저자: ChamathPalihapitiya—대담진행CNBC측Sorkin제외 · 외국인
요약 보기
AI 토큰 가격이 '오일 배럴' 비유처럼 급락할 것이므로 고가 LLM(Anthropic·OpenAI) 중심 과잉투자와 CEO 비용 충격이 올 수 있다. 하드웨어·메모리·반도체 수요는 수년간 유지되어 나스닥이 S&P를 상회할 가능성이 크다. 반면 소프트웨어 및 고가 모델 기업은 수익화 불확실성이 크다. 각 LLM 기업의 신뢰·사업 리스크 및 프라이버시(ZDR) 문제도 지적했다.
다듬은 글
Good morning, Mr. Sorkin. You're even in a suit, wow. That's right. For our sake, so this is... Yes. My current job is selling enterprise software. This is what I do. This is what you do. So, before we get into anything, and I want to discuss everything (밸리의 상황, AI, 모든 것), I want to go back and talk about SPAC, because people have been asking.
I know, these are your favorite topics. How do you assess the situation with IBM? Does it have broader implications, or does it only concern IBM itself?
I think we know that there are several companies whose revenue is growing by a billion dollars a day. Sooner or later, the ecosystem that consumes this product must also start making money, and the end buyer of these tokens should also make a profit. So sooner or later the question arises: if two or three companies generate 200, 300, 400 billion dollars a year, and it doubles and triples every year, where will the rest of the money be made, and who will market it in a way that generates those profits? I think you're starting to see how the process is failing.
At a very basic level, if you think of AI as something like oil, let's use this analogy for simplicity. A barrel of oil costs, what, 86 dollars now? Unfortunately, the price has increased. WTI is worth 80, WTI, 81, okay. Let's say a barrel of intelligence, that's a million tokens. You can buy this barrel for 26 dollars from a very good model from Anthropic. You can buy it from OpenAI for 26. Anthropic's newest model will set you back 56. Elon sells you a barrel of intelligence for a dollar. Zuckerberg is going to sell it for a dollar and a half. Demis and Sundar are trying to sell it to you for a dollar. The Chinese will sell it to you for 50 cents.
So this rationalization must happen. We have the same resource with such an insane value. If you bet too early on those who sell extremely expensive "barrels of intelligence" and try to pass those costs on to others, you may face difficulties in the future, and this should eventually manifest itself.
And do you think that's reflected in the IBM story? Do you think it's just more of a general comment? I think Arvind did a great job. It's impossible to deny the trajectory of business development; he was able to change course and orient it towards cloud technologies, and he deserves a lot of respect for that. But the reality is that everyone in this ecosystem, including me, is trying to figure out how to price this so that the end buyer makes more money and grows faster. This is still an open question.
I think you're getting to the point Alex Karp made here about a week and a half ago about "token maximization." Some companies (this applies to Anthropic and OpenAI, and I've talked to their CEOs as well) say the math doesn't add up here. In my opinion, CEOs and CFOs have no idea how much token maximization is happening within their organizations. I suspect that one day there will be an unforeseen downturn: earnings per share will be down a few cents, and the CEO will ask the CFO, "What happened? Where did these additional operating costs come from?" And they will track them to a barrel of intelligence at 50 dollars instead of one dollar or 50 cents. This has not happened yet.
This is interesting. There are software promotions. It is also interesting that IBM is included in the Dow Jones index. The Dow fell 500 points. Let's count: multiply 50 points by whatever divisor it is. The Nasdaq is up more than 100, and it's still up. So if it's a question of technology in general, you're not seeing it in the Nasdaq yet, and the Dow's 500-point drop is driven by... I think there's a lot more to the Nasdaq that's not software: all the hardware, all the memory, all the chips. This whole complex will still make a lot of money. True. The question again is what is ultimately true. If you buy oil, you need to pour it into a new engine. This engine should make you go faster, right? It should consume less.
You seem to be implying that the high-end LLM complex, essentially Anthropic and OpenAI, given their pricing (which is still in the private markets), is going to collapse. Is that what you're trying to say? I think that's what you're trying to get me to say. No, I don't do that, not at all. Everything is clear. Hardware and memory infrastructure will be in demand for several more years, as there are huge limitations in this ecosystem. There is a colossal deficit, and these things will continue to show good results. The Nasdaq will likely outperform the S&P. Of course. If you look at the only other niche where there is real growth, as you said, it's the LLM complex. I don't know what the long-term assessment will be, but if a number of public companies in the coming quarters don't meet targets due to operating expense overruns, and they link it to uncontrolled costs they didn't even suspect within their organizations, it will "let some air out" of that particular category.
So we just talked about four big modelers, maybe five: OpenAI, Anthropic, you mentioned Gemini, and we talked about Grok and Meta. How do you see their future in the long term? I think we are seeing convergence. It used to be that when a new model came out, it was far superior to everything else. You thought, "Oh my God, we went from kerosene to jet fuel." So of course I'll switch to jet fuel. Now the analogy is this: we're seeing the nth version of the iPhone. We all keep updating to it, and we still use it for roughly the same tasks. "My elbow hurts, what should I do?" And you ask yourself, "Why am I spending so much money?" I think that's where we are right now. So if you're going to direct all this activity, why not spend 50 cents or a dollar? That's a fairly rational question.
