Is there an AI bubble? The global race to build data centers

In a few years, artificial intelligence has shifted from a software feature to a race to build facilities, buy accelerators and secure electricity. Major technology companies are announcing budgets measured in hundreds of billions of dollars; around them, specialised data centres, networks, substations and power-generation projects are growing.

The scale raises a fair question: is capacity being built too quickly for the revenue AI can generate? But stating that a bubble exists as fact would require knowing how much of those facilities will be occupied, which customers will pay for them, and what it will cost to operate them for years. Those data are incomplete and not reported on a comparable basis. The honest answer remains open.

Why are people talking about an AI bubble?

A financial bubble is often associated with prices or investment that depend on increasingly optimistic expectations while observable demand or revenue falls short of supporting them. In AI, the discussion is less about a single valuation and more about a chain of connected bets: chipmakers, cloud providers, model labs, data-center operators and companies financing construction.

The parallel does not prove that a crash will follow. A technology can be useful and reshape industries while particular projects, companies or investments fail. The economic question is more specific: how much capacity is needed, where, and what recurring revenue will pay for the hardware, buildings, power and financing?

How much are the big technology companies spending?

There is no audited, comparable figure for “global AI spending” that captures the whole phenomenon. The International Energy Agency's (IEA) broadest figure says capital expenditure by five major technology companies exceeded USD 400 billion in 2025 and estimates it will rise another 75% in 2026. That is combined company spending, not an account exclusively for AI; the 2026 figure is an estimate, not a final result. (Source: IEA, *Key Questions on Energy and AI*, executive summary (April 2026).)

• On its Q2 2026 earnings call, Amazon raised its 2026 cash capex outlook to about USD 220 billion. It covers AI, chips, robotics, satellites and other areas; it is not an AI-only or data-center-only figure.
• Alphabet raised its calendar-year 2026 capex guidance in July to USD 195–205 billion. This is a company-wide figure with no exact AI-only split. It reported USD 44.9 billion of quarterly capex in Q2.
• Meta estimated 2026 capex of USD 130–145 billion, including principal payments on finance leases. It reported USD 31.08 billion under that definition in Q2. This is not an AI-only line item either.
• Microsoft reported USD 41 billion of capex in fiscal Q4 2026; roughly two-thirds was short-lived assets, primarily CPUs and GPUs for both AI and non-AI workloads. Cash purchases of property and equipment were USD 35.8 billion. For calendar 2026, it put reported capex expectations at about USD 175 billion after shifting some future data-centre leases from finance to operating leases; it said its overall investment expectations were unchanged. This is company-wide, not AI-only.
• Oracle reported USD 28.499 billion of capex in the first quarter of fiscal 2027, which is not a full calendar year. Revenue that quarter was USD 19.3 billion and cloud revenue USD 11.6 billion. Its USD 664 billion RPO is the value of remaining performance obligations under contracts: it is not cash collected, quarterly revenue or capex.

These figures should not simply be added: they mix fiscal and calendar periods, quarterly results and annual guidance, cash payments and leases, and spending on businesses beyond AI. They show the scale and direction of spending, not a single pool of AI investment. The IEA presents its figure as total capital spending by those companies, without breaking out how much goes to data centres or AI, and public disclosures do not precisely separate all AI spending either.

Why does AI need so much infrastructure?

Training an advanced model and answering millions of requests takes many calculations in parallel. GPUs and other accelerators perform that work; a data center groups them into clusters linked by high-speed networks. Memory to feed the processors, storage for data and models, backup systems and software to distribute the workload are also needed.

The physical facility is more than a server room. Buildings must be built or adapted, supplied with electricity, fitted with transformers and substations, connected to the grid, and equipped to remove heat from the hardware. The internal network matters too: if accelerators cannot exchange data quickly enough, some purchased capacity is poorly used. A delay in any one of these parts can leave the rest expensive and idle.

Demand also continues after a model is trained. Inference—producing answers whenever a person or application queries a model—repeats in daily use. The IEA observes growing use of more intensive tasks, such as video, reasoning and agents that chain actions together. At the same time, models and hardware are becoming more efficient: according to the IEA, energy per AI task has fallen by at least an order of magnitude annually in recent years. Total demand depends on which trend moves faster. (Source: IEA, *Key Questions on Energy and AI*, executive summary (April 2026).)

Electricity is becoming a bottleneck

The best available distinction is between all data centers and those dedicated specifically to AI. The IEA estimates that data centers worldwide used 485 TWh of electricity in 2025. Electricity use at AI-focused centers rose 50% that year, compared with 17% growth for all data centers. These are estimates of 2025 consumption, not 2026 figures or a share of all global electricity attributable to AI alone. (Source: IEA, *Key Questions on Energy and AI*, executive summary (April 2026).)

