To date, merger inquiries appear to have been the CMA’s preferred tool for screening AI partnerships, with a view to guarding against risks to competition in the rapidly developing AI sector. However, as explored below, this approach presents challenges from a jurisdictional perspective.

On 23 July 2024, competition authorities from the EU, the US and the UK published a joint statement on competition in generative artificial intelligence (AI) foundation models¹ (FMs) and AI products (the Joint Statement). The Joint Statement suggested that the rapid evolution of generative AI in recent years means we could be at a “technological inflection point”, which requires vigilance from competition authorities in order to safeguard ”against tactics that could undermine fair competition”.

The Joint Statement is part of a backdrop in which regulators across the globe (and not only in the competition law sphere) are seeking to grapple with the regulatory challenges posed by AI. For its part, the CMA published an initial report on competition in AI Foundation Models in September 2023², and an update paper focusing on potential competition risk posed by Foundation Models in April 2024³.

What risks to competition might AI represent?

The Joint Statement notes that competition authorities have ”learnt from (their) experience with digital markets”, suggesting that competition authorities are keen to avoid the mistakes of the past – their perceived inaction on Big Tech mergers in particular, which may have contributed to the concentration of market power in the hands of a few very large players.

The Joint Statement identified three key risks to competition posed by the rapid growth in the use and power of AI.

1. Concentrated control of key inputs: The risk that a small group of undertakings might enjoy outsized influence over the future development of AI if they are the only suppliers of the industry’s “critical ingredients” (such as specialised chips and large scale datasets). This could limit the scope of disruptive innovation, or allow companies to shape innovation to their own advantage, at the expense of fair competition.
2. Entrenching or extending market power in AI-related markets. In particular, the Joint Statement highlighted the risk posed by firms with substantial market power at multiple levels related to the AI stack.
3. Arrangements (such as partnerships and investments) between key players in the AI space being used “to undermine or co-opt competitive threats and steer market outcomes in their favour”.

There are parallels here with some of the risks identified by the CMA in its reports into AI FMs, with the CMA noting in its April 2024 Update Paper that certain firms “may try to use partnerships and investments to quash competitive threats, even where it is uncertain whether those threats will materialise”. The CMA also recognised that such partnerships may not always fall within the scope of the merger control rules, but that it is important that competition agencies are “suitably equipped to take action where these arrangements give rise to competition concerns”.

Consequently, one of the key questions for the CMA (and other competition authorities) has been: what are the most effective tools to manage these risks?

What is the CMA doing to manage these risks?

To date, the primary tool the CMA has used in its attempts to intervene (rather than simply understand) the AI sector has been the opening of merger inquiries into the kind of “arrangements” identified in point 3 above. 

At present there are several ongoing CMA merger inquiries in the sector, including Microsoft/Open AI⁴, and Alphabet (Google LLC)/Anthropic⁵. These arrangements pose a jurisdictional challenge for the CMA (and other competition authorities seeking to assert jurisdiction), as they are not always structured as traditional acquisitions or mergers. 

Microsoft/Mistral⁶ involved a partnership between Microsoft and Mistral, a French-based FM developer founded in 2023.  Microsoft had invested €15 million into Mistral by way of convertible bonds (convertible to equity in the event of a future funding round) and entered into various partnership agreements with Mistral related to research and development and the use of Microsoft’s Azure cloud infrastructure for Mistral’s compute requirements and distribution of Mistral’s model. 

In order for a “relevant merger situation to have been created” under the Enterprise Act 2002 (EA02), two enterprises must either have ceased to be distinct, or will cease to be distinct as result of the transaction. In other words, control over a business (or part of it) must be acquired by one or more parties who did not previously have control. The ability to exercise “material influence” is the lowest level of control that may give rise to a relevant merger situation. In determining whether material influence has been acquired, the CMA focuses on the acquirer’s ability to influence materially the policy of the target business and its strategic direction. The CMA has previously been able to apply the concept of material influence to assert jurisdiction to acquisitions of minority stakes as low as 16% (as in Amazon/Deliveroo⁷), in light of factors suggesting that the acquirer had obtained the ability to influence materially the target’s strategic decision making.

