An organisational and architectural model of artificial intelligence in which training data, compute, model weights, and inference services are concentrated under a single operator in large data centres, with users accessing capability remotely through APIs; the contrast class to decentralised AI, edge inference, and multi-agent approaches that distribute computation, control, or decision-making across many independent nodes.
Semantic Classification
Content
Definition
Centralised AI names the dominant deployment pattern of the frontier era: a small number of laboratories train very large models on proprietary data using massive, co-located GPU clusters, keep the weights private, and serve inference from their own or hyperscale cloud infrastructure. The economics of scaling laws drive this concentration — pre-training runs costing hundreds of millions of pounds are only feasible for organisations that can aggregate compute, energy, data, and specialised talent in one place — and the resulting capability is delivered to everyone else as a metered API.
The model has genuine strengths. Centralised operators can enforce uniform safety mitigations, patch vulnerabilities in one place, monitor misuse across their whole user base, and offer a level of model quality that distributed alternatives have not matched. It is also the configuration that compute governance regimes implicitly assume: chokepoints in the supply of advanced chips and data-centre capacity are only effective levers because training is concentrated.
Its weaknesses mirror those of any centralised system: single points of failure and control, opaque decision-making about model behaviour, data gravity that pulls user information into a few silos, latency and connectivity dependence, and geopolitical concentration of a strategically important capability. These are precisely the pressures that motivate the contrast classes — Edge Inference moves computation onto local devices for latency, privacy, and resilience; Decentralised AI approaches such as federated and distributed low-communication training spread the training itself across independent participants; and Multi-Agent System architectures distribute decision-making authority rather than concentrating it in one monolithic model.
Current Landscape
The tension between centralised and distributed AI is now a live architectural and policy question rather than a settled outcome. On-device model deployments from major platform vendors, increasingly capable open-weight models, and training methods tolerant of slow interconnects (exemplified by DiLoCo-style distributed optimisation) are eroding the assumption that capability must live in a single data centre. At the same time, frontier capability continues to concentrate: the largest training runs, the deepest safety teams, and the majority of paid inference all sit with a handful of centralised providers.
-
Distributed training has crossed from theory to practice: Prime Intellect’s INTELLECT-1 (late 2024) trained a 10-billion-parameter model across five countries and three continents on up to 112 H100 GPUs, using DiLoCo plus an int8 all-reduce to cut communication bandwidth roughly 400-fold versus standard data-parallel training.
-
Decentralised RL and larger runs (2025): INTELLECT-2 (32B) was trained by globally decentralised, permissionless reinforcement learning on heterogeneous contributed compute, and Nous Research’s Consilience began pretraining a 40B model on ~20 trillion tokens over its Psyche network — evidence the ceiling on distributed scale keeps rising.
-
Method lineage: These build on DeepMind’s DiLoCo and Prime Intellect’s OpenDiLoCo (2024), whose inner/outer optimisation synchronises pseudo-gradients only every few hundred steps, making training viable over slow or intermittent interconnects.
-
Centralisation still dominant: Despite this, the largest frontier runs, deepest safety teams, and most paid inference remain with a handful of centralised providers, and most production systems are hybrids — routing between a large centralised model and small local ones, or coordinating edge agents against a central orchestrator.
Sources:
-
https://www.galaxy.com/insights/research/decentralized-ai-training