What are the Economics of an AI Native Internet?

What are the Economics of an AI Native Internet?

The Internet is responsible for about $16 trillion of the global $108 trillion annual GDP. Click the link to see my Perplexity space with the numbers.

The Internet is far bigger than the web. The Internet is made up of all activity on top of TCP/IP and the Domain Name System not only web traffic. This includes all apps, messaging services, online games, payments systems, devices and so on. Only about 40% of the $16 trillion is web related value.

AI is broadly anticipated to be the primary engine of digital and non-web economic growth in the coming decade, with consensus estimates ranging from a modest 1-2% boost to global GDP to transformational impacts of up to $20 trillion, much of which will stem from non-web, embedded, and invisible AI-driven flow

At the same time AI is transitioning to an agentic model where invisible browsers like Browserbase act as the 'reader' of web based content, subsumed under an AI interface. So even the 40% currently accounted for by the web will no longer be driven by direct consumer browser use.

This has implications for every stakeholder in the Internet economy.

An AI Native Internet is an Internet where the web, apps, payment systems are all interfaces to a new AI canvas or 'front door' accessed via devices - mobile devices, AR devices and new yet to appear ones.

The $16 trillion of current value and the potential $20 trillion new value will all migrate to the AI Native internet, or disappear due to failing to make the transition.

For that reason the question of what the business model of the AI Native internet will be is crucial.

The dream scenario is that AI is universal and free to users. Like web search today. Of course that is already somewhat true but the subscription revenues of OpenAI, Anthropic, Google Gemini and others represent the lions share of revenue today, alongside API triggered token use where developers pay per million tokens consumed for inputs and outputs.

The domination of subscription and token based revenues is a function of the current state of AI. It is hard to impossible for AI to inherit the marketing spend that currently goes to Google and other advertising platforms for reasons we discussed last week.

And it is currently hard for content owners to get the traffic from AI, and hence the revenues, that they get from the web and app economy today. Again we discussed why last week, but broadly AI does not send much traffic to content owners.

This week Matthew Prince from Cloudflare was interviewed about a pay per crawl proposal using an http 402 technology that would enforce payment to content owners for a crawl. He suggests $1 per year per MAU(Monthly Active User), from the LLMs to a pool. Then this would be distributed to the content owners. I assume there would be some math creating a pro-rata payout according to some measure of value exchanged.

There is a lot to grasp here, but let's assume that happens and all 4 billion humans with smart phones are in the MAU, that would create a $4 bn pool annually. It is only a small proportion of $16 trillion. So pay per crawl is definitely a feature of a business model for an AI Native Internet, but it is not the full model. Web advertising alone is a $667 billion portion of current Internet revenue, so one can see that $4 billion is good start, but small.

The logical first question is, how much of the $667 billion from advertising can migrate to this new AI native internet where the web is subsumed under AI and its agents?

For AI that is a pressing question as this week's articles show that the cost of running AI, especially with reasoning and agents is rising, not falling.

The model we (all builders) should build is an AI Native Internet that is free for users, paid for publishers, and yes, paid for AI's in return for the traffic they send and the value they create from that. It seems it might collectively be a $16-40 trillion economy in the years ahead.

The virtuous loop we laid out last week moves closer to this: identity for agents, attribution for links, settlement for outcomes. Cloudflare's pay-per-crawl plus a neutral clearing function (accounting, registries, enforcement) can sit beside an AI-era CPC/CPA/CPM model.

Importantly, AIs shouldn't only pay - they should also get paid. When they route high-intent traffic that converts on content owner's sites. This aligns incentives: better answers send better clicks, with auditable receipts. Indeed AIs would earn a lot more than they pay because the crawls revenues they pay out would dwarf the share of the entire economy they are paid.

Tools like Anthropic's MCP and unified metadata layers make those receipts credible by stitching permissions, provenance, and business context into the workflow.

I believe this needs a single trusted third party being paid by the ecosystem to run such a set of services just as the domain name system has a root server run by ICANN, the AI Native Internet would benefit from a root monetization system.

The .com system is managed by Verisign. Its revenue annually is $1.56 bn of the $16 trillion economy. Add Google's share of advertising and the share of others providing all of the services, and it is hard to get infrastructure to be even 10% of the $16 trillion. A monetization infrastructure for AI would combine several roles and would represent a tiny fraction of the total value created.

What to watch next

Whether Google offers a true opt-out for AI Overviews - and how fast regulators push interoperability beyond search data.

Adoption of "402 Payment Required"-style signals and pay-per-crawl at scale, turning rights into programmable rails.

If DOJ's ad-tech remedies force structural change this fall although the transition to AI will render this irrelevant.

Whether chip moves (OpenAI-Broadcom) and model routing meaningfully bend inference COGS for all, reducing AI costs (against the current trend).

The long arc still points to the same equilibrium: free (or near-free) AI for users; real money to publishers for training and traffic; and revenue for AIs that create measurable value in the handoff. Build the identity, registry, and clearinghouse - and everybody wins.

And a final note. The Anthropic settlement with book publishers this week ($1.5 billion) is being widely misinterpreted. Anthropic was not sued for reading publicly available books. And it was not sued for reading books it purchased. The suit was due to the fact that Anthropic read stolen books, and hence breached copyright in the act of stealing. Had it paid for the books no suit would have succeeded. The case does not indicate that copyright protects legally acquired content from being a source of training data. Legally acquired content can be used for training unless the law changes, and there is no indication of that.

Essay

AI and jobs, again

Noahpinion • August 30, 2025

Essay•AI•LaborMarket