Are AI clouds building too much capacity?
In our updated market reports, you will find everything you need
SUMMARY
AI clouds are building too much capacity in parts of the future pipeline, but not in the top-tier market customers need today.
The apparent contradiction is the whole story. Modern, well-connected accelerator clusters remain scarce, while hundreds of billions of dollars are already being committed to sites, chips and power that will arrive after today's shortages have eased.
The biggest risk is economic overcapacity rather than rows of empty servers. A GPU cluster can stay busy and still disappoint its owner if rental prices fall faster than electricity, depreciation, interest and maintenance costs.
Headline capital expenditure exaggerates the amount of usable AI compute arriving immediately. The totals mix short-lived chips with data-center shells, networks, power systems, warehouses, robotics and projects that may take years to open.
Cloud revenue is catching up faster than skeptics expected. AWS, Azure, Google Cloud and Oracle Infrastructure are all growing rapidly, and leading providers still report supply constraints. That makes a near-term collapse in demand unlikely.
Backlogs offer real protection, but the combined totals flatter the market. Definitions differ, contracts stretch across many years, and the same model developer can support commitments at several infrastructure providers at once.
Power shortages will filter weak proposals and delay construction, although they will not rescue a bad investment once a builder has signed long-term electricity, financing and lease commitments. Several delayed projects could also arrive together after grid upgrades are completed.
Inference is the strongest argument for continued expansion. Training serves a small group of model developers, while inference can spread across coding, advertising, search, customer service and business workflows. The buildout works only if valuable paid usage grows faster than efficiency improves.
Hardware useful life and commercial life are different. Older accelerators may keep running for years, but their premium pricing can disappear after a newer generation arrives. That is manageable for a hyperscaler with many internal workloads and much harder for a leveraged specialist cloud.
The first glut will probably be patchy. Older chips, remote sites, weak networking and projects tied to one customer should face price pressure before large current-generation clusters in major regions do.
CoreWeave captures both sides of the boom: enormous contracted demand, rapid revenue growth and a large operating footprint, alongside heavy interest expense, losses, leases and customer concentration. Sold capacity is not automatically safe capacity.
The likely outcome is a selective shakeout rather than a global AI-compute crash. Most of the infrastructure should eventually find a use, but some owners will refinance, merge or sell assets after discovering that useful hardware can still be an overpriced investment.
Why are people worried about AI cloud overbuilding now?
The worry is justified because AI infrastructure spending has reached a scale where even strong cloud growth may struggle to produce good returns on every new dollar.
Amazon expects about $200 billion of capital expenditure this year. Alphabet has raised its guidance to $195 billion-$205 billion. Meta now expects $125 billion-$145 billion. Microsoft spent $31.9 billion in its latest reported quarter alone, with roughly two-thirds going into shorter-lived assets such as GPUs and CPUs.
The comparison needs a few adjustments. Amazon also spends on warehouses, robotics and satellites. Meta mainly builds for its own products. Higher chip and memory prices have pushed budgets upward without creating the same increase in computing power. Even after those adjustments, the four companies are running toward roughly $650 billion-$675 billion of annual investment when we use Microsoft's latest quarterly pace.
The cash pressure is already visible. Amazon's trailing free cash flow fell to $1.2 billion after property and equipment purchases increased by $59.3 billion, mainly because of AI investment. Alphabet says its expanded buildout will keep pressuring free cash flow through depreciation, energy and data-center operating costs. Microsoft Cloud's gross margin has slipped to 66% as AI infrastructure and usage costs rise.
Demand has accelerated across every major provider lately. This is a real boom, financed at a speed that leaves little room for forecasting mistakes.
| Builder | Current spending level | What makes it less comparable |
|---|---|---|
| Amazon | About $200B this year | Includes logistics, robotics and satellites |
| Alphabet | $195B-$205B this year | Covers servers, networks and long-lived data centers |
| Meta | $125B-$145B this year | Mostly supports Meta's own products |
| Microsoft | $31.9B in the latest quarter | About two-thirds went to shorter-lived chips and servers |
If you want more recent data on this point, please see our latest GPU cloud market report.
