Coming Soon to the Internet Near You

For 28 years the internet has run on one primary undisputed currency, and that's human eyeballs clicking on digital advertisements. There's a looming business model shift, or more accurately a fundamental crossing of the Rubicon taking place as traditional digital advertising, the "currency of eyeballs", collapses because AI agents do not engage with visual marketing or banner ads. In terms of how digital traffic is generated, how it is routed, and ultimately how it's consumed, the internet is literally no longer a purely human to human network.

Human clicks on marketing and ads was that single action that built the entire free web. It funded the search engines, the social networks, the local news blogs, the video platforms. But what exactly happens to that multi-billion dollar perfectly balanced ecosystem when the machines take over the browsing?

If you've been paying attention ever since the Google AI summaries turned up in our lives, how the internet actually makes money tomorrow, and how you will be shopping in 5 years, should be activating your curiosity.

Matthew Prince, the CEO and co-founder of Cloudflare, ought to know. Cloudflare isn't just a website. They are essentially the wiring of the internet. They provide the security, the performance routing, the bot mitigation for a large chunk of all global internet traffic. When data moves literally anywhere on Earth, Cloudflare usually sees it passing through their pipes and can watch the traffic patterns shift in real time over the independent websites we love to read with morning coffee. Will they even exist next year? And what about the user interface of e-commerce, the experience of online shopping, from paper towels to a new car. It's about looking at the market demographics of the web today, something we can't talk about until we understand who or what is actually generating the traffic in 2026.

When the Machines Outnumbered Us

Prince's metric, which is kind of the diagnosis that sets this entire conversation in motion, states that in the first half of 2026, bot and AI agent traffic officially surpassed human traffic online. In other words, more machines are browsing the web right now than living, breathing humans, and we're unlikely to go back.

Cloudflare Radar, which is an internal public facing tracking system, showed that by early June 2026, 57% of all HTML requests were coming from bots compared to roughly 43% from humans. But to be safe let's not just take Cloudflare's word for it. The independent Thales and Imperva 2025 bad bot report already saw automated traffic hitting 51% way back in 2024. Furthermore, Human Security, a premier bot mitigation firm, noted a trend where AI-driven traffic was growing eight times faster than human traffic across the entirety of 2025. The baseline reality accurate, but Prince additionally projects that within 5 years, bot traffic will outnumber human traffic 1,000 to one. His reasoning is that these AI agents won't just be scraping data to learn. They'll be actively doing more tasks for us like shop, and the number of them doing the tasks are likely at least steeply increase, if not explode. Take for example shopping for a digital camera. A human being might open their laptop, visit five websites, read a few reviews, check a couple of YouTube videos, and make a choice.

The human tasked AI agent, on the other hand, might instantly hit 5,000 websites, cross-referencing every spec, every price drop, every shipping time, and every single customer review on the entire global web in a fraction of a second; it traverses the web on a completely different intensity scale. That said, we're equating a machine making 5,000 invisible instantaneous HTTP requests to actual valuable human eyeball attention. The volume of movement, the sheer number of visited touch points are exponentially higher, but the economic value being extracted and the conscious intent behind the action is a distinction. Depending on what we're measuring, taking that 100x extrapolation as a measure of impending economic value is not equivalent to actual engagement like in prolonged usage of interactive application. An HTTP request generated by a bot is just a momentary ping to a server. It is not a human reading a long form article or watching a pre-roll advertisement before video. Bots aren't watching a Netflix stream. However, while the economic value of those 5,000 robotic pings might be debatable, the physical toll they take on the internet's infrastructure is undeniably real, because the server processing the request doesn't know or care if it's a single human visit or machine. It still takes electricity to open the door and serve the data either way.

The Hidden CPU Crisis

An additional layer to this is the hidden CPU crisis. If you were to run just one AI agent per knowledge worker on Earth using traditional container architectures, it would require roughly 50% of the entire global CPU capacity. That's just to give white collar workers one single digital assistant. We have billions of processors in the world. I know it sounds like complete hyperbole until you understand the underlying mechanics of how these agents actually operate. We need to define what is meant by container architectures.

When you run an AI agent, you can't just let it run loose on a shared server where it might accidentally interact with another person's agent or access private memory. You have to isolate it for security reasons. A container is essentially a virtual, highly secure mini computer carved out within a larger server. It gives the AI its own isolated sandbox to execute logic securely. But spinning up millions of these highly secure isolated containers requires immense processing overhead. And specifically, it requires CPUs, not GPUs. We constantly hear about Nvidia and GPUs, graphics processing units driving the AI boom. GPUs are incredible at doing thousands of simple math problems simultaneously. That makes them perfect for the training phase of AI, where the model is ingesting billions of pages of text and just guessing the next word in a sequence, which is the pattern recognition work.

