TOPIC #3
AI: An Update
AI and token usage grow while data center focus increases on power intensity, efficiency, and reliability.
Productivity Gains from AI Top CFO Priorities
Executives say the motivation for using AI is to make employees more productive, improve operations, and support faster business decisions, above reducing headcount or costs.
- A CFO survey from the Richmond Fed shows that improving production efficiency is the highest-ranked motivation for AI investment, followed by improving labor productivity and decision-making.
- Reducing labor and non-labor costs ranked much lower comparatively.
Analysis from the Atlanta Fed suggests productivity gains are most closely tied to innovation and demand-side uses, including new or improved offerings and better customer reach, pointing to growth and service improvement, not just automation.
AI is showing up less as immediate savings and more as better workflows, faster analysis, and higher-quality output. Cost savings may come later, but the near-term story is productivity and better execution.
FIGURE 1
Federal Reserve Survey: Mean Likert Value of CFO AI Investment Motivations
Note: The mean likert value for each category is calculated by averaging ratings of importance for motivations to invest in AI: “Not at all” as 0, “Slightly” as 1, “Moderately” as 2, “Very” as 3, “Extremely” as 4.
Source: Atlanta Fed, Richmond Fed
Key Takeaways
AI adoption is shifting from cost reduction to productivity, faster decisions, improved workflows, and innovation-led growth.
Enterprise token use and agentic AI are expanding computing demand, creating cost, governance, and load implications.
Data center growth is driving larger, denser, more power-intensive facilities, increasing pressure on power requirements, efficiency needs, and reliability.
Reported Productivity Gains Are Running Ahead of Measured Gains
Research from the Richmond Fed points to an AI “productivity paradox”: companies report that AI is improving output per worker, but those gains are only partly showing up in revenue-based productivity measures.
- Reported AI productivity gains from CFOs ranged from 1.1% to 2.2% across sectors in 2025 and are expected to rise to 2.1% to 3.7% in 2026.
- When measuring AI-attributed changes in revenue relative to AI-driven employment changes, there were smaller gains, ranging from 0.4% to 0.8% in 2025 and 1.1% to 2.2% in 2026.
- High-skill services and finance showed the largest gains under both reported and implied measures.
AI may already be improving workflows, task efficiency, decision making, and output quality. As with computer adoption in the 1980s and 1990s, practical improvements may appear before the full economic impact is easy to measure.
Firms expect AI productivity benefits to continue growing in 2026, with AI-driven productivity gains becoming more visible as AI tools are more fully integrated into business processes.
FIGURE 2
Reported and Implied Productivity Growth from AI
Source: Atlanta Fed, Richmond Fed
Assessing Growing Token Usage
As AI adoption grows, token consumption is becoming an increasingly important measure of AI utilization and a meaningful component of enterprise operating costs.
Falling token costs could make broader AI use economically attractive, especially for coding, analysis, customer support, and agentic workflows. Potential business benefits include faster experimentation, broader employee adoption, synthetic data generation, and quicker software prototyping.
But usage volume is not the same as productivity. Measuring tokens consumed can show adoption but not necessarily value. Companies will need governance that compares AI usage costs against avoided labor costs, improved speed, better quality, and actual business outcomes.
For utilities, continued growth in token consumption still matters because more token usage ultimately translates into more computing demand and more sustained data center load.
Interconnection Delays Are Pushing Data Centers Toward Self-Supply
Data center demand is rising faster than grid planning and interconnection can absorb, leading to increasing data center grid interconnection timelines. This is pushing more developers toward behind-the-meter (BTM) or “bring your own power” strategies.
- Data center demand is projected to reach 134.4 GW in 2030, up from 75.8 GW in 2026.
- These strategies could reduce the timeline to full operation for a 500 MW data center by three to five years.
A recent data center survey found 56% of developers exploring on-site or co-located power, while Cleanview’s project tracker has identified 59 data centers with 90 GW of BTM projects (see Fig. 3).
- Bloom Energy, a BTM fuel cell supplier, estimates 38% of facilities may use some on-site primary power by 2030.
- Fully BTM facilities could rise to 27% by 2030, up from 1% in 2024.
For AI data centers, power has become a front-end development decision. Even with data centers bringing their own power, many jurisdictions are evaluating or approving utility proposals to mitigate growing data center power demands. The Database of Emerging Large Load Tariffs has identified more than 75 such proposals as of late June 2026.
FIGURE 3
Cumulative Capacity of Announced Behind-the-Meter Data Center Power Projects (GW)
Source: Cleanview
AI Compute Growth Is Driving a Step Change in Data Center Power Requirements
AI developers are deploying larger facilities and more power-dense computing systems to support increasingly complex AI models that require many more watts of power than standard IT services.