What does “too much AI cloud capacity” actually mean?
AI clouds have built too much capacity when the hardware and data centers cannot earn an acceptable return before they become cheaper, outdated or financially burdensome.
Low utilization during a new site's first months tells us very little. Buildings open before every server arrives, customers ramp gradually, and cloud operators need spare room for demand spikes and failures. A cloud running every accelerator at full load would probably be unable to serve the next customer.
The better test follows the money over the asset's useful life. Revenue must cover the servers, networking, buildings, electricity, cooling, maintenance and financing, while still leaving a worthwhile return. A GPU cluster can stay busy and still lose money when rental prices fall faster than its costs.
Three different failures get mixed together in this debate. Sometimes live servers sit idle. Sometimes customers use them while the owner earns too little. And sometimes every rival builds defensively because being left behind looks more dangerous than wasting capital.
Economic overcapacity is the main risk. The newest clusters remain scarce, while the cost of owning them is rising and the price of a unit of compute keeps falling. The market can have a shortage of top-tier capacity and still create disappointing returns for many builders.
How much new AI capacity are the clouds really creating?
The spending surge is enormous, although the headline dollar totals overstate how much finished AI compute will come online in the near term.
Microsoft's latest quarter gives us a useful breakdown. Roughly two-thirds of its $31.9 billion in capital expenditure went into shorter-lived assets, mainly GPUs and CPUs. The remaining third funded data centers and other infrastructure designed to last 15 years or longer. Alphabet has described a similar mix in recent quarters, with spending divided between servers and the buildings, networking and power systems around them.
That split is important. A power connection, fiber network or data-center shell can support several generations of chips. A GPU bought at a premium price has a much shorter period in which it can command premium rental rates.
The physical buildout is still historic. Microsoft added more than two gigawatts of capacity over a 12-month period. Amazon says it has landed 1.4 million Trainium2 chips, including more than 500,000 in Project Rainier for Anthropic. Alphabet has accelerated deliveries again because its own supply has fallen behind demand.
Announced spending mixes current chips, future buildings, higher component prices and projects that may take years to become usable. Only part of the total becomes working AI compute quickly.
Is there already too much usable AI compute today?
There is no convincing evidence of a broad surplus in the large, modern AI clusters that customers value most.
Microsoft expects broad customer demand to keep exceeding supply even as it brings more capacity online. Alphabet has started using additional third-party infrastructure as a temporary bridge because its internal fleet cannot meet demand quickly enough. Paying another provider is usually less profitable than using your own data centers, so Alphabet is behaving like a company facing a shortage, not a glut.
Amazon's custom chips tell a similar story from another angle. Trainium2 is fully subscribed, and nearly all expected Trainium3 supply was already expected to be committed around the middle of the year. Oracle's infrastructure revenue nearly doubled in its latest fiscal quarter, while CoreWeave says its recent revenue is overwhelmingly tied to committed customer contracts.
Scarcity is uneven. A small cluster of older GPUs in the wrong region may be easy to rent. Tens of thousands of current-generation accelerators connected through fast networking are much harder to secure. Frontier training, large-scale post-training and heavy inference workloads cannot simply use any empty server rack.
The market is tight where the highest-paying demand sits. The overbuilding risk lies mostly in capacity scheduled for later, after several years of rapid construction and efficiency gains.
If you want more recent data on this point, please see our latest GPU cloud market report.
Are AI cloud revenues catching up with the spending?
AI cloud revenues are growing fast enough to support the boom for now, while today's buildout still runs ahead of the costs already visible in earnings.
The latest operating results are unusually strong for businesses of this size. AWS generated $37.6 billion of quarterly revenue and $14.2 billion of operating income, both well above the previous year. Google Cloud reached $24.8 billion of quarterly revenue, while its operating margin rose from 20.7% to 35.6%. Azure grew 40%, and Microsoft's AI business passed a $37 billion annual revenue run rate. Oracle Cloud Infrastructure reached $5.8 billion in quarterly revenue after growing 93%.
