AIInfrastructure Academy Europe

From LLMs and RAG to Agents, Inference and Infrastructure

A 45-minute introduction to the architecture of modern AI systems and what it means for the infrastructure underneath.
‍
You will understand LLMs, inference, RAG, AI agents, model calls, context, AI gateways and the basic architecture of an AI application — before following the workload down into compute, memory, networking, storage, power and cooling.

The ideal starting point for the entire Academy.

What makes AI compute different from conventional server workloads?

We examine GPUs, TPUs and specialized AI accelerators, training versus inference, memory bandwidth, HBM, utilization and the increasingly heterogeneous AI compute landscape.

Key question: Which compute architectures are suited to which AI workloads?

What actually happens when a model generates an answer?

Understand tokens, context windows, time to first token (TTFT), tokens per second (TPS), batching, KV cache, quantization and other fundamentals of inference.

Then connect them directly to infrastructure requirements.

Key question: What determines the performance and infrastructure cost of an AI request?

AI systems can create enormous communication requirements between accelerators, servers and clusters.

We examine scale-up versus scale-out networks, Ethernet, InfiniBand, bandwidth, latency, congestion and distributed AI workloads.

Key question: Why is the network becoming part of the AI compute system?

AI is dramatically changing the power profile of the data center.

We examine rack density, power distribution, UPS requirements, utilization, redundancy, grid constraints and the relationship between AI workload growth and facility capacity.

Key question: How do you plan power infrastructure when AI demand can grow faster than the facility?

Increasing compute density creates increasing heat density.

Understand the move from traditional air cooling toward rear-door heat exchangers, direct-to-chip liquid cooling, CDUs and high-density thermal architectures.

Key question: At what point does conventional cooling stop being enough?

AI doesn't operate on models alone. It needs data.

We examine RAG, vector databases, knowledge graphs, embeddings, data pipelines, storage throughput and data locality — and how these components interact with inference infrastructure.

Key question: What happens between enterprise data and the model?

The next generation of AI applications doesn't simply answer questions. It plans, uses tools, calls other systems and delegates work to other agents.

One user request can therefore become many model calls.

We examine agent loops, orchestration, tool use, multi-agent architectures, retries and workload amplification.

Key question: What happens to infrastructure demand when one request becomes 20?

AI performance is only half of the problem. The other half is economics.
‍
We examine accelerator utilization, model size, tokens, batching, caching, quantization, model routing, specialized inference hardware and performance per watt.
‍
Key question: How do you reduce the cost of inference without destroying performance or quality?

Production AI needs much more than accelerators.
‍
We examine orchestration, observability, monitoring, scheduling, reliability, security, guardrails, evaluation, failure recovery and the operational stack around production AI.
‍
Key question: How do you operate AI infrastructure reliably rather than merely get a demo running?

The final masterclass brings the complete Academy together.
‍
Compute. Memory. Networking. Storage. Power. Cooling. Operations. Economics.
‍
We connect them into one system and examine how increasingly powerful inference, agentic AI and new accelerator architectures could influence the next generation of data centers.
‍
Key question: What should data center operators be preparing for now?