Is Investing in AI Infrastructure Actually Safe? | Chamath Palihapitiya - World Governments Summit

요약 보기

AI 인프라 투자의 지속가능성을 논의. 오픈소스 vs 클로즈드 모델의 수익화 문제, 중국 Alibaba Qwen의 클라우드 기반 수익화 사례 언급. 에너지·칩 인프라 투자의 안전성은 TBD로 불확실하다고 봄. 수요 측면에서는 스케일링 법칙, 추론 연산 증가, 멀티모달 수요로 AI 버블이 아니라는 입장이나, 일부 섹터(메모리)의 과잉투자와 ROI 설명 곤란성을 지적.

다듬은 글

※ 자동 자막 원문(다듬기 전)

01. People talk about the AI button. Is it sustainable? Depending on the solution that you use, you're going to need a lot of natural resources, specialty chemicals, you're going to need battery systems. Is it safe bet investing in energy infrastructure? Is it safe bet investing in chips? All of the things that we've used to underwrite productivity and GDP will get re-questioned.

02. Can China sustain making all these models open source, especially the frontier ones? You're spending billions of dollars to build this model and you if you open source them, you're not really monetizing it in the way that you can potentially monetize it. So, that's why most of the companies are focusing on closed source to try to generate as much profit from the model and then they open source the older versions.

03. We've seen that with Grok, for example. We've seen that with even Lama and other let's say models in the US where the frontier ones are you know, closed source until they're able to generate some revenue or at least some advantage, whether it's getting market share or otherwise. Do you think that is the main reason or is there a more of an ideological shift?

04. If you look at China, look at the example of Alibaba. We have our Qwen model is open source, but we also run our cloud business and today we're monetizing because people are training and also developing their running inference on their models on our infrastructure and thereby generating the economics.

05. But I I I think it remains to be seen how in the future how can you get especially consumers pay for a monthly subscription and generate economics there. That remains to be seen. That's the biggest challenge in in the Chinese market. As Joe said, all of these other things um you're going to need power.

06. You're going to need infrastructure. Depending on the solution that you use, you're going to need a lot of natural resources, specialty chemicals, you're going to need battery systems. And then you're going to need all the know-how. And then on top of all that, you're going to need the capital and the financing.

07. And so I think the reality is that the business model becomes more at the governmental level almost because if you take an open-source alternative, but then the infrastructure of the government can come in and help you with all of these other things, you're going to take it. People talk about the AI bubble.

08. And the bubble in the sense of is it sustainable? Can we keep, you know, seeing the the number that we're seeing? We've recently seen the issue with memory and memory availability. And I'm not saying it is a bubble, but let's assume that there was a set of bubbles along the way of the progress of artificial intelligence.

09. Do you think that energy infrastructure is that? So is the safe bet investing in energy infrastructure? Is the safe bet investing in chips? I think it's very much a TBD. And the reason is that the capability of these models can actually disrupt the underlying asset that you may think is safe. So I think the honest answer is this is why it's such a profoundly interesting moment.

10. All of the things that we've used to underwrite productivity and GDP will get re-questioned because this super intelligence allows us to see things that you know, the frailty of the human condition we can't quite see at today. And I think we should just be open-minded and prepared to deal with that as a as a community cuz I think it's going to be complicated.

11. So there's no bubble in that context. But here's the thing. I think a few things are driving it. Number one, people still believe in the scaling law. And now also uh the idea that uh inference time scaling is important because more computer resources will be allocated to inference. And inference is where, you know, everybody that uses AI will touch influence as opposed to the few model companies that only do training, right?

12. So, a lot of CAPEX that's going toward inference. And then finally, you have people making multimodal models, deal in generative rich media, videos, images. Uh that will burn a lot of GPU resources. So, people seeing this scaling law and uh multimodal world, uh and we haven't even talked about world models yet.

13. And that's I think the the validation that there will need to be more compute uh that need to be invested in. I don't think these trends are going to reverse or uh veer off. In the future, people will still continue to You know, for example, are people going to consume less videos? I don't think so.

14. And uh so, in that sense, there is no bubble. But having said that, people are investing anticipating future demand as opposed to Yeah. Her I I think there's could be you know, in certain sectors you see you know, like memory, you see memory you see shortages today. The massive amounts of investment, it's kind of hard to think about ROI.

15. I hate to be the CFO of one of these companies because they have to they're faced with, you know, having to explain ROI to investors. They can't They can't do that. Yeah.