The grid builds on the existing telecom backbone. In its current conception (see Diagram 1), it seeks to leverage existing network assets — including the central offices and regional points of presence that support today’s voice, data and internet traffic — as a distributed compute fabric for AI. These sites can be progressively enhanced to support lightweight inference capabilities alongside existing network functions. In aggregate, this footprint of network assets represents a substantial pool of latent power, with estimates suggesting more than 100gw of compute capacity embedded within existing networks.
Another key aspect of the grid is the control plane: an intelligent orchestration system that allocates workloads dynamically across the network in real time (see Diagram 2). Workloads are routed to where they are most efficient from a cost, latency and availability perspective, improving the overall utilisation and economics of compute resources. In doing so, the system may also favour locations with more favourable energy footprints, introducing an indirect link to energy efficiency.
See also: The great data delusion — where to invest for AI winners
The shift towards decentralisation
The AI grid marks a shift towards a decentralised system, where workloads are orchestrated across a geographically dispersed infrastructure layer. This, in turn, is a significant departure from the prevailing cloud computing model (see Diagram 3).
See also: AI or algorithm
Under the status quo, workloads are routed to large-scale data centres, where they are pooled together to improve resource utilisation. This reflects a scale-driven model (see table for comparison) dominated by centralised infrastructure and ever-larger facilities optimised for unit economics. Accordingly, rapid cloud adoption in the 2010s underpinned the rise of hyperscalers — so named for their business model of operating immense, standardised data centre fleets where scale has delivered superior efficiency and economics.
Today, hyperscalers continue to host the majority of compute workloads — AI or otherwise. Current data centre developments suggest, however, a shift away from exclusive concentration around hyperscalers towards a more decentralised model. This evolution is driven by the differing compute requirements of AI workloads.
AI’s evolving workload requirements
Over the past year, the term “compute” has surged into public consciousness (see Chart 1), but the underlying concept is neither new nor unique to AI.
Compute simply refers to a computing system’s capacity to carry out workloads, encompassing everything from data processing to mathematical and logical operations.
For more stories about where money flows, click here for Capital Section
Traditionally, central processing units (CPUs) handle general-purpose workloads, which tend to be sequential, logic-heavy and often “bursty” — characterised by short spikes in usage followed by periods of low activity. As a result, the CPU often sits idle for extended periods between short bursts of demand. The hyperscaler cloud model emerged to address this inefficiency, by pooling large volumes of such workloads, improving overall CPU utilisation through aggregation.
AI workloads, however, are structurally different. They are highly parallel in nature, requiring many computations to run simultaneously. Graphics processing units (GPUs) excel at such workloads, operating in tightly coupled clusters to execute large-scale matrix operations efficiently. As a result, traditional hyperscaler efficiency gains from workload pooling are less pronounced in AI, where performance is increasingly shaped by other constraints.
Nevertheless, large-scale data centres still retain a crucial role in the AI era. When it comes to AI, a more meaningful distinction is between training and inferencing workloads. Training workloads continue to favour large-scale “AI factories” because of their reliance on tightly coupled GPU clusters, high interconnect bandwidth, and throughput-optimised execution, all of which benefit from economies of scale in power and infrastructure. These behemoth facilities are typically located in remote regions where abundant land and lower electricity costs support industrial-scale compute clusters.
Inference workloads, by contrast, naturally gravitate towards lightweight edge facilities (physical or digital infrastructure located closer to where data is created or used) in proximity to end-users, where latency is lower while real-time system responsiveness is higher. As AI deployment scales, a growing share of compute is expected to shift towards edge environments. By 2030, industry estimates place this figure at 80% of total inference workloads being processed along the edge.
In short, the AI data centre landscape appears to be increasingly defined by two diverging demands — throughput and latency — which give rise to distinct compute architectures: large-scale training facilities on the one hand, and distributed edge inference on the other.
In this light, the AI grid can also be understood as an attempt to manage this tension. By acting as a coordination layer to link these opposing architectures, it bridges different stages of the AI lifecycle: from the creation of intelligence during training to its distribution via inference.
Companies at the forefront
The most mature articulation of the AI grid vision comes, unsurprisingly, from Nvidia. At its GPU Technology Conference (GTC) in March, the company released a blueprint for implementing the AI grid, built around its own product stack. A range of industry partnerships aligned in this direction has since been announced. Accordingly, the most concrete implementation of the AI grid so far positions Nvidia as both the de facto system architect as well as its primary enabler.
Alongside Nvidia, a broader set of players is also emerging across the AI grid universe.
This includes network infrastructure providers — telcos, multisystem operators (MSOs) and content delivery networks (CDNs) — that form the physical connectivity layer of the grid. These players have begun deploying infrastructure to support AI workloads. T-Mobile, for example, has begun piloting AI-RAN (Artificial Intelligence-Radio Access Network) in select cell sites, enabling portions of its network to handle AI traffic and lightweight inference workloads.
Equally critical are the AI infrastructure original equipment manufacturers (OEMs) such as Hewlett Packard Enterprise (HPE) and Dell. These companies operate as system integrators, translating Nvidia components into fully assembled systems that are ready for deployment within data centres and edge environments. In effect, they supply the “building blocks” for nodes in the grid. Coinciding with the unveiling of Nvidia’s AI grid at this year’s GTC, HPE introduced networking and server products for edge deployments, reinforcing the push towards an AI Grid architecture.
Above this sits the compute layer, where training and inference workloads are actually executed. This layer is dominated by hyperscalers, alongside a fast-emerging class of “neoclouds” built specifically around high-performance, AI-optimised infrastructure. Prominent neoclouds include CoreWeave, Nebius, Crusoe and Lambda.
At a broader level, however, the AI compute landscape remains overwhelmingly dominated by hyperscalers. Case in point: Amazon, Google, Meta, Microsoft and Oracle collectively accounted for 71% of global AI compute capacity in 2025. Consequently, mainstream success of the AI grid remains contingent on hyperscaler participation.
Capital behind the grid
Capital flows similarly reinforce the scale of the shift that is underway. The capital already committed to compute infrastructure points to an emerging industrial-scale buildout. The OpenAI-Oracle-Softbank-backed Stargate Project is a case in point, centred on a staggering US$500 billion capital commitment to build out 10gw of AI compute infrastructure in the US over the next four years.
Within the industry itself, Nvidia has emerged as a prominent strategic investor across the expanding AI ecosystem. In 2025 alone, the company took positions in more than 50 tech companies, a notable portion of which are tied to compute infrastructure. CoreWeave, for instance, has received substantial funding from Nvidia, including a recent US$2 billion equity investment in January. The same month, CoreWeave also secured US$1 billion from Jane Street, alongside a US$6 billion commitment from the trading firm to use CoreWeave’s compute capacity.
Indeed, beyond industry participants, institutional and venture capital interest has also been substantial. Within the neocloud segment, CoreWeave’s peers such as Crusoe and Lambda have also attracted significant investor backing, including ARK Invest, Tiger Global, Franklin Templeton and Mubadala Capital.
The emerging picture
Taken together, these converging developments point to a clear conclusion: The AI grid is already under construction. The demand is real, and the capital backing it is substantial. The industrial vision of AI is now rapidly taking shape.