The customers are real. The timing creates the risk: current revenue partly comes from equipment installed earlier, while this year's spending will create depreciation, power bills and financing costs for years.
Microsoft shows the tension clearly. Azure is still growing around 40%, yet Microsoft Cloud's gross margin has fallen from 69% to 66% over the past year as infrastructure and AI usage costs rose. Google Cloud's margins improved sharply, but Alphabet has warned that its new capacity will increase depreciation and energy expenses. Amazon's AWS profit is climbing while company-wide free cash flow has almost disappeared under the weight of investment.
The economics work when revenue keeps compounding quickly after the new facilities open. The harder test begins when today's construction wave reaches full depreciation and the easiest AI customers are already signed.
Do the giant AI cloud backlogs guarantee future demand?
The giant AI cloud backlogs provide strong near-term protection, although they give a more flattering picture of demand than a simple total suggests.
Microsoft reports $627 billion of commercial remaining performance obligations. Google Cloud's backlog has reached $514 billion after increasing by more than $50 billion in one quarter. Oracle reports $638 billion, and CoreWeave nearly $100 billion. Adding them produces more than $1.8 trillion, but that total is not a clean measure of independent AI demand.
The definitions and time periods differ. Microsoft's number includes software and non-AI cloud contracts. Google says most of its backlog comes from normal GCP agreements across a broad customer base. Oracle's figure contains very large AI infrastructure commitments. CoreWeave expects its contracts to run as long as seven years, with only 36% of unsatisfied obligations due to become revenue during the first 24 months.
The same end customer can also appear in several backlogs. A model developer may reserve compute from Microsoft, Oracle and CoreWeave while buying chips or services elsewhere. All those contracts can be real, even though one AI company's future cash generation ultimately supports several infrastructure providers.
Contract quality varies too. Prepayments, customer-supplied GPUs and strong counterparties reduce risk. Long commitments from a highly leveraged AI company shift the question toward whether that customer can keep paying.
These backlogs make a sudden demand collapse unlikely. They cannot tell us what renewal prices will look like once supply improves or whether the end users of AI will generate enough value to support every layer of contracts.
| Provider | Reported backlog or RPO | Useful reading | Main caution |
|---|---|---|---|
| Microsoft | $627B | Broad commercial visibility | Includes software and OpenAI-related obligations |
| Google Cloud | $514B | More than half expected within 24 months | Covers standard GCP and AI contracts |
| Oracle | $638B | Rapidly expanding AI infrastructure commitments | Large contracts may share the same end demand |
| CoreWeave | About $99B | Long contracted runway | Revenue remains concentrated among a few customers |
Are all those announced AI data centers actually going to open?
A large share of announced AI data-center capacity will arrive later than advertised, shrink before construction or never become revenue-producing capacity.
The market often mixes four very different stages. A proposed site may only have land. A power-secured site still needs permits, equipment and construction. An energized building may be waiting for servers. A fully fitted cluster may need months of testing before customers can use it.
CoreWeave's disclosures show why those stages matter. It has more than one gigawatt of active power, more than 3.5 gigawatts contracted and a goal of more than eight gigawatts by 2030. Only the first figure describes capacity that can support workloads now. The rest depends on financing, construction, electricity delivery and customer schedules.
Utilities have become more skeptical of headline pipelines. PJM's latest forecasting method treats near-term projects as firm only when they have stronger service or construction commitments, while less certain requests are discounted. EPRI says forecasting remains difficult because public data are thin and many announced projects are speculative.
Lead times add another filter. Large transformers, switchgear, turbines, cooling equipment and high-bandwidth networking cannot be ordered and installed overnight. In many major markets, access to utility power can take several years.
The announced pipeline exaggerates the next wave of supply. Press releases make a future glut look closer and larger than the physical buildout suggests.
