Matthew Prince, co-founder and chief executive of Cloudflare, told The Verge's Decoder podcast in an episode published on September 26, 2026 that automated traffic overtook human traffic across his company's network in May, more than a year ahead of the forecast he gave at Web Summit in November 2025, and that machines will generate roughly a thousand times the human volume by around 2031. The interview, available at https://www.youtube.com/watch?v=4t8RhKbkb70, also set out how Cloudflare intends to route money from AI companies back to the people who publish, why Prince sorted his workforce into builders, sellers and measurers before cutting about a fifth of it, and what he expects Google to concede on crawling.

In Short

Cloudflare's boss says most of the traffic on the web is now produced by software rather than people, and that within a few years machines will produce almost all of it. That matters because advertising, which has paid for most of the web for 30 years, only works when a person sees the ad, and a crawler never does. If his plan works, you would see websites charging AI companies a fraction of a cent per page instead of relying on ad impressions, with local news and unique communities such as Reddit commanding the highest prices.

The forecast that kept arriving early

Prince told host Nilay Patel that the question of when non-human traffic would pass human traffic was first put to him at Web Summit in November 2025. Cloudflare pulled its own data and answered: the second half of 2027. Asked again at South by Southwest in March 2026, the company re-ran the numbers and moved the date forward to the first half of 2027. Then, in May, his teams came back with a different message. "You won't believe it, but automated traffic is now past human traffic online," Prince recalled them saying.

The crossover had been public since June 3, 2026, when Prince posted Cloudflare Radar figures showing automated systems at 57.4% of HTTP requests for web content against 42.6% from people, a threshold PPC Land examined at the time. Within that majority, training-related crawlers alone accounted for 50.6% of all traffic, while search-related bots made up 10.7%. A later Decodo analysis built on the same telemetry found that six of the top 60 traffic-source countries already send more machine requests than human ones.

What the interview adds is the extrapolation. Prince said Cloudflare now expects automated traffic to run at a thousand times the human level roughly five years from now, not because human traffic will fall but because the machine side keeps compounding. He attached a qualifier of his own: "with the giant caveat that I've gone wrong so far in every prediction that I've made on this." The episode's own description from The Verge phrased the same forecast as bots that "could outnumber humans a thousand to one within five years."

Cloudflare's vantage point rests on scale. Prince described equipment running in more than 350 cities and thousands of data centers, and put the company in front of more than 20% of the web. Those figures are the reason its Radar dataset is treated as a proxy for the internet as a whole, and the reason its defaults carry weight. According to Prince, Cloudflare takes five minutes to sign up for and, in his words, "It takes 30 seconds to leave," which he offered as the check on that power.

Why advertising cannot carry a machine audience

The economic argument runs through one sentence. "Bots don't click on ads," Prince said. For at least 30 years, according to him, advertising has been the business model that funded the growth of the web, and Google has defined that model for 28 of them, buying DoubleClick and building Google Analytics along the way. A crawler responds to none of it. Somebody still has to pay for the bandwidth and servers that a thousandfold increase in requests would consume.

Prince's answer is a status code. HTTP 402, reserved for "payment required" since the earliest versions of the protocol, would let a site return a price rather than a page when a bot arrives. He initially cited it as 403 before correcting himself to 402. Cloudflare has been working with Coinbase and Stripe on the settlement layer, he said, so that a crawler can pay "a thousandth of a penny" or a fraction of a penny per access, with the money ultimately coming from the subscription fees people pay for AI assistants and agents. The comparison he reached for was Spotify or Apple Music: a pooled subscription that flows back to rights holders according to usage.

The mechanism is not new to PPC Land readers. Cloudflare opened pay per crawl in private beta on July 1, 2025 using 402 responses, expanded it with customisable payment messages on August 28, 2025, and updated an open source x402 proxy template with per-route pricing in January 2026. By August 2026, Cloudflare customers were issuing more than one billion 402 responses a day, according to Decodo figures reported by PPC Land.

One tension sits inside the interview. On July 1, 2026, Cloudflare moved away from charging per crawl toward paying publishers when their content contributes to an answer, after finding that more than half of the crawl traffic from bots it classifies as legitimate went toward re-fetching pages that had not changed. Prince did not mention that shift on Decoder, describing instead a per-access micropayment and a premium on top of it for publishers and academics who believe their material is worth more.