The question is whether a Meta model, or Grok, is good enough for what you need to do. In most cases, the answer is a resounding yes. For some very narrow tasks, no. But in these narrow use cases, I would say spend 50 dollars per barrel of oil. Spend it, that makes sense. If you are a cybersecurity company like Palo Alto Networks and can generate billions of dollars in additional revenue by protecting the infrastructure of large corporations, use Mythos.
But this weekend there was panic, even inside Anthropic. We were talking about this yesterday morning. Anthropic, as you know, was planning to end free use of Mythos for some paid users. They were going to cancel this extension, so you would have to renew your subscription or pay extra for access. And then suddenly they said to themselves: we can't do that, because if we do, people will cancel their subscriptions and move to OpenAI. That's right. And the OpenAI model is great. Sol is wonderful. True. Mythos and Fable are great when you can use them. All these players face the same limitation: they are massively limited in capacity, massively limited by data center resources, and therefore limited in use. Meanwhile, the companies that have all this power (Google, SpaceX, and Meta) are finally getting back on their feet and starting to consistently release models one after another that are 80 to 95 percent as good. This is a very complex dynamic for leading laboratories.
Do you see any changes, by the way? For a long time, OpenAI and Sam have been under a barrage of criticism. There were a lot of questions about trust and things like that, and perhaps even more questions given the situation with Apple on Friday. But it seems that Anthropic has also taken some blows, and that may be the case for Alex Karp and others who think either the costs don't make mathematical sense, or who look at the history with Figma and how they're going into different markets anxiously. Look, let's be frank. There are three leaders of lab companies who have been in the spotlight for 20 years. Elon, you've seen every facet of his personality, and you can judge 20 years of his behavior. Sundar and Demis, 20 years. Zach, 20 years. These people have proven that they run a relatively simple and straightforward business. There are no tricks here. Maybe there was something on Facebook in the past; it looks like they fixed it. Google certainly never did that, and neither did Elon. You may not like them for other reasons. Suddenly you have these two guys who are, without a doubt, incredible entrepreneurs, but the amount of data that can be collected, and the instances where they have had to make difficult decisions, are still relatively limited, because they have only been under such close public scrutiny for a few years. If you look at the early years of Facebook, it's a mess. Maybe the early years of Google tell us about you and your 20s. I was portrayed as the person on Facebook who made everyone cry. But if you look back, this is a company with a capitalization of 2 trillion dollars, so these are crocodile tears.
Can I just say, IBM says they just couldn't close a few big deals. They always say things like this. But it's interesting: the clients they were trying to attract were distracted by industry cybersecurity concerns, and then they'll say the weather was bad. I have a real problem with this, and I've said it several times already. You can look at the graph of financial needs of these companies, and you can tell exactly when they oscillate between two different messages. Message one: we've created a supergod, and venture capitalists, like lemmings, run around saying, "Well, I have to be on this supergod's side." Message two: this is a real weapon, let's stop the world, and why don't you regulate everything in a row? They just go here and there. We've already played this game. The problem is that with a trillion-dollar market cap, this game has a ripple effect that affects everyone in the industry.
That's why I think what Alex Karp did deserves a medal. This is an incredible person who had the courage to come here, and everyone watched this clip. All these CEOs of S&P 500 companies called me after that video and said, "Help me understand this." He was on the right side of history, because he shows you: wait a minute. You need a much more predictable way of doing business. If you succumb to the volatility of private company funding cycles, or the desire of private companies to meet trillion-dollar valuations, you're taking on a level of risk you probably didn't even realize you had. I think this is a valid criticism.
We need to take a little break. Last question, as it relates to the topic of Alex Karp. Essentially, he's saying that Palantir needs some kind of middle ground anyway. I think so too, partly to protect you. If I understand correctly, he was in a sense arguing with the LLMs themselves. Here's the thing: LLMs do a great job of a very superficial form of privacy called "zero data retention" (ZDR). They say: turn this on and everything will be fine. But if you look at the fine print, say Joe enters Kentucky Fried Chicken's secret formula into the model and asks for help optimizing it. That falls under ZDR. But if he simply clicks the "like" button, is there a guarantee that the information he liked will not be stored in some way? The honest technical answer is no, because we don't know how to do it. Another interesting thing: when Alex was on the air, he said something extremely pertinent. I think this applies to companies, and it also applies to countries, so sooner or later we should talk about it. Anthropic never said, "He is absolutely wrong, here are the exact technical details, and there's nothing to look at here." I thought, why didn't they do that right after the video went viral, to be very specific and refute this claim? I find it funny that here we are ten days later, and not a word has been said, except for Alex's version, because we all know that's the real version.
Comments 0