Will power shortages stop AI clouds from overbuilding?
Power shortages will slow AI cloud expansion without protecting investors from bad projects that eventually secure electricity.
The newest US estimate from Lawrence Berkeley National Laboratory puts data centers at a possible 9.5% to 15.3% of national electricity use by 2030, with a central estimate of 11.8%. EPRI's updated scenarios are similarly wide, at roughly 9% to 17%. Those ranges show how uncertain AI adoption, hardware efficiency and project delivery still are.
Local pressure will be much sharper than the national average. EPRI estimates that data centers could consume 39% to 57% of Virginia's electricity by 2030 under its scenarios. Large AI campuses regularly request hundreds of megawatts, sometimes approaching the output of a major power plant.
Scarce electricity forces developers to prove that projects are serious. Utilities and regulators are less willing to reserve grid capacity for vague proposals, and projects with weak financing or customers are more likely to fall away.
But once a builder signs long-term power, generation or transmission commitments, scarcity stops being protection. A delayed data center can keep accumulating financing and contractual costs before it earns revenue. Several delayed projects can also reach completion together after grid upgrades arrive, creating a sudden jump in supply.
Power slows the construction schedule. The finished cluster still has to make money.
Can inference demand fill all these new AI data centers?
Inference can absorb far more AI cloud capacity than training alone, and current usage suggests that this shift is already underway.
Training demand comes from a limited group of model developers running very large jobs. Inference is generated whenever someone uses a chatbot, coding assistant, search tool, recommendation system, video model or agent. That creates a much broader pool of recurring workloads.
The latest company data show the expansion. Microsoft's AI business has passed a $37 billion annual revenue run rate, and its cloud cost growth now reflects rising GitHub Copilot usage as well as infrastructure investment. Amazon says Bedrock serves more than 100,000 companies and that Trainium2 handles most of the service's inference. Google attributes its latest Cloud acceleration to core GCP, AI solutions and AI infrastructure, rather than one isolated training program.
Agents could push consumption much higher. A conventional software request may call one service. An agent can plan, search, retrieve documents, call several models, use tools, check the answer and try again. One human request can generate many model operations.
Still, inference only supports the buildout when customers keep finding valuable reasons to run it. Millions of cheap novelty queries will not cover hundreds of billions in infrastructure. Coding, advertising, search, customer service and business workflows have a better chance because their value can be measured against labor, sales or conversion rates.
We see enough real inference growth to extend the capacity cycle. Whether it becomes a mass business workload quickly enough to fill the campuses now being planned is still the central bet.
If you want more recent data on this point, please see our latest GPU cloud market report.
Will cheaper AI models cancel the need for more capacity?
Cheaper models will lower the compute needed for each task, but total capacity demand can still rise when usage grows faster than efficiency.
The industry is improving on several fronts at once. New accelerators deliver more output per watt. Quantization, caching, speculative decoding, distillation and mixture-of-experts designs reduce serving costs. Smaller models now handle tasks that previously required the largest systems. Cloud operators are also getting better at keeping expensive hardware busy.
These gains pull in opposite directions. A company can run the same workload with fewer chips, reducing demand. Lower prices can also make AI economical in products that were previously too expensive, increasing the number of workloads.
The numbers currently lean toward the second effect. Azure, AWS, Google Cloud and Oracle Infrastructure have all accelerated despite rapid improvements in price-performance. Amazon's custom-chip business has passed a $20 billion revenue run rate while growing at triple-digit rates. Microsoft says AI product usage is rising quickly enough to pressure gross margins even as Azure becomes more efficient.
That relationship can change. Suppose the cost of a useful AI task falls tenfold while the number of paid tasks grows only threefold. Total compute demand drops sharply. The buildout requires usage to keep outrunning those efficiency gains for several years.
Efficiency is the biggest long-term threat to the capacity forecast, especially for projects designed around today's revenue per GPU. So far, it has enlarged the market faster than it has reduced electricity and hardware needs.
Can today's expensive AI chips stay valuable long enough?
