Context: Governmental agencies are increasingly research the field of Generative Artificial Intelligence (March 13, 2024 ai fray article).
What’s new: The UK’s Competition & Market Authority (CMA) released an “update paper” on AI Foundation Models (PDF) on Thursday, just in time for a speech by CMA CEO Sarah Cardell at the American Bar Association’s (ABA) big antitrust event (speaker’s notes on CMA website). The UK’s competition regulator has outlined certain principles of a competitive market, voiced fears over “winner takes all” dynamics previously seen in technology markets, and further to its report, the agency is now examining certain dynamics more closely. Without stating a specific competition issue relating to it, the CMA is taking an interest in Microsoft’s partnership with OpenAI, for example.
Direct impact: No antitrust fines or similar measures are likely to be ordered in the near term. The CMA has primarily reinforced its “we are watching you” message to industry. While the CMA’s general concerns over dynamics previously seen in technology markets are legitimate, the update paper, on top of generally recognizing the enormous benefits of AI innovation, summarizes trends that could also be interpreted as signs of fierce competition among large technology corporations that creates opportunities for newcomers.
Wider ramifications: In the current climate, the world’s competition regulators almost appear to be engaged in a race to the first major AI case they could bring. This is definitely the time for antitrust authorities to familiarize themselves with AI and to keep an eye on the market, but there is now a risk of hyperactivity clouding judgment.
The most basic questions to ask oneself with respect to fears over a concentration of market power are these:
- What is the ideal number of parties competing in a given market? It’s certainly never one and hardly two, but the answer is not “the more, the merrier” either.
- Is fierce competition among large players the realistic best-case scenario or would it be realistic (and, in that case, obviously desirable) to have even little companies compete in certain fields?
In this respect, AI is not a single market, but as the CMA and other government agencies have explained, there is a value chain and there are various adjacent markets. Cloud computing resources and chips are the two most commonly cited key inputs that enable AI services.
The CMA deserves credit for focusing not only on inputs but also mentioning “access points” and, in that context, the mobile duopoly (Apple and Google). The launch of the next iPhone may already shift some people’s focus to that aspect of competition in AI.
There are also other good things to be said about the CMA’s paper. But there are also some questions to be raised:
- The CMA notes that according to a Stanford paper, the number of Foundation Models (FMs) has grown by more than 120 to over 330 just between the CMA’s initial AI FM report in September 2023 and March 2024. The CMA takes note of that development, but doesn’t appear to ask itself the question of whether this means that there actually is a lot of innovation and competition, and that this may be a time where so much is in flux that regulators could easily pick the wrong target(s).
- Figure 5 on page 18 of the CMA’s update report got a lot of attention on social media. It’s a chart that shows relationships between “GAMMAN” (Google, Amazon, Meta, Microsoft, Apple, Nvidia; though Apple is going it alone so far) and more than a dozen FM developers. The lines connecting the four large players that have formed many such partnerships and the FM companies show that many of the latter have multiple such alliances. Antrophic works with Google and Amazon; Cohere with Google, Amazon and Nvidia; Falcon with Google, Microsoft and Amazon; Mistral with Google, Amazon, Microsoft and Nvidia; and so forth. That, too, could be viewed as a sign of positive dynamics. Clearly, FM makers have choice.
- There are only three cloud hyperscalers in the U.S. and Europe: Amazon, Microsoft, Google. The CMA voices concern over that, but the update paper does not talk about whether intense competition between those three may be sufficient to ensure that neither one can make supracompetitive demands. There is also no discussion of whether various cloud service providers that are smaller than those three, yet sizeable (such as Germany’s IONOS and France’s OVHcloud) may also be capable of meeting the needs of at least a significant number of FM companies, which would mean even more competition than just among the three largest ones.
The CMA’s paper is not the first one and likely won’t be the last of its kind to suggest between the lines that traditional competition (including, but not limited to, merger control) frameworks may not be fit to safeguard competition in AI-related markets. It is possible that some regulators will go to lawmakers in the near term and ask for additional tools.
One cannot blame the CMA for recognizing what would have been a past mistake in light of its current concerns, but it’s a fact that Google got to acquire a UK company named DeepMind ten years ago that is now at the heart of Google’s AI strategy.