Agent traffic already beats humans 60/40, and nobody is paying

1. Post-training as the way in — ▶ 5:40

   Nous formed in 2023 out of a Discord, with four founders and a pseudonymous co-founder, and still runs on the same Discord. Its first product, a graph-based agent, failed because "the models were so shitty that the product was shitty." Post-training turned out to be the resource-light part of the equation, so the team pioneered synthetic data when many doubted it was additive. It matched closed-lab performance "with very little." The thesis is thousands of specialised labs, not one general superintelligence, in the way you wouldn't hire the friend who'd ace Jeopardy.

2. Neutral like a grid — ▶ 7:07

   Prince told Anthropic he could not join its board because he cannot pick a side in rivalries that are "really intense." He recalls the Modi conference in India, where Sam Altman and Dario Amodei refused to hold hands. OpenAI, Anthropic and Chinese labs all run on Cloudflare, which he likens to an electrical grid for them. Traffic swings toward whichever model is best at a given moment. OpenAI also came to Cloudflare to act as a clearinghouse, so its models are used for security by the right people.

3. Pausing research is untenable — ▶ 9:06

   Prince calls the idea that big labs should pause research over security fears "pretty untenable," since rivals in China keep building. The Nous co-founder says fear is memetic and hard to decide from, citing a pediatrician sister alarmed by the paperclip analogy. The sharper worry is government choosing who accesses the most powerful technology. Regulation is also a very effective way to stop distillation, so incentives are tangled. Prince adds that distillation generates real revenue for some model companies, and Cloudflare could stop a lot of it if asked.

4. Guardrails, not prompts — ▶ 13:13

   Prince runs a public company, so giving an agent his email and calendar risks leaking material non-public information into a model others can query. Nous co-designed OpenShell with Nvidia, a sandbox where policies, not prompts, set the guardrails. An agent holding sensitive data might have no internet access at all, and a communication channel should only talk to its owner. The tradeoff is a spectrum: let the agent do everything and it is very powerful, let it do nothing and it is very safe but inept.

5. Labs may sell a commodity — ▶ 17:18

   Prince ranks what is scarce: talent is easing as PhD programs spread, compute follows since "once you make one, you can make many," and data is what remains. Google sees 2x as much of the web as OpenAI, 4x Microsoft and 6x Anthropic. If OpenAI reaches AGI first, Anthropic follows the next day and Google the day after, and consumers switch easily. "I would not trade places with Sam or Dario right now for anything." They may spend trillions on a commodity. The exits are specialised, locked-up models or vertical integration.

6. The web is growing again — ▶ 20:44

   Prince rejects the dead-internet theory. The amount of the web shrank or plateaued from 2012 to mid-2025, then began accelerating at the fastest rate since the early 2000s. Some of that is generated content, but much comes from widening who can create: Lovable, Base44 and Replit let people build without computer science skills. The Nous co-founder agrees, saying design is diversifying rather than converging. Prince frames it as technology at its best, expanding a creative task from a small guild to a far larger group.

7. Agent traffic outruns forecasts — ▶ 22:38

   Prince projected in November 2025 that automated traffic would pass human traffic at the end of 2027. By March 2026 he had moved it to early 2027, and in May his team reported it had already happened, now at 60/40. Extrapolated, that is 1,000x more agent than human traffic in 5 years, and he'd take the under. Agents don't click ads, so advertising can't pay. His lunch query scanned 1,000 menus and benefited one restaurant. COVID doubled traffic in two weeks and nearly broke Europe. One agent per knowledge worker needs 40x annual CPU production.

8. WordPress cost cliff — ▶ 26:05

   A large WordPress host spends about $2.30 a month serving the average site and charges about $12. If traffic rises 1,000x, so does cost, and nobody will pay $2,000 a month. Cloudflare rebuilt WordPress in a modern way, open sourced it and called it M-Dash. The Nous co-founder shows the other side: a Hermes codebase that had ballooned to a million lines was refactored by 1,200 sub-agents, 250 at once, over 11 hours, driven by one person. The same technology that creates the load makes rewriting for efficiency cheap.

9. GPUs sit mostly idle — ▶ 28:04

   Throwing hardware at the problem is politically hard, though Prince notes the US holds 92% of the world's data centers and some hostility traces back to rivals. He argues efficiency is the better lever: rented GPUs at hyperscalers typically run at 5-8% utilization. He wants the VMware of GPUs, using virtualization and containers to reach 50-80%, as happened with CPUs. His claim is that GPUs will speed-run 30-40 years of CPU history in 5. Cloudflare's own contribution is a lightweight agent sandbox that uses only the resources a task needs.

10. Inference leaves the device — ▶ 30:57

    Prince used to expect 50% of inference on device and 50% in the network or cloud. Long-running agents have tilted his view toward more off device, because the binding constraint is not silicon but "the most precious power in the world," the phone battery. Work should hand off to the nearest cell tower base station rather than Ashburn, Virginia, keeping the benefits of staying local. That needs clear rails so the user never knows where inference runs, chosen by battery, task length, resources needed and how private the task is.

11. Open models import bias — ▶ 33:25

    Prince says Chinese models exploded once the government relaxed rules, such as refusing to role-play a student at Tiananmen Square. The Nous co-founder used DeepSeek's reasoning traces for synthetic data, and a model post-trained from Meta's Llama inherited the same biases. Nous built "refusal bench" to train them out. Good American open-source options are scarce. His hypothesis is that being open invites scrutiny over data provenance, and "the juice isn't worth the squeeze." He cites Jensen Huang's open letter urging clear legal rules defining distillation.

