Data Centers and the R&D Tax Credit: Documenting Innovation in the Age of AI

Recent New York Times reporting on Meta’s use of AI data centers to claim billions in federal tax benefits has sparked renewed debate over whether large-scale data center development activities qualify for the research credit under IRC §41.

According to the reporting, Meta has characterized certain AI-focused data center activities as experimental or pilot-model development efforts, resulting in billions of dollars of claimed research tax benefits. Critics have questioned whether deploying commercially available chips and infrastructure constitutes qualified research.

Shawn Marchant, R&D Tax Principal at Tanner, contributed his perspective to the New York Times article, expressing skepticism whether all costs associated with building a data center, including, in this case, billions spent on NVidia chips qualify for R&D tax credits. His concern underscores an important distinction: While significant innovation is taking place within AI infrastructure, every investment in technology or data center development may not constitute qualified research.

The controversy raises an important question for data center operators, cloud providers, hyperscalers, and supercomputing companies:

What types of data center activities truly qualify for the R&D credit, and how should those efforts be documented?

What the NYT article failed to mention was Shawn’s comment that while skepticism is warranted, there may, in fact, be a reason to include the costs. For example, Meta would need to demonstrate what uncertainty they were trying to solve with the 10,000^th Nvidia chip that they couldn’t have answered with the 165^th chip and then perform the related experimentation to answer the question.  While merely purchasing and operating servers is generally not research, many modern data centers are no longer simple facilities housing commodity equipment. Today’s AI infrastructure requires companies to solve unprecedented engineering challenges involving power delivery, thermal management, cooling technologies, equipment longevity, and system optimization. These development efforts often involve significant technical uncertainty and extensive experimentation, making them strong candidates for R&D credit consideration.

Why Data Centers Are Different Today

For much of the industry’s history, data centers were built around relatively standardized technologies. Servers generated manageable heat loads, cooling systems relied primarily on traditional air handling, and power requirements were predictable.

The AI revolution has changed that reality.

Modern hyperscale facilities are deploying racks with power densities that can exceed 135 kW per rack, compared to approximately 20 kW per rack in traditional data center environments. This dramatic increase in power consumption creates new challenges in heat removal, electrical distribution, equipment reliability, and facility design.

As a result, many data center operators have found themselves engaged in genuine engineering development rather than routine facility operation.

Understanding Qualified Research in a Data Center Environment

To qualify under IRC §41, an activity generally must pursue a permitted purpose, address technical uncertainty, rely on a process of experimentation, and be technological in nature.

For data centers, the key distinction is whether a company is merely implementing known solutions or developing new methods to overcome technical limitations. The strongest claims typically arise when engineering teams are attempting to improve:

• Energy efficiency
• Cooling effectiveness
• Compute density
• Equipment reliability
• GPU lifespan
• Water consumption
• Power conversion efficiency
• Infrastructure automation
• System performance

When these efforts involve unresolved technical questions and systematic testing, they begin to resemble the type of research Congress intended to encourage.

Example 1: Advanced Cooling System Development

Perhaps the most significant area of innovation today is cooling technology.

As AI clusters have grown in size and power density, traditional air-cooling systems have become increasingly inadequate. Engineering teams are now developing direct-to-chip liquid cooling systems, immersion cooling technologies, and hybrid thermal management solutions capable of supporting next-generation GPU deployments.

Related reading: NVIDIA Blackwell Platform Boosts Water Efficiency by Over 300x.

Consider a company attempting to design a cooling system capable of supporting AI racks operating at substantially higher thermal loads than existing facilities.

At the start of the project, engineers may not know:

• Which coolant design will perform adequately
• Whether desired operating temperatures can be maintained
• How flow rates will impact reliability
• Whether the solution will increase equipment life or reduce it

To answer these questions, teams often conduct thermal simulations, computational fluid dynamics (CFD) modeling, pilot deployments, prototype testing, and iterative redesigns.

These activities represent classic examples of technical uncertainty addressed through experimentation.

Documentation to Maintain

Strong supporting records may include:

• Thermal modeling designs and reports
• CFD simulation outputs
• Engineering design documentation
• Test protocols
• Sensor data
• Temperature analysis reports
• Design review presentations
• Failure analyses
• Prototype comparison studies

Example 2: Extending GPU and Accelerator Lifespan

One of the most valuable development areas emerging within AI infrastructure involves maximizing the useful life of high-performance computing equipment.

Modern AI accelerators represent enormous capital investments. As organizations deploy increasingly expensive GPUs, extending operational life by even modest percentages can generate significant economic benefits.

Supercomputing and hyperscale operators frequently undertake projects focused on understanding how thermal stress, workload patterns, cooling strategies, and operating environments affect long-term hardware reliability. These projects often include accelerated stress testing, predictive maintenance modeling, thermal profiling, and degradation analysis.

