저자: ChamathPalihapitiya외사회자/패널 · 외국인
요약 보기
구글 실적(매출 약 965억 달러, 클라우드 +30%, 유튜브 광고 +14%)과 2025년 설비투자 750억 달러 발표 후 주가 7% 하락. 패널들은 AI 모델 품질은 구글이 최상위라 보나 투자 대비 수익률(ROIC) 공개가 부족한 것이 문제라고 봄. 감가상각 기간 연장 이력 등을 근거로 ROIC 관점에서 750억 달러 투자가 긍정적이라는 입장. 일자리 대체 문제는 생산성 증가와 새 직업 창출로 상쇄될 것이라는 낙관론이 우세. 후반부는 GLP-1 약물의 심혈관·간질환 등 효과에 대한 VA 데이터 연구 소개.
다듬은 글
out overtime let's wrap up on Google Google dropped 7% after reporting earnings on Tuesday chth Revenue up 12% year-over year to 96.5 billion wow Cloud revenue is up 30% year-over year 12 billion YouTube ads surging 14% to 10.5 billion net profit was around 26.5 billion up 28% year over year so they are really focusing on profitability obviously full year numbers this is extraordinary total revenue 350 billion with 100 billion in net profit cloud and YouTube finished 2024 with a combined run rate of $110 billion doll YouTube is basically Netflix inside of uh inside of Google and there Google cloud is essentially AWS inside of Google the thing that made investors get concerned shth is that Google said it would invest 75 billion in capex in 2025 42 % jump over 2024 29% more than analysts expect obviously this has to do with data center servers and the AI buildout what do you think about that number 75 billion obviously in relation to what we saw with deep seek doing it with maybe a little bit less maybe they're lying is this just an absolute waste of money or gargantuan number or is it something they can easily with that amount of profitability and cash they have absorb and use in the future I think I should stipulate that Google's models are probably the best of all the models across a broad base of capabilities if you test for them and so let's start there which is whatever they're doing is working the thing that they need to do is be able to translate those models now into better products and I think that that will happen slowly like for example if you look at Deep research most of the people online that are evaluating deep research would now say that open AI is both faster and just better on the margins all of these things can be improved I don't think that's a comment on the base model I think it's a comment on post training and how they're attempting to productize these things so the other thing that Google has is a money machine that directly benefits from these AI driven optimizations on ad targeting the only other company that has anywhere close to the same credibility is meta so I think what both of these two companies need to do is do a better job of explaining how that 75 billion gets segregated how much of that goes to these AI enabled models that actually do better ads optimization there was a there was a really interesting discussion by a former meta machine learning engineer on X about how they did it it's pretty amazing it's staggering but if they could say half the money goes to that and the other half goes to more speculative you know pre-training and posttraining I think the market would have eaten it up so it's probably more of a disclosure issue for Google because I would say right now their model quality is the best you have thoughts on this buildout Antonio obviously I think you've been involved with xai uh and obviously you're involved with Twitter and X previously and colossus's colossal buildout that was extraordinary to watch in such a short period of time do you have concerns that like some people do that this buildout is too expensive and there would be too hard to monetize all that expense or is it uh maybe a little bit of hand ringing and the opportunity in AI is so obviously huge that you just got to take the leap of faith and and and if you build it the revenue will come yeah I think choth has the right framework here which is return of us a capital and what he's saying by you know if Google had said hey half the money is for ads and half the money is for PL post training the market would have seen that half the money has a high returning capital and goog's returning capital is actually going up U after the impation of of models into their into their company and I I think that sort of abstracts the entire Market which is people are waking up that return invested capital in data c will matter that the models are basic Commodities and super competitive in the best case it's kind of a a land warn Asia it's a melee in in a worst case just total Commodities what does matter is the return on Capital data center which is influenced by how good the data centers are so you mentioned the xcii data center it's 100,000 GPU cluster it's the most dense coherent cluster in the world and it will just train faster and better than other clusters and it's also built for the most cheaply and the fastest so XI have the highest train on Capital and I also think the best train data and therefore will win I think Google will also win because they have um their tensorflow chips they make their own some of their own chips they do focus on roic and they have a great Monopoly to kind of fund it all so I don't think it's overblown I think that I think this is going to be as you guys have said before in the podcast you know bigger than industri Revolution but uh it's also true that you need to have a good roic and if you don't you're not transparent about it you can see what happens I I would argue Google's probably been the most frugal thoughtful and well-managed Computing infrastructure investor of all time you know the 98 to 2005 era Google it was all about just cheap throwaway racks that that was the big Advantage they had is they weren't using the expensive Oracle servers and they had a two to threee kind of depreciation timeline on those but they were super cheap and so the the roic was quite good then they got into Energy Efficiency which they realized was a big driving cost they started to build systems that