A Click From OpenAI Is 750 Times Harder Than From Google a Decade Ago

Sarah Guo opens with the scale: a $66 billion market cap, about $1.8 billion in trailing revenue, and a company widely treated as the largest CDN on earth, now stacked with security products. Matthew Prince, Cloudflare's co-founder and CEO, declines the CDN label. The original thesis was a firewall in the cloud: servers, software, and then networking gear were all moving there, and the objection everyone raised was speed. The team worked only to get back to parity, overshot, and everything got faster — which is how they ended up competing in CDN anyway.

What Cloudflare is trying to be, Prince says, is what the internet should have been had anyone in the 1960s through 1980s known how important it would become: faster, more reliable, more secure, more efficient, more private. The company launched in September 2010 and was coming up on fifteen years. Asked how it became so dominant, he refuses a grand strategy. The story, as he tells it, is that Cloudflare has been customer zero the entire way.

A cloud firewall needed data, so they shipped a free service. The free tier produced millions of customers and every kind of weird traffic, which meant they got attacked from every direction and had to build a public-policy team. Someone nearly stole their domain as a way in, so they built a registrar. Start with a simple idea, make it broadly available, then solve the problems that appear once you have done that. Guo calls the resulting position the internet's traffic cop: a seat from which Prince can see how traffic, crawlers, and AI companies actually behave.

Prince's starting claim is historical. For thirty years the web's dominant value-creation model has been search. Whole industries grew up around it. You could sell a product or a subscription, sell ads against content, or — the piece he thinks people skip — create for the ego of being read. The old line is that people make content to get rich or get famous. Wikipedia, he notes, runs on the second motive.

The interface is now shifting from search to AI. Google usage and Cloudflare's own logs both show people turning to assistants where they used to turn to Google, and Google's own results page is morphing into an AI product in public. The natural consequence is that users consume derivatives instead of original pages. A Pew Research study found that an AI overview at the top of search makes people much less likely to click through. Prince calls that a duh. Cloudflare's data is harsher.

Compared with ten years ago, the same piece of content is ten times harder to get a click from Google, because of the answer box, AI overviews, and the redesigned search interface — and that, he jokes, is the good news. From OpenAI it is 750 times harder than the Google of old; from Anthropic, 30,000 times harder. If the web's value model was traffic, the new interface does not send it. Lose the product sale, the ad, and even the knowledge that someone read you, and people stop creating, starving the web first and then the AI companies that use the same content as fuel.

Gil pushes the next step: agents will take actions in the background, so time spent on the open web falls further. Prince splits the problem. Agentic commerce is the easy part — plenty of companies will *want* an agent to buy a widget. Aggregators of information are the threatened middle, because agents can disintermediate them. Content is the hard case, because the default assumption has been that you get it for free.

Prince, a recovering law professor, thinks creators are looking in the wrong place. In copyright, the more derivative a use is versus a direct copy, the safer it tends to be under fair use. Two California cases landed within a week of each other, one treating AI use of content as fair use and one not. The more sensible reading, he argues, is that derivative generation is less likely to be a violation — which is also when a reader is least likely to go back to the source. Creators are betting on a body of law that may protect the very uses that kill their traffic.

The business-model conversation with the labs is more encouraging. About 80 percent of the major AI companies are Cloudflare customers, so Prince talks to them constantly. They agree they should pay for content. The sticking point is a level playing field. Each lab believes its technology is the best and that it will win on equal terms. None of them will sign up if Google still gets the corpus free, or if they pay while a rival does not.

Markets require scarcity, and content currently has none because it is given away. After a year of talking across print, video, audio, music, and film — from the Associated Press to Ziff Davis — publishers told him the last six months had made the old model astronomically harder. On July 1 Cloudflare announced Content Independence Day: for free, across paid and unpaid customers, it began blocking AI training crawlers by default. The point of treating Google like everyone else was to manufacture the scarcity a market needs before anyone can set a price.

The next step is rules of the road. Cloudflare is working with the IETF and other standards bodies so a crawler has to announce what it is doing. The aim is fine-grained permission: humans free, robots pay. Today's `robots.txt` is a blunt street sign, allow or disallow, directory or whole site. Plenty of prominent companies say they follow the speed limit, then, when blocked, behave not unlike Russian or Iranian hackers to get around it. Cloudflare is good at spotting that, Prince says, and good at embarrassing the firms that do it. Watch the blog: some prominent AI companies are going to get called out.

