Blog

Co-authored by Dr Kenny Huang¹ and Charles Mok²

AI services runs over the Internet. But do they all control their own infrastructures in the same ways?

Abstract

This study analyzes the differences in domain name and IP address strategies among a number of current mainstream artificial intelligence (AI) service providers. We find that these technical choices not only reflect deployment decisions but also deep-seated corporate knowledge and capabilities in Internet infrastructure service provision, as well as brand positioning and market strategies. Upon analysis, the authors categorize these AI service providers into two main categories. The first, including Google’s Gemini and Microsoft’s Copilot, leverages and enhances its existing product functionalities to expand its service portfolios with AI services. These “Internet incumbants” possess robust proprietary IP address infrastructure, enabling them to integrate AI into existing products or expand product categories more seamlessly to achieve stability and flexibility. The second group of “AI startup” including OpenAI’s ChatGPT, DeepSeek, and Anthropic’s Claude, tends to use dedicated domain names to establish independent and innovative AI brands to create a new ecosystems, attracting users specifically to their AI services. Some of them lack extensive IP address infrastructure, resulting in limitations in network routing control and facing greater challenges in future expansion and service quality optimization. The network governance of AI services will be crucial for their future development. For these startup AI providers, in addition to actively investing in the hardware and systems infrastructure, such as datacenters and hyperscale computing, they will also need to take further ownership of their own IP address infrastructure to gain better network autonomy and to enhance network resilience. After all, AI services still run on the Internet, so, in order for the startup AI service providers to compete with the incumbent Internet service providers on the same infrastructure footings, they must invest in further understanding and taking control of the rules, policies, standards and practices of the IP connectivity infrastructure, that is, what we call “Internet governance.”

1. Introduction

With the explosive growth of AI technology, its governance issues are also receiving increasing attention. However, to most people, that would mean the impacts of AI technologies and the services provided by large-language models (LLMs) on issues such as safety, transparency, data governance, environmental impact, or how AI would affect human, social or business activities. Relatively little attention is being paid to its relationship to the traditional “Internet governance” of how effective control of the critical resources of Internet infrastructure would facilitate the delivery of these current and future AI services.

In the “traditional” online world, network application providers all aspire to become the “destinations” for internet users, and domain names and proprietary IP addresses are the indispensable foundations for their becoming such destinations. Throughout the previous decades of Internet development, familiar brand names such as Google, Instagram, or Facebook have transformed from domain names into verbs that represents certain product functions or user behavior. Even though in the age of the mobile apps, domain names have taken a back seat in that users often do not input or refer to domain names in order to access certain services, they are still very much present in the background, along with the proprietary IP infrastructures that support the delivery of these application services.

This is exactly where Internet governance has played a vital role in Internet development, by setting up and maintaining the rules, policies, standards and practices for the Internet. Through a multistakeholder model of Internet governance, critical internet resources can be managed and used more fairly and effectively, and without conflicts. IP addresses and domain names are the most common examples of critical internet resources. By examining the use of domain names and IP addresses, we can observe the landscape of the emerging AI service providers from an Internet governance perspective.

In this paper, we analyze five major AI service providers—Open AI’s ChatGPT, Google’s Gemini, DeepSeek, Microsoft’s Copilot, and Anthropic’s Claude—from an Internet governance perspective, exploring their strategies for adopting domain names and IP addresses. We will then dissect how these strategies reflect their brand positioning, service development direction, and potential challenges they may face at the network infrastructure level in the future.

This study uses the APNIC WHOIS Database to query the domain names, IP addresses, and Autonomous Service Numbers (ASN) of these five services, as shown in Table 1.