Today's AI chips will remain useful for years, although many may lose their premium pricing well before their accounting life ends.
Microsoft says about two-thirds of its latest quarterly capital spending went into shorter-lived assets such as GPUs and CPUs. CoreWeave generally depreciates technology equipment over six years, while major platforms often assume several years of useful life for servers and networking gear.
Physical usefulness can last that long. Older accelerators can move from frontier training into fine-tuning, batch inference, simulation and less urgent workloads. Buildings, cooling, fiber and power connections can support several chip generations.
Commercial value moves faster. Customers pay the most for the newest chips, the densest clusters and the best networking. When a more efficient generation arrives, the older fleet usually needs a lower rental price to stay full. Electricity and interest do not fall nearly as fast, so that fleet can stay busy while its returns collapse.
A cash-rich hyperscaler can move older chips into internal products or cheaper cloud tiers. A leveraged specialist may depend on the original premium price to cover debt.
The risk is already visible in the release cadence. Amazon is deploying Trainium3 while describing Trainium4 for the following year, and NVIDIA's architecture cycle keeps compressing the time available to earn top-tier rates. Most current GPUs will find work. Some owners will discover that useful hardware can still be an overpriced asset.
Which AI clouds are taking the biggest overcapacity risk?
Specialist AI clouds and highly leveraged data-center developers face much greater overcapacity risk than Amazon, Microsoft, Alphabet or Meta.
The hyperscalers fund construction from large profitable businesses. They also have many places to use spare compute. Google can allocate capacity across Search, Gemini, YouTube and Cloud. Microsoft can use it for Azure, GitHub, Microsoft 365 and internal research. Meta can improve advertising and recommendations even without renting the chips to an outside customer.
Specialist clouds need external customers to keep paying. Debt, leases and equipment financing continue whether a deployment is late or a renewal price falls. Their smaller customer base also gives one contract much more influence over the whole business.
Custom chips widen the gap. Amazon, Google and Microsoft can reduce their dependence on the most expensive merchant GPUs and tune hardware for their own workloads. A neocloud built mainly around NVIDIA accelerators has less control over equipment cost and faces more direct price comparison.
The hyperscalers can still waste extraordinary amounts of shareholder money. Internal workloads may make poor infrastructure returns harder to see, and the strategic need to stay competitive weakens spending discipline. Their likely failure mode is lower free cash flow and weaker returns on capital.
A specialist operator can face a refinancing crisis before AI demand reaches its long-term potential. That is where the first serious shakeout is likely to begin.
If you want more recent data on this point, please see our latest GPU cloud market report.
Is CoreWeave proving the demand or exposing the danger?
CoreWeave proves that customers are signing enormous AI compute contracts, while its finances show how quickly sold capacity can turn into a balance-sheet test.
The operating growth is real. Quarterly revenue reached $2.08 billion, more than double the previous year, and backlog reached $99.4 billion. CoreWeave has passed one gigawatt of active power, with more than 3.5 gigawatts contracted. Committed contracts generated 98% of its latest quarterly revenue.
The cost of that growth is just as striking. CoreWeave reported a $144 million operating loss, $536 million of net interest expense and a $740 million net loss in the quarter. It spent $7.7 billion on property and equipment during those three months, compared with $1.4 billion a year earlier.
Its debt stack has expanded with the fleet. At the end of the quarter, CoreWeave had $11.8 billion outstanding under delayed-draw term loans and another $6.4 billion of notes, before counting large lease obligations. Executed leases that had not yet started carried an estimated $40.7 billion of future undiscounted payments.
Customer concentration has improved but remains high. Its largest customer produced 45% of quarterly revenue and the second largest 20%. Several major customers are also building their own clouds or custom chips.
CoreWeave can make this model work because many loans are tied to contracted cash flows from strong counterparties. The fragility appears when capacity is delayed, customers renegotiate, or newer hardware pushes renewal prices down. Interest expense arrives on schedule even when the data center does not.
