저자: OlivierBlanchard · 외국인
요약 보기
DeepSeek 충격은 AI 효율화(모델·칩)의 연장선이며 AI 워크로드 확대 테마는 유지된다고 주장. 엔비디아는 여전히 유리한 포지션이며 Stargate 투자 기대치는 다소 과장됐으나 2025년 인프라 수요는 충분하다는 입장. AMD·Qualcomm·Broadcom·MediaTek을 주목하고, 인텔은 리더십 변화와 뒤처진 부분 때문에 관망(3개월 지켜보자)이라고 평가.
다듬은 글
With us is Olivier Blanchard, Research Director at Futurum Group. Thank you for being with us. I mean, we can discuss specific companies whose stocks have fallen significantly. What do you think about the last 24 hours? Hi Nicole, thanks for inviting me again, and Happy New Year. First time in January, nice to see you.
What happened yesterday was certainly a huge shock to the markets, but I don't think it's just DeepSeek. This is something we've been talking about for a few months now, and I think the big story here is the expansion of AI workloads, both training and inference, from cloud services to edge computing.
And what we see is, in fact, a combination of two things. First, models are becoming much more efficient, which means they are faster to train, and it requires less computing power, less energy, less water— all resources that we typically waste on large AI workloads. So , thanks to these efficiency improvements, we see less and less need for these costs.
And secondly, chips are also becoming much more efficient, which is the basis of Nvidia Blackwell's idea. These are much better chips. They are still a little more expensive up front, but in the long run, in terms of operating costs, you get a much greater return on investment because these chips perform better.
So, better software, better training, more efficient models, and also more efficient chips. And what DeepSeek did was essentially demonstrate the possibilities of what could be achieved with this. I'm not sure if we'll even be mentioning DeepSeek in 6 months, but we'll definitely be talking about this broader theme: improving the efficiency of models and chips as a cheaper alternative to how we currently plan AI training and infrastructure investments.
Interestingly, I read about the 40- year-old founder of DeepSeek in China, and it was noted that at one point the models were trained on less than 10,000 Nvidia A100 GPUs. And the question is, isn't that what Elon Musk and others were talking about when they were criticizing? Then I read today that they supposedly didn't use Nvidia chips, but you're talking about training and conclusions.
Is there any way to find out if Nvidia helped create Deep Seek, because that could help Nvidia itself, right? I mean, it could mean more demand for Nvidia, and I know we're still getting to the point where maybe we're looking at it from the wrong angle , but it's just an interesting fact. That is an important fact.
Do you know anything about this? Uh, well, no, not yet. And I, I don't want to speculate on the topic of , uh, what chips were used. This, this is actually not very relevant. Uh, I mean, it's, it's going to be interesting whether it's Nvidia chips or not. Uh, maybe it's a mix of Nvidia chips and other chips.
Uh, they, they're not the only players in the market. Uh, but I think what's interesting is that this is a confirmation of the increasing efficiency of the models themselves and the increased efficiency of, uh, the chips themselves. And Nvidia here , uh, ironically, I, I, you know, you're showing, uh, uh, some of the, uh, losses from yesterday.
Nvidia is very well- positioned to take advantage of these more efficient models. And I think if we go back to where we started, if we, if we rewind to yesterday's announcements, uh, and, and the sale, and go back to the previous week, which is to the announcements last week about the Stargate infrastructure investment project, and go back to what Jensen said at CES about what Nvidia was working on and what the product and revenue roadmap was.
I think it's still, it's still within the goals. Nothing has changed. Just last week we raised our expectations for Stargate in terms of investment, and then Deep Seek basically knocked the ground out from under us, saying, "No, you don't need to invest that much ." But what was said two weeks ago is still essentially true.
We are still seeing this expansion of AI learning and inference. You have this extension in the cloud, at the edge, and on end devices. So, we're entering a more hybrid model of AI training and workloads that still benefits Nvidia. It's just, you know, I think last week's numbers were a little inflated.