Scarcity, lunch menus and a tragedy of the commons

Prince framed the problem as one of missing constraints. Ask an AI agent where to eat lunch, he said, and it scans every menu in the area rather than the two or three a person would open. Only one restaurant gets the order. Extend that behaviour across the web and an agent will read every resource available before returning a recommendation, at a cost borne entirely by the sites it visited. A charge, even a tiny one, forces the agent to be selective.

Patel, a former copyright lawyer, pushed the point further: physical media once imposed scarcity on information, the internet removed it, and attention became the scarce thing instead. Prince agreed that markets need constrained supply, comparing open access to air on the surface and paid access to air underwater. The agentic AI systems generating the traffic are, on that reading, the first buyers who will accept a per-page price because their operators already pay per token for everything else.

He put the demand side in numbers of his own. Prince claimed that Anthropic had booked $10 billion of revenue in a single month, and argued that AI companies need three things: researchers, chips and data. Universities and chipmakers will ease the first two constraints over time, he said, while data is the input moving the other way, from freely available toward gated. According to Prince, the large AI companies are willing to pay and the large content owners are eager to sell; what has been missing is the technology connecting them.

The Spotify analogy and a Danish songwriter

The longest exchange concerned music. Prince said he flew to Stockholm to meet Spotify founder Daniel Ek, and described the arc from Napster to iTunes at 99 cents a song to streaming. He put the music industry's value at roughly $8 billion to $9 billion around 23 years ago and said Spotify sent something like $12 billion back to the music ecosystem last year.

That figure does not match Spotify's own disclosures. Its Loud and Clear report, covered by PPC Land in April 2026, put 2025 payouts at more than $11 billion, following $10 billion in 2024. Prince may have been rounding or citing a different measure; the interview does not say.

Patel disputed where the streaming money lands, arguing that touring, sponsorship and sync licensing grew while per-stream income stayed small, and that private equity rather than musicians now collects on catalogues sold for lump sums. Prince countered that catalogue buyers pay because streaming income is lucrative, and that more money now flows into music creation than at any point in history. Neither side conceded.

The detail Prince wanted to land was a feedback loop. Spotify, he said, publishes the searches it cannot answer well back to creators, and one songwriter in Denmark earns about EUR 40 million a year writing songs for those unfulfilled queries. Large language models, in Prince's description, make the same map at the scale of human knowledge: "we've built a mathematical model of human knowledge," he said, and the model's gaps show where new information is worth paying for. Executives at the big AI labs, he added, do not want another story about the White House. "They want new true knowledge."

Local news and the Reddit price

Prince and his wife own the Park Record, a newspaper in Park City, Utah. "I think we will make more money off AI licensing deals this year than we do off digital advertising," he said of it. His reasoning was that the media economy built on volume and scale hollowed out local reporting, while an AI assistant that wants to be a competent travel planner needs to know which restaurant just opened on Main Street, and only the local paper knows.

He cited a pricing gap to make the case. The New York Times and Reddit hold roughly similar volumes of content measured in tokens, according to Prince, yet Reddit gets at least seven times more money per token in its deals, and by some measures 14 times. His explanation was substitutability: without the Times an AI company could license the Wall Street Journal and have a model rewrite it in a different register, while nothing substitutes for Reddit. He called the comparison unfair to the Times even as he made it.

The public record supports the shape of the gap without confirming the multiple. Reddit's Google deal, signed in February 2024, has been reported at about $60 million a year and entered renewal talks in July 2026 as several publishers weighed cutting Google's crawler off. The Times, meanwhile, disclosed in August 2026 that it spends close to $2 billion a year producing about half a million works while suing three AI companies and licensing to a fourth.

Prince's second concern was recognition rather than revenue. Musicians, he said, want fame and money, and "Fame probably beats rich." Wikipedia's contributor numbers are falling, in his account, because people no longer see readers arrive at the source. He has pitched the large AI labs on an annual award, measured mathematically, for the people whose work contributed most to a field. PPC Land's own reporting on Wikimedia data found human traffic to Wikipedia down 8% over a year even as crawler activity pushed total visits up.