that were more energy efficient as a result they lasted longer and then their their depreciation schedule moved up to three to four years meaning they could kind of write down the value of the servers over three to four years and then from 2010 to 2015 era their Hardware system for cloud allowed them to kind of extend through repurposing the utilization and they increased their depreciation from to four to five years and then in 2021 I don't know if you guys remember this they made this kind of big change to their depreciation schedule on data center infrastructure to six years so when they invest capex in a data center they would write down the servers over a six-year cycle because of AI optimization on maintenance so they started using AI just for internal management of the infrastructure so I would view the $75 billion capex actually as a very positive signal for the company I think that it means that they have a really strong line of sight on how they're going to have full utilization and great return on this if you do the math roic math on this $75 billion assume a 20% roic you've got to be generating call it roughly an incremental $15 billion of profit a year plus the amortization of the 75 billion so take the 75 billion divide it by 6 that's 12 billion plus 15 billion so what they're so basically they would need to make an incremental 27 billion of operating profit a year on the 75 billion for this to meet their roic performance that doesn't seem crazy because that's just under 20% of their in of their annual operating profit this is a very kind of I think important point which is Google doesn't just do this to to build out AI in the future they have a really strong line of sight on how this can kind of increment and you don't have a huge hurdle for them for this to pay back so I don't know I would kind of say Antonio's point do this as a positive I think if you use their historical ability to manage infrastructure and make predictions on Investments as an indicator of the future this is a strong and positive indicator and I do think that for all the naysayers out there that think that search is going to evolve to chat you could look at this as being a really important proof point that Google has the confidence that they're going to be able to move from search to chat and as jamat points out they've got great performative models and I would view this more of as a positive than a negative if they were underinvestigated see that make this degree of investment highlights the confidence in the strategy Antony I don't know if you've been watching the agent space or deep research that came out from Google and then closed AI I mean open AI launch their copycat version of this product sorry short feelings on it I'm curious your thoughts on job displacement we're looking at self-driving obviously wh Mo's got cars on the road obviously uh Tesla which you were previously on the board of and I know you were the First Institutional investor in Tesla self-driving is pretty good I only get like one or two disengagements per hour and they tend to be the ones where I just want to take the turn a little sharper kind of thing so that's getting pretty close byd is very close you got a A lot of job displacement that a car millions and millions of drivers in 10 years could lose their jobs researchers working at Gartner Group whatever Boston Consulting Group like it feels like there could be massive job displacement do you have concerns about that in the American economy and globally there's a lot of Hing about this and it's real in Prior moments of large disruption that there have been uh job displacements because there's you know generous people need get retrained for something else to can't retrained so you know I think uh there was studies M fre did of the steel industry in Pittsburgh in the 70s that the cost of each steel worker job loss was about a million dollars to the economy because these people could be retrained and that might happen here but I'm going I have a more benign Outlook personally what I think is going to happen is that you will have job loss uh but the amount of productivity that will be released in the US economy is going to be extraordinary right so so GDP growth is a function of the noral people working times productivity it's very simple economists want to make it more complex than that but the number of people working in t productivity if you go productivity goes up to 5 6 7 8% you get a massive boost in GDP growth and then what happens well people can get retra Tred they can get different jobs different Services happen people start companies the application layer of LMS is just starting the the barrier to start a company is quite low things like launch probably explode because there's all these people who have don't have jobs anymore that were accountants and want to start something it's hard to know what happens you know I believe in American agility I believe in this country I believe that we will figure out figure out a way to make this all work and if there's enough productivity and money in the economy that's flowing people will find new jobs they will find new businesses to start they will find new things to do we have to get out of the way take regulation down and let Americans be creative and unleash the American productivity machine Let's make that happen yeah that's I think really strong thoughts and I mean that's the game we're seeing on the field how many companies are we seeing hit a million dollars in Revenue with five employees where that used to take yeah 25 so the efficiency is there chath your your outlook on this issue since it seems to be coming up again specifically with truck drivers and Uber drivers and all kinds of other research jobs that seem to be be doing pretty well um or can be done pretty well by AI where