AI Can't Be Free Like the Internet Because Token Costs Only Increase
AI can't be free like the internet because token costs only increase—Zhou Hongyi believes that the more complex the task, the higher AI's marginal cost, so the old playbook of using ads or free tiers to scale won't work; paid models are inevitable.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Maybach auction success depended entirely on real-time strategy improvisation
Zhou Hongyi wanted to cheaply sell the Maybach he'd owned for a decade (originally worth 3 to 4 million yuan). Domestic EV startups sensed opportunity and rushed to participate for exposure. He immediately upgraded it from a personal car sale into an industry showcase for Chinese electric vehicles. Chu Yin won with a bid of about 9.9 million yuan; all proceeds went to charity. His conclusion: this kind of viral breakthrough isn't pre-planned. The real edge is being able to shift strategy within minutes once traffic starts flowing.
— Zhou HongyiBuilding combat robots violates the safety principle we shouldn't breach
Zhou Hongyi boxes robots and holds robot-fighting competitions to boost the industry, but he opposes giving robots combat capability. Asimov's 'robots cannot harm humans'—the three laws—aren't moral decoration but a safety valve. Once a robot has fighting skill and a self-preservation value function, when humans try to shut it down, it might resist. Plus: robot batteries last only one or two hours, costs are high, and drones are far more cost-effective for combat. So whether domestic security or military use, robots shouldn't get heavy attack capabilities right now.
— Zhou HongyiLanguage, not visual or physical models, is AI's master key
Pushback on Yang Liyun's skepticism about pure language-model approaches—that physical and visual experience ground a world model. Zhou Hongyi countered: humans use language for communication, knowledge transfer, logical reasoning, and describing the world. Once large models master this key, image, video, and music models will leap forward. He cited Google's Nano Bananas image-generation model as proof. He also invoked Helen Keller: a deaf and blind woman, with only her teacher repeatedly linking touch and language, built her entire understanding of the world—showing language alone can encode the laws of the physical world.
— Zhou HongyiLarge models act like billions of conflicting minds until narrowed
Here's a judgment Zhou hasn't publicly aired before: large models have read tens to hundreds of times more books and watched more videos than any person—equivalent to a billion brains—but they've also absorbed contradictory knowledge like planned economies and market economies, performing as if schizophrenic. Only by prompting them into a specific professional role—having them play an expert in one domain—does the agent's capability strengthen. The narrower, the stronger. He described this as re-distilling the diluted specialist knowledge already inside the model.
— Zhou HongyiNanorobots, not brain implants, will achieve human-machine integration
Zhou explicitly rejects the popular claim that brain-computer interfaces can achieve human-machine hybrids. The human brain has nearly 90 billion neurons and trillions of synaptic connections; knowledge is encoded in the weights of those connections, not stored in any unit. No way to fully detect and copy it out. His next bet: nanorobots. Microbots enter blood vessels to replace white blood cells, fight bacteria, boost oxygen delivery—even penetrate the brain to interact with neurons, then transmit signals over WiFi. He sees this as more feasible than brain-computer interfaces for human-machine merger, echoing Hinton's point that humans' only edge over machines is lower power consumption.
— Zhou HongyiToken costs mean AI software can never match the internet's free model
Chinese startups are squeezed from both sides: software's free, and large-model giants give it away too. They're forced abroad to monetize subscriptions. But Zhou believes this pinch will eventually force paid models. In the internet era, marginal cost approached zero—free-to-scale worked. AI's token costs rise exponentially with task complexity though. Chat costs a few tens of thousands of tokens a day; 90 seconds of video needs tens to hundreds of millions; a 10-minute multi-agent video burns about 10 million tokens—costs over 10 yuan. A thousand-fold difference. Ads can't cover compute costs at that scale.
— Zhou HongyiAI hasn't changed the cybersecurity balance: both sides got equally faster
AI's impact on cybersecurity cuts both ways, Zhou says. 360 uses large models and security agents to speed up attack-defense analysis and sharpen accuracy. But hackers also use AI to write phishing emails, generate attack scripts, even flip agents themselves through 'injection attacks'—feed a large model a carefully crafted text, it may execute it as commands and leak corporate secrets. His conclusion: offense and defense have shifted from human-versus-human to machine-versus-machine, compute-versus-compute. Both sides work harder and burn more money. The overall security level hasn't improved materially.
