The expansion of artificial intelligence agents could trigger a much deeper change to the internet than simply increasing traffic. Zachary Smith, co-founder and CEO of Datum, argues that applications, companies and even individual users will need private networks capable of precisely controlling which services each agent can communicate with. His vision points toward an internet that is less open by default and, in certain areas, more similar to controlled networks such as the one used by Visa.
The future of the internet with AI agents in 30 seconds
- Zachary Smith believes AI agents will increase the number of applications and connections between cloud services.
- His proposal involves limiting which systems each agent can reach through private connectivity.
- Fragmentation among cloud providers and digital sovereignty requirements are adding complexity to the network.
- Datum is developing a networking platform primarily aimed at alternative cloud providers.
- The comparison with Visa describes a network where participants and connections are controlled, not a literal replacement for the public internet.
Smith’s argument is particularly interesting given his background. Before Datum, he co-founded Packet with his brother, an automated bare-metal infrastructure platform that Equinix acquired in 2020 for $335 million. Earlier, he helped build Voxel, which was sold in 2011. His route into the industry was also unconventional: he studied classical double bass at Juilliard and began working with computers while paying his way through school by repairing PCs and connecting customers to the internet.
He now believes internet architecture is once again entering one of those periods in which several major trends converge. AI is one of them, but it is not the only one.
From a handful of clouds to applications spread across dozens of services
For much of the past decade, talking about cloud computing has largely meant talking about Amazon Web Services (AWS), Microsoft Azure and Google Cloud, alongside other major providers such as Oracle.
Artificial intelligence is partially changing that concentration.
On the one hand, so-called neoclouds have emerged, providers primarily specialising in accelerated infrastructure for AI. Smith mentions companies such as CoreWeave and Lambda as examples of this new category.
But another, less visible shift is also taking place.
Modern applications are increasingly built by combining specialised services. A company might run part of an application with one provider, use a managed database from another, outsource authentication to a third party, consume several AI models and rely on yet another platform to send emails.
The result is that an application that appears to operate as a single unit can actually depend on dozens of external services.
AI agents could accelerate this trend because they reduce the difficulty of developing software and integrating services. Smith believes AI-assisted programming is substantially increasing the number of people capable of creating applications, although during the interview he acknowledges that he does not have specific figures to quantify that expansion.
This creates a paradox: developing certain applications becomes easier while the infrastructure underneath them can become more complex.
APIs and protocols such as Model Context Protocol (MCP) allow agents and applications to discover capabilities and interact with different systems. The more services they can use automatically, the greater the need to control those communications.
That is where the problem Smith considers particularly important begins to emerge.
Why an AI agent should not have access to the entire internet
An agent does not necessarily just generate text.
It can query databases, call APIs, access storage, execute code or use enterprise tools. Depending on its permissions, it can also perform actions.
Smith proposes applying a simple principle to these systems: an agent should only be able to communicate with the resources it needs.
Instead of placing all these connections over the public internet, his approach involves building private networks between authorised applications and providers.
The example used during the conversation is straightforward. If an application needs to communicate with a specific database, it should be possible to allow that connection without giving the application unrestricted access to the rest of the internet.
This is not an entirely new concept.
Large companies have been using private links, virtual private networks (VPNs), direct interconnections and dedicated connectivity services for years to connect data centres, providers and cloud platforms without relying exclusively on the public internet.
Large interconnection data centres play an important role in this model.
Their well-known meet-me rooms allow telecommunications operators, content providers, cloud companies and enterprises to physically connect their networks. Instead of necessarily sending traffic through the public internet, two organisations can establish private connections between their infrastructures.
Smith knows this market well from his time at Equinix and believes a similar concept could now move from networks down to individual applications.
The difference would be making that connectivity simple enough to be managed by developers and even software agents without requiring in-depth knowledge of BGP, autonomous systems, IP addressing or the physical architecture underneath.
An internet that looks more like the Visa network
Smith uses Visa as an analogy to explain how part of the internet could evolve.
A payment network connects banks, processors and millions of merchants, but that does not mean any device can freely join and communicate with every participant. There are controls governing who can connect and under what conditions.
