Compute for Gen AI: Assessment of competitive conditions
Some competition authorities and policymakers, such as the European Commission and the UK’s Competition and Markets Authority (CMA), have raised concerns about competitive dynamics in the provision of compute for Gen AI. In particular, there is a concern that a large share of compute capacity is concentrated in the hands of a few vertically integrated providers, potentially limiting access for third parties and hindering Gen AI adoption. To date, concerns have been framed as forward-looking risks that warrant continued monitoring as the market develops.
Against this backdrop, Amazon Web Services (AWS) has commissioned Copenhagen Economics to examine, based on available evidence, (i) whether providers of compute for Gen AI operate in a competitive environment and (ii) how expected developments in the sector may affect competition going forward.
In this report, we provide a preliminary assessment based on publicly available information. We assess whether market outcomes, such as trends relating to market entry, growth, investment, and pricing, are broadly consistent with a well-functioning market. Although available evidence is not granular enough to capture potential differences across workloads or buyer types, it is sufficient to support a high-level assessment. The research underlying this report was carried out between April and June 2025. Given the fast-moving nature of this market, subsequent developments may not be fully reflected.
This report was commissioned by Amazon Web Services.
In brief
Nordic productivity growth has fallen behind the global frontier: a 14 per cent lead over the US two decades ago has narrowed to near parity, and on current trends the Nordics risk falling behind within the next decade.
100+ providersCompute providers identified for Gen AI, including AI-specialised entrants such as CoreWeave, Cerebras, and Lambda. A further 48 on-premises providers had integrated NVIDIA accelerators as of May 2025.
93%The share of AI developers and deployers who multi-source their Gen AI workloads, either across cloud providers or by combining on-premises with cloud.
74%The fall in the inflation-adjusted price per FP32 FLOP since 2019. GPU rental prices commonly used for Gen AI fell from peaks of USD 8 in 2023 to under USD 2 in 2024.
USD 1 trillionThe forecast annual worldwide data centre investment by 2029, around half of it supporting AI workloads.
39% a yearEstimated average annual growth in demand for Gen AI compute between 2023 and 2030.
Commissioned by Amazon Web Services. Published August 2025. Research conducted April to June 2025. Global, with European and US policy context.
contact us
Click below to read more about our TMT & Digital service.
related Links
Download the Executive Visual Summary here.
About this analysis
Copenhagen Economics examined whether providers of compute for generative AI operate in a competitive environment, and how expected developments may affect competition in future. The assessment works from market outcomes, namely entry, growth, investment and pricing, and asks whether these are broadly consistent with a well-functioning market. It responds to concerns raised by the European Commission and the UK Competition and Markets Authority about compute capacity concentrated among a few vertically integrated providers.
The authors describe this as a preliminary assessment based on publicly available information. The evidence is not granular enough to capture differences between workload types, namely pre-training, fine-tuning and inference, or between buyer types such as developers and deployers, and the report states that further research would be needed to do so. Research was carried out between April and June 2025, and the report notes that subsequent developments in a fast-moving market may not be reflected. Revenue growth rates across providers are not directly comparable, given differing baselines and service portfolios. Copenhagen Economics states that the client chooses the research question and that findings and conclusions are its own.