Deloitte’s enterprise AI infrastructure survey: A 2028 outlook

Complexity is also rising alongside use cases: Nearly all respondents (96%) rate their AI workloads as medium or high complexity, yet Deloitte’s broader research shows that most organizations are still moving beyond basic automation and single-agent solutions, and haven’t yet tackled the business process transformation or technical complexity of multi-agent systems.² More advanced capabilities—such as contextual reasoning, multi-turn logic, and multimodal generative workloads—add further complexity even for advanced organizations. Despite the groundwork still needed, confidence in execution is high: Ninety-seven percent of respondents say they’re confident or very confident that they can scale AI workloads in the next three years.

To run these workloads, businesses are choosing among four models.

• Closed: Proprietary and application programming interface-accessed
• Open: Publicly accessible and adaptable
• Packaged software-as-a-service (SaaS): Cloud-delivered and embedded in enterprise apps
• Derivative: Fine-tuned or customized

Closed models are the most used among those surveyed (figure 1), with slightly higher adoption by respondents from innovation-forward industries like technology, media, and telecommunications (34%), and mid-sized organizations (35%). However, as respondents look to 2028, there’s no clear direction on which model or mix of models will win out. Model strategy is evolving so quickly that point-in-time distributions may reflect market conditions when the survey was fielded, not a fixed endpoint.

Beyond the survey, and based on Deloitte’s client experience, AI-first leaders are increasingly bullish on closed models, with usage among market leaders estimated at around 60% to 70%. Closed models provide differentiation via new capabilities for running sophisticated workloads, protecting sensitive data, and enabling monetization.³ Moreover, recent market moves—such as a major technology company shifting a previously open-source model to a proprietary one⁴—continue to shape strategies and reinforce this tilt toward closed solutions.

At the same time, open-source and derivative models continue to power low-cost experimentation and innovation through strong developer and data science communities. NVIDIA, for example, expanded its open-source AI resources by releasing new open models, training tools, and data sets in 2026.⁵ The performance of open-source models is often near frontier models and may be better suited for smaller, specialized use cases.⁶

We can infer that agentic SaaS (platforms with AI agents to autonomously execute workflows) is here to stay,⁷ pointing to a more integrated and “agentified” system-of-record solutions (the system where a business domain’s records are created and governed), along with new AI-adjusted pricing models.

Each of the four models has its merits, and adoption may hinge on two factors: which new models become available and what the AI is being used for—to create competitive advantage (differentiation) or to improve standard capabilities faster and cheaper (democratization).

Token consumption is surging, a sign of growth, but also a risk signal

High complexity, bigger scale, and rising enthusiasm eventually compound to an explosion in token consumption. (An AI token is a unit of both compute and cost of AI applications and infrastructure.) Today, the largest proportion of respondents (37%) are consuming one billion to 10 billion tokens per month on average, and 30% are consuming more than 10 billion tokens per month. By 2028, 61% of respondents expect to consume more than 10 billion tokens per month on average, roughly doubling token consumption in about two years (figure 2). This increase is primarily driven by the rise in large token footprints of more than 100 billion tokens per month, reflecting a tripling of token use from 2026 to 2028.

The surge in token consumption can also be attributed to the shift toward advanced reasoning models, which generate “thinking tokens” as an output before returning an answer.⁸

While token consumption growth is a function of increased application adoption, it doesn’t necessarily indicate effective adoption—especially in early stages. In many organizations, token growth can signal inefficient solution patterns (like oversized prompts, weak context management, or limited reuse), reflecting the need to assess the strategic value of AI applications and the underlying cost dynamics. Strong solution design, engineering, and architecture principles can help manage input and output token volumes as AI scales.

As strategic applications scale in both number of users and complexity, token consumption growth may become the most practical early-warning indicator of shifting infrastructure supply-and-demand dynamics.

While the growth trajectory is undeniable, 51% of respondents reported that economic uncertainty could be the leading factor that limits long-term investments and changes AI factory plans. This stands to reason, given that AI factories require capital investment and often involve multiyear commitments. Others might also include organizational business challenges (48%), regulatory pressures (48%), and talent and skill gaps (40%) that could delay or slow planned AI factory investments.

Despite these hurdles, the vast majority (71%) of respondents expect that AI factories will support innovation, risk management (64%), and token optimization objectives (59%) (figure 4).

New, increasingly hybrid infrastructure approaches come with the need for new frameworks and platforms for managing AI workloads. Respondents believe that vendor-specific AI frameworks (55%), AI network fabrics (which can enable network virtualization) (50%), and foundational models (47%) can support increased consumption of graphics processing units (GPUs) and hybrid approaches. And as hybrid AI stacks mature, the conversation is shifting from architecture choices to the economics of running AI at scale.

In today’s market, the rapid run-up in high‑bandwidth memory¹⁰ and other nonvolatile memory components is expected to materially increase the cost of AI factories, aggravated by longer procurement times as demand outpaces supply.¹¹ Wafer costs are also expected to rise by 20%,¹² likely pushing GPU and central processing unit (CPU) prices higher.

As AI factories scale to utility-grade capacity, power generation and grid interconnections¹³ also become strategic constraints and potential areas of investment. Scaling, however, could vary by context—edge deployments might only use CPUs and small language models in limited spaces, while centralized AI factories can range from modest GPU clusters to purpose-built, campus-scale infrastructure.

Forty-seven percent to 48% of the surveyed leaders expect to address their power needs through a mix of power from the grid, self-built power generation on or near the AI factory site, and third-party–built generation near the AI factory—potentially reducing single-point dependency and improving resilience and contractual flexibility. Cooling strategies could require a range of approaches, given how much power is being generated and whether the organization plans to invest in traditional air cooling or liquid cooling.

Taken together, these cost and capacity pressures shift the focus to who makes the calls and how trade-offs get made—especially as decision authority, accountability, and budget ownership are often distributed across the organization.

This also puts CIOs and CTOs in the driver’s seat to help business leaders understand AI consumption patterns and their underlying cost implications. To address misalignment risks across teams, they can establish shared decision rights and value-based governance that link platform investments to execution accountability and measurable outcomes. In parallel, they can bring upstream critical insights—such as token consumption patterns and functional AI strategies—that enable more informed C-suite discussions on the economics of AI, from both the supply and demand sides.

While many AI programs have the funding needed to scale, at least for now, talent remains an issue. First, there’s a skills gap: While many respondents (more than 50%) noted that their organizations have data engineers, security and compliance specialists, AI and machine learning (ML) and agent operations engineers, and data scientists, the talent needed to scale and innovate with AI isn’t fully met (figure 7). Organizations may need more security and compliance specialists and AI and ML and agent operations engineers, as well as change management experts, because success often lies in AI’s reliable adoption in day-to-day workflows rather than mere deployment. In the future, many organizations may be looking to scale skills like robotics systems engineering and carbon and energy monitoring, signaling a move toward physical automation and justifiable energy performance.

Second, there’s a capability divide: Eighty-one percent of respondents believe IT teams have the technical and financial acumen to navigate scaling, compared with 65% for business and product teams (a 16-percentage-point gap). Continued investment in skill development in these strategic areas may be necessary for the vision and execution of AI programs in the future.