Asked what a training marketplace would do to winners and losers, Prince takes a tangent he knows will land badly. A lot of what is wrong with the world, he says, is ultimately Google's fault — then walks it back a half step. Google has been a net force for good, believes in ecosystems, and is trying to do the right thing. What it taught every creator was to worship a deity called traffic. Traffic became the proxy for value.

That religion produced Facebook, then TikTok, then the Huffington Post A/B-testing headlines for whichever one generated the largest cortisol spike. It produced Demand Media and BuzzFeed and a generation of me-too stories with a slightly different bend, optimized to stir people into a click so someone could sell a subscription or an ad. Prince does not think it advanced human knowledge. It taught the web to copy itself.

If content can actually be sold, what is it worth? Taken together, Prince says, the AI companies are a decent approximation of the sum of human knowledge — not perfect, but the best we have had. In aggregate they look like a giant block of Swiss cheese: a lot of cheese, and holes. Their own algorithms prune whatever already sits in the meaty part. Content that lands in a hole is extremely valuable to them.

A market that rewarded filling those holes, rather than stimulating the most cortisol, would actually advance knowledge. Guo asks whether that is just Mercor, Surge, or Scale — expert labeling, distributed. Prince thinks it can go further. For the first time, the pruning math of the models can point at the gaps and publish a bounty. Write the missing article about an obscure animal, get paid. That beats another piece on whether San Francisco is in decline, which he calls rage bait.

He flew to Stockholm to see Daniel Ek because almost nobody has compensated creators at this scale. The day before iTunes launched, the music industry was an $8 to $9 billion business. Spotify alone now pays the industry more than $10 billion a year. Done right, the model expands the pie. The detail that stuck: Spotify publishes search queries it cannot answer — someone wants a disco song about dancing with your dog — and several musicians now make tens of millions of dollars a year writing for that unmet demand.

Guo cites Google's Med-PaLM 2, already a couple of years old and already outperforming the average physician on rated answers, and asks when good human content runs out. Prince rejects the premise. People will keep running experiments and making discoveries. A future of robots in the lab is distant. In the meantime the market, not a labeling-company budget, can decide whether a unit of content is worth more or less than another unit of GPU time.

Prince's dystopia is not mass unemployment of journalists, scientists, or researchers. The work still has to be done. What he fears is a return to something like the Medici era. Five big AI companies — a conservative one, a liberal one, a European one, a Chinese one — each hire their own newsroom, academics, and experts to fill holes in *their* cheese. Knowledge becomes a private asset behind each lab's wall. The web, historically a distributor of value and knowledge, would shrink into five silos. The outcome he wants is independent creators paid, and that knowledge shared across models.

Gil asks whether the labs will actually pay enough, given that labeling spend is still a rounding error next to a web that started from free. Prince will not treat labeling firms as the right comparison; employee cost and GPU spend are going up too. Content's contribution is somewhere between zero and 100 percent of the value, and a market with scarcity can price it against GPU-hours. The mistake already being made is deals that do not scale with the labs.

A $20 million check for all of your content looks great on day one and is, he says, an incredibly naive contract. What you want is a percentage of the model's subscription fee, or of future ad revenue if the product is ad-supported, so that as the AI company grows — and as open-web ad dollars shrink — you share the upside. Advertising, subscriptions, and tentpole content that people still have to consume will not vanish. Creators should still get their work into the systems, and get paid for it.

As the web shifts from serving pages to running models, Guo asks how Cloudflare thinks about inference. In 2020 the company partnered with Nvidia to put GPUs at the edge. The market returned crickets: not a single sales inquiry. They apologized to Nvidia and moved on. Four years later they reissued the same press release and the product took off. Prince still believes most inference will run on the end device. Models that are too big or too hungry belong at the next-best place: inside the network, at the edge.

Because Cloudflare already sits in front of so much of the internet, MCP and whatever protocol connects agents to services have to pass through it. The company is investing in the security, rails, and payments those protocols need. Prince's hunch is that the same payment rails built for content become a model for agent-to-agent settlement. Asked whether they will build a neocloud, he laughs that he does not even know what a neocloud is. The team follows customer zero and the 80 percent of AI companies already on the network.

