This blog is the latest Hub & Spoke coverage in a series on Artificial Intelligence and related innovative technologies. You can read more of our coverage here, here, here, here, here, here, here, here, here, and here.
In AI world, the antitrust spotlight has long focused on the cloud, with a handful of companies controlling access to centralized computing. But a quieter infrastructure battle is taking shape at the network’s edge, one that could prove equally consequential for the future of AI competition. As AI inference moves closer to the physical world, including in factories, vehicles, hospitals, and consumer devices, control over edge infrastructure is fast becoming a decisive competitive advantage. And where decisive competitive advantages concentrate, antitrust enforcers tend to follow.
The Stack Problem
To understand edge computing’s antitrust risks, start from the top. AI competition is increasingly a story about who controls the full vertical stack, including chips, cloud, software, and now edge infrastructure. At each layer, a small number of companies exercise significant control.
The chip layer is the clearest example. Until relatively recently, Nvidia dominated AI training infrastructure through its GPU chips and, critically, through CUDA, the software development platform that many complained locked developers into Nvidia’s ecosystem. In 2024, the DOJ reportedly probed whether Nvidia offers preferential supply and pricing to customers who buy exclusively from its stack, or purchase entire integrated systems rather than components. In Europe, enforcers have focused on tying chips to networking equipment and discounting practices that suppress standalone purchases. Separately, the FTC previously investigated SoftBank’s proposed acquisition of Ampere, an Arm-based chip designer, on the theory that a combined SoftBank-Arm-Ampere entity might adjust Arm’s licensing terms to disadvantage rivals like Qualcomm or Amazon’s Graviton. (The FTC ultimately ended the investigation with no action). The concern in all of these investigations is the same: using dominance at one layer of the stack to foreclose competition at others.
The cloud layer has drawn parallel scrutiny. In 2023, the FTC conducted a sector-wide study of cloud providers including Microsoft, Google, CoreWeave, and Oracle, and UK enforcers have opened (and then closed) investigations into cloud positions, with particular focus on customer lock-in. Google has itself complained to European regulators about Microsoft’s Azure licensing practices, alleging that they trap enterprise customers in its ecosystem, a reminder that antitrust dynamics cut across competitors and markets alike, and that today’s complainant is often tomorrow’s defendant.
Edge: The Next Frontier
Edge computing, i.e. processing AI workloads locally, closer to where data is generated, rather than routing everything through distant data centers, is emerging as the next critical input layer. The appeal is reduced delay: autonomous vehicles, industrial automation, and real-time AI inference all require processing speeds that centralized cloud infrastructure simply cannot deliver. That makes edge deployment not a convenience but a necessity for a growing class of AI applications. And like chips and cloud before it, edge infrastructure is attracting both intense commercial interest and early enforcer attention.
The concern specific to edge is vertical foreclosure. Nvidia, already dominant in chips and deeply integrated into cloud infrastructure through its DGX systems and Mellanox networking hardware, has been moving into turnkey edge solutions, offering hardware, software, and networking as a bundled package for industrial and enterprise deployments. The same CUDA ecosystem and SDK bundling that allegedly created lock-in at the data center level could, if extended downstack, replicate that dynamic at the edge. Industrial edge AI, defect detection, sensor fusion, robotics are all markets where a dominant firm offering a fully integrated hardware-software solution could crowd out independent edge platform developers before the market has a chance to mature. The question for antitrust purposes is whether customers who might prefer to mix and match components could find, in practice, that they cannot, either for contractual reasons or because the technical integration makes switching prohibitively costly.
Tying, Bundling, and the Foreclosure Risk
Perhaps the most immediate antitrust theory in edge computing is tying. A provider who controls edge hardware, edge software, and cloud connectivity has strong incentives (and potentially the ability) to bundle them in ways that foreclose competition at any individual layer.