Are the same few AI companies supporting too many data centers?
A small group of model developers and technology companies still supports a large share of the biggest AI infrastructure commitments, which makes the market less diversified than the headline numbers suggest.
OpenAI, Anthropic, Meta and a few other large buyers appear across multiple cloud, chip and data-center agreements. One model company may reserve capacity from several providers while each provider counts the contract in its own backlog. The contracts remain legally separate, but their ability to pay can trace back to the same API, subscription or enterprise customers.
The chain can become long. A chipmaker sells accelerators to a cloud. The cloud finances a data center around a model developer's contract. The developer expects future product revenue from enterprises and consumers. Energy suppliers and landlords add more commitments beneath the same expected demand.
This concentration becomes less dangerous as inference spreads across thousands of companies. Google says most of its $514 billion backlog comes from typical GCP contracts across a broad customer mix. AWS serves enterprises through Bedrock and its wider cloud platform. Microsoft's AI revenue now spans Azure, GitHub and productivity software.
Yet the largest new campuses are often justified by a handful of exceptional contracts. CoreWeave's largest two customers produced 65% of its latest quarterly revenue. Oracle's backlog has grown by hundreds of billions largely because of AI infrastructure agreements.
The market is much broader than one laboratory, but the biggest projects still depend on too few independent sources of cash.
Where will AI cloud overcapacity appear first?
AI cloud overcapacity will probably appear first in older chips, poorly connected sites and heavily financed projects, with no single global glut.
Location matters. Frontier training can use remote sites when power and networking are excellent. Real-time inference often needs lower latency, local data handling or a particular cloud region. A megawatt in a remote market cannot always replace a megawatt near major customers.
Cluster design matters even more. Ten thousand accelerators spread across small facilities have different value from ten thousand accelerators connected inside one high-speed training system. A data center with power but delayed networking may look finished while remaining commercially weak.
Hardware generations create another split. Customers can compete for the newest systems while older GPUs become widely available at lower prices. The same facility may move from premium training to cheaper inference and still stay physically busy.
Reuse will soften the damage. Older chips can serve less demanding workloads, and data-center shells can be refitted. Hyperscalers have the broadest options because they control both infrastructure and applications. Specialist operators may have to cut prices to find replacement customers.
The first glut should look like a patchwork. One region may have a queue for current-generation clusters while another discounts older machines. Some buildings will remain valuable after their original customer leaves, while the first owner still loses money because refinancing and upgrades cost too much.
What would prove that AI clouds have finally overbuilt?
We should call the AI cloud market overbuilt when availability, pricing, contracts and financial returns weaken together for several quarters.
The first clear change would be broad availability of current-generation large clusters. A few discounts on older GPUs would be noise; waiting times would need to disappear across Microsoft, Amazon, Google, Oracle and the major specialist clouds.
Pricing would then need to fall faster than hardware efficiency improves. Lower prices are normal when each new chip can do more work. They become dangerous when an owner earns less per dollar invested after adjusting for that better performance.
Contract evidence would provide the next test. Slower backlog growth, delayed customer drawdowns, renegotiated commitments and much lower renewal rates would show that reserved capacity exceeded durable usage. We would also expect more projects to be cancelled after securing power and financing, rather than merely postponed before construction.
The financial proof would appear in rising depreciation, energy and interest costs alongside slower cloud revenue. Specialist clouds would struggle to refinance, sell assets or issue shares repeatedly. Hyperscalers would report weaker returns on invested capital even if their broader businesses remained profitable.
That full pattern is absent today. Margin pressure, heavy financing and uncertainty around future projects are already appearing around the edges.
| Test | Evidence of real overbuilding | Current reading |
|---|---|---|
| Availability | Modern large clusters broadly available without queues | Leading providers still report constraints |
| Pricing | Repeated cuts beyond efficiency gains | Pressure exists, broad collapse does not |
| Contracts | Delays, renegotiations and weak renewals across providers | Backlogs are still expanding |
| Construction | Firm, financed projects cancelled at scale | Speculative projects are being filtered first |
| Returns | Revenue slows while depreciation, power and interest keep rising | Margin pressure is visible, cloud growth remains strong |
| Financing | Forced asset sales and failed refinancing | Risk is concentrated in leveraged specialists |
Are AI clouds building too much capacity?
