Everything About AI
NextFin News - Gary Marcus’s warning that the industry is “probably overbuilding data centers” lands at the most uncomfortable point in the AI boom: demand is real, but the timing of supply may be wrong. Cloud revenue and AI usage are accelerating, while utilities, developers and hyperscalers are committing to power and buildings before the economics of every workload are clear. The likely risk is not that artificial intelligence needs no infrastructure. It is that too much infrastructure arrives at once, in the wrong places, for the wrong generation of hardware.
That distinction matters because an overbuild can damage returns without leaving empty buildings everywhere. A facility can be leased and still earn poor returns if power costs, depreciation, financing expense or obsolete accelerators absorb the revenue. The first-order story is demand. The second-order story is utilization and capital recovery.
The Boom Has Real Demand Behind It
The strongest argument against Marcus’s view is visible in operating results. Amazon said AWS grew 39.3% year over year in the third quarter of 2026, its sixth consecutive quarter of acceleration. Amazon also said Fortune 500 spending on Bedrock was nearly seven times higher than a year earlier, while total Bedrock token volume rose nearly 17 times. Microsoft said its cloud business surpassed $50 billion in quarterly revenue, up 26% year over year, in its fiscal second quarter. Microsoft’s figure covers the broader cloud business, not AI alone, but it still shows that the infrastructure platform is expanding rather than contracting.
Alphabet’s third-quarter management commentary likewise described strong demand for AI infrastructure and AI solutions as a driver of cloud growth. Taken together, these disclosures do not look like a demand collapse. They show that customers are paying for cloud capacity and that AI usage is moving beyond demonstrations.
The International Energy Agency’s updated analysis points in the same direction. Global data-center electricity consumption is projected to rise from about 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030, while electricity use at AI-focused data centers is projected to triple over the same period. The U.S. Department of Energy’s 2025 update, based on Lawrence Berkeley National Laboratory analysis, puts reference-case U.S. data-center electricity use at 649 terawatt-hours in 2030, equal to 11.8% of forecast national electricity consumption. The same update puts 2028 use at 464 terawatt-hours.
The industry is not building for a static market. But a growing market can still be overbuilt. Demand can expand while capacity expands faster. That produces growth and excess simultaneously. It also creates a distinction between contracted capacity and profitable capacity: a customer may reserve power, a developer may announce a campus, and a cloud operator may buy accelerators before the final workload has a durable economic return.
That is the central question behind Marcus’s claim: not whether AI needs compute, but whether the industry can convert the next wave of fixed investment into revenue before the technology and economics change again.
The Mechanism Is a Timing Mismatch, Not a Demand Mirage
The most defensible reading is cyclical: the immediate overbuild risk comes from a timing mismatch that should eventually mean-revert, even if the long-term demand trend remains structural. Data centers require lengthy permitting, power interconnection and construction, while AI software demand can change much faster. Capital therefore responds to a shortage that may already be easing by the time a project reaches operation.
Power makes the lag especially severe. The IEA says more than 2,500 gigawatts of projects, including large-load projects such as data centers, remain stalled in grid-connection queues worldwide. The bottleneck encourages early commitments and aggressive capacity planning. When power becomes available in clusters, several projects can reach completion close together. The result can be scarcity first and oversupply later.
Hardware depreciation adds a second clock. A facility may last decades, but the accelerators inside it face faster performance cycles. If newer chips deliver more useful output per watt and per dollar, customers can satisfy the same workload with fewer older machines. The building remains an asset; the revenue attached to its original hardware may not. This is why the risk is not captured by vacancy alone.
Efficiency strengthens both sides of the argument. The IEA says power consumption per AI task is declining rapidly, even as total demand rises because usage expands and more energy-intensive agentic workloads appear. Falling cost per task can increase adoption, but it can also reduce the amount of hardware required for a fixed volume of work. The industry can therefore experience a rebound in total usage while individual facilities face pricing pressure.
“You have to bet big when you find major inflection points.” — Andy Jassy, Amazon chief executive, in an official Amazon interview about the company’s AI infrastructure investment approach.
