arrangement is currently transitioning slightly to use more of the public transmission grid rather than staying entirely "behind the meter.") 2. Under Construction / Active Reactivation (The "Resurrection" Approach)Right now, the most active "construction" sites are actually efforts to turn defunct nuclear reactors back on.Microsoft & Three Mile Island (1 Project): Microsoft signed a 20-year, $1.6 billion agreement with Constellation Energy to restart Unit 1 of the Three Mile Island nuclear plant in Pennsylvania (which was shut down in 2019). The project has already hired hundreds of workers for turbine and cooling system upgrades. It is currently targeting a restart in late 2027. Google & Kairos Power (1 Project): Google is backing Kairos Power to build the first commercial advanced Small Modular Reactors (SMRs). Kairos received a construction permit from the Nuclear Regulatory Commission in late 2024 for its Hermes 2 reactor in Tennessee. Construction is active, with a target operational date of 2030.3. In Plan / Proposed (The "Future Pipeline")This is where the numbers explode. Because grid connections are so slow, hyperscalers and data center developers have proposed an absolute mountain of on-site, dedicated power generation. The SMR Plans (Multiple): * Oracle announced they have secured building permits to design a massive 1-gigawatt data center powered by three custom Small Modular Reactors, though construction timelines and locations remain heavily guarded. Amazon has invested heavily in X-Energy, backing plans to eventually build 5 gigawatts of SMR capacity.The Natural Gas Pipeline (Hundreds of Projects): According to industry tracking in early 2026, there is currently over 130 gigawatts of behind-the-meter, on-site power proposed for US data centers. Over 80% of this planned capacity is natural gas.Because permitting and supply chain realities will kill many of these proposals, energy analysts project that only 22 to 32 gigawatts of this on-site generation will actually be built and deployed by 2030. The Takeaway: Hyperscalers have realized that building a power plant takes a decade. To bridge the gap, they are buying the power plants that already exist, writing blank checks to restart the ones that were abandoned, and laying the groundwork for a massive wave of SMRs and gas turbines in the 2030s.
+May 12, 20260if they build their own power stations, which of these problems are mitigated and which ones still exists or are amplified ?If a hyperscaler decides to bypass the utility company entirely and build its own power station directly on-site—a strategy known in the industry as "behind-the-meter" generation or "co-location"—it dramatically changes the bottleneck landscape.However, it is not a magic bullet. Building their own power stations essentially means these tech giants are trading one set of grid-level problems for a completely new set of heavy-industry problems.Here is a breakdown of how the bottlenecks shift if hyperscalers become their own power companies.🟢 What is Mitigated (The Wins)By generating their own power on-site, hyperscalers successfully eliminate the friction of dealing with public utility infrastructure.The Interconnection Queue: Highly Mitigated. If a data center is entirely "islanded" (off-grid) or only uses the grid for a tiny amount of backup power, they can largely bypass the 2-to-4 year RTO/ISO utility queue. They don't have to wait for a grid operator to study how their massive load will impact the surrounding city, because they aren't drawing from the city.Long-Distance Transmission: Mitigated. Because the power generation (e.g., a natural gas plant or small modular reactor) is built immediately adjacent to the data center, there is no need to permit, secure land rights for, and build 100-mile high-voltage transmission lines.The Baseload Problem: Mitigated. This is the primary reason they are exploring this route. By building their own natural gas turbines, geothermal plants, or nuclear facilities, hyperscalers guarantee themselves the uninterrupted, 24/7 power that AI requires, entirely insulated from public grid brownouts or the unreliability of wind and solar.🟡 What Still Exists (The Persistent Bottlenecks)Certain physical realities of handling electricity cannot be avoided, regardless of who owns the power plant.The Transformer Shortage: Still Exists. Even if you generate your own power 500 feet away from your servers, you still have to step that power up and down to make it usable for the data halls. The global 3-to-4 year backlog for high-voltage transformers, switchgear, and commercial breakers will still throttle their timeline.Water Consumption Constraints: Still Exists. Both thermal power plants (like nuclear and natural gas) and AI data centers require massive amounts of water for cooling. Co-locating them doubles the localized strain on local water tables, which is already a major point of friction with local municipalities.🔴 What is Amplified (The New Headwinds)When a software company decides to build a power plant, they are stepping into one of the most heavily regulated, capital-intensive industries on earth.Permitting & NIMBYism: Massively Amplified. If local communities fight tooth-and-nail against a relatively quiet data center being built in their town, the opposition to a dedicated natural gas plant or a nuclear reactor is exponentially higher. Instead of just fighting local zoning boards, hyperscalers now have to clear the EPA, state environmental regulators, and, in the case of nuclear, the heavily bureaucratic Nuclear Regulatory Commission (NRC).Supply Chain Constraints: Amplified. Hyperscalers are no longer just competing for GPUs and cooling racks. They now have to compete in the supply chain for power generation equipment. The waitlists for large-scale natural gas turbines (from companies like GE or Siemens) are years long. The supply chain for Small Modular Nuclear Reactors (SMRs) practically doesn't exist yet at commercial scale.Capital Expenditure (CapEx): Amplified. Building a gigawatt-scale power plant costs billions of dollars. Hyperscalers are now taking on the capital risk of building utility infrastructure on top of the capital risk of building the AI infrastructure.Operational Risk: Amplified. Running a cloud network is very different from operating a high-pressure steam turbine or handling nuclear material. Tech companies will have to acquire, train, and manage entirely new workforces of heavy-industrial engineers, or form complex, expensive joint ventures with existing energy companies (like Microsoft’s recent deal to restart Three Mile Island with Constellation Energy).SummaryBuilding their own power stations gives hyperscalers control at the expense of complexity. It frees them from the sluggish pace of public grid upgrades and transmission lines, but forces them into the brutally slow realities of environmental permitting, heavy-industrial supply chains, and power plant construction.
