Summary

Cloudflare’s thesis is broader than CDN: Prince describes rebuilding the network for an AI-shaped internet. The host cites a roughly $66 billion market cap and $1.8 billion in trailing revenue. Prince says 80% of major AI companies are Cloudflare customers, and sees its network position extending into security, edge inference, crawler permissions, and payments.
AI is breaking the traffic bargain that financed three decades of online content. Prince says the same content is ten times less likely to receive a Google click than ten years ago; compared with “the Google of old,” referral is 750 times harder from OpenAI and 30,000 times harder from Anthropic. As users consume derivatives instead of originals, creators lose product sales, advertising, and even the ego reward of knowing they were read—potentially starving both the web and the models trained on it.
A content market cannot emerge until publishers manufacture scarcity and can technically enforce it. Cloudflare’s July 1 “Content Independence Day” made AI training blocked by default, free for paid and unpaid customers alike. Cloudflare is working with the IETF and other standards bodies on crawler declarations and granular permissions such as allowing human access while requiring robots to pay. Robots.txt is only “the street signs”; Prince says some prominent companies evade blocks using tactics resembling Russian or Iranian hackers and suggests some misbehaving AI companies may soon be called out.
Prince wants AI to reward information gain rather than Google-era engagement arbitrage. Search taught publishers to worship traffic, producing A/B-tested headlines designed for “the largest cortisol response”; models instead resemble Swiss cheese, where redundant content is pruned and unfilled holes are valuable. Spotify is his proof that a new distribution system can expand the pie: the music industry was roughly $8–$9 billion before iTunes, while Spotify alone now pays more than $10 billion annually.
Google is the gating factor because other model providers fear its privileged access to free content. Prince wants ordinary search indexing separated from transformations into answer boxes, AI Overviews, and Gemini: “That’s a different deal.” He calls a flat $20 million license for an entire archive “incredibly naive” and favors compensation tied to AI subscription or advertising revenue, so creators participate in the upside as their legacy economics erode.
Inference economics favor Cloudflare if models become dramatically more efficient and workloads spread from hyperscale data centers to devices and the edge. Cloudflare’s 2020 NVIDIA edge-GPU launch produced “crickets,” yet the same proposition took off four years later. Prince expects the industry to “speed run the last 30 years of CPU efficiency gains” in GPUs over five to ten years; a 100-fold inference-efficiency gain would benefit Cloudflare’s usage-based model while hurting hyperscalers renting whole GPUs.
Agentic traffic could turn Cloudflare’s network position into identity, permissioning, and payment infrastructure. Cloudflare handles roughly 15 trillion requests a day, making per-access micropayments too large for Bitcoin or even Solana as described. Prince imagines cryptographically signed identity and permissions distinguishing a human, a human-directed agent, an autonomous agent, or a browser, including restrictions on how retrieved data may be reused.

Deep dive

1. Cloudflare was built as the missing layer of the internet

The host opens with the operating scale: a roughly $66 billion market capitalization, about $1.8 billion in trailing revenue, and the largest CDN footprint by far. Prince immediately corrects the category: “We’ve never really thought of ourselves as a CDN.” The founding question was whether a firewall could move into the cloud without making the internet slower.
Cloudflare launched in September 2010 and was approaching its fifteenth year. Prince’s larger description is architectural: the company is “what the network should have been, what the internet should have been” had designers in the 1960s, 1970s, and 1980s known how central it would become—faster, more reliable, secure, efficient, and private.
Customer zero explains much of the product expansion. A free firewall supplied the data it needed, but serving millions of customers brought unusual sites, attacks from every direction, public-policy problems, and internal security demands. After someone nearly stole Cloudflare’s domain, the company built its own registrar: start with one broad service, then “solve all the problems that become sort of inherent.”

