AI Is Short of Capacity, Not Demand
IDC ·
Talk of an AI correction usually starts with a change in investor sentiment. Coupled with rocky macroeconomic headwinds, this impacts company valuations, thereby straining financing for AI-specific buildouts. In short, it affects supply. IDC tracks markets, not balance sheets, so we start with the demand side, and it is strong. 95% of enterprises already run […] The post AI Is Short of Capacity, Not Demand appeared first on IDC .
Talk of an AI correction usually starts with a change in investor sentiment. Coupled with rocky macroeconomic headwinds, this impacts company valuations, thereby straining financing for AI-specific buildouts. In short, it affects supply. IDC tracks markets, not balance sheets, so we start with the demand side, and it is strong.
95% of enterprises already run AI agents in production with inferencing-specific actions, and IDC expects 2.5 billion agents by 2030.
The constraint is capacity. Power, memory, packaging, and site buildouts have not kept pace with demand for AI adoption; as a result, supply-side constraints will set the pace of growth through 2027 and into 2028. The IDC view is that this is not a demand bubble . A bubble is oversupply versus demand, and today, memory, power, and packaging are all in short supply.
The inferencing economy has arrived and is rapidly gaining traction, driven by agentic AI. We have witnessed agentic AI move from pilot programs into production. 95.3% of enterprises run at least one agent workflow in production, with an average of about 11, according to IDC’s Future Enterprise Resiliency and Spending (FERS) survey (Wave 4, July 2026) . By 2030, we forecast there will be 2.5 billion agents, with the number of active agents nearly tripling next year, according to IDC’s Worldwide Agentic AI Adoption Guide (June 2026). Enterprises also expect AI-enabled applications to account for over half of enterprise workloads within two years, according to IDC’s June 2026 Digital Infrastructure Sentiment Study .
Enterprises are also spending faster on AI initiatives than planned. In the past 12 months, 67% exceeded their agent budgets, and 24% overshot by 25% or more, according to the IDC FERS survey. We view these overruns as a demand signal. Even though organizations are consuming more tokens than they forecast, they continue to expand agent use across more workflows. Rather than reducing agent use to align with an existing budget, they are increasing budgets and pursuing technology enhancements, such as model routing and improved inference memory management, to reduce per-action costs.
A common industry statement is that falling token prices will flatten demand. But cheaper tokens do not mean cheaper agents.
IDC predicts that the cost per agent action will drop by about 95% from 2025 to 2030, while total token delivery costs will rise 487-fold.
We believe that Jevons paradox is real, and we are in the midst of experiencing it today.
Efficiency gains are real. Agents confirm context, chain steps, call tools, and pass messages to other agents. In parallel, agent-driven inference delivers two to four times more throughput per GPU. Model routing cuts average cost per token by 30% to 60%. These gains will lower unit token costs. They will not slow total consumption. Volume outpaces price for structural reasons. A single business outcome can consume tens or hundreds of thousands of tokens. History shows that sharply cutting the cost of a basic unit accelerates innovation, because teams can take on a wider range of workflows.
With strong AI demand for the foreseeable future, the question is whether enough AI infrastructure will be available to meet inference demand. Two primary groups buy that infrastructure: service providers and non-service providers (aka enterprises and governments). IDC’s AI Infrastructure Tracker projects revenue from service provider (including Hyperscalers and cloud service providers) procurement to grow from $401 billion in 2025 to $1.45 trillion annually in 2030. To keep up, datacenter facilities capex (excluding the cost of IT infrastructure) is forecast to rise from $276 billion to $655 billion over the same period. Increasingly, we see specialty cloud service providers (aka Neoclouds), frontier model makers, and extremely large hybrid enterprise/service providers emerging as a secondary source of infrastructure buyers.
Across the two principal buyer types, IDC is seeing a consistent trend toward distributed, purpose-built AI infrastructure, or “AI factories,” that are optimized for AI workloads, with specific components optimized for complex training, inference, and agent environments. Tech leaders view AI factories as a way to operationalize AI workloads by turning data into tokens and actionable intelligence at predictable costs. IDC’s AI Strategies survey shows that they prioritize security and governance controls, as well as built-in operational intelligence, over rapid deployment. These specialized systems can be deployed in various locations, including at the edge, in the datacenter, in the cloud, or on service provider infrastructure, enabling flexible deployment that scales with AI use cases.
As agentic AI and AI inference workloads expand, especially in mainstream enterprises, capacity challenges are about more than just the strain on total available compute. They also raise the question of whether the right infrastructure can be deployed where needed, with sufficient power, cooling, and connectivity, at an economically viable cost. Shortages in these areas can coexist with underused capacity elsewhere. Two constraints stand out.