However, in Microsoft/Mistral, the CMA concluded that the material influence threshold was not met as, among other things, Microsoft’s potential shareholding would be less than 1%, it did not have board representation, and the distribution and compute arrangements were non-exclusive agreements which did not lock Microsoft in as cloud computing supplier to Mistral or otherwise enable Microsoft to restrict Mistral’s commercial freedom (e.g. (by allowing Microsoft to determine or influence how Mistral exploits its intellectual property).

Why merger inquiries might be the CMA’s preferred route for screening of AI partnerships

In Microsoft/Mistral the CMA suggested that, in some circumstances, material influence could be acquired through a compute agreement or distribution agreement alone. It would, therefore, have been interesting to see what difference it would have made to the question of jurisdiction in that case had the compute and distribution agreements with Microsoft been exclusive. An assertion of jurisdiction by the CMA on this basis would be unprecedented, particularly in view of Microsoft’s de minimis equity investment and complete lack of presence at board level. 

Aside from merger enquiries, another option open to the CMA to address concerns about exclusive arrangements in the AI sector is the opening of a Chapter I or (where one or more of the companies concerned may be dominant in a relevant market) Chapter II investigation under the Competition Act 1998 (CA98). However, exclusivity in and of itself is not necessarily likely to harm competition, so any case of this nature would be unlikely to be straightforward for the CMA, not least because of the difficulties of demonstrating a potential foreclosure effect and harm to competition in (nascent) AI-related markets⁸. While such difficulties would also be present to some extent in a merger inquiry under the EA02, the CMA arguably has greater scope to deal with threats to future competition in merger inquiries though the application of concepts like risks to dynamic competition. 

CMA merger decisions are also less vulnerable to appeal than those under the CA98 - as merger appeals are subject to judicial review principles, rather than a review on the merits - and don’t carry the same adverse reputational impact for companies as CA98 investigations.

CA98 investigations may therefore be a somewhat inappropriate and/or heavy-handed tool for arrangements where the CMA is yet to form a strong view on whether there is likely to be a risk to competition, but simply wants to have the means to “screen” them for potential acquisitions or entrenchments of market power. One can see why the use of merger inquiries may be viewed by the CMA as the more attractive route to review these types of arrangements.

Microsoft/Inflection, and the concept of enterprises ceasing to be distinct 

In the absence of more conventional acquisitions or the taking of significant minority stakes, the CMA may seek to satisfy the control element of its jurisdictional test through a different route. Its recent decision in Microsoft/Inflection⁹ is arguably an example of this.

On 19 March 2024, Microsoft announced that it had hired several former Inflection employees, amounting to almost all of Inflection’s core team, including two of its co-founders: Mustafa Suleyman and Karén Simonyan. In addition, Microsoft also entered into a series of arrangements with Inflection including a non-exclusive licence in respect of Inflection’s intellectual property (the IP Licence). Unlike the Microsoft/Mistral case, there was no capital injection akin to an equity investment by Microsoft in Inflection itself. 

The CMA, in a summary clearance decision published on 4 September 2024, confirmed that it had jurisdiction to review the transaction on the basis that the above arrangements had resulted in two enterprises (Microsoft and at least part of Inflection's business) ceasing to be distinct, with Microsoft acquiring control over Inflection’s activities. According to the CMA, since the vast majority of Inflection’s core team had been hired by Microsoft, Microsoft had ”with them, acquired the team’s collective know-how of Inflection’s activities pre-Transaction to support and grow Microsoft’s AI activities.“ Additionally, the IP Licence entered into by the parties was “key to the value of the transaction” since, together with the hiring of the Inflection team, it meant that “Microsoft had substantively acquired (Inflection’s) pre-Transaction FM and chatbot development capabilities”.

In the context of businesses where certain staff are “at the core” of the business model, the CMA added that it considers that acquiring a team with relevant know-how, even without any other assets, may fall within the CMA’s jurisdiction. Whilst the CMA’s jurisdictional guidance states that a “transfer of … employees alone may be sufficient to constitute an enterprise”¹⁰, these types of situations have not attracted the attention of the CMA in the past. This intervention by the CMA is, therefore, potentially quite significant, and could signify increased scrutiny of transactions that may previously have been seen (rightly or wrongly) as falling outside the scope of the UK’s merger control regime.