In its central scenario, the IEA projects that electricity use by all data centers will nearly double to about 950 TWh in 2030, close to 3% of global electricity demand. It also projects that consumption at AI-focused centers will grow much faster. A projection depends on construction, efficiency, which uses become popular and which connections are secured; it is not an outcome that has already happened. (Source: IEA, *Key Questions on Energy and AI*, executive summary (April 2026).)

The challenge is not just generating electricity. More than 2,500 GW of projects are in grid-connection queues worldwide, according to the IEA, but that figure combines renewable generation, storage and large electricity users, including data centers. The agency estimates that planning and building new grids can take 5–15 years, compared with 1–3 years for a data center. Substations, transformers and lines can take longer than the buildings waiting to connect. (Source: IEA, *Electricity 2026*, grids chapter.)

The energy mix also varies by region. For global electricity-demand growth between 2026 and 2030, the IEA expects renewables, gas and nuclear together to cover the entire increase. That is a projection for the global power system, not a recipe for each facility. Solar and wind need grids, storage and flexible generation; some companies are exploring nuclear contracts or on-site generation, while local gas is proposed in areas with slow connections. No single technology solves every cost, availability, emissions and timing problem. (Source: IEA, *Electricity 2026*, executive summary.)

Cooling is another important physical component, although design and climate matter. We do not include a global water figure: available sources do not allow sites to be compared rigorously. Water withdrawal and water consumption—the portion that does not return to the basin in the same period—are also different measures. A figure from one facility should not become a worldwide claim. For a closer look at how water linked to AI infrastructure is measured, see our explainer on AI water consumption.

How do companies expect to earn these investments back?

Infrastructure is sustainable only if someone pays to use it. Cloud providers expect to sell training and inference capacity to businesses and labs; manufacturers sell accelerators, networking systems and software; model companies charge for subscriptions, APIs and services; and adopting businesses hope to lower costs or sell more. Agents could increase usage if they become part of work processes, but that still has to turn into lasting revenue.

There are signs of commercial demand, though they do not answer the profitability question by themselves. Alphabet said Google Cloud generated USD 24.8 billion in Q2 2026, up 82% year on year, with growth partly attributed to AI infrastructure and solutions. NVIDIA reported USD 89 billion in data-center revenue in its second fiscal quarter of 2027. These are revenues from particular segments; they do not reveal each AI customer's final profit or what share of cloud is due to AI alone.

The transparency gap matters. Amazon, Alphabet, Microsoft and Meta do not publish a full account isolating revenue and profit from each AI data center. A company may be selling capacity and still face chip, electricity, depreciation, staffing and interest costs that reduce its return. Revenue growth, a large contract or an order backlog therefore do not automatically equal a profitable investment.

Why some people see bubble risk

• Annual spending is rising quickly, and multiyear projects are announced before all capacity is built or operational. If future demand is revised down, some assets could be underused.
• Companies need recurring revenue to cover electricity, cooling, operations, connectivity and financing as well as construction. An AI model's popularity or technical power alone does not justify the investment.
• Accelerators are expensive and age faster than buildings. Microsoft said roughly two-thirds of the cited quarterly capex was short-lived assets, mainly CPUs and GPUs. Obsolescence and depreciation could pressure margins if capacity is not monetised in time.
• Some growth relies on outside financing and long contracts. Oracle issued shares to fund its investment programme; CoreWeave reported USD 35.6 billion of debt as of 30 June 2026. Those examples show different structures and do not describe every participant equally.
• Order or commitment figures can look more certain than they are. RPO includes contracted revenue for future periods and can be subject to cancellation or conditions; announced capacity still needs power, permits, suppliers and customers before it becomes a service.

These risks are more visible at young or highly leveraged operators, but they do not prove that large platforms will lose money on their investments. Nor does the fact that a profitable company can finance a project make the capital free: it has a cost and competes with other uses.

Why demand may justify the buildout

Current demand is not hypothetical in every segment. The IEA estimates that capacity at facilities dedicated to AI more than tripled in the 18 months before its April 2026 report; electricity use at those centers rose 50% in 2025. The agency also reported that major model providers said their active users tripled and revenue quintupled over the past year, although there is no complete global usage statistic. (Source: IEA, *Key Questions on Energy and AI*, executive summary (April 2026).)

• Traditional cloud is growing alongside AI. Not every new server room is an “AI factory”; storage, databases, commerce and cloud software workloads also support infrastructure.
• Some regions face power limits and pending connections. Local scarcity can coexist with excess capacity or cancelled projects in other markets.
• Training models is only one workload. Inference, agents, video and other tasks may keep facilities busy after they are built. Profitability will depend on customers accepting the price and finding measurable value in the service.
• Components do not all become obsolete at the same time. An accelerator may depreciate quickly, while buildings, substations, networks, cooling and grid connections could serve new chip generations or other computing loads. Reuse does not guarantee recovery of the original cost, but it may limit physical waste.