The web is about to be asked to carry a thousand times its current load, and nothing yet pays for the extra.

Matthew Prince's example is lunch. He asked an agent where to eat in Manhattan. It scanned 1,000 menus from 1,000 restaurants and returned one answer. One restaurant gained a customer. The other 999 paid for the bandwidth and got nothing. Advertising cannot close that gap, because agents do not click ads. Prince calls it a tragedy of the commons and thinks it is a bigger problem than anyone is admitting.

The physical limit is real. In April 2020 internet traffic doubled over two weeks and the European internet almost fell over. Prince's reading is blunt: if it almost broke at 2x, "at 1000x it really breaks." Give every knowledge worker in the world a single agent, which he calls conservative, and the world needs 40 times its annual CPU production. That is before GPUs.

The cost structure underneath is just as exposed. One of the largest WordPress hosts told Prince it spends about $2.30 a month serving the average site and charges about $12. If traffic rises 1,000x, so does the cost. Nobody will pay $2,000 a month for a WordPress site. So Cloudflare rebuilt WordPress in a modern way, open sourced it and named it M-Dash. Agents may make that kind of rewrite feasible. The Nous co-founder described pointing Hermes at a codebase that had ballooned to a million lines. It spun up 1,200 sub-agents, 250 at a time, and returned a refactor after 11 hours, driven by one person. The technology that creates the load also makes rewriting for efficiency cheap.

Prince's answer is efficiency, not more hardware. Hardware is a political problem. Asking for 1,000x more data centers runs into a hostile mood, even though the US holds 92% of the world's data centers today. He thinks some of the hostility traces back to rivals who want that advantage gone. The waste is the better target. Across the hyperscalers, rented GPUs typically run at 5-8% utilization. He wants the VMware of GPUs: virtualization and containers that lift that to 50-80%, as happened with CPUs. His claim is that the industry will speed-run 30-40 years of CPU history in 5. Cloudflare's own version is a lightweight sandbox for agents that pulls in only what a task needs, not a whole operating system and toolchain.

Placement matters as much as utilization. Prince used to expect half of inference to run on device. Now he leans toward more running off device, because the binding constraint is not silicon but the phone battery. Work should hop to the nearest cell tower base station rather than Ashburn, Virginia. It needs rails that choose by battery, task length, resources and privacy, invisibly to the user.

The same abundance logic hits the labs, and it is worse for them. Prince walks through what is scarce. Talent is easing as universities stand up AI programs. Compute is a lithography problem: make one chip, make many, and more power is coming online. That leaves data. Google sees 2x as much of the web as OpenAI, 4x Microsoft and 6x Anthropic, and it had the talent and the silicon as well. If OpenAI reaches AGI first, Anthropic gets there the next day and Google the day after, and consumers switch easily. Hence: "I would not trade places with Sam or Dario right now for anything." They may be spending hundreds of billions, if not trillions, to make a commodity. Distillation sharpens the risk, because anyone who calls your model can extract what you paid to teach it.

Prince sees two exits. One is specialised systems that lock up a field, such as all the best cancer research in one model. The other is vertical integration, with model makers owning the harness and the user experience. He concedes that is bad for users who want to move around. Nous builds on the first. Its co-founder argues nobody wants one superintelligence, and that post-training lets a small team shape weights cheaply. Nous used synthetic data, when many doubted it, to match closed labs with very little.

Neutrality is the equivalent bet on the infrastructure side. Prince told Anthropic he could not join its board because he cannot pick a side. OpenAI, Anthropic and Chinese labs all use Cloudflare, which he likens to the electrical grid, and traffic shifts toward whichever model is best that week.

Open has its own costs. Nous used DeepSeek to generate reasoning traces for synthetic data. The resulting model, post-trained from Meta's Llama, picked up the same Chinese-shaped evasions on subjects like Tiananmen Square. Nous built a tool called refusal bench to train them out. There is no good American open-source alternative. The co-founder's hypothesis is that being open invites scrutiny over where the data came from, so for large American firms it is not worth the trouble. Both men want distillation defined, and both admit the line is blurry. A blog post Cloudflare co-authors could end up in a model through a chain of generated data.

Regulation is where incentives tangle. Prince calls pausing research untenable, since rivals in China keep going. The Nous co-founder worries about government choosing who gets the most powerful technology, and notes that regulation is also an effective way to stop distillation, which protects the labs' revenue. Prince adds that distillation generates real revenue for some model companies, and Cloudflare could stop a lot of it if asked. Fear spreads because it is memetic, and it is hard to make "good decisions out of a place of hysteria."

The obvious objection is to Prince's own number. He forecast in November 2025 that automated traffic would pass human traffic at the end of 2027. By March 2026 he had moved it to the start of 2027. In May his team told him it had already happened, at 60/40. He admits he has been wrong on every prediction so far, and he would take the under on 1,000x in 5 years. A straight line through a series that keeps surprising you is a thin base.

There is also a rosier reading, which he offers himself. The web shrank or plateaued from 2012 to mid-2025, then began growing at its fastest rate since the early 2000s. More people can build with tools like Lovable, Base44 and Replit. More agents and more creators are good news until the bill arrives.

The efficiency half of the problem has engineers, product names and numbers. The payment half has a wish list: those who want their content in models get in, and those who want to be paid can insist. Prince offers no mechanism for it. Until someone does, the 999 restaurants keep paying.