The speed at which hardware, software, and networking are changing is making artificial intelligence (AI) models evolve at a fast rate. We’ll see the pace of innovation that happened in the entirety of 2024 now taking place within each quarter in 2025.

This shift is compounding – AI is evolving from monolithic models to multi-models, redefining enterprise AI strategy.

What does this mean?

Where single large-scale AI systems were trained to handle a broad range of tasks – now modular, multi-model architectures are breaking down tasks into specialised AI components. AI is becoming task-specific, dynamic, and optimized. This results in more efficient outcomes and reduces costs.

To understand the impact of AI’s innovation and evolution and how an AI-first ecosystem empowers first-movers, this blog will highlight:

• Tracking AI’s evolution
• How AI is impacting metro infrastructure
• How multi-model AI is powering speed to market
• How Digital Realty powers AI’s next phase

Tracking AI’s evolution: Durability, disruption, and deployment

As AI adoption grows, ensuring workload durability, efficiency, and scalability is critical. The shift from monolithic AI models to multi-model architectures is causing disruption, but it’s important to understand the history of AI’s evolution to understand why it’s taking place.

Five years ago, AI infrastructure primarily centred around cloud-based training, with companies relying on public clouds and hyperscalers for training. AI inference at the edge was minimal due to latency and bandwidth constraints, and networking bottlenecks limited AI scalability. Most traffic had to route through centralised cloud data centres.

Today, AI is shifting towards distributed inference, requiring metro data centres for real-time decision-making.

AI’s impact on metro infrastructure and why proximity matters more than ever

AI workloads increasingly demand low-latency (≤50ms) inference. This is because AI inference – the process of running trained AI models on real-world data – requires low latency for applications that require real-time decision-making. This makes strategic metro compute infrastructure critical.

AI traffic doesn’t simply route to the nearest edge location – it must consider:

• Inference capabilities of available compute resources
• GPU utilization and hardware availability
• Network efficiency and workload performance

As AI becomes instantly responsive, it supports enterprise goals for their end-user experience. Low-latency inference powers:

• Human-like responsiveness in AI-powered customer interactions (think chatbots and voice assistants)
• AI-driven automation powers self-driving cars, industrial automation, and robotic systems that need to be processed in real time to work safely
• Personalised recommendations via streaming services, e-commerce platforms, and ad technology
• AI in financial trading, where high-frequency algorithms must process and act on market data within milliseconds to capitalise on price movements

So, to achieve sub-50ms inference latency, AI workloads must be deployed closer to end users. This drives demand for metro AI infrastructure that minimizes the physical distance between compute resources and data sources.

How multi-model AI is powering speed to market

Traditional AI models have had problems with high costs of computation, poor efficiency, and problems scaling up. Today, enterprises are embracing multi-model AI architectures that route tasks to specialised AI components based on their strengths.

Key innovations driving this shift include:

• Mixture of Experts (MoE): An AI framework that dynamically selects specialised models (or “experts”) to process different parts of a task, improving efficiency and accuracy
• Multi-token prediction: A technique where AI models predict multiple works or elements at once, making responses faster and more coherent
• Reasoning models as orchestrators: AI models designed to analyse and break down complex queries, then route tasks to the best-suited models
• AI interconnection: The ability to seamlessly connect AI workloads across different environments (cloud, edge, and private infrastructure) for optimal performance and efficiency

Having a strong infrastructure foundation is critical to meeting ongoing AI innovation needs. Digital Realty began evaluating the future impact of AI on the data centres more than seven years ago. We’ve seen repeatedly with our customers that the key to success is to plan with the end in mind as much as possible. Having the right partners along the way is critical to success.

How Digital Realty powers AI’s next phase

As innovations in software, hardware, and networks take place, Digital Realty welcomes the transformation through three key products and solutions ready to meet AI demand:

1. ServiceFabric makes it easy to connect AI inference across metro environments. It has low-latency, high bandwidth and dynamically routes AI workloads to optimal compute locations.
2. Private AI Exchange (AIPx) is an enterprise-grade AI interconnection. It securely orchestrates AI workloads across multi-cloud and private environments and supports compliance-driven AI deployments without compromising performance.
3. PlatformDIGITAL, our global data centre platform, provides metro AI-ready data centres positioned for:
   1. Proximity to key AI hubs for ultra-low-latency inference
   2. Optimised GPU utilisation for scalable AI workloads
   3. Seamless integration with AI-native networking

The next wave of AI innovation requires more than just powerful models. It needs a well-designed, connected system to make AI decisions in real time.

For enterprises, the focus must be on:

• Multi-model AI architecture for efficiency and accuracy
• AI-native infrastructure to support real-time AI inference at metro scale
• Seamless interconnection to maximise AI’s business impact

At Digital Realty, we’re not just supporting AI adoption, we’re shaping the future of AI-powered business.

Empower your IT strategy with future-proofed, AI-ready infrastructure. Contact us today to get started.