The competitiveness of the artificial intelligence market at first glance masks how investment arrangements and partnerships between the largest players risks undermining their incentives to compete. Regulators must continue to monitor these arrangements for anticompetitive effects, writes Shishene Jing.
The artificial intelligence market, defined by large language models (LLMs) and the chatbots built on them, is a fast-moving market with a coterie of large players: OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, xAI’s Grok, Meta’s Llama, Microsoft’s Copilot, and China’s Deepseek. Economists such as Jason Furman have characterized the LLM market as “fiercely competitive.”
However, the activity of the market hides how these and ancillary companies connected to the AI boom, like Nvidia, have entered into a knot of investment arrangements bringing their operations together. The agreements risk the benefits of an otherwise competitive market.
In their report on these partnerships, the Federal Trade Commission found that Microsoft’s investment in OpenAI and Amazon’s investment in Anthropic—to name just two— were in some sense de facto “tying” capital investments to commitments from the LLM developers to spend on the investors’ own cloud products, Azure and AWS, respectively. This effectively made it difficult for competing cloud service providers without the giant balance sheets of Microsoft and Amazon to compete for the LLM providers’ business. OpenAI even threatened antitrust action for what it perceived to be too-restrictive cloud spend requirements. OpenAI has since signed a cloud deal with Amazon, which has similarly agreed to invest billions in the company
As the development of AI requires increasing amounts of capital, cloud providers that combine financing with exclusivity rights may reshape market structure without engaging in traditional mergers or explicit tying agreements. Failure to prevent anticompetitive tying arrangements may enable dominant cloud platforms to entrench their position while raising barriers to entry for smaller competitors like Coreweave, who do not necessarily provide lower quality or more expensive services but simply lack the capital to enter these deals.
These deals have a trickle-down effect. Applications built on OpenAI and Anthropic may develop their products to align with the cloud infrastructure of their choice LLM. Foreclosing smaller cloud providers from access to OpenAI and Anthropic may mean that all AI markets are effectively tied to the big cloud providers, strengthening their hold on cloud compute (processing power) and the AI markets. These concerns are now exacerbated by SpaceX’s purchase of Cursor which competes head to head with Claude Code.
Two recent arrangements illustrate these concerns. Nvidia recently became a major investor in various LLMs, including OpenAI and Grok. Nvidia is both an investor and a producer of a critical input, AI chips, which power the training of LLMs. This dual role poses a conflict of interest, as it could incentivize Nvidia to prioritize providing chips to customers in which it has invested over those customers’ competitors. This could make it harder for smaller LLMs, including new, non-profit, open-source competitors, to develop and compete. Second, SpaceX’s most recent compute deal with Anthropic poses questions concerning potential lessening of competition when a competitor provides a critical input to a rival.
The case of Nvidia
Nvidia is simultaneously the dominant supplier of the most critical input to LLM development—graphics processing units, or GPUs—and a substantial equity investor in multiple frontier AI laboratories. Nvidia has taken investment positions in OpenAI, xAI, and Coreweave. This dual role creates potential conflicts of interest. As an investor, Nvidia has financial interests in the success of its portfolio companies, and those interests may diverge from its interests as a chip supplier when portfolio companies compete with non-portfolio companies who also are Nvidia customers. The question, then, is whether Nvidia has the incentive and the ability to act on these interests by selectively favoring portfolio companies in chip supply allocation, prioritization, or pricing. While no public evidence suggests this has occurred, the legal duty to maximize Nvidia’s shareholder’s interests would make such conduct not just incentivized but required.
This concern is sharpened by current market conditions. The demand for high-end AI-training GPUs—particularly Nvidia’s H100 and its successors in the Hopper and Blackwell architectures—substantially exceeds supply. In a supply-constrained market, decisions on whom to prioritize in the sale of chips carry significant consequences for competition. A company that can train its models faster and with greater access to compute will produce better-performing systems. In turn, those systems are more likely to achieve commercial adoption and generate revenue.
A decision to exclude non-portfolio customers could violate the essential facilities doctrine and, more directly, the refusal to deal cases under Section 2 of the Clayton Act. Per Aspen Skiing Co. v. Aspen Highlands Skiing Corp., 472 US 585 (1985) (“Aspen Skiing”), a monopolist may incur antitrust liability for terminating a prior practice of business with a competitor where that termination reflects a sacrifice of short-term profits in service of a long-term strategy to exclude the competitor from the market. Under the essential facilities doctrine, a monopolist who denies access to an essential market input it controls may constitute illegal monopolization.