2. Domain Name Strategy: Brand Positioning and User Entry Point

From Table 1, all five AI service providers have set up dedicated and specific domain names for their AI services, demonstrating their fundamental desire to establish their respective identifiable online presence. However, their choice of domain name type reveals different brand strategies:

2.1. Dedicated Domain Name Strategy

Services like ChatGPT, DeepSeek, and Claude opt for dedicated domain names based on the names of their products. As their respective owners—OpenAI, DeepSeek and Anthropic—can be considered as “pure AI startups,” they clearly intend to build their AI services as independent brands, and to become a direct destination for users accessing the AI services. This strategy also means that their AI and LLM services are to be considered their primary services for their users, even though as they establish their AI service ecosystems, more and more services or features may be added in future.

2.2. Subdomain Strategy: Branding and Marketing

For the other two leading AI services we surveyed, Google’s Gemini and Microsoft’s Copilot, they have chosen to operate under a subdomain of their parent companies’ primary domains. Being established “big tech” companies, google.com is already a destination for its flagship search engine service, and microsoft.com is the brand entry portal of the company’s portfolio of software and online services.

On deeper dive, while a user can input “copilot.com” to access Copilot’s AI service, the domain name will revert back to “copilot.microsoft.com.” This is in contrast to Google which did not even own the “gemini.com” domain, which is owned by a credit card company. From a pure branding and non-technical point of view, this may be somewhat of a disadvantage for Google.

However, the use of subdomains for these companies is consistent with them being “Internet incumbents” strategically looking to enhance their existing services—from search engines to productivity software—with their new AI offerings. While they must still aspire to build up their new AI service brands, that is not their primary objective, as they view AI as not only a new offering but also a transformative technology to enrich their existing products.

Secondly, as these big tech companies have done in the past, using subdomains for new services helps brand strategy flexibility and substitutability and new services, including AI, can be more easily integrated with the parent companies’ existing services, allowing greater brand strategy flexibility to more quickly adapt to new product positioning or market changes.

However, the existence and use of a subdomain may still create certain limitation to reaching a full potential of a standalone brand, which has its own advantages. In the case of Google’s Gemini, its decision to brand the product with a more commonly used word—gemini—with the corresponding domain name already taken by another product with the same name albeit in another industry may still lead to confusion in the future, and limit its reaching a full brand potential. However, this is a marketing and technical decision, not a technical issue.

3. IP Address Strategy: Network Infrastructure and Governance Capability

In the use of IP addresses, we also observe two distinct patterns between the AI startups and the Internet incumbents, directly relating to the providers’ autonomy and governance capabilities at the network infrastructure level.

3.1. Proprietary Corporate IP Infrastructure

For Google and Microsoft, they already possess vast and well-developed proprietary IP infrastructure, due to their previous investment into owning their own Internet infrastructure, including critical Internet resources, from its earlier days in the Internet development. Their new AI services directly utilize the IP addresses and infrastructures held by their own respective corporations. This grants them greater control and flexibility over network routing, allowing them to more effectively optimize service delivery paths and directly deliver content to target users, ensuring a more efficient and stable service experience, as they have been doing for their existing products. They can manage the end-to-end network performance of their AI services, which is a significant advantage in terms of network governance and the guarantee of quality of service delivery.

3.2. Reliance on Third Party ISP-Provided IP Addresses

OpenAI and DeepSeek are examples of AI startups that still rely on IP addresses provided by other Internet Service Providers (ISPs). This means they must connect to the internet through other ISPs’ network to deliver their AI services to their endusers. This dependence may lead to the following potential challenges:

1. Limitations in Network Routing Control: Lacking their own IP addresses and Autonomous System Numbers (ASNs), these providers cannot directly control the routing paths of their traffic across the internet. They have to defer to their ISPs’ routing policies, and their traffic policies that may or may not align with the AI service providers’ optimal traffic strategy.
2. Service Quality and Target User Matching Issues: An ISP’s network is typically designed to serves specific geographical user groups, and its routing optimization may not match the ongoing and changing needs of AI services with unspecified target users. This can lead to potential bottlenecks or uncertainties in service performance and user experience.
3. Future Network Infrastructure Challenges: As the scale and complexity of AI services continue to increase, the demands on network infrastructure will also rise. AI service providers lacking proprietary IP addresses may face greater challenges and limitations in the future when expanding network capacity, optimizing global service distribution, and addressing network security challenges—even if they own their own datacenters. This also suggests that their network governance capabilities may still be relatively weak, with greater dependence on their ISPs also potentially mean high future switching costs.