The importance of the transfer of employees in this case appears to be that the transfer of the core team of employees, with know-how, meant that there was a high degree of economic continuity in the activities of the “business’” being acquired. A lot may depend on the nature of the sector in question, the number of employees who move over, whether their know-how is essential to the activities of the business they have left, and whether there is anything (such as intellectual property rights falling outside the transaction perimeter) preventing the hired employees from deploying that know-how. 

In the event, the CMA decided not to refer Microsoft/Inflection to Phase 2. The theories of harm examined by the CMA were essentially quite traditional horizontal unilateral effects theories, and its reasoning in this regard – which centred on Inflection’s marginal presence on the markets for FMs and chatbots – was unremarkable. The case is therefore notable only for its potential implications as to jurisdiction, in particular the expansive approach taken by the CMA as to what constitutes an enterprise.

Microsoft/Inflection approach likely not a panacea 

While Microsoft/Inflection could prove significant as regards the CMA’s approach to jurisdiction, the facts of the case – namely the hiring of a team of core personnel – will certainly not be present in every AI partnership. Despite its flexibility, particularly when compared to the “decisive influence” test that applies at EU level¹¹, the CMA is still likely to be restricted by the limits of the material influence test when looking at more “traditional” partnership arrangements – as well as the other elements needed to establish jurisdiction under the EA02.

Looking ahead, it will be interesting to see how the CMA approaches the question of jurisdiction in Google/Anthropic, in respect of which publicly available information suggests that Alphabet have agreed to invest up to $2 billion in Anthropic, increasing its existing circa 10% stake in the company. 

The CMA has recently decided that Amazon’s arrangements with Anthropic did not qualify as a relevant merger situation, on the basis that neither the share of supply test nor the turnover test under the EA02 were met¹². The CMA did not therefore need to reach a conclusion on whether material influence had been acquired by Amazon as a result of its reported $4 billion equity investment, alongside agreements relating to the use of Amazon Web Services’ cloud platform and the licensing of Anthropic’s models. However, had the CMA been examining the merger after the Digital Markets, Competition and Consumers Act (which introduces a new, “acquirer-focused” threshold¹³) enters into force towards the end of this year or early next year, the question of material influence would have been decisive.

In sum, it would appear that the key jurisdictional question of whether the material influence threshold has been met will continue to be of paramount importance, and there is no one-size-fits-all approach for the CMA as it continues to grapple with how to manage the risks to competition posed by the increasing importance of AI.

¹ The CMA has defined foundation models as “large, machine learning models trained on vast amounts of data”. 
² Available at AI Foundation Models: Initial report 
³ Available at AI Foundation Models: Update paper 
⁴ Microsoft/OpenAI partnership merger inquiry 
⁵ Alphabet Inc. (Google LLC)/Anthropic merger inquiry
⁶ Microsoft/Mistral, ME/7102/24
⁷ Amazon/Deliveroo merger inquiry 
⁸ After initially reviewing the Microsoft/Open AI partnership under EU merger control rules, the European Commission dropped that probe on the grounds that the EU control test had not been met. We understand that the Commission is currently still reviewing the arrangements, though, as at the time of writing, it has not opened an official investigation under Regulation 1/2003.
⁹ Summary of phase 1 decision 
¹⁰ Paragraph 4.9, Mergers guidance on the CMA’s jurisdiction and procedure (2024)
¹¹ Albeit it should be noted that the European Commission confirmed that it considered the Microsoft/Inflection transaction to constitute a concentration, and invited Member States to refer the transaction to it under Article 22 of the EU Merger Regulation. That referral recently had to be withdrawn by the relevant Member States, on the basis the Commission lacked the jurisdiction to review the transaction in light of the recent Court of Justice decision in Illumina/Grail (on which see our recent article)
¹² See Summary of phase 1 decision 
¹³ See our previous article for more information