Rapid advances in artificial intelligence (AI) are reshaping economies and societies worldwide. While AI carries the potential for substantial productivity and welfare gains, it also presents significant risks, including job displacement, misinformation, cybersecurity, privacy breaches, and market concentration (Ben-Ishai et al. 2024, Comunale and Manera 2024). The dynamism of AI markets and the accessibility of cheaper and better-performing AI models to AI-adopting firms are critical conditions for the broad-based adoption of AI and long-term productivity gains (Acemoglu 2025, Filippucci et al. 2024, Aghion et al. 2024).
With the success of ChatGPT, many observers have feared that the excessive concentration of the technology within one or two leading US companies could reduce market dynamism (CMA 2024, Coeuré 2024, Cottier et al. 2023). This view was challenged by the release by the Chinese developer Deepseek’s of the R1 model in early 2025. This ‘Sputnik moment’ in AI (Acemoglu, 2025) stunned the world and highlighted several unexpected developments that have contributed to making AI more easily and more broadly accessible. It demonstrated that an almost unknown startup (The Guardian 2025) could train an AI model at the very top of AI capabilities at a fraction of the development costs of other leading models, publish it in ‘open-weight’,
and offer ten times cheaper access to users (Artificial Analysis 2025).
In this column, we provide novel empirical evidence to inform the debate on the state of competition in AI markets, primarily building on our paper (André et al. 2025) that collects novel data on generative AI prices and performance. The paper shows that during the past two years, the dominant position of digital incumbents in the AI supply chain (chips, models, distribution, and application) has neither curbed innovation nor prevented potential AI users from accessing better and cheaper AI models (Korinek and Vipra 2025, Hagiu and Wright 2025). The trends of better quality and more accessible AI create strong preconditions for adoption across various sectors (Bick et al. 2024), with significant implications for the economy and policymakers. While current trends give reasons for optimism about the diffusion of the technology, concerns remain about future concentration and market power, as happened with traditional digital markets.
Key trends in AI markets
Recent empirical evidence indicates a dynamic and competitive AI market landscape (André et al. 2025), benefiting companies that adopt AI technologies as an input in production, complementing and sometimes substituting for labour in the production of goods and services across several domains and a diverse set of cognitive tasks. The key trends in AI markets we document here column include:
1. A growing number of market players: The AI market has seen a surge in new entrants, including tech incumbents and specialised AI startups in model development (training of foundation models), provision (cloud infrastructure for inference), and applications (consumer-facing services).
2. A dynamic market environment: The pace of innovation in AI is extremely rapid, with continuous improvements in models that are more capable and reliable in an increasing number of tasks (including professional drafting, research assistance, software development, creation of image and videos, or real-time audio translations) and domains (science, law, finance, customer interactions, etc.). The current supply of AI offers a wide array of models, capabilities, and prices, allowing companies to choose their preferred quality and trade-off performance, costs, and privacy in their business operations.
3. Declining quality-adjusted prices: Advances in hardware and algorithms, as well as competitive pressure, have made AI models much more accessible. The best model in March 2023 (GPT-4) is now 1,000 times cheaper to access than two years ago. Our calculations show that the quality-adjusted AI price has decreased by 80% over the past two years. These trends include well-known text models (large language models) as well as specialised image or audio AI models.
New data and indicators to monitor AI markets
Our findings rely on an extensive data collection of AI foundation models available as-a-service on the market (accessible from the cloud to companies using AI in their production processes; see Bergemann et al. 2025 and Zhong 2025). In the period January 2023 to 2025, there were several indications of market dynamism in three segments of the AI value chain that we examined: AI model development, AI model provision from the cloud, and AI downstream applications.
First, the number of available AI foundation models has been rising exponentially (Figure 1), developed by an increasing number of companies and offering several interaction modalities (text, image, audio, or video). These results hold both for the number of cloud providers and downstream applications.
Figure 1 AI supply has been rising fast
Note: Developers refer to companies that train and optimise foundation AI models. The number of models refers to the number of active foundation models every month.
Source: André et al. (2025).
The rapid pace of innovation has pushed out the economic frontier of AI
AI has improved and become cheaper at a similar pace as earlier general-purpose technologies (Filippucci et al, 2025). Using common industry benchmarks
to evaluate the performance of AI models and collecting prices of AI from cloud providers, we constructed an ‘AI economic frontier’, identifying the best models each month in terms of the price-performance ratio (Figure 2). Results suggest that in the last two years, this AI economic frontier has shifted continuously towards lower prices and higher quality. Moreover, the developers and models that reach the frontier have been changing, with five to six players alternating at the frontier (OpenAI, Google, Meta, DeepSeek, Anthropic, Mistral, etc.) and around ten others following closely. So far, leading positions in AI development are oligopolistic but highly contestable.
Figure 2 The AI economic frontier shows the continuous improvements of AI
Note: Performance is defined by a normalised weighted performance index on industry benchmarks. Each dot represents the model with the best available price-performance trade-off within Text-to-Text models.
Source: André et al. (2025).
This variety of models at the frontier allows a broad range of demand to be served. Many users may not always need the best available models and would rather pay an order of magnitude less to access models that are ‘good enough’ and more easily specialised for specific tasks or preferences. In addition to the offer of closed models directly from the cloud, ‘open-weight’ models offer an option for cheaper (with no license fee), more transparent, and easily customisable models used outside of the public cloud environment.