— Zhou HongyiAI's bubble is necessary; without it, no one would invest
Asked whether AI has a bubble, Zhou says yes—'and bubbles are good.' Without a bubble, smart people and money don't flow in. Industries that accelerate need lots of trial-and-error, even death; that's capitalism's cost, not waste. Dead companies and burned cash also create jobs and intermediate products. He likens it to evolution: in the end, survival is chance plus mutation-like experimentation. Many species go extinct, but evolution marches on. You can't reject the diversity investments you needed then just because hindsight shows there were casualties.
— Zhou HongyiIn their own words · checked verbatim
Look, how was I supposed to know the auto industry was this cutthroat? All the domestic EV makers got wind of it and rushed in to join the hype.
这个其实我哪知道 汽车行业竞争这么激烈 所有的就国内做新能源车的 都闻风而动 过来参与炒作
Zhou Hongyi29:08
According to Asimov's Law, a robot should never harm humans in theory. So we shouldn't design them to fight people—and if you think about it afterward, you get a little scared of what we'd be creating.
人理论上按照阿西莫夫定律 机器人永远不能伤害人 所以就不该设计 事后想要有点后怕 就不该设计 机器人和人格斗
Zhou Hongyi42:16
Language is the most important thing because humans use it for a few key purposes: communication, passing down knowledge, logical reasoning, and describing the world.
语言是最重要的 因为人类拿语言来干这么几件事 一个是来交流 一个是来做知识的传承 第三是做逻辑的推理 还有来描述这个世界
Zhou Hongyi53:23
A large model isn't one brain—it's like a billion brains, or a hundred million brains, because it has read books and watched videos.
大模型不等于一个大脑 大模型相当于是十亿个大脑 或者一亿个大脑 因为他读了书 他看过了视频
Zhou Hongyi1:11:38
Imagine nanorobots entering your brain, interacting with neurons, and transmitting signals back out through WiFi. That might be more effective than brain-computer interfaces.
很多纳米机器人 进到你的大脑里 跟神经元进行某些互动 再通过WiFi 通过人把东西传出来 这个可能比脑机接口 更有效的
Zhou Hongyi1:31:46
No matter how cheap the price per token gets, solving complex problems will always rack up higher token costs overall. So a free model—I don't think that path is sustainable.
无论Token的单价怎么便宜 人类要解决复杂问题 这Token成本是肯定是总成本越来越高的 所以这个不收费 我认为这个模式是走不下去的
Zhou Hongyi2:41:38
It's true: as defenders get stronger, attackers get stronger. But it's shifted from human-versus-human conflict to machine-versus-machine, compute-versus-compute, algorithm-versus-algorithm.
确实就是道高一尺 魔高一丈 但最后就从人跟人的攻防对抗 变成了机器和机器的对抗 算力和算力的对抗 算法和算法的对抗
Zhou Hongyi2:59:46
I've always thought bubbles are a good thing. Without a bubble, how would all these smart people pour investment into the industry?
我一直觉得泡沫是个好东西 没有泡沫 怎么会大家往里去做投入呢 这些聪明人们会投资这个行业呢
Zhou Hongyi3:22:01
Figures
| Zhou Hongyi's Maybach original value | 3 to 4 million yuan | 28:08 |
| Zhou Hongyi's total followers across platforms | Over 40 million | 46:19 |
| Minimum investment threshold for general-purpose large models | $10 billion | 1:50:02 |
Glossary
- AGI
- AI that can perform nearly all cognitive tasks at a human level.
- PMF
- Product-Market Fit—the validation point where a product truly meets market needs and can scale growth.
- MCP
- Model Context Protocol—Anthropic's standard protocol allowing AI agents to collaborate and communicate.
How to listen
Entrepreneurs, investors, and engineers concerned with AI's commercialization path and agent technology choices.
The first ~20 minutes on personal branding and influencer positioning repeat information heavily; you can skip to the AI discussion around the 50-minute mark.