The first constraint is physical capacity, i.e., datacenter facilities. Our data shows that power, grid access, and datacenter capacity could limit near-term growth. IDC’s Datacenter Facilities Tracker shows that the number of provisioned racks will rise from 25 million to 58 million from 2026 to 2030, leading to a corresponding increase in energy consumption from 519 TWh in 2025 to 1,308 TWh. This represents an increase of about 2.5-fold and 2.3-fold, respectively (while AI infrastructure spending rises 3.35-fold, leading to a supply shortage). We find that just in the U.S., almost 7 GW of planned datacenter capacity for 2026 has been delayed or canceled due to a lack of power. U.S. colocation vacancy is at record lows, and IDC projects a 45 GW power shortfall in the U.S. by 2028.
The second constraint is the semiconductor and enabling technologies supply chain. Take the global memory shortage, for example. IDC describes it as structural because demand for AI has shifted fab capacity toward high-bandwidth memory (HBM). IDC forecasts supply growth of just 16% for DRAM and 17% for NAND in 2026, well below historical norms. HBM4 output is not expected to meet demand until mid-2027.
Another example is TSMC’s wafer-level system integration platform (CoWoS). It remains the most critical bottleneck for GPUs, CPUs, and custom ASICs. IDC expects industry wafer capacity to grow by about 80% in 2027, while demand grows close to 2x. Semiconductor packaging (ABF substrate suppliers) remains near full capacity, with no meaningful additions expected before the second half of 2027. We also find that supply shortages and long lead times have spread beyond memory, chips, and packaging to optical components, power parts, liquid cooling, and basic datacenter power equipment. We expect these shortages to persist through 2027, with memory remaining tight into 2028.
We see AI demand hitting real-world constraints; the problem is a mismatch in time horizons. AI-ready datacenters take five to seven years to site and build, while backers want a return on investment and customers need capacity now. Investments in datacenter facilities last for decades. However, the same cannot be said for the IT infrastructure housed in them: we see a new generation of accelerators (e.g., GPUs) every 12 to 18 months.
Growing AI usage is shifting risk from demand to supply, rather than eliminating it. When demand outruns supply, history has consistently shown that scarce market capacity leads to persistently inflated per-unit prices.
IDC’s tracker data, outlined above, shows memory, power, and packaging shortages persisting through 2027 and into 2028. That is why supply, not demand, will set the pace.
Demand that cannot be served or pushed out does not disappear altogether. It spreads, shifting from segment to segment. In the past 12 months, 31% of companies have shifted at least some workloads from proprietary to open-weight models, according to the IDC AI Economics and Cost Management Survey. We expect easily adoptable open-weight models to shift inference to private, sovereign, and edge environments, diverting some inference demand away from public cloud providers. More than half of organizations globally already use sovereign cloud for AI, with another third considering it over the next 12 months, according to IDC’s Digital Sovereignty study. IDC’s HPC/AI Index shows non-service-provider spending on AI infrastructure steadily rising from $28.6 billion in 2025 to $96.4 billion in 2030, a 27.5% CAGR. This is in addition to their IT infrastructure spending for non-AI workloads. The likely outcome is a hybrid market that combines on-device processing, sovereign and private clouds, AI factories, and a marketplace of open-weight and proprietary models. This decentralized AI market also makes managing the creation, deployment, and tracking of agents, models, and data a critical challenge for enterprises, reinforcing the need for unified AI workload management and agent governance, whether delivered within AI factory solutions or as standalone platforms.
As the market matures, we expect some consolidation among specialty cloud providers. The durable assets that change hands will be newly built datacenter sites and upgraded grid connections, which will be the foundation for the next round of AI-native businesses. Accelerators age faster, and as they are refreshed and depreciated, prior-generation systems will enter the secondary market at lower prices over time, making AI infrastructure more accessible to enterprises and supporting further demand growth. Specialty cloud providers that succeed will do so by rapidly moving up the stack to model serving, sovereign AI, or full agentic infrastructure with storage, data management, and governance add-ons that improve margin and create workload stickiness. The result will be an AI ecosystem that is more distributed, dynamic, and focused on meeting the actual demands of customers, both enterprise and consumer.
Four signals would change our view:
Beyond 2028, the question is whether supply catches up. We plan to track how supply constraints may affect enterprise and service provider spending.
We are also watching whether circular financing, in which vendors invest in the specialty cloud service providers that buy their products, holds up. A pullback in that funding, or in private equity backing, would be a macro warning sign. If that pullback happens, it will accelerate the consolidation described above, with well-capitalized Hyperscalers and profitable cloud service providers best positioned to acquire datacenter sites and grid connections at a discount.
Demand is not the risk in this market. The real risk is whether enough capacity can be sited, powered, and supplied in time, and who ends up owning it.
The post AI Is Short of Capacity, Not Demand appeared first on IDC .