2. AI destroys the referral economics that search created

Prince reduces the old web to three sources of value: sell a product or subscription, sell advertising against content, or gain recognition from being read. “There are only two reasons why people create content: to get rich or to get famous.” Wikipedia and much of the open web depended on that third, nonfinancial incentive.
Search was the dominant value-creation model for 30 years, but AI is becoming the new interface—even inside Google itself. The structural change is that users consume derivatives rather than the original work, while answer boxes and AI Overviews resolve the query before a click becomes necessary.
Cloudflare’s data makes the deterioration explicit: compared with ten years earlier, earning the same Google click has become ten times harder. Against “the Google of old,” Prince puts OpenAI at 750 times harder and Anthropic at 30,000 times harder. The Pew finding that AI Overviews reduce link clicks is, in his words, “sort of like a duh.”
A host separates agentic commerce from content. Prince expects commerce to be comparatively tractable because merchants want agents to buy their widgets, although information aggregators may be disintermediated. Content begins from a worse default—models take it for free—so disappearing traffic could starve both independent publishing and “the fuel for their engines.”

3. Copyright cannot solve the derivative-content paradox

Prince, a “recovering law professor,” sees an uncomfortable inversion in copyright doctrine: the more derivative an AI output becomes, the more likely it may qualify as fair use. Yet that same transformation makes users less likely to visit the source. Two California decisions within one week reached opposite conclusions, reinforcing his warning that today’s law may not rescue publishers.
His conversations with major AI companies are more encouraging in principle: “You’re absolutely right. We should be paying for content.” The objection is competitive symmetry. No lab wants to pay while Google or another rival keeps obtaining the same material free, especially when each believes its technology will win on a level playing field.
Prince’s sequence is categorical: “In order to have an economy, you have to have a market. In order to have a market, you have to have scarcity.” After discussions spanning print, audio, video, music, and film—from the Associated Press to Ziff Davis—Cloudflare made AI-training access blocked by default on July 1, as a free service for all customers, paid or unpaid.
Robots.txt remains too blunt and too voluntary: Prince compares it to roadside signs that drivers can ignore. Cloudflare is working with the IETF and other standards bodies on crawler declarations and granular permissions such as “humans can get my content for free, but robots have to pay.” He also alleges some prominent companies circumvent blocks with hacker-like tactics and suggests that some may soon be called out.

4. A better market would pay creators to fill knowledge gaps

Prince blames Google—not as the worst actor, but as a broadly beneficial company with damaging incentives—for teaching creators to worship traffic as a proxy for value. Huffington Post-style headline testing sought “the largest cortisol response,” while Demand Media and BuzzFeed optimized rage, repetition, and slight variations on stories rather than advances in knowledge.
His alternative metaphor is a “giant block of Swiss cheese.” Collectively, AI models approximate human knowledge, but their algorithms prune information already represented in the solid cheese; the holes remain unusually valuable. A content market could therefore reward whoever fills those holes, rather than whoever stimulates the most cortisol.
Daniel Ek and Spotify supply Prince’s strongest commercial example. The music industry was approximately $8–$9 billion the day before iTunes launched; Spotify alone now pays more than $10 billion annually. Spotify also publishes unanswered searches to musicians, and Prince says some earn “literally tens of millions of dollars a year” producing music for that identified unmet demand.
A host asks whether expert-data firms such as Mercor, Surge, or Scale already perform this function, and cites Med-PaLM 2 outperforming the average physician. Prince rejects the idea that useful human content runs out: experiments and discoveries continue. His darker scenario is five political or geographic AI silos employing their own journalists and scientists instead of compensating independent creators whose knowledge all models can share.