+May 12, 20260what is taking so much time for power?The delay in securing power for new data centers boils down to a collision between the lightning-fast speed of software development and the painfully slow reality of heavy industry, physics, and bureaucracy.You cannot simply plug a 500-megawatt artificial intelligence data center into the existing grid the way you plug in a new appliance. Here is exactly what is causing the massive 4-to-5 year delays.1. The Interconnection Queue (Bureaucracy & Physics)Before a utility allows a hyperscaler to draw massive amounts of power, the grid operator must conduct exhaustive engineering studies to ensure that adding this new, colossal load will not destabilize the local grid, cause brownouts, or fry existing infrastructure.The Backlog: Grid operators in major tech hubs (like PJM Interconnection in the US, which covers Virginia's "Data Center Alley", or EirGrid in Ireland) are drowning in applications.The Timeline: These grid-impact studies, combined with the administrative approval process, can take two to four years before a single shovel hits the ground.2. The Transformer Shortage (Supply Chain)Even if you have the money and the permit, you still need the physical hardware to step down the high-voltage electricity from the grid into usable power for the servers.The Bottleneck: There is a severe global shortage of heavy electrical equipment, particularly large power transformers, switchgear, and specialized commercial generators.The Lead Times: A few years ago, you could order a high-voltage transformer and have it in a few months. Today, the lead time for these massive, custom-built pieces of equipment is between 120 and 200 weeks (roughly 2.5 to 4 years). There are only a handful of manufacturers in the world capable of building them.3. Transmission and Permitting (The NIMBY Problem)If the power exists, but it is 100 miles away from the data center site, the utility has to build high-voltage transmission lines to connect them. This is often the slowest part of the entire process.Land Rights: Building new transmission lines requires acquiring rights-of-way across private land, state borders, and protected environments.Lawsuits: These projects almost always face "Not In My Back Yard" (NIMBY) opposition, environmental lawsuits, and local political battles.The Timeline: Successfully permitting and building a new major transmission line often takes seven to ten years.4. The "Baseload" Generation ProblemAI data centers run at near 100% utilization, 24 hours a day, 7 days a week. They require "baseload" power—electricity that never stops flowing.The Grid Transition: At the exact moment hyperscalers are demanding record amounts of baseload power, the world's power grids are actively retiring their traditional baseload plants (like coal) in favor of intermittent renewables (like solar and wind, which only generate power when the sun shines or the wind blows).The Mismatch: You cannot power a 24/7 AI cluster solely with a solar farm. Hyperscalers need nuclear or natural gas for that constant stability. Building a new natural gas plant takes several years of permitting and construction; restarting an old nuclear plant or building a new Small Modular Reactor (SMR) takes the better part of a decade.In short, the tech industry moves in months, but the utility industry moves in decades. The money is there, but you cannot buy your way out of a four-year manufacturing queue for a transformer, and you cannot fast-track a decade's worth of grid infrastructure planning.