3.3. The AI Startup Which Has Taken a Step to IP Infrastructure Ownership

Compared to OpenAI and DeepSeek, the other two AI startup examples in this study, Anthropic’s Claude stands out and has taken a noticeable lead and significant step in owning its IP and ASN registration and infrastructure. This enables the company and its products and services to be in better control of its own network routing, service quality and future planning. It should be recommended that other pure AI startups should also adopt similar strategies, and we predict that many of them will in the near future.

4. Further Observations

Summarizing from the above comparisons, the group of Internet incumbents and AI startups may hold different advantages or deficiencies when we compare their marketing and branding strategies with their IP infrastructure and governance capabilities.

While the AI startups have an advantage in its flexibility to create their new and innovative AI product branding, the Internet incumbents may yet enjoy the benefits from leveraging its existing product offerings to help capture users to its AI services, such as enhancing Google’s search engines with AI, or adding AI features to Microsoft’s Office 365, which incidentally is already co-branded as Microsoft 365 Copilot.

Just as Google started from search engines and its Chrome browsers, and Microsoft began with their Windows operating system, Office software and its Internet Explorer browser (now Edge), OpenAI likely will take on a similar path to develop its product universe, beginning with ChatGPT, and now it has already announced its AI Agent called Operator, which operates as a “virtual browser” that can interact with websites and will run in the cloud on OpenAI’s servers. Conceptually, OpenAI is taking on a similar path to create its service universe with AI agents, similar to Google did with its suite of services and apps.

In terms of network infrastructure and Internet governance capabilities, the advantage clearly goes to the Internet incumbents, over the group of AI startups, as they have already undergone the process of achieving their own independent control of their own infrastructures, critical Internet resources, routing and quality of service. As the AI startups grow and expand, they can be expected to attempt to acquire more capabilities to control their own infrastructures, including their domain name and IP address strategies. In particular, as OpenAI has already aggressively expanded into hyperscale datacenters, it may actually be a natural next step for their strategic expansion, through internal development or acquisitions. Other pure AI startups should also at least adopt the same strategy as Anthropic to start to put the destiny of its own network routing and quality control in its own hands, as soon as possible.

For the case of DeepSeek, it is also particularly interesting to observe that their global network infrastructure depends on Cloudflare, a US company. On the other hand, its relative competitive position and control of its own IP infrastructure, compared with its leading domestic Chinese AI rivals, which are China’s own Internet incumbents, such as Baidu, Alibaba and Tencent, would be an interesting future subject to study.

5. Conclusion and Future Outlook

From an Internet governance perspective, AI service providers’ choices regarding domain names and IP addresses are not merely technical decisions and capital allocation. They also reflect their deep-seated brand strategies, service positioning, and vision for future autonomy.

Intenrnet incumbents focused on enhancing existing services can more seamlessly and flexibly integrate AI into their existing ecosystems, thanks to their parent companies’ prior investment on their existing network infrastructure and brand influence. Conversely, pure AI startup providers demonstrate the ambition of independent brands and new services, but their network infrastructure dependence may pose potential challenges for future expansion and service quality improvement.

As AI technology continues to evolve and its application scenarios become increasingly widespread, and especially with the anticipated rollouts of AI agent products, the issue of AI service network governance will become even more critical. For emerging AI startup providers, in addition to technical innovation, active investment in network infrastructure, including managing their own IP addresses, will be a crucial step to enhance their network resilience, gain better network autonomy, optimize user experience, and secure an advantage in an increasingly complex and competitive market.