Google, September 15 and what changed after the recording

The most consequential exchange concerned Google. Prince described a company that struck deals to index the whole web when it sent traffic in return and now trains on the same pages while sending less. "That's a different give to get," he said, comparing the arc to a Marvel film in which "the hero of yesterday becomes the villain of tomorrow."

Then he named a date. Cloudflare would, as of September 15, set the defaults across its free customers so that Google was blocked for AI training. "But if they won't differentiate between AI training and search engine training, we're just going to block them across the board," he said, adding that any site owner could turn the default off. Patel, who coined the phrase Google Zero for the end of search referrals and said "I don't think he loves the fact that I am the person who keeps calling it Google Zero" of Sundar Pichai, asked whether the screws had to be turned that far. Prince answered that Google's crawl is the more exposed party: Googlebot builds its ranking from a tree of links, and removing 20% of the web from that tree breaks the tree for everything.

The interview was recorded before September 15. Prince spoke of the default in the future tense, and referred to a Justice Department filing in support of OpenAI's fair use defence and a music industry suit against Anthropic as having happened "today"; the Department's statement of interest in the New York Times case was filed on September 3, 2026, and Sony Music Publishing and Warner Chappell sued Anthropic in the Northern District of California on August 28, 2026, according to Music Business Worldwide.

What Cloudflare shipped on September 15 differs from the ultimatum. According to a blog post by Bryan Becker, the company's director of product security, Cloudflare added a Disallow AI Training setting that publishes a no-training preference in robots.txt while allowing Googlebot, Applebot and Bingbot to keep crawling for search. Apple, Google and Microsoft received an Accountable designation on the strength of four commitments: an opt-out from training, an opt-out from AI summaries, URL-level visibility into which pages were made available for training, and an assurance that opting out will not affect search rankings. Microsoft's support for the robots.txt preference is targeted for early 2027. The Block option now stops all three crawlers entirely, search included, which is the outcome Prince threatened, but as a choice rather than a default.

Cloudflare's post also disclosed the demand behind the split: fewer than 1% of its sites block search bots, while 17% have enabled some mechanism to block training. New ad-monetised domains are offered Disallow AI Training and a block on agents for pages carrying ads by default, a policy first set out on July 1, 2026 alongside four disclosure conditions for mixed-purpose crawlers. Cloudflare aims to let site owners control how much of their content appears in AI summaries by early next year.

Prince's account of Google's position matched that direction. He said Cloudflare had pushed Google to split its crawler, that Google resisted on the grounds that crawling twice doubles the load on every site, and that Google instead committed to announcing what its crawler is doing on each visit and letting publishers refuse specific uses. He attributed the shift largely to a ruling in the United Kingdom. The Competition and Markets Authority imposed its first Publisher Conduct Requirement on Google on June 3, 2026, the same day Google began testing a Search Console opt-out toggle, and Prince said Google had signalled it would apply the procedures beyond the UK. Four months earlier a Google executive had called a separate AI Overviews opt-out a huge engineering project.

The interview extends a line Prince has run for over a year. In July 2025 he said Cloudflare would get Google to provide ways to block AI Overviews without losing search indexing, and in January 2026 he argued that Google's crawler reaches 3.2 times as much of the web as OpenAI's. On Decoder he said Google's tone had changed over the past few months and credited people inside the company who want the ecosystem to survive.

Law as a patchwork, technology as the lever

Prince was dismissive of litigation as the primary fix. The best case for AI companies, in his telling, is the Anthropic books decision, where the judge found training to be fair use while faulting the company for acquiring pirated copies; a result valuable enough, he said, that Anthropic settled the rest for $2 billion. The court-approved figure was $1.5 billion, reported by PPC Land in September 2025, and Prince himself had then described the payment as the cost of keeping a favourable precedent dominant. The $2 billion figure in the interview is unexplained.

His preferred instrument is identity. Cloudflare spends its days, he said, against state-backed attackers who hide well, whereas AI companies are registered corporations that increasingly declare who they are on arrival so that site owners can set terms. That was not always the case: in August 2025 Cloudflare documented Perplexity fetching pages while disguising its bot identity, and Prince at the time compared its behaviour to that of state hackers. On Decoder he said the deals large publishers now sign are substantially better than two years ago, and that the open question is extending that to the rest of the web before a commons problem drains it.

