AI Inference vs. Training: infrastructure, rack density, and operational requirements
TL;DR In 2026, AI inference workloads represent approximately two-thirds of global AI compute, compared with one-third in 2023 (Deloitte). Yet many colocation operators still offer general-purpose infrastructure that does not differentiate between the two workloads. They diverge across three structural dimensions: physical constraints (rack density, cooling, latency), optimal data center geography, and long-term operational requirements (24/7 permanence, location, CADA sovereign compliance). A colocation operator designed for high-density inference addresses requirements that general-purpose infrastructure cannot meet.
In 2026, AI inference workloads represent approximately two-thirds of all global AI compute, according to Deloitte’s TMT Predictions 2026 report. This figure was one-third in 2023 and approximately half in 2025, meaning that the shift occurred in less than three years. At the same time, McKinsey projects that inference-dedicated infrastructure capacity will grow at a CAGR of 35% between 2025 and 2030, compared with 22% for training, in its 2025 analysis of hyperscalers. These dynamics have a direct impact on infrastructure decisions made by scaling neoclouds: AI inference and training do not share the same physical constraints, the same optimal geography, or the same long-term operational requirements. Yet many colocation operators continue to offer general-purpose infrastructure that treats both workloads as interchangeable. This article examines the three areas where the divergence is structural.
The technical constraints that distinguish AI inference from training
Rack density and cooling: why native DLC is non-negotiable for high-density inference
Training operates in batch mode on distributed clusters, with rack densities generally below 30 kW. High-density AI inference follows a different physical model: Nvidia’s B200 and GB300 GPUs have thermal power ratings of up to 1,000 W per GPU, and NVLink rack configurations reach 50 to 100 kW per rack. Conventional air cooling systems effectively cap at 20–30 kW per rack under the best conditions, making them structurally incompatible with these densities. This is why a data center that adds direct liquid cooling (DLC) retrospectively to an existing air-cooled infrastructure cannot guarantee the same long-term thermal performance as an architecture natively designed for DLC. For high-density inference, native DLC is a technical prerequisite, not a configuration option.
Network topology and latency: requirements incompatible with Nordic geographies
Training is a batch workload, meaning network latency is not a decisive operational criterion. A training cluster can operate in Iceland or Finland because access to low-cost energy takes precedence over geographical proximity. AI inference follows the opposite logic: it processes real-time user requests for APIs, agents, and chatbots that require P99 response latency below 10 ms. This requirement imposes a location close to end-user markets, namely European metropolitan areas, rather than Nordic regions where distance creates unavoidable latency. Furthermore, inference processes user data in real time at large scale, making it subject to GDPR and the Cloud and AI Development Act (CADA, effective August 4, 2026). Level 4 of this framework conditions sovereign qualification on the operator’s capital ownership nationality, beyond the sole physical location of the data.
These two constraints, thermal performance and latency, require selecting a colocation operator according to criteria that differ from those historically applied to training.
The geographical consequences of the shift toward inference in Europe
57% of European neocloud AI commitments signed in Nordic countries in 2025
According to CBRE (Q3 2025 data), 57% of new colocation contracts signed by European neoclouds for AI workloads were located in Nordic countries. This figure reflects market constraints during the 2024–2025 period: access to electrical power was the primary limiting factor for any GPU deployment in Europe, due to high-voltage grid connection lead times exceeding several years in Western Europe. Nordic countries offer abundant energy resources and shorter grid connection timelines, which directed infrastructure decisions toward these geographies. However, this movement is consistent with training workloads, whose batch nature allows them to tolerate distance. For production AI inference, it creates a structural shortage of sovereign capacity in markets with high user density: France, Germany, and the Netherlands.
France facing a shortage of sovereign high-density inference capacity
ABI Research projects that global infrastructure capacity dedicated to AI inference will reach 46 GW by 2035, with a CAGR of 42% over the 2025–2035 period, according to a May 2026 publication. This volume will need to be geographically distributed to meet the latency requirements of production inference. Due to investments concentrated in Nordic geographies since 2023, the Paris region and secondary French markets remain underserved in terms of high-density colocation capacity designed for inference. Fast-growing neoclouds relying on geographies unsuited to inference expose themselves to latency constraints that engineering solutions cannot fully overcome.
Production AI inference: a permanent operational asset by 2030–2035
From a one-off asset to production infrastructure: the operational horizon of inference
Training is deployed through distinct campaigns — runs lasting 3 to 6 months, with resources allocated and then released. Production AI inference operates continuously, 24/7, at the core of neocloud commercial services. This permanent nature changes infrastructure planning parameters: available capacity, density per block, geographical location, and operator operational quality must be defined over a timeframe corresponding to the lifespan of a production service, namely several years. ABI Research projects that neoclouds and hyperscalers will reach inference capacity parity around 2035, at 15 GW and 16 GW respectively. To reach these volumes, colocation operators capable of delivering high-density blocks within predictable timelines and in latency-optimized locations represent the industry’s bottleneck.
CADA Level 4 compliance: why operator nationality has become a selection criterion
The Cloud and AI Development Act (CADA), which entered into force on August 4, 2026, introduces a four-level sovereign qualification framework for cloud and data center services used by public entities and critical infrastructure operators. Level 4 conditions qualification on the operator’s capital ownership nationality: a provider subject to the US Cloud Act does not meet this level, regardless of the physical location of its servers. For neoclouds serving customers in these categories, or anticipating doing so, choosing a European-nationality colocation operator is a regulatory prerequisite. Due to CADA coming into force at precisely the moment when inference demand is accelerating, infrastructure decisions made in 2026 and 2027 will determine sovereign qualification capabilities over the following decade.
Voltekko: high-density, sovereign AI inference colocation in Western Europe
Voltekko designs and operates colocation data centers exclusively engineered for high-density AI inference workloads. The 6 to 7 MW IT blocks are delivered within 6 to 9 months across two sites (the Paris region and Alcochete, Portugal), geographically positioned in markets with high end-user density, in line with the latency requirements of production inference. Direct liquid cooling is native, using a DLC 70/30 configuration, integrated from the initial block design phase. The blocks are dedicated, with no infrastructure-layer sharing between customers.
Operations are managed by EQUANS, a Bouygues subsidiary specializing in industrial technical services with twenty years of data center experience. Long-term financial stability is guaranteed by REED, backed by the Société Générale Group. Given the CADA Level 4 requirements in force since August 2026, Voltekko addresses capital ownership nationality, data residency, and long-term operational compliance throughout deployment lifecycles.
Frequently asked questions about AI inference infrastructure
What is the main technical difference between inference infrastructure and training infrastructure?
Training tolerates rack densities below 30 kW, remote geographical locations, and temporary deployments. High-density inference requires 50 to 100 kW rack configurations with native liquid cooling, a location close to end-user markets to meet P99 latency constraints (below 10 ms), and permanent operational availability over a multi-year horizon.
Why are Nordic countries less suited to production AI inference?
Nordic countries offer genuine energy advantages for batch training workloads. However, the distance from end-user markets creates latency that is incompatible with the requirements of production AI inference applications. In addition, CADA Level 4 requirements regarding sovereign capital ownership restrict the qualification of operators for workloads covered by the regulation.