Last week we discussed the CMA’s views that the foundation model (FM) development and distribution market (‘the upstream market’) could tip either way into the established pattern of Big Tech dominance or into an innovative, competitive market. This week we look at the CMA’s views about the potential competitive state of downstream markets for consumer facing applications, products and services which incorporate, are built with or are ‘powered by’ FMs.
How FMs are used downstream
AI can be used in a wide variety of ways in downstream markets, with competitive overlap between them:
- offered on a standalone basis: for example, chatbots can be used to search the web as an alternative to search engines, such as Perplexity AI (“where knowledge begins). Equally, existing search engines also are incorporating AI, such as Microsoft’s Bing.
- integrated into existing products: Google, Microsoft, Adobe, and Slack are all integrating FM-based features into their existing productivity software services. Equally, new entrants are offering their own competing ‘AI-powered personal research assistants’ such Notion AI.
- as an add-on feature to existing product: for example, Duolingo, one of the world’s leading online language learning platforms, uses ChatGPT to create immersive role-playing experiences: the waiter in the Paris café on which you test your French will have their own personality and backstory that you can draw out in your chat.
How downstream firms access FMs
While ‘build your own’ FM is an option for the better resourced downstream firms, most will access a third party FM developed and distributed by an upstream FM provider in one of the following ways:
- partner with a FM provider to enhance an existing FM – usually to train a generalised FM model to a specific domain, such as in health care. The advantage for the downstream provider is that it may own the IP rights to the domains-specific version of the FM, but this can be expensive option: BCG estimates domain finetuning costs range from $10k-$100k.
- buy API access to a third party FM and FM deployment tools. This is the business model used by Duolingo to backend ChatGPT that brings you that immersive Paris Café experience. This avoids the development costs of an FM, but can create deep dependence on the FM provider.
- build a third party plug-in to an FM. A downstream firm can develop a plug-in (the plug-in provider) to augment an FM (plug-in host). For example, OpenTable’s ChatGPT plug-in allows a user to ask ChatGPT “tell m the best Italian restaurant in Chelsea for 6 people next Tuesday”, ChatGPT ‘interrogates’ Open Table’s plug-in, which responds to the user through ChatGPT with recommendations, reviews and table availability etc and then can make the reservation with the chosen eatery.
What are the prospects for FM downstream competition
The CMA made some interesting observations about the “potentially distinctive” characteristics of FMs in its competitive impact which could mean that the competitive outcomes could go either way.
The CMA noted that it currently appeared fairly easy for downstream firms to switch between the multiple options for access to FMs outlined above:
“These options are generally available through flexible pricing or ‘try before you buy’ schemes, making it relatively easy for downstream firms to experiment with alternative solutions before committing to one…Stakeholders told us that downstream firms currently find it relatively easy to switch (all deployment options) and multi-home (e.g., plug-ins) between different FM providers. These factors should, other things being equal, lead to more intense competition between rival FM providers at the upstream level. This upstream competition should, in turn, ensure a wide range of easy and affordable FM deployment options for downstream firms, which can also drive competition in downstream FM services.”
But this could all change if there was concentration in upstream FM development: e.g. if open sourced models fell away as an alternative to closed-source models.
Incumbency advantage with consumers
The CMA noted that in technology markets consumer choices can be strongly influenced by how services are presented (called ‘choice architecture’). This could be a tool of incumbency to embed their own AI as the primary or the default choice across their large customer bases. But the CMA noted that it was unclear how choice architecture would be used in FM-driven products. How relevant is ranking etc when the way users and FM interact is specific answers to specific questions?
The CMA also posited that Incumbents also might be able to exploit concerns over AI errors, such as ‘hallucinations’, to leverage brand recognition and consumer trust. But equally, the CMA notes that ChatGPT was the fastest growing app in history, despite the fact that Open AI had little existing market presence.
As such, “it is unclear whether any incumbency advantages are sufficiently strong to outweigh consumers’ apparent willingness to try out and eventually switch to FM services as a result of the FM ‘hype’.”
Economies of scale
Access to data from their large, embedded customer bases (called ‘data feedback loops’) has been identified as a major advantage for incumbents for building and refining their products in other technology markets.
However, the CMA considered that “[t]he significance of these effects will also depend on the specific situation in which FMs are used.” As in ‘traditional’ search, AI search is likely to benefit from data feedback loops because ‘outputs one consumer finds helpful is a good indication that other consumers using a similar query may also find it helpful.” But given that the ‘magic’ of AI is its ability to produce specific answers in response to specific (even idiosyncratic) questions from indiviudal users, aggregating data from many users may not be of much benefit.
More realistically, the CMA thought there was a risk of a two-sided network effect emerging with plug-ins where the more plug-ins a particular plug-in host can offer, the more consumers may be drawn to that host, which in turn attracts more plug-ins. As result, “plug-in hosts could become ‘app store’ style platforms with market power with the ability to foreclose plug-in customers.”
Bundling and tying
The CMA noted that consumers are likely to be able to access FMs in different ways, some standalone and some bundled:
“In future, consumers could choose their preferred FM at the point at which they buy a new phone or computer (either pre-installed or as a prompted choice), as part of their browser (as search engines are currently distributed), or as a standalone application (in which case app stores may be important).”
But it is also possible consumers may prefer an integrated ecosystem of FM services and non-FM services. Not only may this be more convenient, but “the ecosystem may offer a highly customised service, based on rich data on how the consumer has interacted with the ecosystem over time.”
The CMA identified a number of ways that upstream FM developers could restrict competition downstream:
- FM developers such as Google and Open AI use their own FMs in their customer facing apps, and also supply both complete FMs and developer tools to third party downstream competitors.
- Plug-in providers need to build to the interfaces provided by the plug-in hosts, such as a natural language interface (e.g. achatbot) and these interfaces can determine the user-friendliness of their plug-in. At the same time, the plug-in host may compete with its own plug-in, and could access a better, ‘internal’ interface.
- Cloud computing providers, on the one hand, can give downstream FM providers with commercial access to needed computing power otherwise beyond their reach but, on the other hand, these cloud providers can be competing against their customers with their own downstream FMs.
- Partnerships may be used by a vertically integrated firm with market power to impose restrictive terms in a partnership agreement that prevented that other firm from competing effectively with it in downstream markets.
The CMA thought that the significance of vertical integration for FM markets would depend on a number of factors:
- FM monetisation strategies are still evolving: ‘[i]f most monetisation is at the upstream level, for example, there could be limited incentive to foreclose rivals in downstream FM services.”
- the availability of alternative ways to fund FM development e.g. through VC funding.
- if smaller upstream FMs become competitive with larger ones, downstream firms could find it easier to build models from scratch or fine-tune existing models for specific applications.
- if generative FMs does not generalise to the point where significant adaptation is not required for each use case or each domain (e.g. medical vs CRM use).
Much as the CMA said of upstream markets, while a risk, it is not a foregone conclusion that Big Tech will dominate downstream AI markets in the same way it came to dominate earlier technology markets. The CMA said that, for example, it is difficult to know whether web-inquiry chatbots will be a substitute or a compliment for traditional search engines:
“New FM services have the potential to disrupt even long-standing market positions. For example, Open Ai’s ChatGPT chatbot has the potential to disrupt Google’s position in search. But it is also possible that things go the other way, and FMs reinforce existing incumbent firms’ market positions. For instance, the adoption of FMs by leading search engine providers could strengthen their positions in online search, as they may be best placed to develop and implement these new technologies effectively.”
Next week we examine the CMA’s views on the consumer protection issues arising from FMs.
Read more: AI Foundation Models: Initial Report