As we prepare to turn the page on 2026, the data center (DC) buildout continues at full throttle. The big five hyperscalers—Amazon, Google, Meta, Microsoft, and Oracle—are on track to have allotted over $800 billion in capex in 2026, about 15% higher than forecast a year ago and over 75% higher than the 2025 total.
Yet even spending at this rate is not enough to meet the growing need for DC supply, brought about by the rising adoption of agentic AI (such as in coding agents, personal assistants, and enterprise AI systems) and the increasing affordability of AI models. Despite growing hurdles around deployment (with mounting concerns particularly around permitting and energization), BCG still expects the high rate of spending to continue through 2031. About 180 GW of DC supply is anticipated to come online globally over the next five years (versus 100 GW to 110 GW by the end of 2026), but BCG’s latest research finds that supply is likely to fall short of demand by more than 20 GW, the equivalent of roughly $1 trillion in capex at current prices.
Bottlenecks in critical supply chains are the principal cause of the DC supply shortfall. These supply chains are at or near their limits, restricting the availability of essential components, such as advanced-node wafers and memory, especially high-bandwidth memory (HBM). The availability of electric power is also a constraint, particularly in certain regions such as the US. The bottlenecks tend to ease or tighten in lockstep because the supply chains depend on each other. So, solving one constraint, such as chip shortages, cannot alone correct the shortfall.
To manage in this environment, chip OEMs need to innovate by leveraging new architectures, partnerships, and software to win share among leading offtakers (customers with long-term contracts). AI labs and model companies need to balance their efforts to push the frontier of intelligence with two additional endeavors: meeting the growing enterprise demand for more cost-efficient models (whether open-weight or closed) and providing support to ease deployment frictions. And compute providers (traditional and neocloud) need to keep building broader value-added offerings (such as model routing and inference as a service) that can differentiate them in an increasingly crowded space.
While the first phase of the DC buildout has been a boon to companies across the ecosystem, we don’t expect the rising tide to continue lifting all boats. Mounting delays and permitting concerns, along with local opposition, mean scarce capital is increasingly dependent on DC developers’ ability to deliver to specifications and on time. Those that falter may face issues securing financing and offtakers.
We therefore expect the next chapter of the DC buildout to be increasingly defined by a shrinking set of winners that can lead the way in overcoming the scaling pains and deployment hurdles the AI industry currently faces.
Demand for AI Is Still in the Early Stages
Although AI use is increasingly widespread, with north of 1 billion active individual users globally across AI tools, high-volume usage to date continues to be concentrated among a few companies that were early adopters. In fact, only 6% of the more than 600 US public companies with a market cap over $5 billion are categorized as AI-mature users, as per our 2026 Applied AI Index. However, BCG’s bottom-up assessment of over 18,000 discrete human tasks and workflows—and AI’s ability to augment them—implies the technology’s true potential remains materially higher. (See “About Our Research.”)
About Our Research
We then constructed a proprietary global DC model that takes a bottom-up, first-principles approach to estimating the supply and demand for compute. We derived supply estimates through primary research on capacity across key bottlenecks. We then triangulated these estimates with revenue forecasts for relevant industries and site-level tracking data, which we collected and verified using techniques such as satellite-imagery, which can show the progress at DC sites. We assessed demand estimates through task-level modeling (across more than 18,000 tasks) to determine the applicability of AI. We then engaged with more than 400 CIOs, CFOs, and CEOs to track the enterprise adoption of AI for these tasks.
We continuously verified assumptions and outputs through discussions with semiconductor and equipment suppliers, leading-edge chipmakers, OEMs, telecom and network infrastructure providers, utilities, DC operators and developers, hyperscalers, neocloud providers, AI labs, and other technical experts. We also assessed the results against market developments and company disclosures. (See the exhibit.) Our analysis was conducted in September 2026.
Companies that are classified as AI leaders outperform their own industry’s total shareholder returns by 9 percentage points (pp), with outperformance driven by both revenue growth and margin expansion—setting a benchmark that others will want to match. In fact, the BCG AI Radar 2026 survey found that 50% of CEOs believe their job stability depends on getting their AI strategy right. And 82% of the same CEOs are more optimistic about AI’s ROI potential than they were a year ago, further propelling investment.