Prince also described a countervailing trend. The web as a whole plateaued around 2012 and shrank through 2025, in his account, until something flipped in October 2025 as vibe coding platforms let anyone stand up an application. The growth of the web, he said, is "faster today than it has been at any time since the early 2000s." Patel's rejoinder was that the web is thriving as an application platform while suffering as an information platform, "something happening to the web as an information platform that is devastating," and that the incentive to publish new information on it is in permanent decline.

Builders, sellers and measurers

The second half of the episode turned to Cloudflare's own workforce. Prince laid off 1,100 people, about 20% of the company, and defended the decision in a Wall Street Journal op-ed that ran under the headline "How I choose which Cloudflare employees to replace with AI." He said he did not pick the headline but wrote everything beneath it. Cloudflare now employs between 4,500 and 5,000 people, he said, more than it did two years ago.

The dating shifted within the conversation. Patel placed the layoffs in May; Prince later asked whether it would be easier to find a job "back in April when we did this" or a year on when everyone is doing it. PPC Land's reporting in June 2026 recorded that Cloudflare had let go of more than 1,100 employees the previous month, which points to May.

The rubric came from Peter Drucker. Every firm, Prince said, has builders who create products, sellers who close deals, and measurers who count what the other two do. The majority of those cut were measurers: internal audit, finance functions and a layer of middle management. His analogy was an electric screwdriver arriving at a company that had hired the best manual screwdriver operators. The most senior staff adopted AI tools out of confidence, the most junior because they were native to them, and the group in between, trained to succeed under the old rules, resisted until Cloudflare began retraining everyone in mid-2025.

Two examples carried the argument. An investor relations team of about 20 people used to spend two weeks after each quarterly close assembling earnings documents. "We took what used to take 2 weeks and we reduced it down to 3 minutes," Prince said, adding that investors found the output contained fewer errors. On management, he cited the business school convention of six direct reports per manager and said tooling pushed the right number toward 12, which flattens hierarchy. Cloudflare is not there yet, he conceded; the true figure two years ago was closer to six than the eight he claimed at the time, and the company now sits around eight.

Prince said the decision cost him: the share price fell and he received death threats from people opposed to AI, not from employees. He argued that acting first was kinder, because those let go could be placed elsewhere before the rest of the industry followed. "I think we were early," he said, and cited Uber as having just made cuts on the "exact same rationale." Uber announced on September 2, 2026 that it would cut about 3,300 roles, roughly 10% of its workforce, and reduce its manager count by 20%, according to Bloomberg, which first reported the memo.

The internal tooling behind the change, which Cloudflare calls Cloudflare OS, pulls data from Salesforce, Workday and other systems of record into one interface; Prince said the company has open sourced it because customers asked to run their own versions.

Training models on the top performers

The passage that drew Patel's sharpest questions concerned talent identification. Prince said Cloudflare knew who its high performers were, trained models on what they did, ingested signals from across the organisation including code commits, and traced individual actions through to revenue, cost or collaboration outcomes. "AI has bias, but it's uncorrelated to all the rest of the bias in the organization," he said, arguing that humans who eat in the same cafeteria develop correlated blind spots while a model does not.

The models surfaced junior employees who were outperforming and had never been noticed, according to Prince, and he now calls them directly. The models also produced errors, flagging people as stars who were not. Patel asked whether being managed in a quantified way dehumanises a workforce and noted that management by robot can get dark. Prince said a promotion decided by an unaccountable AI system would be wrong, that a system telling a chief executive to call a support agent who is quietly excellent is humanising, and that the outcome depends on leadership. "Don't be a bad organization," he said. He rejected the framing that management had been automated: "I don't think we're replacing managers," he said, describing the tools as surfacing data so managers can do their jobs better.

He also questioned whether separate manager and individual contributor tracks make sense when individual contributors now run fleets of agents, and said senior managers are leaving large roles to become individual contributors at AI labs because one person can get far more done there.