Jassy’s point captures the rational behavior that can create irrational aggregate supply. Each hyperscaler has a reason to secure capacity ahead of rivals. A cloud provider that waits for confirmed demand may lose customers to a competitor with available GPUs. Yet when providers make the same defensive decision, the industry’s combined build-out can exceed the amount customers ultimately use at attractive prices.
This is a coordination problem. The individual investment decision can be sensible while the collective outcome is poor. The closest analogy is a commodity cycle: producers invest when prices signal shortage, projects arrive with a lag, and returns fall when supply catches up. The difference is that AI hardware can also improve rapidly, making old capacity economically less competitive even if physical demand continues to rise.
Why the Market Can Miss the Second-Order Risk
The consensus narrative treats revenue growth as proof that capacity spending is safe. That is too simple. The relevant test is not whether cloud revenue rises; it is whether revenue grows faster than the full cost of serving it, including depreciation, electricity, networking, financing and idle capacity.
Amazon’s own commentary illustrates the pressure. While AWS growth accelerated, the company also highlighted heightened component costs and significant depreciation from the capital-expenditure cycle. That combination matters. Revenue can rise quickly while free cash flow and returns on invested capital weaken. If the industry responds to strong demand by committing to still more capacity, the first-order effect is additional supply. The second-order effect is lower utilization or lower pricing when that supply arrives. The third-order effect is that investors value infrastructure companies on cash conversion rather than announced megawatts.
Power can transmit the shock across sectors. A data-center boom lifts demand for electricity generation, transmission equipment, cooling systems and construction. But if projects are delayed or canceled, utilities can be left with infrastructure plans based on load forecasts that no longer materialize on schedule. Ratepayers and regulators then become part of the adjustment. The debate shifts from whether the grid can serve AI to who pays for capacity built before AI demand is firm.
Financing creates another cross-market channel. Data-center projects depend on long-lived assets and often require debt, private credit or joint ventures. If utilization disappoints, the damage does not stop at a developer’s earnings report. Higher risk premiums can raise the cost of capital for the wider infrastructure chain, including power projects and equipment suppliers. A demand miss can become a larger valuation event because the sector’s fixed costs are high.
The market may already understand that AI infrastructure is capital intensive. The less-priced risk is the expectation gap around duration. Investors may be modeling years of high utilization, while the operational reality is a series of shorter cycles: one chip generation, one model architecture, one customer contract, one power-delivery date. The more the industry finances long-lived assets against short-lived technology assumptions, the more sensitive returns become to delays.
That is why a capacity glut can appear without a dramatic collapse in AI usage. If the industry’s output per machine rises, customers may obtain more inference and training for each dollar even as total AI activity increases. Revenue growth can remain positive, but pricing per unit of compute falls. Suppliers with the strongest balance sheets and most flexible facilities benefit; highly leveraged operators with older equipment face the adjustment.
The Strongest Counter-Thesis Is That Shortage Will Outrun Supply
The serious case against the overbuild thesis is not promotional optimism. It is that demand is still so far ahead of available power and compute that new supply will be absorbed before it becomes excessive. The IEA’s projections support that possibility. Data-center electricity use is expected to nearly double globally by 2030, and U.S. electricity demand through 2030 is expected to grow nearly 2% annually, with around half of the increase driven by data centers. Amazon’s AWS results show that demand remained strong as the company added capacity.
There is also a strategic reason hyperscalers may rationally over-order. AI is becoming a platform layer across search, cloud, software, advertising and enterprise workflows. If demand compounds faster than expected, a provider that underbuilds may lose customers, developer ecosystems and model leadership. The option value of capacity can be high. A reserved site or power contract may be worth more than its immediate accounting return because it protects a strategic position.
This counter-thesis also challenges the assumption that falling compute costs destroy infrastructure economics. Lower prices can expand the market. A business that cannot afford a large language model today may use one when inference becomes cheaper. Agentic systems may run more tasks, more frequently, and across more corporate processes. The result could be a rebound in total demand large enough to absorb today’s projects.
The evidence is meaningful, but it does not eliminate the overbuild risk. It changes its location. The likely excess would be concentrated in projects that lack firm customers, depend on uncertain grid connections, use less competitive hardware or assume pricing that falls faster than utilization rises. A shortage at premium, powered sites can coexist with excess at speculative campuses.