But an AI agent navigating the internet isn't just guessing words. It's executing complex branching sequential logic. It has to navigate a website's layout, click specific buttons, input a password, wait for a two-factor authentication code, and verify a transaction. GPUs can't do that. GPUs are terrible at that kind of sequential step-by-step logic. You need a CPU, a central processing unit to act as the hands and feet of the AI. The GPU is the brain that figures out what you want. But the CPU actually has to go out into the internet and execute the task. So, as the AI industry shifts from just creating chat bots that talk to us to creating agents that actually do things for us, the hardware demand completely flips.

To prove this bottleneck is profoundly real and happening right now, data from AMD, Intel, and Nvidia confirm a massive structural shift in data centre design. Historically, data centres had a CPU to GPU ratio of roughly 1:8 for heavy AI workloads. The GPUs did all the heavy lifting. Now to support agentic deployments, that ratio is shifting rapidly toward 1:1. The market reaction to that is getting chaotic. Server CPU prices have spiked 10 to 20% in just a few quarters. Intel is literally altering their global fabrication capacity, pulling resources away from consumer chips to desperately try and meet this new server CPU demand just for these agents.

The Energy Bill

And ARM estimates that the demand for CPU core per gigawatt of power will quadruple in the AI agent era. If we connect this to the bigger picture, there are tangible energy implications. It is not just about silicon. It is about electricity. The International Energy Agency projects that data centre electricity consumption will hit 945 terawatt hours by 2030.

To put 945 terawatt hours in perspective, in 2024, global data centre power consumption was roughly 415 terawatt hours. We are talking about more than doubling it in 6 years. Or adding the entire annual energy grid of a country like Japan just to power our digital assistants.

Imagine the physical infrastructure required for that. The power plants, the massive cooling towers, the copper wiring, the strain on local municipal grids. This physical reality might actually place a hard speed limit on the AI revolution's pace.

Compute itself is the new ultimate bottleneck. And if the infrastructure is groaning under the weight of this massive agentic shift, the most basic economic question emerges, who pays the power bill for all this electricity, these new data centres, and these expensive GPUs?

Who Pays for the Web?

For almost three decades, the answer was beautifully simple. Advertising.

Advertisers paid for the internet so that we could use it for free. But bots don't look at banner ads, and that's a problem underpinning the major rejigging of the internet business model. Prince, the Cloudflare CEO, makes a sweeping definitive claim, declaring that the 28-year-old advertising business model is dead. His logic is straightforward and on its face compelling. If the vast majority of web traffic is increasingly agent-driven and AI agents are immune to visual marketing, emotional persuasion, and banner ads, then the fundamental economic foundation of the open web collapses. But let's not wholesale buy the death of advertising narrative; there's data that shows the advertising industry is not only surviving, it is booming. According to Dentsu's global ad spend forecast, global ad spend is projected to rise 5.1% in 2026, surpassing $1.04 trillion for the very first time in human history. That growth rate is actually outpacing the projected expansion of the global economy itself. The Interactive Advertising Bureau, the IAB, backs this up, forecasting US ad spend to be up 9.5% in 2026. So we need to square the circle on these two realities. Prince says advertising is dying because bots make up 57% of traffic and bots don't click ads. But the financial data proves advertising has literally never been bigger, crossing a trillion dollars.

Whose forecast is right here? Or can both be true at the same time?

Advertising, Redistributed

Maybe advertising isn't dying, rather it's being redistributed at colossal scale. The money is poised to flee the open web and the independent blogs, the midsize news sites, the niche forums. The ad dollars are consolidating into what the industry calls 'walled gardens', meaning platforms that control the entire environment from top to bottom like social media apps, connected TV ecosystems.

The IAB data shows double-digit gains for social media platforms and connected TV networks. In a walled garden, the platform forces the user to log in, proving they are a human, and then controls exactly what ads they see. But even more importantly, the ad money is flowing into new, highly sophisticated agentic commerce ad revenue. Google is bringing sponsored results directly into its Gemini AI interface. Amazon for example is building massive back-end advertising tools specifically designed to influence AI agents. If an AI agent is scanning Amazon for the best paper towels, a brand can pay Amazon to ensure their product is weighted heavily in the agent's decision matrix. The advertising money hasn't disappeared. It's just retreated behind the fortified walls of the mega platforms that have the infrastructure to control the AI experience. Now here's where it gets really interesting. Because while Prince might be wrong about the total death of advertising globally, where he is absolutely correct is the devastation of the open web publisher.