While optimism around AI’s ROI is getting stronger, broad-based measurement of these returns is still challenging. Coding in areas such as software engineering and data analytics continues to be the most mature AI use case. Large enterprises are seeing up to a 50% improvement in productivity across their developer groups, and leading developers are releasing many times more pieces of production code by using AI assistants.
Other industry-specific use cases also continue to show promise but are earlier in their maturity, with inadequate data to tangibly prove impact.
- Large pharmaceutical firms are using AI to accelerate drug discovery and target identification.
- Semiconductor companies are exploring AI to optimize chip designs.
- Freight companies are using AI to optimize load matching.
- Advertising firms are leveraging AI for campaign testing and optimization.
- Insurance companies are using AI to automate low-risk underwriting.
New AI software vendors are catalyzing broader enterprise adoption by launching horizontal and vertical AI tools. For example, Harvy AI and Legora are enabling the adoption of AI for legal work, and Decagon and Sierra are facilitating AI in customer service operations. Such companies are scaling faster than any other startup wave in history. Many have reached $100 million in annual recurring revenue in under three years—swift growth compared with the historical software-as-a-service average of more than seven years.
The rapid growth in AI usage has been compounded by the development of smarter and less costly models. In recent years, AI models have reached or surpassed median human performance across many domains, such as speech, image recognition, and reading comprehension. In addition, less expensive, open-weight or smaller models have become a credible alternative for workloads that do not require frontier intelligence. These advances, along with the growing efficiency of AI hardware, have made it economically feasible to adopt AI at a time when budgets continue to be top of mind for the C-suite.
Supply Chain Bottlenecks Pace the Buildout
We estimate that the global market is already experiencing an approximate 13 GW shortfall in DC supply, a gap that is likely to persist as supply chain constraints continue to prevent capacity from catching up. The shortfall is visible in compute prices: rental rates in 2026 have routinely reached more than $20 million to $30 million per megawatt per year, and rental rates for three-year-old hardware with Nvidia’s Hopper architecture have been within 20% of their launch rates.
The limited availability of scaled power envelopes is expected to be a clear constraint over the next few years, with lead times for grid interconnection or behind-the-meter permitting slowing the ability to add powered (or energized) land across the US and Europe. The reduced production of HBM, particularly HBM3 and HBM4 (which, outside of China, is manufactured at-scale by only Micron Technology, Samsung, and SK Hynix) poses another constraint to increasing the base of installed compute. And while power and memory are front and center to most conversations, another constraint is close behind—the limited supply of advanced-node wafer capacity (3 nanometers and below), which is critical for the next generation of frontier AI hardware.
While supply chains remain constrained, various companies are making advancements to relieve these critical bottlenecks. DC developers are looking to try to bring facilities online faster by using behind-the-meter power, such as onsite gas-fired power plants that use aeroderivative gas turbines or reciprocating engines. In addition, AI hardware companies are bringing new architectures and designs online to reduce the dependency on HBM. For example, Cerebras uses materially more on-chip SRAM, Qualcomm Technologies uses a stacked memory and logic compute architecture, and SambaNova has a three-tier memory design.
Meanwhile, companies including Intel and Samsung, as well as SpaceX through its Terafab project, are actively investing in the manufacture of advanced-node wafer capacity, joining Taiwan Semiconductor Manufacturing, which is itself investing in new capacity in Japan, Taiwan, and the US. However, a fabrication plant can take up to five years to build and bring online, which will greatly limit the impact of these investments over the next two to three years.
Nonetheless, the breadth of existing bottlenecks far exceeds those described here, which makes eliminating the DC supply shortfall particularly challenging. Not only will chip component and power constraints be hard to alleviate in a timely manner (gas turbines now face multiyear lead times, for example, and permitting is becoming increasingly difficult and expensive), but shortages in skilled labor, logistical support, and cooling equipment are part of a long tail of hurdles that we expect will contribute to continued tightness in the compute market.
The Next Wave of DCs Will Be Unlike the Last
As the DC ecosystem scales over the next five years, we expect the market for DC supply to evolve across several dimensions.
A CPU Market Uptick. CPU demand is growing faster than many prior forecasts anticipated, driven by orchestration needs for agentic workflows. Standalone CPU capacity (the CPU chips that are not part of AI racks) is expected to grow at a CAGR of about 11% as a result, reaching 53 GW to 60 GW of related DC supply in 2031. This growth is expected to be further compounded by the number of CPUs embedded in AI codesigned systems, such as Nvidia’s Vera CPUs in its Vera Rubin racks.