Functional Objective

A typical project objective might be:

Increase GPU operational lifespan while maintaining required performance levels and reducing failure rates.

Technical Uncertainty

At project inception, engineers may not know:

• The relationship between thermal cycling and hardware degradation
• Optimal operating temperatures
• Effects of specific workloads on reliability
• Whether alternative cooling approaches improve lifespan

Potential Documentation

• Reliability studies
• Mean-time-between-failure analyses
• Environmental monitoring reports
• Failure investigations
• Equipment health monitoring logs
• Predictive maintenance models
• Engineering test results

Example 3: AI-Driven Data Center Optimization

Artificial intelligence is increasingly being used to operate the facilities that power artificial intelligence itself.

One of the best-known examples comes from Google DeepMind, which applied machine learning models to optimize data center cooling operations. The company reported reductions in cooling energy consumption of up to 40% through the use of predictive algorithms and automated control systems.

Related reading: DeepMind AI Reduces Google Data Centre Cooling Bill by 40%.

It is critical to highlight the technical uncertainties specific to your data center and experimentation performed to resolve those uncertainties. Projects of this nature frequently involve substantial technical uncertainty.

Engineers may not know:

• Whether machine learning models can accurately predict thermal behavior
• Which control variables most influence efficiency
• How automated decision-making affects reliability
• Whether optimization creates unintended operational consequences

As a result, organizations often perform extensive experimentation involving model training, simulation, pilot testing, A/B testing, and iterative parameter adjustments in order to answer those technical questions.

Documentation Examples

• Training datasets
• Source code repositories
• System architecture diagrams
• A/B testing results
• Algorithm performance reports
• Pilot deployment summaries
• ML model validation reports

Example 4: Next-Generation Cooling Technologies

Some of the most exciting recent innovations involve entirely new cooling architectures.

For example, Microsoft has demonstrated microfluidic cooling technology that routes coolant through microscopic channels etched directly into silicon. According to Microsoft’s published results, the technology demonstrated cooling performance up to three times greater than conventional cold-plate approaches.

Related reading: Microsoft’s research on microfluidic cooling technology.

These projects frequently involve:

• Heat transfer modeling
• Silicon design optimization
• Microchannel engineering
• Fluid dynamics research
• Reliability and durability testing

Virtually every phase of such development typically involves technical uncertainty and experimentation.

Building a Defensible R&D Credit Position

Over the last several years, data center operators and infrastructure providers have achieved significant functional advances through engineering innovation. The lesson from recent discussions surrounding AI data centers is not that all data center costs qualify for the research credit.

Rather, qualification depends on demonstrating that engineers were attempting to solve genuine technological problems through a process of experimentation.

This distinction is particularly important given the scrutiny surrounding AI infrastructure investments and the concerns Marchant raised in the New York Times. Companies should carefully evaluate which activities represent genuine research and which reflect the acquisition, installation, or implementation of existing technologies.

The most successful R&D credit studies for data center operators typically focus on clearly documented development projects where teams can identify:

1. A desired functional improvement.
2. A technical uncertainty that existed at the outset.
3. Experimentation performed to resolve that uncertainty.
4. Engineering records supporting the work performed.

Companies should avoid relying solely on broad narratives about innovation or simply assuming that all costs associated with construction of a data center qualify. Instead, they should assemble contemporaneous evidence showing what engineers were trying to accomplish, what obstacles they faced, what alternatives they evaluated, and how decisions evolved throughout the project, in addition to the connected qualified expenses.

Conclusion

The rapid growth of AI has transformed data centers from relatively standardized facilities into some of the most complex engineering environments in the world. From liquid cooling and microfluidics to GPU lifespan optimization and AI-driven facility management, data center operators are increasingly engaged in work that resembles traditional research and development. R&D in the AI and data center space will continue to provide compute efficiencies and reduce power consumption.

While simply purchasing servers, chips, sockets, or constructing facilities generally does not satisfy the requirements of IRC §41, the underlying engineering efforts that enable modern AI infrastructure often involve substantial technical uncertainty, systematic experimentation, and measurable functional improvements.

As tax authorities continue to examine data center R&D claims, the companies with the strongest positions will likely be those that can clearly connect their development activities to documented engineering challenges, documented experimentation, and documented technological advancement.

The foundation of a successful claim is not the size of the investment, but the quality of the evidence demonstrating that genuine research occurred.

How Tanner Can Help

Tanner’s R&D Tax Credit professionals help companies evaluate qualifying research activities, identify eligible expenses, and develop the documentation necessary to support defensible R&D tax credit claims. As innovation accelerates across AI infrastructure and other industries, our team helps businesses understand where legitimate research opportunities exist and how to substantiate them.