Zero-Click Searches and AI Overviews

If you run an independent website, a local news blog, or just a recipe site, you are feeling this existential crisis right now, which brings us to the concept of zero-click searches and AI overviews. This is arguably the most heavily evidenced and socially impactful. A zero-click search is exactly what the name implies. You go to a search engine, you ask a question, and instead of giving you a list of 10 blue links to choose from, the AI generates a comprehensive summary right at the very top of the page. User reads the AI summary answer, is satisfied, and closes the tab without ever clicking a link to visit the website, which is still presented but buried just a bit further down the page. But the site authors actually did the hard work of providing that information, or in the case of news, a publisher paid a journalist to do the research, and the AI scrapes is instantaneous. The Google search engine keeps the user engaged on their own platform and the publisher gets absolutely nothing. No referral traffic, no page views, and therefore no ad revenue. It's a complete severance of the economic feedback loop.

The 2025 study by Pew Research that tracked nearly 69,000 individual search sessions found that users clicked on a traditional search result in only 8% of visits when an AI summary appeared at the top of the screen. When there was no AI summary, that click-through rate almost doubled to 15%. But the most devastating statistic for publishers, only 1% of users actually clicked a citation link embedded inside the AI summary itself.

Now if you're a publisher relying on search traffic to pay your writers, a 1% click-through rate isn't obviously profitable. A major SEO analytics firm found a 58% lower click-through rate for top ranking pages when an AI overview was present. News publisher search traffic fell roughly 33% globally in just a 12-month period. As for Google's defence. Liz Reid, a senior executive overseeing Google Search, has publicly claimed that aggregate clicks to the open web are relatively unaffected. Her argument is that the AI summaries are only removing low value 'bounce clicks'. A bounce click is when someone clicks a link, realises within 2 seconds it wasn't what they wanted, and immediately bounces back to the search page. Google argues they are saving users time by eliminating those useless clicks. That sounds plausible, but Google has steadfastly refused to release the internal data to actually prove that bounce click theory. And every independent study from Pew, IFS, and Similar Web happens to contradict it. Website owners aren't just losing bounce clicks, they are losing the core readership. So, I have to ask the obvious questions. If this trend continues, if AI summarises everything and no human ever clicks through to the original article, aren't we just strip mining the open web and extracting all the value until there's nothing left but AI models, talking to other AI models. And this will in turn lead to rampant hallucinating based on outdated data because all the human writers went bankrupt 5 years ago? The AI has got to be going to eat its own tail. Well at least that's the existential dread hanging over the publishing and media industries right now. If the open web is being stripped for training data and traditional display ads don't work for independent creators anymore, what replaces it? How does a specialised blog, a local investigative newspaper, or an independent tech creator actually pay their rent and keep the servers running in 2026?

Reviving Error Code 402

This is where Cloudflare's Matthew Prince proposes cures. He diagnoses the disease perfectly and then he proposes a radical pivot away from advertising entirely, moving towards a system of pay per crawl and machine to machine micro payments. To appreciate this solution we need to revive some old internet lore that most people don't actually know. We all know the 404 not found error code. You click a broken link, you get a 404. But the original architects of the internet, the pioneers who built the protocols, also built in an error code called 402: payment required.

They envisioned a web where digital transactions were native and frictionless. But 402 was basically never fully utilised because the advertising model took over completely. It was easier to show you an ad than to ask you for a penny.

Now, Cloudflare has helped launch a consortium called the x402 Foundation to bring this dormant code back to life. Let's walk through how this would actually work. The idea is that every single time an AI agent or a data scraping bot hits your website to read an article, train its model, or grab a recipe to answer a user's question, the machine hits a 402 error, and realises payment is required. The machine then automatically instantly pays a fraction of a cent to access the data. To sketch out the math required for this, Cloudflare processes about half a billion requests a second across their network. If even 1 to 10% of those requests are monetisable through these microtransactions, the world needs a financial system capable of handling 10 to 100 million transactions per second. And to provide a sense of scale, we could contrast this with something like Visa. Visa processes about 18,000 transactions per second globally, and Visa's stated maximum capability is roughly 65,000 transactions per second. So to do web scale micro payments, to have bots paying publishers a tenth of a penny every time they read an article, you need a financial network orders of magnitude larger than Visa, dealing in microscopic fractions of a currency. That may be a scalable technological hurdle, but the history of human behaviour on the internet makes me a bit skeptical of this entire cure.