A Broader Chipmaker Ecosystem. Purpose-built accelerators (or application-specific integrated circuits) are expected to play a more meaningful role going forward as both independent chip OEMs and hyperscalers invest in custom chips that they expect can more efficiently manage specific workloads. In 2026, Google, for example, agreed to sell its tensor processing units to third parties, including Anthropic, while Amazon Web Services partnered with Cerebras to pair its systems with high-performance AWS Trainium chips for optimized inference performance.
A Credible Alternative to Frontier Intelligence. Less expensive AI models (including open-weight models, whose core components are released publicly) are becoming a credible alternative to most advanced AI models, addressing the pressure to reduce the cost of inference. We estimate that about 70% or more of workloads do not require frontier intelligence, and businesses are already pushing for more cost discipline, achievable through more sophisticated model routing (whether in-house or through third-party tools, such as Together AI, Fireworks AI, or Baseten inference clouds).
A Range of DC Locations. As some locations become saturated or unfeasible, new locations will likely emerge for DC construction both in the US and outside of it, such as in the Middle East, India, and Malaysia. The availability of electric power and the supportiveness of regulatory environments globally will continue to dictate where future locations may be.
Key Players Face New Questions—and Opportunities
The buildout of DCs is still in the early stages. AI is complex to deploy in enterprise contexts, broad ROI signals are still too early to detect and hard to measure, and budgets often struggle to keep up with the breadth of use cases. As we look to the future, several factors are likely to affect how the DC market evolves, including:
- Improvements in the efficiency and performance of AI hardware, software, and models that further reduce the cost of adoption
- The continued growth of AI adoption across enterprises and the budgets to support it
- The amount invested in the broader ecosystem—across value chains and labor pools—to support the DC buildout
- The advent of new form factors for DC deployments, such as nonterrestrial DCs
We also expect the evolution of DCs to have several impacts and opportunities for key stakeholders.
Investors. DCs are an established and attractive market for private equity. Chip financing models are also likely to become increasingly interesting—and necessary—as supply constraints continue to bolster the value of DC supply and a secondary market for hardware emerges. However, given the size of potential outlays (renting 1 GW of capacity on the latest Nvidia hardware can cost as much as $40 billion for a year), investors will be increasingly diligent about both counterparty risks (which companies will be the most resilient offtakers) and collateral risks (which chip types will retain the most value).
As the first rounds of investment in the buildout start to bear fruit, investors will also need to think about the next set of important businesses—ones that don’t just enable the buildout but also support the ongoing operations, maintenance, and, ultimately, the retrofit of the rapidly growing fleet of installed DCs.
DC Developers. Investors are increasingly rewarding DC developers that can generate confidence in delivery timelines for large-scale projects by building the necessary partnerships, local relationships, supply chains, and site architectures that can accelerate delivery and increase cost efficiency. Securing powered land is the primary bottleneck for DC developers, with interconnection and permitting queues pushing them—or their tenants—toward power-secured sites and smaller, more modular builds. Addressing key local and state-level concerns around environmental and economic impact is also increasingly necessary and a growing priority for forward-thinking developers.
Chipmakers. The growth in AI inference—and the heterogenous workloads that underpin it—are creating new opportunities for novel, highly tailored architectures and chip designs. For example, ultra-low-latency workloads (those that require a time to first token of under 500 milliseconds and can make up as much as 10% to 15% of overall inference demand) make up a new segment in which next-generation chipmakers can challenge incumbent chip designs.
Even with a strong design, however, these chipmakers need to think about the practicalities of their go-to-market strategy—pinpointing the right customers and partners, determining the best sales motion (such as third-party sales or selling installed compute), and getting supply chain allocations in a market where allocations are heavily weighted to incumbents.
The DC ecosystem continues to evolve, facing new frictions as it scales and constantly innovating to overcome them. These growing pains will undoubtedly create new, unforeseen challenges, but they will also present opportunities for forward-thinking investors and operators alike.
This article focuses on the supply and demand dynamics for compute and the implications for investors, developers, and chipmakers. It does not address the broader debate on data center siting, community impact, or environmental considerations.