An AI startup seeking to deploy inference models at the edge may find, in practice, that choosing a dominant provider’s hardware stack quietly mandates its cloud services, its data pipelines, and its AI tooling as well. Customers who might prefer to source components independently may face technical incompatibilities, withheld software licenses, or pricing structures that make standalone purchases economically irrational. That is tying in effect, even if not in name, which is precisely the kind of conduct that antitrust law, applied carefully, is designed to address.
Google’s Role
Google’s position at the edge is distinct from Nvidia’s but structurally similar. Through Google Distributed Cloud, it has deployed edge hardware and software into telecommunications networks, retail environments, and government systems. Its TPU chips and Android-based device ecosystem round out a vertically integrated stack that spans from cloud to device. And as the Google Search litigation established at length, Google has a playbook of using default arrangements and exclusive dealing (in Search, in mobile, and allegedly emerging in AI) to entrench its position across adjacent markets.
The concern in edge, is the same as AI and Search: the flywheel. In Search, proprietary access to query data allowed Google to continuously improve its product while rivals were starved of the same inputs. As Judge Mehta described it, this steered users to Google and away from rivals, creating a continuous feedback loop that allowed Google (but not its competitors) to improve its product, attracting ever more users, generating ever more data. At the edge, privileged access to real-time inference data from deployed hardware could produce the same dynamic in AI model development. A firm that controls the edge layer sits at the point where AI models meet the physical world, may generate streams of proprietary operational data that can be used to train and refine those models in ways competitors simply cannot replicate.
The market has not yet tipped. firms face genuine competition, both from Big Tech firms and independent players like Cloudflare and Fastly. But, as the Search litigation demonstrated at painful length, waiting until dominance is fully established before acting makes remedies both harder to design and harder to sustain. By the time a market has tipped, defaults are sticky, switching costs are prohibitive, and the behavioral remedies courts are willing to impose rarely go far enough to restore competition.
The Distribution Layer: Telecommunications as Chokepoint
Edge computing is also deeply intertwined with 5G and telecommunications infrastructure, where carriers like AT&T, Verizon, and T-Mobile control the physical network layer. This creates a layered bottleneck: dominant providers seek partnerships with dominant carriers that could lock out independent edge platform developers or AI startups who lack the capital to secure comparable agreements.
History offers a useful analogy. The original AT&T case was not only about telephone service, it was about the company’s control over the physical network layer, which it leveraged to exclude rival devices and services from connecting to its infrastructure at all. The modem, now a relic, was once a contested product market precisely because AT&T’s monopoly over the network was ultimately broken, preventing it from being able to exclude competition among ancillary devices. Today’s edge infrastructure may present the same dynamic in a different technological register: whoever controls the last mile of connectivity to edge deployment locations may exercise meaningful influence over which AI platforms can reach end users, and on what terms. Big firms who can pay the toll to the telecom carrier could get privileged access, risking lock-out of upstart competitors.
Enforcers Are Watching
The DOJ, FTC, and UK CMA identified three AI competition risks in their 2024 joint statement: concentration of key inputs, entrenchment of existing market power, and arrangements among key players. Edge computing implicates all three, and the current enforcement posture (information-gathering, merger review, sector studies) suggests that enforcers are keeping watch.
The specific theories worth monitoring are familiar from the cases already underway: exclusive dealing that deprives rivals of deployment opportunities and training data; tying that bundles edge hardware with cloud services and AI tooling in ways customers cannot practically separate; default arrangements that, even without formal exclusivity, produce foreclosure effects by making switching costly; and acquisition strategies that consolidate control over adjacent layers of the stack before independent competitors can establish themselves. Each of these theories has already found purchase in active litigation or investigation. The task for enforcers and practitioners is to recognize when they are being applied to a new infrastructure layer before that layer becomes the foundation of the next walled garden.
For companies operating in this space, whether as platform developers, enterprise customers, or potential complainants, the lesson from the broader AI antitrust landscape is clear: enforcers are watching, the legal frameworks are in place, and the window for early intervention is open. The time to understand your exposure, and your opportunities, is now.