Marcus would be wrong in the stronger sense if the industry’s utilization rate keeps rising despite the next wave of delivered capacity. The cleanest falsifying signal is operational: if major cloud providers report AI infrastructure revenue growth above 30% year over year for four consecutive quarters after the current construction pipeline comes online, while depreciation and power costs grow more slowly than revenue, the overbuild thesis would lose force. A second signal would be persistent scarcity pricing for powered capacity rather than falling rental rates.
The counter-thesis therefore wins on demand but not necessarily on returns. Capacity is not the same thing as profitable capacity.
What the Debate Means for Assets and Industries
The short-term beneficiaries are the suppliers that sell into the build before the final AI economics are tested. Power equipment, cooling, networking and construction can receive orders during the investment phase. Their revenue depends on projects being funded and built, not on whether every future token generates an attractive margin. This creates a timing asymmetry: suppliers may be paid while data-center owners are still waiting for utilization.
Hyperscalers occupy a different position. Their scale and diversified cash flows let them absorb a period of weaker returns, and their own cloud platforms capture more of the value chain. But they also carry the largest aggregate capital burden. Amazon’s plan to invest up to $50 billion in government AI and high-performance computing infrastructure, designed to add nearly 1.3 gigawatts of capacity, shows how quickly strategic commitments can become balance-sheet commitments.
Colocation companies and specialized data-center developers face more direct sensitivity to occupancy, contract quality and financing. A long-term contract with an investment-grade customer can protect cash flow, but headline capacity agreements may still include expansion options, staged delivery or conditions tied to power availability. The market will increasingly distinguish binding demand from announced demand.
Utilities face a two-sided exposure. In the short run, data centers can support load growth and justify investment in generation and transmission. In the medium term, a delayed or reduced build-out can leave utilities with underused assets or political pressure over tariffs. The IEA’s grid-queue data shows why this risk is structural at the planning level even if individual projects are cyclical.
Chip suppliers are exposed to a different version of the cycle. A new accelerator generation can stimulate another spending wave, but it can also accelerate the obsolescence of installed equipment. The winners will be those with continuing demand across training, inference and general-purpose workloads, not merely those benefiting from one construction peak.
AI adoption can remain strong while the value of a particular data center falls. That is the second-order conclusion: the secular trend can be right and the capital allocation can still be wrong.
Three Horizons for the Next Data-Center Cycle
Over the short term, sentiment and liquidity will dominate. Announcements of large power contracts, new campuses or strategic AI partnerships can sustain investment momentum. The warning sign would be a shift from securing capacity to renegotiating delivery dates, combined with rising financing costs or weaker pricing for unpowered sites. In that scenario, equipment suppliers may still report healthy orders while developers begin to cut projects.
Over the medium term, fundamentals will turn on utilization, price per unit of compute and depreciation. The base case is a selective oversupply: premium sites remain scarce, but marginal projects are delayed, repriced or repurposed. The upside case is that agentic workloads and cheaper inference cause demand to outpace delivered capacity, keeping utilization high and supporting returns. The downside case is that model efficiency improves faster than end-user monetization, causing prices to fall before the new supply reaches steady utilization.
Over the long term, the structural thesis remains intact but becomes more capital selective. Electricity consumption from data centers is projected to rise materially through 2030, and AI applications are likely to expand beyond today’s workloads. Yet the industry’s physical footprint may not grow in a straight line. Better chips, more efficient software, liquid cooling, workload scheduling and distributed computing can all increase output without requiring a proportional increase in buildings.
The most important metrics are therefore not announcements alone. They are contracted and delivered megawatts, utilization by hardware generation, cloud AI revenue relative to depreciation, inference prices, and the share of projects that reach power-on within their original schedules. If those measures improve together, the industry is growing into its capacity. If revenue rises while utilization, pricing and cash conversion weaken, Marcus’s warning will look less like a contrarian opinion and more like a balance-sheet diagnosis.
Data cutoff: 2026-11-11 UTC.
The AI infrastructure boom is structurally real, but its next leg is cyclical: the industry can run out of power and still build too much of it. The winners will be determined less by who announces the most megawatts than by who converts each delivered megawatt into durable cash flow.
Explore more exclusive insights at nextfin.ai.