where do you stand on this issue today I think it's a process Buffett wrote about this in an annual letter and what he described was the changing nature of jobs during The Agrarian Revolution and what you saw was a large cohort of people who supported themselves through farming and then the total Quantum of those jobs shrank by 90% when you had industrialized farming and tractors and whatnot but the economy to Antonio's Point grew around that business and added other kinds of businesses that didn't make sense in that moment and I think what you see is that when economies get more and more evolved you see the growth of services businesses that sort of are these things that can only happen when you have excess right the person that you pay for closet organizing would not have had a job in the turn of the agrarian Revolution or the Industrial Revolution but they can they can exist in 2020 something and frankly they can make a good job the life coach there are all these jobs that just kind of come out of nowhere podcaster influencer all of that Venture venture capitalist Capital allocator absolutely mutant strawberry Creator yeah so I think that I think that you're cut that waiting for this next turn of creativity like the big problem that we all have maybe I'll take the more glass half empty version of what Antonio said is we really haven't been unlocking people's creativity over the last 15 years we've been trundling around except safe of a few companies which we all know and we can just repeat them endlessly but we know which ones they are that are truly innovating and at the edge everybody else is kind of diddling around Naval casing so the real problem is that we have not had a lot of reps in being creative I'll give you just a very simple example Nick can you find this even extremely state industries did you see the byd clip of the car that and I thought myself Park yes with a little swipe it's how tragic is it that you look at this and you're blown away yeah and you're blown away for two reasons one is a I didn't think that that was possible and B why doesn't exist in America but the reality underneath the hood is ex extremely benign how this is implemented and so I think that the problem is we spent so much time losing the script I think Antonio is right when you unshackle people to not focus on the mental load of getting the pronouns right or the this or the that you won't be so overwhelmed by things like this because you will have already been pushing the boundaries of human creativity we need to get back to that we need to let these creative people cook and I hope I hope that that happens and I think that that will happen freeberg can you do a signs corner on the FDA approval for the nonopioid painkiller I not today no no I know but could you do it like in a couple weeks yeah one the one I wanted to do is that new macro study on gp1s which I think is super interesting oh tell us you guys remember I talked a while ago about how they were able to mine VA data so the VA right they they take care of veterans and they have all the medical records and on an anonymized basis they can make that data available to researchers and so if you guys remember this this is how they identified that the Epstein bar virus or the virus that causes mono as being statistically certain to be the trigger for multiple sclerosis in the cohort of hundreds of thousands of patients that were in this this data set no one got Ms that didn't get Epstein bar virus if you did not get Epstein bar virus you did not get Ms I don't know if you guys remember I did the science Corner a while ago anyway so the data set that you can mine at the VA is incredible so they pulled all the data from everyone that's been on the glp1 agonists and they identified all of the health effects across multiple indications the statistical difference between the cohorts okay so this research team out of St Louis pulled all the data from the VA database and basically they looked at you know 1.2 million people with diabetes that didn't take anything compared to 215,000 that took the glp1 receptor agonists and another 600,000 people that took other drugs for diabetes so basically this cohort segmentation allows them to isolate the effect of the glp1 drugs um and as you can see here this shows across hundreds of thousands of patients the effect of the gp1 on a hazard ratio which means like How likely are you to have the following health condition versus the population that's not taking the GP ones and then uh on the right if you scroll to the right Nick are the increase in risk and on the left are the things that go down so the increase the only thing that increased is like you know 8% or 10% increase in nausea and vomiting y confirm musculoskeletal complications grd which is you know gastric reflux from sleep indigestion yeah and sleep disturbance related indigestion so it's all abdominal stuff now if you go over here to the benefit side so the benefit side is what conditions did you see a decrease in so you have a decrease in shock a decrease in uh he hepatic failure so liver failure respiratory failure cardiac arrest in fact on cardiac arrest you see a 30% decrease in the probability of having cardiac arrest from the cohort that's on the gp1 verus Bia wow schizophr so so this goes to the point if you guys remember the interview I did a couple months ago with the CEO of Eli Lily that they have all these clinical trials going on right now for different indications for the gp1 receptor agonists that they're seeing that there's health benefits Beyond just the weight loss in reducing things like kidney disease obviously liver problems mental problems and we do we know why and and if we don't know why do you think it's because this is suppressing the food and it's the the lack of food or the change
Comments 0