저자: DanIves · 외국인
요약 보기
AI 혁명으로 미국이 기술 경쟁에서 중국을 앞서며 향후 수년간 3~4조 달러가 AI에 투입된다고 주장. Oracle은 현재 페널티박스지만 1년 뒤 매수 기회로 평가. 엔비디아 실적은 훌륭했으나 하이퍼스케일러 경쟁 심화 우려 존재. 엔비디아의 Grok 200억 달러 비독점 라이선스 딜은 추론(inference) 시장 대응 목적으로 해석되며, 규제 리스크도 언급. Megaspeed 관련 중국 칩 우회 의혹 조사는 진행 중이며 H200 중국 판매 재개 가능성이 논의됨.
다듬은 글
I disagree that Oracle is a lo loser just given my view when you look at the AI revolution. I mean for the first time in 30 years US is ahead of China when it comes to tech. I think what you're seeing here play out is like yeah there's some white knuckles but the reality is you have 3 to four trillion that's going to be spent next few years.
Oracle is going to play a huge role in that. I think right now in the penalty box, I think we sit here a year from now and it was a huge buying opportunity just like the view if you go back a year ago, New York City cab driver is bearish on Google. Um, that clearly spooked the market. Um, I know you don't think we're in an AI bubble, but how do you reconcile that with the reality versus expectations dynamics of what forms a stock price where Nvidia's price action suggests that expectations may now be ahead of reality?
Yeah, I think well a couple points there and you'll notice that we did buy on that setback and you're right, it's not a huge position. It's about 1 and 8.5%. Um, and one of the reasons for that, we're watching price action. That's that's true. But I also think there was a crypto AI relationship here.
The kinds of investors interested in crypto are also interested in AI. uh and you know there there was a selloff and you're going to have some investors speculators certainly a follow in the momentum uh and that's all they're doing you know and they'll ask questions later uh in terms of the Nvidia report it was spectacular was a reaceleration in topline growth was an increase in margins uh but we do see other signs of competition emerging out there uh and you know the hyperscalers their cloud strategies probably do not have 75% gross margins.
Uh and they're buying from Nvidia and now Nvidia is telegraphing our margins are higher than our customers. So that that might be a unsettling to them and in pushing them pursue a strategy involving AMD or involving Google's TPU um or involving Amazon's tranium or inferentia shares of Nvidia sinking among the broader tech pullback here sinking down 1.3%.
The company announcing last week a $20 billion non-exclusive licensing agreement with chipmaking startup Grock for Gro's inference technology. Joining me now to discuss what a deal like this means for future the AI race and the whole ecosystem is CNBC contributor Plexo Capitals Low Tony. It's great to have you low.
What what do you think of the other structure of this deal that Nvidia is doing? It's not an allout acquisition but it's not that dissimilar. Yeah, it's interesting because some of these headlines frame the deal as a strategic acquisition, but really to me this is more clearly an architectural admission.
You know, most people think that Nvidia bought Grock, but they didn't. They really captured a capability and that's the architecture, specifically the LPU without the regulatory friction of buying a whole firm. And they had to do that because of inference economics breaking the old math. What does that mean?
What are they actually getting? Yeah. So when you think about Nvidia, the GPU dominates training. But training is very episodic with big spikes. So inference is when people are actually using the models to execute chats with GPT, execute aenic workflows. And that is a different math because inference already costs 15x more than training over a model's lifetime.
So that's really the math driving this deal. It's like the training attracted attention, but the inference determines outcomes. What do you make of what they're what they're paying here? I mean, this was grow was last valued at $6.9 billion in a September funding round. Nvidia has, you know, according to our David Faber $20 billion price tag on this deal.
You know, some could argue it's a drop in the bucket. Some could argue it's overpriced. I think when we look at Nvidia and where they are right now with the GPU, they were built for flexibility, but they ran into the problem and they knew this problem was coming that inference is now starting to dominate and so it would have been tough to retrofit the GPU.
So this deal does make sense. It is expensive but they have the money. Is it is it a defensive move to the holdoff competition here in Grock? Yeah, you know, I think look at it as you know it's it is slightly defensive. that they're not necessarily late, but they are reacting. And I think we should go back to a comment that your prior guest made about Google and the fact that it's vertically integrated.
