Why Amazon Partnering With OpenAI and Anthropic Could Raise Antitrust Concerns

Last updated 17 seconds ago. Our resources are updated regularly but please keep in mind that links, programs, policies, and contact information do change.

Amazon holds investments in OpenAI and Anthropic, while both artificial intelligence (AI) developers have committed to buy large amounts of computing services from Amazon Web Services (AWS). Those relationships could raise antitrust concerns if they reduce the developers’ ability to change suppliers or compete independently.

Amazon’s Bedrock service offers models from Anthropic and OpenAI alongside its own Nova models. That makes the investor and supplier a competitor as well, even as its financing helps developers build capacity. The competition question turns on what the deals let the companies do, rather than the size of the investment alone.

What Amazon is buying, and what the developers are buying back

Amazon announced its OpenAI investment and commercial agreements on February 27, 2026, describing an investment of $50 billion, initially divided into $15 billion and a conditional $35 billion. Amazon’s quarterly filing reports that it invested $28.7 billion in OpenAI preferred stock during the six months ended June 30, 2026, then funded the remaining $21.3 billion after that date. Preferred stock generally provides financial preferences, such as dividend payments before those on common stock.

The commercial agreement expanded an existing $38 billion AWS arrangement by $100 billion over eight years, with OpenAI committing to approximately 2 gigawatts of capacity using Amazon’s Trainium chips. That is a commitment to purchase computing services, separate from Amazon’s purchase of OpenAI stock. Gigawatts are units of electrical power. The power measure describes the scale of the computing infrastructure, not a score for an AI model’s quality.

Anthropic announced its expanded Amazon computing agreement on Apr 20, 2026, committing more than $100 billion over ten years to AWS technologies and describing up to 5 gigawatts of new capacity to train and run Claude. Amazon describes AWS as Anthropic’s primary training and cloud provider for mission-critical workloads. Primary is a description of the supplier’s role, not a claim that every Anthropic workload must use Amazon.

The announced Anthropic investment included $5 billion immediately and up to another $20 billion tied to commercial milestones, in addition to the $8 billion Amazon had previously invested.

Amazon’s quarterly filing records two $5 billion purchases of Anthropic nonvoting preferred stock in the second quarter of 2026, one in Series G and one in Series H. Nonvoting means the shares do not carry ordinary shareholder votes, although separate contractual rights can still affect a partner’s influence.

The filing describes a financing facility initially capped at $20 billion, reduced to $15 billion by the Series H investment. Additional amounts become available for Anthropic to draw at its discretion as AWS meets computing-capacity delivery milestones. Those milestones connect further financing to the infrastructure the supplier has agreed to deliver.

Amazon’s earlier $8 billion in Anthropic convertible notes, invested from the third quarter of 2023 through the fourth quarter of 2025, was subject to an ownership cap that Amazon may elect to waive. Portions of those notes converted to nonvoting preferred stock in the first quarters of 2025 and 2026. The conversion allows financing initially provided as notes to become an ownership stake in shares. These disclosed provisions do not by themselves establish Amazon’s current ownership percentage or control over Anthropic’s board.

The OpenAI announcement names AWS the exclusive third-party cloud distribution provider for OpenAI Frontier, a platform that lets organizations build and manage teams of AI agents. The stated exclusivity attaches to Frontier and that distribution role, not to an assertion that Amazon exclusively supplies every OpenAI model or computing need.

The companies’ strongest argument is useful capacity

In the Federal Trade Commission (FTC) Office of Technology’s January 2025 FTC Staff Report on AI Partnerships & Investments 6(b) Study, staff describes cloud services for AI training and operation as expensive and capital-intensive, with billions spent by major providers. A partnership that supplies this capacity can address an operational constraint rather than merely rearrange ownership.

In its April 20, 2026 announcement, Anthropic said rapid demand had strained reliability and performance for free and paid Claude users, particularly during peak hours. Anthropic said the agreement would expand available capacity within three months and deliver nearly 1 gigawatt by the end of the year. The prediction itself is not a measurement of the improvement delivered.