+May 12, 20260are hyperscalers getting roi on their data center investments?The short answer is: Not yet, at least not at a level that justifies the current scale of spending. Right now, the return on investment (ROI) for hyperscalers (Amazon, Microsoft, Alphabet, Meta, and Oracle) is a massive debate on Wall Street. While these companies are seeing revenue growth from AI, the math between what they are spending and what they are earning back doesn't currently balance.Here is a breakdown of the current reality regarding hyperscaler ROI on data center investments.The Scale of the CapExThe sheer volume of capital expenditure (CapEx) is staggering. The "Big Five" hyperscalers are collectively projected to spend between $660 billion and $690 billion on CapEx in 2026, nearly doubling their 2025 levels. Roughly 70% to 75% of this (around $450 billion) is flowing straight into AI infrastructure—compute clusters, GPUs, optical fabrics, and liquid-cooled data halls. To put this in perspective, hyperscalers are increasingly having to tap into debt and corporate bond markets to fund this buildout, as the costs are beginning to outstrip their internal free cash flows. The Revenue RealityThe AI sector is generating real revenue, but it is a fraction of the infrastructure cost.The Model Builders: OpenAI reportedly ended 2025 with about $20 billion in annual recurring revenue (ARR), and Anthropic hit roughly a $9 billion run rate in early 2026. While impressive, OpenAI's revenue represents just about 3% of the projected 2026 hyperscaler CapEx total. The Cloud Providers: Microsoft, Google, and Amazon are seeing upticks in their cloud divisions driven by enterprise AI workloads and services like Copilot.The ROI Math: According to recent modeling by J.P. Morgan, the AI sector would need to generate approximately $650 billion in annual revenue just to achieve a 10% return on the current infrastructure investments. The industry is nowhere near that top-line figure yet. If the ROI isn't there yet, why are they spending?Hyperscaler management teams are acutely aware of the ROI uncertainty, but they are treating these investments as an existential necessity rather than a standard, near-term yield play. Defensive Compulsion: No major tech giant can afford to lose the platform shift. They are willing to accept deep, upfront margin compression to build a competitive moat and ensure they aren't displaced by a rival's superior compute capabilities. Building for Internal Ecosystems: Hyperscalers aren't just building data centers to rent out to third-party AI startups. Meta, for example, is building immense compute power to train open-source models and improve its own core advertising and social algorithms. Alphabet is protecting its search monopoly. Anticipating the Inference Boom: Right now, the bulk of the power is being used for training massive models. Hyperscalers are betting that as AI matures and agentic workflows become standard in enterprise software, the demand for inference (the daily execution of AI tasks) will consume every unit of capacity they build. The Market ReactionInvestors are starting to lose patience with blanket spending. Earlier in the AI boom, any company aggressively spending on AI was rewarded by the market. Moving into 2026, Wall Street has become much more selective.Markets are actively rewarding the "picks and shovels" of this cycle—Nvidia, TSMC, memory suppliers, power infrastructure, and cooling tech—because their ROI is immediate and tangible. For the hyperscalers, investors are now punishing companies that rely too heavily on debt to fund data centers, demanding a clearer, provable link between CapEx depreciation and actual operating earnings.
+May 12, 20260so if power is constrained why are companies spending so much capex?It is the biggest paradox—and the biggest bottleneck—in the tech industry right now. Microsoft recently admitted on an earnings call that they are sitting on billions in unfulfilled Azure cloud orders simply because they cannot find the electricity to power the GPUs. So, if power is the limiting factor, why are the five major hyperscalers projected to spend nearly $690 billion on CapEx in 2026? Because the power constraint is exactly what is driving a massive portion of that spending. Here is why the money keeps flowing even when the electricity doesn't. 1. Buying "Compute Per Watt" (Hardware Upgrades)If a hyperscaler has a data center capped at 100 megawatts, they cannot physically draw more power from the grid. The only way to increase their AI processing capacity within that facility is to become drastically more power-efficient.This means ripping out older chips and replacing them with the newest generations. A massive chunk of current CapEx isn't going toward expanding the total footprint; it's going toward replacing economically obsolete hardware. The newest GPUs deliver significantly more compute per watt, allowing hyperscalers to squeeze more AI performance out of their existing, strictly limited power envelopes. 2. Hyperscalers Are Becoming Energy CompaniesA growing percentage of that CapEx is no longer going toward servers and networking gear—it is going directly into energy infrastructure.Because local utilities cannot upgrade the grid fast enough to accommodate AI's gigawatt-scale demands, hyperscalers are adopting a "bring your own power" strategy. They are spending billions to build their own microgrids, deploy massive on-site battery storage systems, invest in natural gas plants, and, increasingly, fund nuclear power restarts. They are spending the CapEx to literally build the power supply they are missing. 3. The 4-to-5 Year Waiting RoomGetting a massive data center connected to the power grid is not a quick process; the average wait time for a major grid connection in primary data center markets is now exceeding four years. Hyperscalers are spending billions today to secure strategic land, sign long-term utility commitments, and begin heavy construction. They are deploying capital now to ensure they have the physical space and grid priority secured for 2028 and 2030. If they wait for the power bottleneck to clear before they start spending, they will be five years behind their competitors.4. Hardware HoardingSupply chains for the most advanced chips, cooling systems, and specialized electrical equipment remain tight. Hyperscalers are buying the actual hardware (GPUs, liquid cooling racks, custom networking fabrics) well in advance. They are willing to let billions of dollars worth of servers sit idle or in warehouses so that the exact minute a new power facility comes online, the hardware is ready to be racked and turned on.The Bottom LineIn a power-constrained world, capital intensity actually goes up, not down. Hyperscalers are forced to overspend on the newest chips to maximize efficiency, overspend on real estate to get in line for future grid connections, and overspend on energy infrastructure to build the power the public grid can't provide.