The core of that vertical integration is their TPU chip. Now, the founder of Grock, Jonathan Ross, actually built the original TPU, or what became the TPU during his 20% time at Google because he saw this bottleneck coming a decade ago. So Nvidia is doing the same thing, right? They're they're integrating at the infrastructure layer but under much tighter regulatory constraints.
They can't buy everyone. So they're just licensing the capability, right? So so all in all of your I and I know you invest in early stage startups. So not necessarily a company like an Nvidia, but does it make you like Nvidia more or less? The way I think about it is when we look at Nvidia and the stock and the runeup, I mean that reflects a training boom.
We've seen all of these models execute on their training strategies and now they're starting to really focus on the ability for people to execute these models. So what this deal does is it protects Nvidia stock for the future inference boom. Yeah. Yeah. I mean, it was interesting earlier. We were talking to Jonathan Kirner, the former DOJ antitrust chief, and he was saying it actually raises some red flags even though it's a, you know, non-exclusive licensing deal or whatever they're calling it.
They're trying to take out competition and take out a key hire that a company that could be bigger if it wasn't doing this deal. Yeah. And those are the things that are going to happen. You know, we've seen Alphabet/G Google execute one of these deals as well with Windsurf. I suspect that people will continue to push the envelope because of the regulatory nature of when these deals get scrutinized for being anti-competitive and monopolistic.
We'll see. I think Nvidia will be able to get this deal through. But I think when we start to see more firms, especially the Mag 7, try to execute these. It's going to be a little bit tougher for them to be able to continue with this type of approach. The deep dive that reporters have done to understand whether we really understand how many GPUs are going to mega speed and where they're ending up here.
Talk us through it. Yeah, I mean, there's a lot of evidence. It's important to point out as you already did that Nvidia says there's nothing going on here. Mega Speed itself says there's nothing going on here. But you know our colleagues reporting here Caroline is that there is an investigation going on.
Um and you know Southeast Asian governments and also Washington are looking at this and in still trying to find out whether there has been anything significant going on. I'm the go-ahead. Sorry. Sorry. Well, from what we know, it's a Singapore-based company, Mega Speed, operating fully in compliance, as they say, with applicable laws.
But what ultimately has had to spring up ever since the Biden administration back in 20122 was a restriction on sophisticated chips coming from Nvidia to China. So suddenly, you saw other Asian countries become real areas of focus for importing chips so that Chinese companies could actually do the workload, could do the compute outside of their own country.
Correct. Yeah. No, that's absolutely right. There's nothing illegal with setting up a data center and serving Chinese customers, providing those customers don't have links with a banned entity uh in the US, whether that's a military or some company which the US government has decided serves China's military.
However, there is a suspicion about these links about who's in control of what they're there's a you know, and this is a problem that Nvidia has to face and Nvidia, you know, say look that there's nothing to see here. there's nothing to worry about. But as you'll see from our story, there's a lot of a links between individuals and in China there, you know, there's a lot of lack of clarity in the relationships which I think everybody is trying to work through to make sure that there is nothing unour going on because Bloomberg didn't find evidence of any mega speed Nvidia chips actually being diverted to China.
But there's all these inconsistencies as you say in mega speed, Southeast Asia demand and chip inventory and where overall they all end up Ian. What's so interesting is this is a moment where maybe actually Nvidia will get more access to China. H200s have in theory or at least according to a true social post been allowed back into China.
The question is whether China wants them, but also whether Mega Speed will actually have Chinese demand going forward if indeed we start to see access once again to mainland. Yeah, I mean there's again we're in a kind of transition period where we'll try to find out exactly how this will all play out.
Nvidia wants to do business directly in China. That's absolutely true at the moment. It cannot do that because of these restrictions. Um what we're trying to find out is whether the Chinese want to do business directly with Nvidia and if if that's the case what level of demand that there will be there.
Nvidia has really kind of had a good year in Washington. It's gone from really restrictive rules to a freeing up of some of them. But again, that has to translate into shipments into that Chinese market. And we haven't seen that yet. And we don't know th
Comments 0