AI clouds are probably committing too much capacity in parts of the market, while top-tier compute remains tight today.
The newest results have made the near-term answer clearer. Google Cloud grew 82%, raised its backlog to $514 billion and is renting third-party capacity while it waits for more of its own. Microsoft still says broad demand exceeds supply. AWS grew at its fastest rate in 15 quarters, and Oracle's infrastructure business nearly doubled. These figures come from large operating businesses with paying customers, which gives them far more weight than developers' capacity forecasts.
The financial warning has also become sharper. The biggest builders are spending roughly $650 billion-$675 billion a year, free cash flow is under pressure, cloud margins are absorbing infrastructure costs and specialist operators are borrowing heavily. CoreWeave's latest quarter combined more than $2 billion of revenue with over half a billion dollars of interest expense. Demand can be genuine while the investment returns disappoint.
The most likely outcome is a selective shakeout. Current-generation capacity in the right regions should remain tight in the near term. Older hardware, weak locations and projects tied to one customer will face falling prices first. Some specialist clouds and developers will be forced to refinance, merge or sell assets even as total AI usage continues growing.
Most of the physical infrastructure should find a use eventually. Many owners will still fail to recover what they spent.
So yes, parts of the industry are going too far. AI clouds are building ahead of proven long-term demand, but the market as a whole does not yet have too much top-tier capacity. The weakest projects already depend on a future that looks too generous.
If you want more recent data on this point, please see our latest GPU cloud market report.
OUR METHODOLOGY
This analysis tests whether AI clouds are building more capacity than the market can use profitably. We separate the capacity available today from projects scheduled for later, because current shortages and future overbuilding can exist at the same time.
We examine the question through the factors that determine whether new infrastructure can create durable economic value: current supply conditions, capital expenditure, cloud revenue growth, contractual demand, power availability, hardware depreciation, financing exposure, efficiency gains and customer concentration.
We prioritize reported operating results, regulatory filings, earnings commentary, active power and signed commitments over distant forecasts or headline project announcements. Capital expenditure, backlog, contracted power and operational capacity are kept separate because they measure different stages of the buildout.
We test the evidence in both directions. Supply constraints, accelerating cloud revenue and expanding contracts support continued construction. Falling free cash flow, margin pressure, rapid hardware turnover, concentrated customers and heavy financing raise the risk that some builders will never earn an adequate return.
We do not treat temporary spare capacity at a newly opened site as proof of overbuilding. The stronger test is whether revenue can cover equipment, buildings, electricity, cooling, maintenance and financing over the asset's useful life while still producing a worthwhile return.
Backlog figures are treated as evidence of contracted visibility, not as directly comparable measures of independent AI demand. The definitions differ across providers, contracts run for different periods, and the same end customer may support commitments at several clouds.
Power pipelines are also discounted according to their stage. Active power can support workloads now; contracted power still depends on construction, equipment, grid delivery, customer schedules and financing. Announced projects without firm service or construction commitments receive less weight.
Key company sources include Amazon's Q1 2026 results, Amazon's shareholder letter on 2026 capital expenditure, Microsoft's FY2026 Q3 earnings call, Microsoft's FY2026 Q3 results, Alphabet's earnings materials, Meta's Q1 2026 results, Oracle's FY2026 results, and CoreWeave's Q1 2026 results.
Key power and infrastructure sources include Lawrence Berkeley National Laboratory's 2025 US data-center electricity update, EPRI's 2026 data-center electricity-demand scenarios, PJM's load-forecast development materials, and PJM Load Analysis Subcommittee materials on data-center forecasting.
Related blog posts
- Are we already building too many AI data centers?
- Who will be the next AWS, but for AI?
