行业动态
x.com ·
@andrewchen
Andrew Chen：AI 原生消费应用的推理成本与 ARPU 之争可能比预期更早到来
Andrew Chen 重申其观点：AI 原生消费应用要真正普及，需满足 ARPU 大于单用户平均推理成本。他估计目前月 ARPU 约 2-5 美元，而 AI 重度应用的 token 成本约 20-50 美元，差距超过 10 倍；叠加消费市场全球化、模型能力提升不断推高用户预期，以及小模型或新移动硬件带来的本地推理可能性，他认为这也是大量产品转向生产力与高付费 prosumer 场景的原因。
AI推荐理由 a16z 合伙人 Andrew Chen 延续其此前关于 AI 原生消费应用经济模型的讨论，给出具体量级判断：当前月 ARPU 约 2-5 美元，而 AI 重度应用的 token 成本约 20-50 美元，缺口约 10 倍以上；并指出消费市场全球化、模型能力提升会持续推高用户需求，可能需依赖小模型或新移动硬件实现本地免费推理。属于有具体数字的行业判断，但为个人观点、无数据来源与验证，且为对旧帖的延伸，故质量分中等。
62
浏览 1,250
点赞 2
评论 0
转发 0
正文
I prev wrote about the battle of inference cost vs ARPU here - it might end up sooner than expected!
andrew chen@andrewchen
for AI-native consumer apps to be truly ubiquitous we need:
ARPU > Average Inference Cost Per User.
How far away are we from that? Ideally AI native apps can hit APIs on every screen yet also pay for just by throwing some ads on it
But I think we’re a while away:
- unfort we’re probably >10x off right now. Monthly ARPU is $2-5? Token costs for an AI heavy app might be $20-50 of cost
- much of consumer is global. Even if we hit the US/EU it’ll be a while before we can serve the broad base
- more importantly every time AI improves, consumers demand more. No one wants to talk to a last gen AI character. If video gets good they’ll want videos everywhere
- same issue for LLMs. We started with short replies back. Now we want agents who can do entire tasks for us
- it might be we need major innovations in small models or new mobile hardware so that we have free local inference. But that will still not be as good as SOTA cloud LLMs obv, but might solve some use cases
No wonder so many products focus on productivity, and on prosumers who can pay $100s or $1000s on work related tasks. This is where you can have huge ARPUs and benefit from being SOTA. And it seems as though there’s no limit for tokens… so why do the low end?
[🔗 View Quoted Tweet](https://x.com/andrewchen/status/2037204260300218503)