Why Micropayments Always Fail

Behavioural economists have a lengthy well documented track record explaining exactly why micro payments always fail. The seminal work of Clay Shirky coined the concept of mental transaction costs. Historically, human beings absolutely hate deciding whether to spend small amounts of money for individual pieces of content. It's plainly annoying. If you're browsing the web and a pop-up asks you to pay two cents to read an article, the mental friction, the sheer annoyance of having to make a purchasing decision is vastly higher than the actual financial cost of the two cents. We saw platforms like DigiCash try this in the 1990s. We saw Blendle try it much more recently with journalism. They all failed. People hate it. Consumers vastly prefer flat rate subscriptions like Netflix or ad supported free models because it entirely removes the mental transaction cost. We just want to browse in peace. But Prince has a theoretical twist to this, which is that machines don't have mental transaction costs. He argues that a bot doesn't feel annoyance. If I tell my AI agent, "Go find me the most durable full hiking boots. You have a budget of $5 for research fees," the agent doesn't sit there agonising over spending a tenth of a cent to read a review on an independent outdoors blog; it just pays it. It instantly executes the 402 payment, grabs the data, processes it, and moves on to the next site. Agent to agent payments bypass the human psychological barrier entirely. In pure economic theory, yes, the machine to machine element is the new variable that makes some economists pause and say, "Okay, maybe this time's different."

But stay with me, I'm not finished. The early real world data on the execution of the 402 protocol, which Cloudflare heavily promotes and co-governs, has not seen the promised explosive adoption. 402 transaction volume actually collapsed by 77% from November 2025 to May 2026.

The average transaction size on the network was hovering around 20 to 52 cents. That is standard e-commerce territory. That's not the subset volume needed for a true web scale pay per crawl ecosystem. Coindesk even published a scathing investigative report calling the entire agent micropayments boom a mirage.

The Elephant in the Room

And we have to point out the unavoidable elephant in the room. Matthew Prince has a multi-billion dollar conflict of interest in steering things in this general direction. Oh, massive. He is diagnosing a terrible disease, the bot takeover, and the death of open web publishers, and then conveniently selling the exact cure. Every single pillar of his proposed prescription maps perfectly to a Cloudflare enterprise product. They after all sell the bot management tools to monitor the traffic. They sell the AI crawl control to let publishers block non-paying scrapers. They sell the highly profitable workers platform to handle the massive CPU efficiency demands. And they co-founded the x402 Foundation for micro payments. This is a tech CEO's pitch.

The trend data Prince cites, the 57% bot volume, the zero-click crisis devastating publishers, the massive global CPU shortage, that is all independently corroborated. But the idea that micro payments will save the web and that we will suddenly see 100 million transactions per second, those should be treated as highly speculative corporate bets with fairly poor historical odds, not inevitable futures.

We're still not sure where this is going. If traditional display ads are retreating behind walled gardens, how do high quality creators actually get compensated? And on the flip side, how do massive consumer businesses make sure these AI agents actually find their products to buy them?

The Swiss Cheese Model

Well, to address the content side first, Prince introduces a metaphor he calls the Swiss cheese model. He describes large language models, the massive neural networks that act as the brains behind these AI agents, as giant blocks of Swiss cheese. There is a vast amount of knowledge there representing the solid cheese. But there are also distinct holes. These holes represent missing, hyper local, temporarily recent, or completely novel information. Like an AI model trained on the internet up to January 2025 doesn't know what happened yesterday at a local city council meeting. Prince argues that the major AI companies, OpenAI, Google, Anthropic, are desperate for high reputable human generated sources to fill those holes with net new knowledge. They don't need a thousand identical AI rewritten articles aggregating a US presidential press conference. They already have that data. They need the local investigative reporter on the ground finding something new. To facilitate this, Prince envisions what he calls a 'Spotify' model for the internet, (I call it the user telemetry advantage). It's based on how Spotify operates internally. Algorithms constantly analyse user search histories that return zero good results.

For example, maybe 10,000 people search for a fast-paced song with a disco beat about how much fun it is to dance with my cat. If Spotify has no songs that fit that exact criteria, they publish those unfulfilled queries to their network of creators. Prince mentions a specific Danish artist who makes 40 million euros a year, doing nothing but writing hyper niche, highly specific songs to fill these algorithmic lacunas. And Prince believes this exact model will map to text-based information. On the open web, creators will be compensated directly by the massive AI companies for generating specific unfulfilled queries to fill the holes in the models.