Content moderation and the rage economy

Patel connected the crawler blocking to Cloudflare's earlier decisions to drop the Daily Stormer and Kiwi Farms, asking whether the same infrastructure power now feels more comfortable because it carries an economic rationale. Prince said each case was a moment in time. Kicking off the Daily Stormer, he acknowledged, used a site as a means to an end to shape a policy debate about regulating protocols such as DNS and TLS; another such moment arrives roughly every five years, and one is probably due.

His broader claim was that the attention economy is the root cause. Google begat Facebook, which begat TikTok, in his phrasing, and the result is a media landscape optimised for a mammalian response, with headlines A/B tested for dopamine or cortisol. Neo-Nazi content, he noted, is one of the most effective ways to provoke clicks. Rewarding knowledge creation rather than rage creation would, he argued, remove the incentive for much of what he otherwise ends up moderating one site at a time. He allowed that he might be too much of an optimist.

The episode had drawn 17,978 views and 72 comments by September 28. Several of the most-liked comments disputed the Spotify comparison and the account of where streaming income lands.

Why this matters for the marketing community

For media buyers, the crossover Prince described has already reached the trading desk. Digiday research covered by PPC Land on September 8, 2026 found agencies reporting retargeting pools filling with non-humans and CPMs up 20%, while some direct-to-consumer brands want more agent traffic because it converts. A thousandfold increase in automated requests would make every audience built from page visits unreliable unless the visitor's nature is identified at the network edge.

For publishers, the interview is a statement of intent from the operator whose defaults govern a fifth of the web. Cloudflare's attribution dashboard has documented crawl-to-referral ratios from 118 to nearly 50,000, and the July 2026 decision to pay per answer rather than per crawl set the compensation model that Prince's micropayment vision has to fit. His Park Record claim, that AI licensing will outearn digital advertising this year at a small local paper, is the first such figure from a named owner, and it arrives as Reddit, USA Today Co. and Politico weigh cutting Google's crawler off and Time treats bots as a secondary audience.

The gap between what Prince threatened for September 15 and what Cloudflare shipped is the detail that matters most operationally. Google, Apple and Microsoft keep search access on any site that only disallows training, and a publisher who wants those crawlers gone must now select Block deliberately. Prince has predicted three outcomes for the web before, and the interview shows which one he is now building toward: a licensing market with Cloudflare as the toll booth and identity, not copyright, as the enforcement layer.

Timeline

Summary

Who: Matthew Prince, co-founder and chief executive of Cloudflare, interviewed by Nilay Patel, editor-in-chief of The Verge, on the Decoder podcast. The claims concern Google, Anthropic, OpenAI, Reddit, The New York Times, Spotify, Uber and the Park Record, and affect publishers, advertisers and AI companies operating on or through Cloudflare's network.

What: Prince said automated traffic passed human traffic on Cloudflare's network in May 2026, forecast a thousandfold gap by around 2031, described a micropayment model built on HTTP 402 with Coinbase and Stripe, claimed his local newspaper will earn more from AI licensing than digital advertising this year, said Reddit earns seven to 14 times more per token than the Times, defended the layoff of 1,100 employees using a builders-sellers-measurers rubric, and threatened to block Google's crawler by default from September 15 if it would not separate training from search. Cloudflare's actual September 15 change kept Google, Apple and Microsoft in search for sites that only disallow training. The interview also contains figures that differ from the public record on Spotify's payouts and Anthropic's settlement.

When: The episode was published on September 26, 2026. It was recorded before September 15, 2026, and its references to a Justice Department filing and a music industry lawsuit place the recording in the first days of September.

Where: The Verge's Decoder podcast, distributed on YouTube at https://www.youtube.com/watch?v=4t8RhKbkb70 and through podcast platforms. Cloudflare's network, the subject of the traffic claims, runs in more than 350 cities and sits in front of more than 20% of the web.

Why: Advertising has funded the open web for 30 years and depends on a human seeing a page. With machines now generating a majority of requests and Cloudflare projecting a thousandfold imbalance, Prince argues that per-access payments enforced at the network layer, with crawler identity rather than copyright as the lever, are the only way to keep publishing viable, and that Cloudflare's position in front of a fifth of the web lets it set those terms by default.