+May 12, 20260when hyperscalers say that there is immense demand and they don't have enough supply, who is the demand? is the demand paying yet?When hyperscalers talk about insatiable demand outstripping their supply of compute, it sounds like thousands of companies are beating down their doors to buy AI. The reality is much more concentrated—and the way that demand is "paying" is one of the most controversial topics in tech right now. To answer your questions directly: The demand is overwhelmingly driven by just two companies, followed by the hyperscalers' own internal needs. And while they are technically paying, a massive chunk of that money is part of a closed-loop, circular economy. Here is exactly who is taking up all those servers and how the bills are getting paid in 2026.1. The Frontier Labs: OpenAI and Anthropic (The "Whales")If you want to know who is consuming the vast majority of the world's GPU capacity, look no further than OpenAI and Anthropic. Industry estimates suggest that a massive percentage of all AI software and compute revenues flow directly through these two entities. Who they use: OpenAI absorbs Microsoft and Oracle's compute. Anthropic is eating up capacity from Amazon and Google to the point where they are effectively maxing out those hyperscalers' available limits. Are they paying? Yes, to the tune of billions. But there is a catch. The frontier labs are highly unprofitable. The money Anthropic uses to pay Google and Amazon for compute is largely the exact same money Google and Amazon invested into Anthropic (like Google’s massive $40 billion commitment to Anthropic this year). It is a circular loop: the hyperscalers fund the AI labs, and the AI labs hand that money right back to the hyperscalers to rent servers. 2. Internal Hyperscaler Consumption (The "Ghost Demand")A massive portion of the compute being built isn't actually for sale; it's being consumed internally by the tech giants themselves.Meta: Meta is spending billions on compute to train Llama and power the AI algorithms that drive their core social media and advertising feeds.Google: Alphabet uses an immense amount of compute to overhaul Google Search and run its own Gemini models.Are they paying? Not in direct cash. They are absorbing the cost as CapEx. The "payment" comes in the form of protecting their existing monopolies (Search, Ads) and maintaining user engagement.3. The Enterprise Cloud Buyers (The "Real" Economy)This is the sector Wall Street cares about most: actual businesses paying real money to rent AI infrastructure to build their own tools.Who they are: This ranges from well-funded AI startups (like legal AI firm Harvey) to Fortune 500 companies modernizing their logistics, coding workflows, or customer service.Are they paying? Yes, and this is where the demand constraint is very real. For example, in early 2026, Google explicitly admitted to investors that their cloud revenue would have been significantly higher if they simply had more compute to sell. They reported an order backlog of $460 billion—representing hard, committed money from enterprises waiting for server space. The Bottom Line When a hyperscaler says they are "compute-constrained," they usually mean they don't have enough servers to keep up with Anthropic or OpenAI's next training run, while simultaneously trying to fulfill the backlog of enterprise cloud contracts. The demand is paying, but the ecosystem is currently propped up by a massive vendor-financing scheme. The ultimate test of the AI economy over the next few years isn't whether OpenAI can pay Microsoft for compute today; it's whether OpenAI can eventually generate enough organic revenue from end-users to pay for that compute without needing another multi-billion-dollar check from Microsoft to do it.
+May 12, 20260how much hyperscalers spent in 2023, 2024 on data centers, how much of that capacity is operational ? and how much revenue it has already generated?Here is the plain and simple breakdown of the numbers:What Hyperscalers Spent2023: Total global data center spending was roughly $300 billion.2024: Global spending surged by 51% to $455 billion. The biggest hyperscalers (Amazon, Microsoft, Google, Meta) directly accounted for roughly $200 billion of that total just for their own facilities and AI hardware. How Much is OperationalActive Facilities: By the end of 2024, there were exactly 1,136 fully operational hyperscale data centers worldwide. New Capacity: Out of the money spent recently, 137 new hyperscale data centers officially came online and started running in 2024. The Waitlist: Physical construction is lagging behind the massive spending. Right now, there are over 500 hyperscale data centers stuck in the pipeline—either planned, under construction, or waiting for power—that are not yet operational. Revenue Generated2024 Total Revenue: The total cloud market brought in $330 billion in 2024. Market trackers note that Generative AI services running on this new capacity are currently driving about half of all new cloud growth. Current Returns: The spending is already paying off heavily. By the first quarter of 2026, global cloud infrastructure spending hit $129 billion per quarter, putting the industry on track to easily clear $500 billion in revenue for the year.