Amazon CEO Andy Jassy says the company’s custom AI chips offer high performance at significantly lower cost for customers. Amazon also describes access to Anthropic’s Claude Platform through an existing AWS account, using the same access controls and monitoring without additional credentials, contracts or billing relationships. For a business already organized around AWS, that integration offers a different benefit from simply having another model to choose.

OpenAI and Amazon describe a Stateful Runtime Environment, a system designed to retain context and prior work while an AI application uses different software tools and data sources. The companies also describe customized OpenAI models for Amazon’s customer-facing applications, complementing Amazon’s Nova models. Their argument is that matching models to infrastructure can make applications more useful and easier to operate. These are the companies’ explanations of intended benefits, not independent findings that every promised performance gain has occurred.

A cloud commitment can make another supplier harder to choose

FTC staff identifies cloud commitments that require AI developers to spend a large portion of a cloud partner’s investment on that partner’s services. The staff report calls this circular spending and says it can reduce the investor’s potential exposure to losses. For the developer, the relevant distinction is between funding it can spend wherever it chooses and funding accompanied by an obligation to buy from the investor.

A long purchasing commitment can narrow the practical value of shopping around even when another provider is technically available. That possibility does not mean the current Amazon agreements prohibit every alternative, or that a large purchase contract necessarily harms competition. The question is what choice remains under the particular terms, and how much it costs to exercise that choice.

FTC staff says a developer using chips available only through a particular cloud provider may face both chip and cloud switching costs. The report describes cloud migration as potentially time- and capital-intensive, especially when a developer relies on specialized computing resources and managed services. Changing suppliers can consequently require engineering work as well as a new purchasing decision. A lower initial computing price and a higher later switching cost can coexist.

In its January 2025 study, FTC staff identified potential effects on computing-resource and engineering-talent access for both partner developers and developers outside the partnerships. The competitive concern extends beyond whether the funded company receives enough capacity to grow. It also asks whether competing developers have workable sources of the inputs they need, and whether the partnership changes that access.

The Justice Department and FTC’s 2023 Merger Guidelines say limiting rivals’ access becomes more feasible when few alternative suppliers exist, or when alternatives differ in quality, price or other characteristics. Evidence about available substitutes can therefore be more revealing than the size of an announced computing purchase. A contract that brings additional capacity and preserves realistic alternatives presents a different competition question from one that leaves rivals dependent on restricted supply.

Confidential information and investment incentives create another issue

FTC staff identifies sensitive information that cloud partners may obtain, including AI development methods, confidential chip co-design, partner finances, and customer usage and revenue figures. Sharing information can help engineers build an integrated product, while commercially sensitive information can also affect how firms compete. The concern requires attention to the information shared and the protections surrounding it, rather than an assumption that every technical collaboration is improper.

The federal merger guidelines explain that access to a partly owned firm’s nonpublic competitive information can facilitate coordinated or accommodating responses even without control over its decisions. That is a possible mechanism for competitive harm, not a finding that Amazon, Anthropic or OpenAI have exchanged information unlawfully. Public descriptions of a partnership do not answer every question about which teams can see another company’s plans.

The guidelines also explain that acquiring a minority stake in a rival can weaken an investor’s incentive to compete aggressively, because it may receive dividends or other shared revenue even when it loses business to that rival. The guidelines say this effect can arise without direct influence over the rival’s decision-making. Amazon’s own models make this an understandable question to examine, but the investment alone does not prove a reduced incentive in any particular product market.

Backing OpenAI and Anthropic is therefore neither an automatic antitrust violation nor an automatic guarantee that the two developers will compete independently. The analysis turns on the rights obtained, economic incentives and competitive effects of the actual relationships.

What U.S. antitrust law would require

The Department of Justice (DOJ) and FTC enforce federal competition laws that include the Sherman Act, Clayton Act and Federal Trade Commission Act. Section 7 of the Clayton Act addresses stock or asset acquisitions whose effect may substantially lessen competition or tend to create a monopoly. In plain terms, the acquisition question concerns a threat to competition, not whether an investment is impressive or whether a competitor dislikes it.

