In 2024, the UK Competition and Markets Authority (CMA) reviewed no fewer than five partnerships between cloud service providers (CSPs) and GenAI foundational model (FM) developers. Some of these companies are household names like Amazon, Google, and Microsoft, while others are lesser-known companies like OpenAI (ChatGPT), Anthropic (Claude), Mistral, or Inflection.
For those looking for a quick summary, here it is. The regulator didn’t have jurisdiction to review Amazon/Anthropic, Google/Anthropic, and Microsoft/Mistral because the parties didn’t meet the turnover threshold (in other words, the target didn’t make enough revenue in the UK to trigger a notification). Microsoft-Inflection was cleared in phase one, and Microsoft-OpenAI has the dubious honor of being stuck in the longest pre-notification stage ever, since December 8, 2023.
However, the CMA’s decisions provide useful insights into the potential antitrust issues that may arise in the Generative Artificial Intelligence space.
No Jurisdiction Doesn’t Mean No Concerns
In three of the partnerships mentioned above, the CMA did not have jurisdiction, so there was no need for a full analysis of the agreements. However, not all partnerships are created equal, even if, at first glance, they appear similar.
Part of the regulator’s analysis focused on whether Amazon, Google, and Microsoft could exert material influence over the targets—Anthropic and Mistral in this case—to somehow gain control. The CMA examined various arrangements between the parties. For instance, it looked into whether the non-voting rights acquired by the CSP over the targets or the compute capacity provided could result in any type of control. But the CMA concluded that Google and Microsoft did not have control over Anthropic and Mistral.
However, the language used by the regulator was noticeably different when addressing Amazon/Anthropic. Despite examining arrangements that were nearly identical, the CMA concluded that…
“The variety of commercial arrangements entered into by firms makes it difficult to state categorically what will (or will not) constitute material influence.” “The CMA also considers that the other categories of agreements and rights set out in the description of the Partnership above (including those consultation and advice rights in relation to significant Anthropic business issues) are, in principle capable of contributing to a finding of material influence in certain circumstances”.
However, because the CMA lacked jurisdiction, there was no need to make a decision on this point. In fact, this partnership was not subject to further analysis due to the lack of jurisdiction, not because it didn’t constitute a relevant merger situation.
What is more, the CMA went a step further in Amazon/Anthropic to state some concerns.
“The CMA’s mergers intelligence function identified the Partnership as warranting an investigation, having found that there was a reasonable chance that the test for a reference to an in-depth phase 2 investigation would be met (ie a reasonable chance that an investigation would reveal that it is or may be the case that a relevant merger situation has been created, and the creation of that situation has resulted, or may be expected to result, in a substantial lessening of competition”.
The Regulator Eyes Cloud Services and Compute Infrastructure
Without access to the specific arrangements between the parties, it is impossible to determine why the CMA appears to have greater concerns about the Amazon deal compared to those involving Google or Microsoft.
However, some of the language in the decisions may provide a clue. Amazon is Anthropic’s primary cloud provider, supplying compute infrastructure (cloud services and chips) as well as distribution capacity. This highlights one of the regulator’s key concerns: the risk of FM developers becoming dependent on CSPs
“The CMA has previously explained that agreements to provide compute infrastructure to an FM developer as well as distribution agreements between an FM developer and a cloud service provider (CSP) may in certain circumstances result in an acquisition of material influence. This may occur where the conditions are such that the agreement creates a dependency on the compute supplier or CSP such that it enables it to influence materially the commercial policy of the FM developer “
While investors and industry analysts are focused on chips, to identify which companies may emerge as market leaders. For regulators, the focus is just one step downstream, in cloud computing. In the chip market, NVIDIA currently dominates the market, but it is followed by a good number of companies such as AMD, TSMC, Amazon, Qualcomm, Intel, and Google, among others.
However, in the cloud services sector, the market is largely limited to AWS, Microsoft Azure, and Google Cloud. Over the long term, it is likely that FM developers will become more dependent on CSPs than they are on chip manufacturers, as the supply of chips will continue increasing. Additionally, Amazon and Google are developing their own chips and integrating more services for FM developers into their ecosystems—practices that, under certain circumstances, could raise antitrust concerns.
Concerns on the provision of cloud services already led the CMA to open an investigation in 2023 and the findings of this probe are expected on August 2025.
Access to compute infrastructure could also be one of the key issues delaying the pre-notification talks for Microsoft/OpenAI. Microsoft serves as the exclusive cloud provider for OpenAI, which could lead the CMA to suspect that Microsoft might have significant control over OpenAI. The situation is further complicated by the fact that Microsoft hired OpenAI’s top executives, reinforcing the possibility that this partnership could constitute a “relevant merger situation.”
Note for investors
What is the CMA telling us? That cloud services are just as critical to the development of the GenAI space as the chips being used. As seen in other sectors, it’s not always the best product that prevails but rather the company that offers the most compelling package or the strongest ecosystem.
While NVIDIA’s dominance in hardware for AI models is undeniable at present, it is not guaranteed to last. AWS, Google, and Microsoft are fully aware of their critical role in providing cloud services and distributing FMs. The CMA’s decisions suggest that these companies may seek to leverage their positions by integrating their own chips as deeply as possible within their ecosystems.
Amazon serves as a clear example. Its arrangements with Anthropic ensure that the latter uses Amazon’s latest chips in their models, providing Amazon with valuable feedback to continue upgrading its chips and enhancing other services within its ecosystem. Notably, Apple is also utilizing Amazon’s chips.
Amazon’s latest chips, Inferentia2 and Trainium2, are reportedly 30-40% cheaper than NVIDIA’s while maintaining high performance. However, one of their best features, not surprisingly, is not the price, but the seamless integration with the AWS ecosystem.
Google is also developing chips capable of performing at large scale. Google’s TPU v4 and v5, according to analysts, offer impressive performance for large-scale AI workloads while also offering a more cost effective solution than NVIDIA. Not surprisingly, Google is also integrating its chips in its own ecosystem.
