​​​

​

​

​

​

​

​​​

​The Next Major Shifts in AI Workloads and Hyperscaler Strategies

​

Detailed analysis structured around six core themes

​

1. The Shift from Training to Inference as the Dominant AI Workload

In the early stages of AI development, the primary focus was clearly on training large-scale models. These training processes are extremely compute-intensive, require very high power densities, and rely on large, tightly coupled GPU clusters. At the same time, they are relatively insensitive to latency, as training runs often last for hours or even days and do not need to respond to users in real time.

As AI applications mature, however, the focus is increasingly shifting toward inference—the productive deployment of trained models in real-world use cases such as search queries, personalized recommendations, chatbots, industrial automation, and financial analytics. Inference workloads are generally less compute-intensive per transaction but are highly latency-sensitive and volume-driven. Millions or even billions of requests must be processed within very short time frames.

McKinsey (*1) expects inference workloads to account for more than 50 percent of total AI compute demand by 2030. This shift fundamentally changes infrastructure requirements: while training environments are optimized for maximum scale and power availability, inference requires proximity to end users, resilient network connectivity, and high availability.

​

Summary

• AI infrastructure is evolving from training-driven to inference-driven

• Inference will dominate volumes and force new location strategies

• Latency and proximity to users are becoming critical success factors

​​

2. New Data Center Profiles: Centralized Training Facilities vs. Distributed Inference Sites

The separation of training and inference results in two clearly distinct data center profiles.

Training facilities are evolving into highly specialized, large-scale campus environments with extremely high power densities (in some cases exceeding 100 kW per rack), liquid cooling systems, and massive energy demand. These facilities can be located with relatively high geographic flexibility, provided that sufficient power is available.

Inference data centers, by contrast, must be located closer to metropolitan areas, enterprise customers, and digital hubs. They are more tightly integrated into existing network infrastructures and require high-capacity fiber connectivity and strong operational resilience. While power densities are more moderate, scalability and rapid expansion capabilities are critical.

As a result, hyperscalers are no longer pursuing a single data center archetype. Instead, they are building hybrid, workload-specific infrastructure landscapes, optimized differently depending on the use case.

​

Summary

• Emergence of two distinct data center types: training vs. inference

• Training: energy-driven, centralized, ultra-high density

• Inference: user- and network-driven, more distributed, latency-critical

​

3. Campus and Network Strategies as Core Design Principles

To manage the growing complexity of AI workloads, hyperscalers are increasingly adopting campus-based development models. Multiple specialized data center buildings are co-located and interconnected through high-performance internal networks. This enables efficient interaction between different workload types, such as general compute, AI training, inference, and storage.

At the same time, network topology is becoming a decisive factor. AI workloads require ultra-low latency within clusters and high-bandwidth connections between sites. Fiber availability, network redundancy, and interconnection density are turning into hard location requirements—often outweighing traditional real estate considerations.

This dynamic favors locations that already offer dense fiber infrastructure and proximity to internet exchange points, while also pushing hyperscalers to actively develop and upgrade infrastructure in emerging regions.

​

Summary

• Campus models are becoming the standard for AI infrastructure

• Network topology is a key competitive differentiator

• Interconnectivity increasingly outweighs pure site size

​​

4. Energy Transforms from a Cost Factor into a Strategic Bottleneck

The most significant limiting factor for AI infrastructure expansion is power availability. In many regions, high-capacity grid connections require lead times of 24 to 36 months or longer. At the same time, AI-driven power requirements are growing exponentially.

As a result, energy is no longer just an operating expense—it has become a strategic bottleneck. Hyperscalers are increasingly selecting sites based on where power can be delivered quickly, reliably, and at scale. This is accelerating the shift toward Tier 2 and Tier 3 markets, where grid capacity can often be deployed faster than in congested core markets—such as the Frankfurt/Rhine-Main data center region in Germany.

In parallel, hyperscalers are intensifying their involvement across the energy value chain, including long-term power purchase agreements (PPAs), investments in renewable generation, battery storage solutions, and, in some cases, direct collaboration with grid operators.

​

Summary

• Energy is the primary constraint on AI-driven growth

• Site selection is increasingly power-driven

• Hyperscalers are becoming active participants in the energy ecosystem

​​

5. Strategic Adjustments by Hyperscalers: Speed over Optimization

Under intense time pressure, hyperscalers are significantly adapting their development and investment strategies. Speed to market now outweighs perfect cost optimization. Rather than relying solely on self-development, hyperscalers increasingly leverage leasing models, partnerships with developers, and lease-to-own structures.

Construction strategies are also evolving. Modular and prefabricated building approaches are gaining importance as a means to reduce delivery timelines. At the same time, existing data centers are being selectively retrofitted to support AI workloads—even if they were not originally designed for very high power densities.

These approaches demonstrate a growing willingness among hyperscalers to accept higher short-term costs in exchange for faster capacity deployment and market share protection.

​

Summary

• Speed to market outweighs cost optimization

• Increased use of leasing, partnerships, and modular construction

• Retrofitting is becoming a core element of AI infrastructure strategies

​​

6. Implications for Investors, Municipalities, and Site Developers

These developments are also reshaping the role of external stakeholders. Investors benefit from strong demand for AI-capable infrastructure but must navigate increasing complexity related to power, permitting, and technology. Municipalities are becoming key partners, as they influence land availability, permitting processes, and often indirectly power access.

For site developers, traditional real estate expertise is no longer sufficient. Integrated solutions that combine power access, grid connectivity, permitting feasibility, and long-term scalability are essential. Locations capable of meeting these requirements are gaining substantial strategic relevance.

​

Summary

• AI turns infrastructure into a strategic location issue

• Municipalities and energy providers become key stakeholders

• Integrated site concepts determine success or failure

​​

Conclusion

AI marks the transition into a new phase of digital infrastructure. Hyperscalers are evolving from pure cloud service providers into system-level infrastructure players, orchestrating power, networks, locations, and capital. For all market participants, the message is clear: those who can successfully combine energy availability, speed, and scalability will emerge as the winners of the next expansion phase of the AI economy.

​

Sources​

• Koehler Advisory – own research

• (*1) McKinsey & Company, The next big shifts in AI workloads and hyperscaler strategies

        https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-         and-hyperscaler-strategies