His hypothesis is that a growing share of digital applications could adopt a similar model.
An agent would have a defined identity and connectivity. A company would determine which providers are authorised. Communications between services could take place over private networks, while organisations would have greater ability to inspect traffic and reduce the exposed surface.
That does not mean Smith predicts the disappearance of the public internet. The Visa comparison describes how certain communications between applications, agents and cloud services could operate.
Security is one of the reasons.
The same AI tools that allow developers to generate code quickly can also be used to identify vulnerabilities and automate testing against exposed systems. Smith argues that unnecessarily keeping applications and services accessible from the internet increases an attack surface that could be reduced through private connectivity.
His approach also fits a familiar cybersecurity principle: limit communications and privileges to what is strictly necessary.
AI, however, adds scale. If millions of new agents begin automatically interacting with applications, APIs and databases, controlling who can communicate with whom could become as important as controlling which user can access particular information.
Digital sovereignty adds another layer of complexity
There is a third trend behind Datum’s proposition: the regulatory and geopolitical fragmentation of the internet.
Smith argues that the model built during globalisation was based on a relatively open global network. Greater restrictions are now appearing around sovereignty, data location, supply chains, technology regulation and trade relationships between countries.
It is worth separating the broader trend from some of the examples used informally during the podcast. Smith discusses scenarios involving national data-storage and cloud-service requirements that should not automatically be interpreted as an exact legal description of current legislation in each country.
The underlying idea is that an international application may increasingly need to combine different providers and locations depending on the data being processed and the applicable obligations.
That multiplies the number of connections that must be managed.
Smith describes the problem through three forces: applications and data distributed across more locations, a growing number of providers, and increasing regulatory and sovereignty requirements.
Datum is trying to position itself between those pieces.
Datum wants to become a networking layer for the new clouds
Datum describes its product as a network cloud, a platform designed to simplify connectivity between applications and providers.
Its initial market is not primarily companies consuming cloud services, but cloud providers themselves.
Smith identifies several categories: neoclouds building physical AI infrastructure, specialised inference providers, and companies focused on data storage and processing.
Datum also operates its own network across approximately 20 locations, according to Smith during the interview, and is expanding that footprint. Its technology is released as open source under the AGPLv3 licence, although certain uses that do not want to comply with the conditions of that licence may require a commercial agreement.
The strategy deliberately resembles the network effect created by interconnection data centres.
A data centre becomes more attractive when it concentrates many operators and providers because every new participant increases the number of potential connections available to everyone else.
Datum wants to bring a similar dynamic into software.
Smith gives the example of a cloud platform that could directly show the applications and services with which a customer can establish a private connection. The complexity of routers, IP addresses and autonomous systems would remain hidden behind simpler interfaces.
It remains a developing vision. Smith himself acknowledges that Datum is only beginning to work with its first customers and still needs to demonstrate that it can generate this network effect.
But the arrival of AI agents raises a question that goes beyond any single company.
For decades, much of internet security has involved enabling connectivity first and then deciding who can access each resource through authentication, firewalls and other layers of protection.
The proliferation of autonomous software could reinforce the opposite approach: decide which connections should exist first and eliminate everything else.
If agents eventually perform personal and business tasks continuously, the question will no longer be only what permissions each agent has. Organisations and users will also need to decide which parts of the internet an agent is allowed to talk to.
Frequently asked questions
Why might AI agents need private networks?
Because they can automatically communicate with APIs, databases, applications and other services. Smith proposes restricting those connections to authorised systems in order to reduce unnecessary exposure to the public internet.
Is the public internet going to disappear?
That is not what Smith argues. His prediction is that private connectivity will become more common for certain communications between businesses, applications, cloud services and AI agents.
Why does Zachary Smith compare the future of the internet to Visa?
He uses Visa as an analogy for a large but controlled network where participants are identified and connections follow defined rules. He is not proposing that the internet literally become a payment network.
What does Datum do?
Datum is developing a connectivity platform it calls a network cloud. It aims to simplify private connections between applications and cloud providers, initially focusing on neoclouds, inference providers and data platforms.