So, I would, I really want to discuss so many points, and I want people after this interview to say, "Wait, maybe I want to buy this stock because I really like what they're working on, or I like the idea of Stargate in general, or something." So let's try to quickly go over some things. I mean, first of all, all this hype around Stargate.
The money that can be raised for this, where 100 billion turns into 500 billion. I mean, theoretically this dollar could last longer, right? Because if this DeepSeek is more efficient or something, we can use it to make us more efficient at home too. Maybe this is good news. I think this is good news.
Yes, that's nice. I think Stargate was like a kind of “unlimited check” for the AI infrastructure investments we need. First, to expand that space, and second, to be more competitive with China and our other, you know, international or global rivals. But I think this moment with efficiency works really well.
So on the costs themselves: if we have the funds, we can work on other things that also help the broader infrastructure. So whether it's investing in nuclear power, or modernizing our water systems and resource diversion—essentially, upgrading many of our data centers— we're confident in all of this .
But it also means that maybe we won't have to spend as much, and we can put some of that money into R&D. We can spend them on other things. These don't have to be huge data centers, which may not be as big as we think. And I think we're looking at infrastructure spending that is adequate for 2025, but not necessarily for where we'll be in 2030, if you know what I mean.
So yes, as a research director, you do a lot more research than any of us on these particular companies. Qualcomm is a name you've been talking about a lot over the last 12 months. I wonder if they are worth buying? Is Qualcomm worth buying? Nvidia, AMD, Intel—these are other companies you keep a close eye on.
Could you tell me, based on the cost estimate or the creativity of their future developments, which of them are worth buying? So, we've already talked about Nvidia, they have a very good development plan and a strong position, so I'm not worried about them. I think AMD is interesting. I saw, maybe it's an unpopular opinion.
I've seen some financial analysts downgrade their ratings a bit. I believe that they have very good prospects in data centers, as well as in the field of end devices, i.e. client PCs. And I like the direction in which they are developing their technology. And I think on the device side, which is everything that's peripheral and endpoint— mobile, PC, XR, automotive—I really like Qualcomm, always, because they're really good at developing low- power, high- performance AI ARM chips in all of those segments, but MediaTek—that's another company that I'm looking at.
And I believe that Broadcom is somewhat of a "dark horse" in this regard. We don't really talk about Broadcom enough, but I think they have some interesting things coming up next year, especially in AI-powered IoT, mostly in smart cameras and drones. I think Broadcom will have an advantage there as well.
So, these are the five tech companies that I'm paying the most attention to right now and that I think are moving in the right direction . So, it seems like Broadcom, Qualcomm, Nvidia, and AMD are options to buy. About Intel, you didn't necessarily say it was a buy. Although they have great opportunities, you seem to be a little aloof about this company so far.
Well, you know, Intel is going through some changes right now, right? There is a change of leadership there. I have complete confidence in their current leadership team. I think they are in very good shape, but they still need to make a few decisions. They are a bit behind in several aspects. There were a few setbacks that they are still trying to make up for.
So for me, Intel is a wait-and- see situation. Never bet against Intel. I think this is a very strong company. They are vital to this ecosystem. They do truly wonderful things. I think they caught up with the competition in the AI-PC race, and they were moving much faster than many expected. But we'll see how it goes.
Let's wait and see what happens with Intel in the next three months. Quick question: do you think the US can become a leader in AI? Of course. This is an interesting, very broad question. I think so . I mean, the US is the leader in AI right now. I believe that China is a very close pursuer. There are areas where China may be a little ahead of the US, and vice versa, but in my opinion, we are now seeing two superpowers in the field of AI.
This is largely a confrontation between the US and China in the field of developing AI solutions, as well as the ability to scale. But yes, I think the US can maintain its lead now. But we'll see how China performs next year. However, it is clear that it is too early to write China off. They're definitely working on some really interesting stuff and this week they lived up to expectations, so well done.
Competition is good. That will keep everyone in good spirits, I think. Those who lost a lot of money yesterday are definitely not happy. Olivier Blanchard, nice to see you. Have a great week for your visit. Futurum Group, thank you very much. It was nice to meet you. Thank you, and Happy New Year to you too .
Comments 0