The US has been pushing out the frontier, but other countries are in the race
Based on simulated market shares constructed using information on the supply of AI models (including price and quality), assumptions about demand for different AI market segments, and various scenarios involving switching costs and reputation, the US holds the leading position in AI development. In January 2025, the US market share for large language models is estimated to be between 86% (high switching cost scenario) and 59% (low switching cost scenario). China has been aggressively catching up since the second quarter of 2024, reaching simulated market shares of 5% (with high switching costs) and 36% (with low switching costs). The rest of the OECD countries combined (including Germany, France, the UK, and Canada) are further behind in AI text models (between 5% and 10%). However, they are in much better positions in specialised AI models (for example, image generation or speech generation), where their market share can jump above 50%, although in Europe and Canada, AI companies tend to be smaller, less well funded, and more specialised startups.
Figure 3 Simulated market shares in AI model development
Note: Simulated market share of AI foundation model revenues from AI as-a-Service per country of origin of the AI developing company under the baseline demand scenario and aggregating all modalities according to the formula in Annex C. In this scenario, 40% of AI demand is addressed to Text-to-Text models, 10% to Audio-to-Text and 50% to Text-to-Image.
Source: André et al. (2025).
AI is getting better, cheaper, and more accessible for firms to adopt
The upward shift of the AI economic frontier depicted in Figure 2 is reflected in our quality-adjusted AI price index, which has declined by an average of 80% over the past two years (Figure 4), primarily due to continued quality increases offered at similar or somewhat lower prices. On average, during the period, 30% of models at the frontier are replaced every month with updated versions. These trends are consistent across all types of models, although the amplitude and timing vary.
In sum, this evidence suggests that the gains from innovation in AI development have been shared with AI users in terms of lower quality-adjusted prices. AI users have also benefited from greater access to AI models via a widespread offer accessible through several cloud providers (for business use) and an increasing number of AI-powered consumer services (consumer-facing applications).
Figure 4 Quality-adjusted AI prices have fallen rapidly
Note: The index for each modality is a weighted sum of the index for each model segment. Each model segment is represented by its respective model at the AI Economic Frontier.
Source: André et al. (2025).
Why these trends matter
Dynamic AI markets are a necessary condition for the diffusion of AI across the economy, facilitating widespread AI adoption in various sectors and serving as a central determinant of long-term productivity gains from AI (Acemoglu 2025, Filippucci et al. 2024, Aghion et al. 2024). Our evidence so far suggests that the supply of AI has been more open than initially expected in various segments of the AI value chain, driving innovation and contributing to favourable conditions for broad AI adoption: lower prices, better quality, and broader accessibility.
Nonetheless, several uncertainties and risks persist about the future dynamism of AI markets. For instance, the capacity of incumbents to leverage existing compute infrastructure and user base in adjacent markets is high. Furthermore, the high concentration of the necessary inputs for AI development – data, compute, and talent – creates additional risks for long-term competition, which warrant continued research on the dynamism of AI markets.
References
Acemoglu, D (2025), “A Sputnik moment for AI?”, Project Syndicate, 4 February.
Comunale, M and A Manera (2024), “The Economic Impacts and the Regulation of AI: A Review of the Academic Literature and Policy Actions”, IMF Working Papers 2024(065).
Acemoglu, D (2025), “The simple macroeconomics of AI”, Economic Policy 40(121): 13-58.
Aghion, P and S Bunel (2024), “AI and Growth: Where do we stand”, unpublished manuscript.
André, C, M Betin P Gal and P Peltier (2025), “Developments in Artificial Intelligence markets: New indicators based on model characteristics, prices and providers”, OECD Economics Department Working Paper.
Artificial Analysis (2025), Q1 2025 state of AI report: Analysis of the AI landscape and the key trends shaping AI.
Ben-Ishai, G, J Dean, J Manyika, R Porat, H Varian and K Walker (2024), “AI and the Opportunity for Shared Prosperity: Lessons from the History of Technology and the Economy”, arXiv preprint arXiv:2401.09718.
Bergemann, D, A Bonatti and A Smolin (2025), “The Economics of Large Language Models: Token Allocation, Fine-Tuning, and Optimal Pricing”, arXiv preprint arXiv:2502.07736.
Bick, A, A Blandinand D J Deming (2024), “The rapid adoption of generative AI”, NBER Working Paper No. w32966.
CMA – Competition and Market Authority (2024), AI foundation models technical update report.
Coeuré, B (2024), “Comments on “The simple macroeconomics of AI” by Daron Acemoglu”.
Cottier, B, T Besiroglu and D Owen (2023), “Who is leading in AI? An analysis of industry AI research”, arXiv preprint arXiv:2312.00043.
Filippucci, F, P Gal and M Schief (2024), “Miracle or Myth? Assessing the macroeconomic productivity gains from Artificial Intelligence”, OECD Artificial Intelligence Papers, No. 29.
Filippucci, F, P Gal, K Laengle and M Schief (2025), “Macroeconomic productivity gains from Artificial Intelligence in G7 economies”, OECD Artificial Intelligence Papers No. 41
Hagiu, A and J Wright (2025), “Artificial intelligence and competition policy”, International Journal of Industrial Organization, 103134.
Korinek, A and J Vipra (2025), “Concentrating intelligence: Scaling and market structure in artificial intelligence”, Economic Policy 40(121): 225-256.
OECD (2024), “Artificial intelligence, data and competition”, OECD Artificial Intelligence Papers No. 18.
The Guardian (2025), “Who is behind Deepseek and how did it achieve its AI Sputnik moment?”, 28 January.
Zhong, H (2025), “Optimal Integration: Human, Machine, and Generative AI”, CEPR Discussion Paper No. 20330.