AI’s global resource race: Challenges and opportunities

Since late 2022, the largest IT and AI firms have leaned toward building ahead of projected demand to ensure adequate capacity, running the risk of empty data centers if projected demand does not materialize. This could happen as a result of various scenarios, including lower-than-expected AI adoption; AI/chip efficiency gains reducing demand for data center power; technology leaps (such as quantum computing, specialist accelerators or software innovations) advancing AI capabilities while using fewer hardware resources; or evolving data center location requirements making current/planned facilities undesirable.

Enterprise demand for AI could falter if the return on GenAI investment is lower than expected. Although enterprises have noted challenges deploying AI and obtaining the expected benefits, AI adoption has increased dramatically in the past three years, and firms continue to expect further take-up, according to S&P Global Market Intelligence 451 Research end-user surveys.

Efficiency gains or technology leaps could also produce a surplus of empty data centers, although to counter this, one could point to the Jevons paradox, whereby greater efficiency reduces the cost of a resource, which in turn increases its overall consumption. If the Jevons paradox applies to AI, it likely will not be due to price signals but to improved speed and quality of services. A technological leap from something like quantum computing could certainly take place, but widespread quantum adoption is still several years out. In the meantime, companies continue to find new uses for AI, while training models to obtain results can take weeks or months. There are endless queues of jobs to run. At this stage, it seems more likely that efficiency gains will simply catalyze innovation and increased usage.

Regarding location, AI providers have been expecting a shift from training very large language models to training and maintaining smaller, more distributed models, while also expecting the industry to shift to monetizing AI via inferencing. Inferencing is the process of using a trained AI model to generate outputs; this need not be done in centralized facilities and can use lower-power hardware. Inferencing is vitally important in the AI ecosystem, as it is where training is put to work to create value in the form of generated text, audio, images, video or other outputs. This trend could lead to some larger, more power-hungry data center facilities sitting empty.  

However, large public cloud providers have a bit of a backup plan. The public cloud model of shared infrastructure should appeal to smaller companies running AI. If an AI use case does not require very low latency inferencing, the underlying infrastructure could remain in centralized data centers. Much will depend on AI use cases and how much inferencing will require near-instantaneous response. This will influence the relative need for distributed capacity in urban centers versus centralized resources in highly efficient data centers. AI firms seem to be hedging their bets by building both large-scale campuses in outlying locations and smaller facilities close to urban centers.

Finally, although there is always potential for overbuilding, the data center industry is increasingly flexible and can respond to overcapacity quickly by reducing new construction. Data center facilities are built more rapidly than in years past, with typical timelines of one to two years for permitting and 12-14 months for construction. The same cannot be said for large-scale power plants, with their much longer time horizons — gas-fired power plants and onshore wind resources, for example, may commonly require around six years from planning to completion. Still, other trends should boost the demand for electricity (e.g., electrification of manufacturing and heating, electric cars and bitcoin mining), helping to ensure that utilities will not be left with overcapacity. 

While demand is strong, the data center industry faces multiple headwinds spanning supply chains, power and energy, sustainability, financing, local resistance and regulatory frameworks.

Energy and environmental constraints

Data centers and utilities have been challenged to provide power for AI growth. Accessing energy is a constraint, as many popular data center metro markets are already flirting with the upper limits on generation and transmission. Northern Virginia, one of the largest data center markets in the world, faces multiyear delays for new grid connections, while markets in places such as Ireland and Singapore have restricted new data center approvals. To provide a sense of scale, Dominion Energy Virginia said in February 2025 that data center firms had requested 40.2 GW of power connections, up from 21.4 GW in July 2024. This dramatic increase in AI power demand reflects the industry's rapid expansion.

Such expanding energy footprints collide with decarbonization goals, leading many AI providers to source renewable energy in addition to grid power. Effective AI energy management strategies are becoming essential as companies seek to balance performance with sustainability. Given the variable nature of renewable generation from wind and solar, ambitious plans have been launched to revive nuclear plants or commission new small modular reactors to provide longer-term low-emissions power. In addition, regulators in various locations are requiring data centers to bring their own power generation facilities (often referred to as "behind-the-meter" solutions) to any new AI data center builds. Cooling AI data centers can also require a significant water supply, raising environmental concerns, especially in drought-prone regions. Managing data center energy usage efficiently has become a priority for operators seeking to minimize environmental impact. 

Hardware and supply chain bottlenecks

The AI industry faces supply chain constraints due to a shortage of GPU chip manufacturing facilities. The chips also require rare earth metals and high-grade silicon, which could constrain their production, particularly given manufacturing locations, component sources and geopolitical tensions.

Data center construction requires copious amounts of materials such as copper, cement and steel, which are subject to global supply chain constraints and tariffs. In addition, lead times are increasing for data center equipment such as diesel generators, power distribution systems, medium-voltage transformers and advanced cooling systems, with delivery sometimes taking several years. Rising costs and import restrictions on essential materials can delay construction, increase costs and create bottlenecks for operators. Geopolitical tensions and protectionist trade policies further exacerbate these challenges, making material sourcing and pricing unpredictable.

Raising the capital required 

GenAI is capital-intensive. We estimate that it could require $200 billion per year globally just to build the data center capacity currently planned, excluding the cost of IT equipment and electrical infrastructure (substations, transmission lines). The resources required to create large-scale AI models provide a competitive advantage to large IT firms with a low cost of capital or those that can find creative new forms of financing, such as OpenAI's partnership with NVIDIA in which the chipmaker provides capital to the AI firm in exchange for shares in the company. Infrastructure-as-a-service options or “neo-clouds,” a new wave of cloud providers focused specifically on GPU-rich infrastructure, can reduce up front capital expenditures and offer an alternative channel for accessing chips. When it comes to energy utilities, a key concern is whether retail power customers will end up paying for some of the infrastructure build-out via higher electricity rates.

Success in the next wave of data center and AI expansion will depend on more than just capacity; it will require innovation across supply chains, financing and technology deployment.

Chip diversification is critical, as the industry is currently dependent on NVIDIA chips, which are mainly manufactured in Taiwan. AMD, Intel and hyperscalers are working to develop competing chips, while semiconductor production is increasing in the US, Japan and the EU. When it comes to data centers, the largest builders and customers have the capital to preorder equipment with long lead times, such as generators and even gas turbines for colocated power plants. These large-scale players will have an advantage if they continue to standardize their designs and order equipment in bulk. Meanwhile, modular approaches to data center builds can accelerate time-to-market and provide access to a centralized store of key components.  

Financial innovation will be required as well: Joint ventures, sustainability-linked bonds and green loans, securitization, sales of leased facilities to real estate investors, new approaches to lending, and other innovations will be required to fund such a capital-intensive industry. 

Efficiency in technology deployment will help determine competitiveness. Liquid cooling and specialized immersion systems are expensive, but they are vital for sustaining high-density GPU clusters and enabling more compact and efficient facility designs. Smaller, domain-specific AI models like those released by DeepSeek can reduce training intensity, while physical AI-powered systems, including robotics and autonomous vehicles, will move inferencing workloads closer to end users, alleviating pressure on hyperscale data center facilities and creating new regional and edge investment opportunities.