The Sherman Act also addresses agreements that unreasonably restrain trade, as well as monopolization and attempts or conspiracies to monopolize. The FTC explains that a partnership agreement can be lawful when its restraint is not unreasonable. Conduct inquiries examine how firms compete and deal with one another, separately from whether purchasing a stake threatens competition.

Obtaining a monopoly through superior products, innovation or business skill is lawful, according to FTC guidance, while exclusionary or predatory conduct can raise antitrust concerns.

The agencies’ Merger Guidelines apply the same competition standard to partial acquisitions as to other acquisitions, examining the parties’ relationship and incentives after the investment. For acquisitions involving an input rivals need, the guidelines examine the risk of limiting access, gaining sensitive information or discouraging rivals from investing. These frameworks identify evidence to investigate, not a shortcut from a headline about an AI partnership to a conclusion of liability.

The guidelines create no independent rights or obligations and say the agencies must apply their frameworks flexibly to each transaction’s facts. A sound assessment would connect specific contractual rights or conduct to effects in a defined area of competition, while considering the capacity and integration benefits the companies describe.

The Hart-Scott-Rodino premerger notification law requires certain voting-stock or asset acquisitions meeting its criteria to be reported before completion, subject to exemptions. A minority acquisition can fall within that process; a majority takeover is not the statute’s starting requirement. The reportability of a particular deal depends on its structure, applicable thresholds and exemptions, so the existence of an investment announcement does not establish that a specific filing was required or made.

FTC guidance describes an ordinary initial waiting period of 30 days and a shorter 15-day period for specified transactions. An agency request for additional information stops the waiting period until both parties substantially comply, after which an additional 30-day period generally runs, with 10-day exceptions for cash tender offers and specified bankruptcies. This is a notification and review procedure, separate from a final determination that an agreement complies with every antitrust law.

The notification statute expressly preserves later legal action even when an agency has acted, or failed to act, under that process. Expiration of a waiting period is therefore not permanent antitrust immunity.

The FTC study and the UK decision answer narrower questions

The FTC study examined Microsoft and OpenAI, Amazon and Anthropic, and Google and Anthropic. The FTC explains that its Section 6(b) authority permits wide-ranging studies without a specific law-enforcement purpose. An industry study can guide enforcement work without itself being an enforcement judgment against the companies it examines.

The staff report drew on respondent information through September 2024 and publicly available information through January 2025. It expressly disclaims any assessment of whether anyone engaged in illegal conduct, and says it is not a formal legal or economic analysis. It therefore cannot be treated as a ruling on the Amazon and OpenAI arrangements announced in February 2026.

In her January 17, 2025 statement responding to the study, then-FTC Commissioner Melissa Holyoak emphasized AI’s potential benefits for consumers and business effectiveness and counseled circumspect enforcement because agency decisions affect market incentives.

Outside the U.S. system, the United Kingdom’s Competition and Markets Authority decided on September 27, 2024 not to refer the Amazon and Anthropic partnership for further merger investigation because neither its turnover nor share-of-supply jurisdictional test was met. That jurisdictional outcome answers whether that partnership met the UK investigation thresholds at the time, rather than whether the later agreements satisfy U.S. antitrust law.

What would change the assessment

FTC staff says it aggregated or anonymized information to protect trade secrets and confidential commercial or financial material. The public report therefore cannot supply a complete account of each company’s contract rights or internal information safeguards. It would be premature to fill those gaps with assumed board control, automatic exclusivity across every model, or an allegation of information misuse.

A consequential new development would be a disclosed contract term that changes switching options, a public agency complaint specifying a competition theory, or a court decision assessing that theory. Operational evidence would matter too: whether the promised capacity becomes available and whether competing developers retain realistic alternatives. Those developments would speak to the difference between an announced benefit, an enforceable restriction and a demonstrated competitive effect.

The DOJ Antitrust Division accepts reports of suspected competitive harm online, by mail or by phone. A business affected by a restriction can contribute concrete evidence about access or conduct, rather than relying on the scale of a funding announcement.

The partnership paradox is resolved by following the choices the agreements leave open. More money and computing power can help an AI developer expand, but competition also depends on whether that developer and its rivals can keep choosing suppliers, developing products and challenging the investor.