However, more recent Supreme Court precedent, particularly Verizon v. Trinko, 540 U.S. 398 (2004) (“Trinko”) substantially cabined both the essential facilities doctrine and the Aspen Skiing line. The Trinko decision expressed skepticism that monopolists have a duty to deal equally with all parties and emphasized the risks of the judiciary regulating prices (Nvidia could, in theory, charge non-portfolio companies more). A refusal-to-deal liability now requires that plaintiffs show that the monopolist’s conduct involves a sacrifice of profitable dealing that can only be explained by the goal of harming competitors.
There is also a nascent concern that Nvidia’s tying arrangement could harm open-source and nonprofit competitors. Big Tech and well-funded startups are not the only players developing LLMs. Academic institutions, nonprofit research organizations, and open-source development communities like the Allen Institute for AI and EleutherAI are also creating AI models based on principles of privacy, transparency, and explicit pro-social benefits. If Nvidia forecloses access to its chips to non-portfolio companies and entities, it could harm these efforts to develop AI not in the hands of Big Tech.
The case of Anthropic and SpaceX
In April, Anthropic entered into a $45 billion compute supply arrangement with SpaceX, the aerospace and satellite company controlled by Elon Musk. The deal involves SpaceX providing compute infrastructure—presumably through its Stargate or similar compute initiatives—to support Anthropic’s model training and inference operations. This arrangement is notable because Musk is the founder and principal owner of xAI, which he recently “dissolved” into SpaceX as SpaceXai, Anthropic’s direct competitor in the development and commercialization of frontier AI models.
The competitive relationship between Anthropic and xAI is direct and acknowledged. Both companies develop LLMs aimed at similar enterprise and consumer markets; both are racing to deploy frontier-class AI systems; and both compete for the same pool of AI researchers, the same enterprise customers, and arguably the same pool of regulatory goodwill in a period when AI governance questions are contested. xAI’s Grok models and Anthropic’s Claude models are, in the clearest sense, competing products.
Against this background, the decision of SpaceX to enter a contract to supply its competitor with an essential input raises the concern that SpaceX can acquire informational advantages and potential leverage over Anthropic. It also raises the possibility of SpaceX losing its incentive to compete with Anthropic if it can benefit from Anthropic’s success as a business partner.
A compute provider necessarily has access to information about its customers’ utilization, training run schedules, model architecture choices, and research priorities. The concern is not just that SpaceX could collect and pass this information on to its xAI subsidiary. If Anthropic becomes dependent on SpaceX’s compute infrastructure—and typically, this is what happens when a company builds upon capital-intensive infrastructure it does not own—SpaceX would acquire leverage over Anthropic’s operations. Degraded service, capacity constraints, or pricing changes could impose competitive costs on Anthropic at critical junctures. The potential for this leverage, even if never exercised, may itself distort Anthropic’s behaviors. This is roughly analogous to merchants relying on Amazon’s platform to sell its products while competing against Amazon’s production and sales of the same products. In the alternative, generating significant revenue from Anthropic as a customer may incentivize SpaceXai to avoid head-to-head competition with a leading lab that would benefit consumers but damage their revenue stream.
Conclusion
The cases of Nvidia’s investments in OpenAI and xAI and Anthropic’s deal with SpaceX reveal how the proclaimed competitiveness of the AI market is riddled with arrangements that jeopardize this dynamism. Analysis of the AI market often treats the LLM providers as the primary arena of competition, but this competition is also competition for access to compute, capital, and talent. Collaborations and deals that tie AI companies, their suppliers, and customers together distort competition.
Antitrust analysis must continue to monitor how these deals between direct competitors and partners along the supply chain risk foreclosing access to essential inputs to the harm of competitors and distorting the behavior of direct competitors.
Author comment: the author thanks Robert Lande, Neil Averitt, John Donohue, and Jed Rakoff for helpful comments.
Author disclosure: The author worked on the 6(b) investigation into the artificial intelligence industry while working for the Federal Trade Commission. While at the FTC, she also helped investigate other Big Tech and AI companies, the specifics of which are not public. In private practice, she has done work on the AI and adjacent markets, including a three-month, non-public, confidential role for a company that sells GPUs. She holds stock in Nvidia and Microsoft. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