5. Google must lose its special status before pricing can clear

The scale challenge is that aggregate payments to labeling companies may be around $10 billion, far below the open web’s advertising economics. Prince says there is value in content, but creators’ mistake was signing deals that do not scale with AI businesses; he favors tying compensation to a percentage of subscription or advertising revenue.
A publisher granting all its material for a flat $20 million makes, in Prince’s view, “an incredibly naive deal.” Compensation should instead scale as AI economics scale: some percentage of subscription fees or future advertising revenue, allowing creators to share the upside as the advertising and referral revenue displaced by AI declines.
The immediate fight is to stop Google being “a special snowflake.” Its historical bargain exchanged indexing rights for traffic, but Prince says Google now takes as much content while returning one-tenth as much. Blocking Google was unthinkable ten years ago, radical six months ago, and actively discussed today; search indexing should remain distinct from answer boxes, AI Overviews, and Gemini transformations.
Large publishers including Condé Nast, Dotdash Meredith, The New York Times, and Reddit can negotiate direct deals. Cloudflare’s prospective role is the long tail: pooled licensing, micropayments, or separate rates for training and search. Prince’s advice to creators is to regain control, use the data on who keeps trying to crawl their material, and propose a “fair exchange of value” rather than continue giving access away.

6. Power efficiency moves inference toward devices and the edge

Cloudflare partnered with NVIDIA in 2020 to put GPUs at its network edge; the launch produced “crickets” and not one sales inquiry. Four years later, essentially the same press release “took off like gangbusters.” Prince expects much inference to run on-device, with models too large or resource-intensive moving to the nearest network edge.
Some local execution is mandatory, not optional. A driverless car seeing a red ball followed by a child cannot make braking depend on network conditions. The binding constraint across phones, cars, and Cloudflare facilities is therefore power efficiency, because edge locations cannot assume the envelope of a new 100-megawatt data center.
Prince recalls telling Intel in 2011 that Cloudflare cared only about “cores per watt,” while Intel urged water cooling. He hears an echo when more modern GPU capacity can require “your own mothballed nuclear power facility.” Apple shows efficient accelerators are possible, and Prince argues there is “no physics reason” for energy consumption to remain this high.
DeepSeek’s lesson, he says, was real scientific progress in training and inference efficiency, obscured in the US by arguments about China. Compression and pruning can remove extremely improbable branches without destroying performance. He expects a current-generation ChatGPT equivalent on a phone before long and GPU systems to speed-run the last 30 years of CPU efficiency gains, including security lessons such as Spectre, within five to ten years.
Prince is generally pro-open models: more open models run on Cloudflare, and the company works closely with Meta on Llama. He calls claims that open-source models will end the world histrionic; for risks such as synthetic pathogens, he would regulate the machine that prints them rather than the model that proposes them.

7. Agents make the network an identity and settlement layer

Prince says 80% of major AI companies are Cloudflare customers, constantly pushing the company to add capabilities. Cloudflare charges for work performed rather than reserved hardware, so a 100-fold inference-efficiency breakthrough would be “great news for us” and “terrible news for the hyperscalers.” Prince is not building a frontier model; he is looking for the “VMware of AI” that can slice expensive GPUs instead of forcing customers to reserve whole machines or commit for a year.
MCP or its successor will connect agents to services, and much of that traffic must cross Cloudflare. Prince expects multiple agent-infrastructure providers connected by standards, with security and payments layered into the network: cryptocurrency still “needs a network,” and AI does too.
Content micropayments could revive blockchain ideas Prince previously doubted, but Cloudflare processes about 15 trillion requests daily—beyond Bitcoin and, he says, even Solana. The eventual system must distinguish humans, human-linked agents, and self-directed agents while supporting different permissions for each.
Browsers create the same governance problem: one that immediately feeds retrieved material into an LLM may deserve narrower access. Prince imagines cryptographically signed statements of identity and permitted use, with Elad Gil pointing to zero-knowledge proofs for demonstrating credentials without revealing them. “All the building blocks” existed earlier, but only in recent months has their combined shape become visible.

Li Yifan’s 11 Years Building Hesai—and Where Opportunity Comes From Vol.70 Momo Started a Company—and Chose What May Be the Most Competitive AI-Companion Market Right Now