In the Netherlands and Ireland, new AI data center projects are now waiting seven to ten years for a grid connection, and some urban areas have paused new connections outright.
AI workload placement across on-premises facilities, regional colocation, and telco edge sites turns on which constraint binds first for a given workload: the grid queue, a data-sovereignty rule, or the latency budget of the inference path.
Quick take
Choose on-premises when data gravity, sovereignty rules, or a massive proprietary training set make moving the data riskier or more expensive than moving compute.
Choose telco edge, or a colocation site near existing grid headroom, when the workload needs sub-10 millisecond response times and a fresh on-site power connection is not realistic.
Choose public cloud or larger colocation for burst training capacity and specialized AI services that are not worth maintaining in-house.
Data gravity is the practical reason on-premises still wins some AI workloads. Prompts, embeddings, vector indexes, and retrieval queries need to sit close to the systems holding the source data, or every response pays a network tax.
Agentic AI raises the stakes further. Once a system can call tools and act on production systems, identity, blast radius, and audit-evidence retention become placement questions, not just model-selection questions.
Why Grid Timelines Set the Floor Before Latency Does
Power, not bandwidth, is the binding constraint across much of Europe right now. Transmission congestion and permitting delays mean some established markets face multi-year queues, and a few have stopped taking new connections entirely.
AI hardware refreshes every three years or less, but grid upgrades routinely take longer than that cycle. A site that finally receives its power connection may already be behind on the hardware plan it was built around.
Developers have responded by placing smaller clusters at medium-voltage connection points near existing industrial or renewable capacity, instead of waiting years for a new high-voltage substation to serve one large campus.
The Data-Gravity Case for Keeping Inference On-Premises
Cisco’s Murali Gandluru frames the split by workload intent: real-time inference where latency is a dealbreaker, such as a factory floor, belongs at the edge, while training on massive proprietary datasets belongs on-premises because moving that data is expensive and risky.
Gandluru also places public cloud in this split. He describes it as the right venue for burst capacity, rapid prototyping, and specialized AI services that are not worth maintaining in-house.
NetApp’s Pravjit Tiwana argues governance should be separated from the infrastructure itself, since AI pipelines increasingly move data and models across on-premises, cloud, and edge boundaries over their lifecycle.
Sovereignty rules push in the same direction as data gravity. The EU’s GDPR and AI Act shape where organizations are willing to host AI workloads, and distributed placement helps enterprises keep sensitive processing inside a chosen jurisdiction.
AI Workload Placement Criteria: Matching the Site to the Constraint
| Site type | Latency fit | Data governance | Power / buildout reality |
|---|---|---|---|
| On-premises | Best for workloads tied to local systems and internal calls; no network hop for retrieval | Full control over where prompts, embeddings, and audit evidence are logged | Limited to existing facility power; no queue, but no elastic headroom either |
| Regional colocation | Depends on carrier density at the site; a general-purpose fit rather than a guaranteed latency floor | Tracked as its own deployment category alongside on-premises and hybrid edge in current market segmentation | Draws on backbone connectivity such as 400G wavelength routes rather than a single dedicated utility feed |
| Telco edge | Built for sub-10 millisecond response times close to 5G and industrial sites | Supports jurisdictional rules such as GDPR by keeping processing inside a region | Placed at medium-voltage connection points where grid headroom already exists, avoiding multi-year queues |
Current market tracking groups on-premises, hybrid edge, and colocation as parallel deployment categories for AI/ML edge inference, but it does not publish a like-for-like latency or cost comparison between a specific colocation facility and a specific edge site.
What Colocation and Long-Haul Connectivity Actually Solve
Colocation and regional data center operators are positioning around the same shift. Digital Realty describes moving “from campus to network” for the inference era and has opened an Innovation Lab in London to test AI and HPC infrastructure under real data center conditions.
None of that matters without the network path between sites. Lumen advertises 400G wavelength connectivity across ~340,000 global fiber route miles reaching ~163,000 on-net buildings, with a 20-day SLA for delivery.
That kind of backbone is what lets a colocation site or telco edge node function as a real extension of on-premises infrastructure rather than an isolated island — the connectivity contract matters as much as the site’s power contract.
The edge data center infrastructure market is being sized on this same shift: from US$ 15.19 billion in 2025 to US$ 106.48 billion by 2035, a 21.5% CAGR, with AI/ML edge inference flagged as one of the most commercially valuable segments and telecom operators repositioning as edge infrastructure operators.
What to Verify Before You Commit to a Site
IDC’s Dave McCarthy describes the goal as a common data plane and a unified security protocol across private facilities, public cloud, and edge endpoints, so workloads can move as the AI lifecycle phase changes.
Skip that groundwork and the cost shows up later. Lenovo’s Robert Daigle warns that ad hoc placement decisions accumulate technical debt: fragmented architectures, duplicated infrastructure, and escalating costs.
Before signing a site contract, confirm three things for that specific location: the actual grid connection timeline, the sovereignty scope of the data that will pass through it, and the wavelength or edge network path that will carry the workload in production.