Closed models run on Cloudflare; more open ones do. Internally they open-source what they can, and they work closely with Meta on Llama. Prince is happy to host either end of the spectrum and calls the claim that open models will end the world histrionic. Synthetic pathogens are worth worrying about; the place to regulate them is the machine that can print them, not the model that can describe them.

For a content marketplace to appear, Prince says the tactical next step is to stop treating Google as a special snowflake. The historic bargain was content in, traffic back; Google still takes as much content and sends about one-tenth the traffic, and the trend is worse. Blocking Google was unthinkable ten years ago, radical six months ago, and is now a live conversation. Other labs fear Google most because they think it has privileged access. Prince's ask is narrow: keep indexing for search, but if the bot turns pages into an answer box, an AI overview, or Gemini, that is a different deal, in the same bucket as every startup.

Direct licenses are already happening at the top — Condé Nast, Dotdash Meredith, the New York Times, Reddit, each sitting down with a large model company. Cloudflare's opening is the cases that do not scale one-off: a big publisher covering every AI startup, or the long tail facing every lab. It might negotiate a Spotify-like pool and distribute it, charge a micropayment per access, or price training differently from search. None of that settles until Google plays by the same rules. The day it does, he thinks the rest of the market happens faster than people expect.

The advice for individual creators is three moves. First, get back control of the content and stop giving it away; Cloudflare made that one-click, and other companies are trying to as well. Second, look at the data: see which AI companies keep crawling you. Third, go start the conversation about a fair exchange. The same day as the taping, news broke that Google was piloting payments to news providers, something it had sworn it would never do. If you cannot sell a thing, sell ads, or even get the ego hit — if an influential piece lands in millions of AI replies and you never know — you are yelling into the void.

Prince already thinks a lot of inference runs on-device, and some of it has to. A driverless car that sees a red ball bounce through a yard, a child running after it, cannot wait on network conditions to decide whether to hit the brakes. The unlock that does not feel like today is power efficiency on the inference side. Apple has shown that relatively efficient GPUs and TPUs are possible. Conversations with Nvidia, he says, feel like talking to Intel in 2010, or Apple in 2005: you are doing it wrong if you care about watts.

In 2011 he sat in Intel's lab outside Portland as a tiny startup using their chips and repeated that the only number he cared about was cores per watt. Intel told him to water-cool. Cloudflare did not have that luxury; it had to fit a limited power envelope in the oldest, most legacy data centers. He sees the same fight in AI. More GPU capacity arriving only if you stand up a mothballed nuclear plant cannot be the solution, and there is no physics reason it needs to be. The company is shopping the ecosystem for whoever can deliver the most tensor units per watt, because the edge of the network lives inside a power envelope that is not a 100-megawatt data center.

From everything Cloudflare can see, most AI companies are still inefficient in utilization. Cloudflare charges for actual work, unlike hyperscalers who rent you a GPU whether you use it or not, so a 100-times inference win is great news for them and terrible news for those clouds. The lesson of DeepSeek, he wishes, had come from students in Hungary rather than a Chinese lab, because the American argument about whether it was real drowned genuine efficiency science: model compression and pruning of vanishingly unlikely branches have barely been touched. He will not train a frontier model; he will run whatever he hosts much more efficiently. Someone will become the VMware of AI, slice GPUs instead of renting whole machines on annual commits, and speed-run thirty years of CPU efficiency — Spectre included — in five to ten years on GPUs, until a current-generation ChatGPT equivalent fits on a phone.

Cloudflare will compete in agentic infrastructure, Prince says, and it will not be one provider; standards will matter, whether that is MCP or Google's more Microsoft-flavored variant. Micropayments on about 15 trillion Cloudflare requests a day are the use case blockchain waited a decade for, and Bitcoin or even Solana will not carry that volume as-is. Identity has to distinguish a human, an agent tied to a human, and a self-directed agent, with cryptographically signed promises about what will be done with the data. A browser that vacuums a page straight into an LLM may get tighter permission than one that does not, on the same rails that police bots, with zero-knowledge proofs for "prove the credential without showing it." Six months earlier he could not see how the blocks fit; in the last few months, the shape of that future has started to come into view.