NVIDIA, for example, reported strong growth in data-center revenue; Oracle said it had delivered 850 MW of additional capacity and booked more AI cloud services. These facts confirm commercial activity and capacity under construction, not that every future project will reach its target occupancy or return. The same evidence can support a reading of real demand or of investment running ahead.

Is this like the dot-com bubble?

There are similarities: enthusiasm for a technology with potential, high growth expectations and infrastructure investment before every use is clear. During the telecom boom of the late 1990s, fiber networks were laid with capacity that went unused. An FCC report from 2000 already described significant amounts of high-capacity dark fiber. That history is a reminder that demand can take time to catch up with supply.

Fiber also shows how financial outcomes and physical value can diverge: an overbuilt network may remain useful as traffic rises, even if the original operator or its investors go bankrupt or lose money. It does not follow that all fiber was recovered or that every AI facility will find customers. An asset's future use is different from the return earned by the people who financed it.

There are also important differences from 1999. Many companies financing infrastructure today have established, profitable businesses in advertising, software, commerce and cloud; customers already pay for models and capacity, and physical centers can host different equipment over time. But AI chips do not last like a fiber trench, the energy scale is different, and server operating costs continue after construction. We are not automatically repeating 1999: risks and fundamentals must be measured separately.

What signals should we watch?

No single indicator settles the debate. Compare spending, connected capacity, utilisation, revenue and costs over several quarters, and distinguish a local problem from a global trend.

• Signals supporting an overinvestment thesis: completed facilities with low utilisation; delayed, cancelled or unfunded projects; customers that do not renew; spending that persistently outpaces related revenue; and margin pressure from depreciation, electricity or interest.
• It would also be a warning if contracts or letters of intent were presented as cash already collected, or if debt-dependent companies could not refinance on manageable terms.
• Signals in the other direction: sustained cluster utilisation; AI capacity entering service and being sold; growing inference and cloud demand from diverse customers; and evidence that businesses repeatedly pay for functions that improve their work.
• It would help to see multiyear contracts convert into revenue and cash collections, alongside margins that cover energy, hardware, cooling and financing. A backlog alone does not prove that outcome.
• The power grid offers another test: approved and completed connections, available new generation and storage, and fewer stalled projects. A large queue of requests can reflect ambition, not capacity that will actually be built.
• Finally, reuse matters: if a facility can host a new accelerator generation or conventional cloud workloads, some of the building retains value. If equipment becomes obsolete, the grid connection does not arrive or there are no paying users, reuse may be limited.

So, is there an AI bubble?

As of 15 October 2026, the evidence shows a real infrastructure expansion, an exceptional rise in spending and energy bottlenecks; it also shows cloud growth and specific sales of services and accelerators. That is not enough to say all the investment is sustainable, or to declare the whole market a bubble. The answer may vary by region, company and layer of the supply chain: one market can lack capacity while another builds more than it can monetise.

The useful question for the coming years is not only how many data centers are announced, but how many connect, how much they are used and what recurring revenue supports their total costs. Early investment can leave valuable infrastructure; it can also destroy capital if demand or financing fails to arrive. For now, the verdict remains open.

Frequently asked questions

What would have to happen for AI to be a bubble?

There is no single threshold. Concern would rise if built capacity remained underused and revenue persistently failed to cover hardware, power, operating and financing costs. High spending alone does not prove that this will happen.

How much is being invested in AI data centers in 2026?

There is no single comparable figure for AI spending. The Amazon, Alphabet and Meta guidance cited here is company-wide capex; Microsoft reports quarterly figures with specific definitions, and Oracle reports by fiscal year. The IEA estimates that capex by five major technology companies will rise 75% in 2026, but that is not a final result or an AI-only total.

Is all data-center growth caused by AI?

No. Data centers also host traditional cloud services, databases, storage, commerce, video and enterprise software. When a source discusses all data centers, its electricity use or investment should not be presented as entirely AI-related.

Is this the same as the dot-com bubble?

There are parallels in expectations, investment and advance infrastructure, and late-1990s fiber left some capacity unused. But today's large technology companies already have profitable businesses and customers, and some physical assets can be reused. The comparison helps explain overcapacity risk; it does not predict the outcome.

How does this relate to the AI models we use?

Demand depends on the model, the type and volume of queries, and where it runs. To compare the tools competing for users and businesses, read our ChatGPT, Claude and Gemini comparison. That comparison does not by itself measure the return each provider earns on its infrastructure.