Andrea Asoni, Chiara Farronato, Matteo Foschi, and Oliver Latham examined recent industry developments and surveyed business leaders to determine whether generative AI markets will remain competitive. Their findings offer managers and policymakers valuable insights with which to navigate the growing GenAI market.
Web 2.0 is generally defined as the period between 2000 and 2010 characterized by a shift from static websites to dynamic, user-driven platforms. See one of the seminal contributions in defining Web 2.0 by O’Reilly, Tim “What is Web 2.0: Design Patterns and Business Models for the Next Generation of Software.” Communication & Strategy No. 1 (2007).
See, for example, the current FTC lawsuit against Meta challenging their acquisitions of Instagram in 2012 and WhatsApp in 2014, or the DOJ lawsuits against Google for monopolizing digital advertising and search.
5.
See Rochet, Jean-Charles, and Jean Tirole. “Two‐sided markets: a progress report.” The RAND journal of economics37.3 (2006) – hereafter and Parker, Geoffrey G., and Marshall W. Van Alstyne. “Two-sided network effects: A theory of information product design.” Management science 51.10 (2005). Network effects and tipping had originally been identified in the literature much earlier. For example, see the classic work of Katz, Michael L, and Carl Shapiro. “Network externalities, competition, and compatibility”, American Economic Review (1985).
6.
See Hagiu, Andrei, and Julian Wright. “Data‐enabled learning, network effects, and competitive advantage.” The RAND Journal of Economics54.4 (2023).
7.
See Bergemann, Dirk, and Alessandro Bonatti. “Markets for information: An introduction.” Annual Review of Economics11.1 (2019).
For example, the UK Competition and Markets Authority (CMA) launched its initial review of the Foundation Model space in May 2023 to better understand how competitive dynamics were unfolding and whether early advantages could lead to long-term market power.
The MMLU is a benchmark that was widely used to assess model’s multitasking capabilities across diverse subject including humanities, STEM, and others. Through 2024, the MMLU benchmark remained one of the most widely used tools to evaluate new models. Over time, this benchmark has become “saturated” (in the sense that almost all current models achieve high scores) and attention has shifted to different measures of success – such as MMLU-Pro, a more demand version of the MMLU benchmark.
Deep Learning is recognized as a pivotal change in the evolution of AI – see for example https://epoch.ai/data/notable-ai-models. See also Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. “Attention is all you need.” Advances in neural information processing systems 30 (2017)
15.
For a related discussion of the risk of tipping in foundation model focusing in particular on the role of data feedback loops see Hagiu, Andrei, and Julian Wright. “Artificial intelligence and competition policy”, International Journal of Industrial Organization. (2025).
16.
See Shumailov, Ilia, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. “The curse of recursion: Training on generated data makes models forget.” arXiv preprint arXiv:2305.17493 (2023).
Data on the compute and application layers is from Pitchbook. Seehttps://pitchbook.com/(accessed May 2025). The criterion to select computing firms is {“ai neocloud” (keyword) AND “cloud” (keyword) AND “SaaS” (vertical)}. The criterion to select applications is {“Artificial Intelligence & Machine Learning” (vertical) AND “artificial intelligence” (keyword) AND “platform” (keyword) AND “developer” OR “Artificial Intelligence & Machine Learning” (vertical) AND “artificial intelligence” (keyword) AND “software” (keyword) AND “developer” (keyword) AND $25 million or more raised or in annual revenue}. This latter data was further filtered to exclude companies that are currently out of business. Data on LLMs and ML hardware is based on Epoch AI,https://epoch.ai/data accessed April 2025.
21.
Consistent with this, a recent event by the startup incubator YCombinator characterized the AI stack as follows: “your start up is just a GPT wrapper, OpenAI is a Nvidia wrapper, Nvidia is a TSMC wrapper, TSMC is an ASML wrapper, ASML is a Zeiss wrapper, Zeiss is a sand wrapper.”
Antón, Miguel, Florian Ederer, Mireia Giné, and Martin Schmalz. “Innovation: the bright side of common ownership?” Management Science 71.5 (2025)
25.
The study was reviewed and deemed exempt from IRB oversight by Harvard’s Institutional Review Board. The survey tools are available upon request. The survey was sent to potential participants (individuals who signed up to the mailing list of the Digital Data Design Institute at Harvard) on April 29, 2025. We exclude respondents who were deemed ineligible and those who did not complete the survey in its entirety. The data described in this paper are all valid answers received by June 3,2025.