The result is a shift from megawatt-scale IT facilities and conventional 10 kW to 40 kW server racks toward gigawatt-scale campuses built around high-density 150+ kW AI server racks. Proposed projects include a 10 GW SoftBank Ohio site, Meta’s 2.2 GW Louisiana site, and OpenAI’s Stargate initiative targeting up to 10 GW.
The increase in power draw is occurring at both the facility and server levels as more computing power is packed into the same space.
- Google expects AI and machine-learning workloads to require more than 500 kW per rack before 2030.
- Nvidia’s Feynman AI chips, which are slated for a 2028 release, would come with a 1 MW power draw per server rack.
These higher-power designs are creating additional complications, requiring developers to rethink how much of each megawatt can be converted into useful compute.
FIGURE 4
Rack Power Demand by Type (kW)
Source: Bloomberg
Reimagining Data Center Design to Meet Growing Power Intensity
According to Nvidia, roughly 30% of data center power is not being used for AI computing. Data from Epoch AI indicates this could be even higher as cumulative AI data center capacity reached 31 GW in Q4 2025.
Figure 5 below shows the split in power demand between computing power and other site needs.
As AI demand continues to grow, these losses become increasingly material. At GW scales, every 1% efficiency gain saves tens of megawatts, reducing grid or BTM power draw or enabling more AI workloads.
FIGURE 5
Cumulative AI Data Center Capacity (2022 Q1 to 2025 Q4) Compared with Selected Illustrative Recent Regional Peak Demand (GW)
Note: *Cooling, distribution, networking, etc.
Source: Epoch AI
Cooling and Distribution Systems Account for Significant Power Losses
Rack densities above 40 to 50 kW make traditional air cooling ineffective. Developers are turning to direct-to-chip liquid cooling, which circulates non-conductive liquid directly to heat-generating components, reducing cooling loads and improving energy efficiency by up to 15%.
Power distribution is an equally important constraint.
- Electricity must be converted from AC to DC and stepped down to the low voltages that chips require, with these conversion steps accounting for roughly a third of power losses.
- Nvidia’s 800-volt DC architecture is being developed to reduce those losses by moving conversion equipment out of racks, consolidating voltage step-downs, and could ultimately reduce distribution losses to under 1%.
With more powerful chips arriving annually, continued efficiency gains will be essential to supporting AI growth and easing power demands.
NERC Alert Highlights Growing Reliability Risks from Data Center Load Drops
On May 4, 2026, NERC issued a Level 3 Essential Actions alert that recommends utilities and grid operators strengthen visibility into how data centers, AI training, cryptocurrency mining, and other computational loads behave at scale across service territories during grid disturbances.
- The alert follows “widespread and unexpected customer-initiated load reduction” of large computational loads that involved several incidents in 2024 and 2025, during which 1,000 MW or more of unexpected large-load output reduction occurred.
- Sensitive protection systems can rapidly reduce or disconnect load, making facility behavior harder for operators to predict.
The Essential Actions are not mandatory (as compared with a reliability standard), but relevant entities must report progress to NERC.
NERC’s recommended actions focus on seven areas: load modeling, stability and system-strength studies, qualified-change review triggers, start-up and testing procedures, fault-response evaluations, dynamic fault recording, and communication protocols between computational load operators and utilities.
For utilities and system planners, the alert underscores that large computational loads may need to be modeled as dynamic grid resources. Their response during disturbances could affect contingency planning, mitigation plans, operating limits, emergency coordination, and broader reliability risk management.
Implications
For the AI industry, power availability, efficiency, and reliability are becoming front-end development priorities, shaping site selection, design, time to market, and self-build considerations.
Growing token use, agentic applications, and higher-density AI racks will intensify compute power needs, requiring stronger cost controls and governance and improved efficiency in cooling, power distribution, and other data center power needs.
Reliability planning must evolve as large computational loads behave dynamically, requiring better modeling, communication protocols, and disturbance-response visibility.
CONTACT OUR EXPERTS
On AI: An Update
Jon Kerner
PARTNER AND INNOVATION & TECHNOLOGY LEADER
jkerner@scottmadden.com 404.814.0020
Luke Martin
PARTNER
lukemartin@scottmadden.com 919.781.4191
RECENT INSIGHTS
Available at scottmadden.com
ScottMadden posts energy and utility industry-relevant content and publications on a regular basis. The list below is a sample of recent insights prepared by our consultants.