This trend is already starting to play out at high levels. We've seen OpenAI sign a landmark $250 million deal over 5 years with News Corp, the publisher of the Wall Street Journal, to ingest their journalism. They signed a similar $16 million annual deal with DoorDash. So currently the AI companies are paying directly for the premium cheese to fill their holes. There's of course a dark pattern-like monopolistic side of this Spotify model. Media economists warn that while these licensing deals sound great in headlines, they overwhelmingly, almost exclusively, favour massive corporate incumbents.

If you are News Corp and you own the Wall Street Journal and the Times of London, you have the leverage to sit across the table from Sam Altman and demand a $250 million cheque. If you are an independent tech blogger, a local recipe creator, or a midsized investigative journalism outfit, you have zero leverage.

Critics argue the system will leave small publishers fighting for what they call 'tip jar money', consolidating media power even further into a few mega corporations while absolving the AI companies of their past unauthorised scraping without building a truly equitable future economy for independent voices.

So what does this all mean for commerce? We've talked about content and journalism, but what happens when you actually want to sell a physical product? What is a brand in an agentic world?

What Is a Brand to a Machine?

Think about it. When a human being driving down the highway sees the golden arches of McDonald's or the mega green Bunnings Warehouse, the logo acts as an emotional, psychological shortcut. You instantly know what to expect and feel a certain way. But an AI agent doesn't care about brand recognition, a logo or a catchy jingle or celebrity endorsement. An AI bot has infinite patience to analyse 5,000 spreadsheets of durability metrics, pricing histories, and material sourcing in a millisecond.

And so to yet another functional disruption to the internet economy story. Agent engine optimisation or AEO. We spent the last 20 years doing SEO, search engine optimisation, trying to make websites appeal to Google's human chasing algorithms so people would click them. Now brands must pivot rapidly to AEO which better optimised for machine legibility. They have to structure their product data, their specs, their pricing and their inventory in highly organised backend formats like JSON or schema markup so that an AI agent can instantly read, compare, and validate it without ever looking at the beautiful, emotionally resonant human facing website. The financial projections on this pivot to AEO rather jaw dropping. Bain and Company released a projection that the US agentic commerce market could reach between 300 and 500 billion by 2030. That would make up roughly 15 to 25% of all e-commerce. It means that within just a few years, a quarter of the entire digital economy might literally be machines buying things from other machines on behalf of humans. Negotiating prices in milliseconds based purely on machine legible data.

Does Emotion Still Sell?

But I have to challenge this vision of the future because it feels like it completely ignores fundamental human psychology. We are purely rational creatures. If I tell my agent, "Buy me the best running shoes." Does Nike's multi-billion dollar deeply emotional marketing campaign completely fail? You'd think so based on this. A brand built on inspiration on Michael Jordan soaring through the air lose out to a generic no-name shoe brand that happens to score 2% higher on a machine legible spreadsheet of heel drop and foam density metrics? If we believe that's how the tectonic displacement of e-commerce will transpire, we can look forward to JSON code substituting all the stuff that appeals to human desire, impulse, and emotion. OK, maybe not ALL the stuff about brands that gets us hooked, maybe there's a hybrid reality, a staged adoption curve on the horizon? In the first stage, agents act purely to assist discovery. They analyse the spreadsheets and narrow the field of thousands of running shoes down to three highly qualified options, but the human still makes the final choice based on brand affinity and aesthetics. Okay, that makes sense. Then we move to agentic shopping, where the human picks the Nike shoe, but the AI scours the web to negotiate the best price. Finally, for low stakes, highly commoditised items like buying paper towels or batteries or dish soap, we move to total autonomous purchase where the brand truly doesn't matter to the consumer, only the specs and the unit price. Because I don't care what logo is on the dish soap. Nike's emotional marketing will still matter, but perhaps only at the final human stage of a funnel heavily curated by machines.

What Is the Human Internet For?

AI automation is no doubt set to reshape the public internet, how we read and how we buy. If the future of the internet is one where AI agents do all the browsing, read all the articles and summarise them, and where AI agents negotiate and buy all the products based on machine legible spreadsheets, what exactly is the human internet for? The next 5 years will define exactly where human beings, our messy curiosity, and our desire for connection actually fit into this new hyper efficient business model.

,,,,