GPT-5.2 vs. Gemini 3 Pro, Larry Ellison’s Big Bet & Risk, Nebius CBO on AI Cloud War | Dec 12, 2025

0:13Welcome, everyone, to the Informations TI TV. My name is Akash Pastrucha. It is Friday, December 12th. We have got a great show lined up for you today. First up, we are diving into OpenAI's GPT 5.2 with a researcher at Harvey who had early access to the model. We'll then dive into Broadcom's earnings. And separately, it is IPO day for Wealthfront. I'm talking to the CEO in just a few minutes. I'm also speaking with the CEO of Rubrik. And we've got another edition of the Editor's Cut. We're going to be talking with two of our top editors about how crazy a week this was for Larry Ellison making big bets.

0:51And finally, we will end with a little trip to the New York AI Summit. I sat down with the co-founder of Nebius to discuss his fast-growing business. It is a big show, so let's get right on into things. OpenAI has released a new version of GPT-5. It's called GPT-5.2. and any kind of release from OpenAI these days is certainly notable given the code red that the company called less than two weeks ago as competition from Google ramps up. Joining me now to break this all down and hear his thoughts on the model is Nico Gruppen, head of applied research at Harvey. Nico, welcome to the show. It's great to have you here.

1:26Thanks for having me. So what did you think of it? 5.2, give us the review. Yeah, so we did indeed have early access to GPT 5.2. And when it comes to evaluating any new model, we're really going to ask three questions. What are the innate reasoning capabilities of the model? We have our own internal benchmark. It's called the Big Law Bench. It measures the ability of LLMs to complete real-world legal tasks. The second question is, how does the model perform in the product? We're going to wire the model up in the product, and we're going to run end-to-end evals across our key product surfaces. And then the third question is, what do the lawyers think?

2:05Right. And this is actually, you know, my favorite part of the sort of new model eval process is it's more unstructured testing. So we'll give the model to our internal lawyers. What they're really probing for are sort of step change improvements in things like legal. You give it to the customers. To our internal lawyers. So at Harvey, we have a legal research team internally. They're all former big law attorneys. Yeah. Got it. OK, so three three vectors. You've got the benchmarking. You've got the lawyers. So how are people thinking about it? Yeah, so these three kind of signals coalesce into our general opinion of the model and sort of what emergent behaviors we've seen.

2:42In terms of overall performance, we're actually quite impressed with the model's performance. So on Big Law Bench, this is our public benchmark. It's the second highest scoring model we've seen to date. So it's inherently a strong legal reasoner. But I think the thing that stood out to us most about this model in particular, was the model's inherent sort of guardrails, right? So this came out in our unstructured testing phase, and we noticed an inherent or sort of emergent behavior where the model displayed stronger capability awareness, right? So it's able to ask itself things like, do I have enough information to answer this question?

3:18What can I do? What can't I do? Should I bring a human into the loop? These sort of behaviors were pretty prominent and noticeable and are actually extremely beneficial for our platform. Who was number one on the benchmark? So that is actually GPT-5.1, right? So what we see is the frontier models from each lab. So GPT-5.1, GPT-5.2, Sonnet 4.5, and of course, Gemini 3 Pro. They're converging to the same sort of mean or average level of performance on our benchmark. And that's what we consider sort of the frontier cohort. How do we understand the calculus here for these model companies in terms of when they decide that this is a decimal point increase, a 5.1, a 5.2 increase versus, hey, this is a brand new model?

4:06What do the deltas here look like and when do they decide that, oh, this is enough of a big step that we should launch a new model altogether? Yeah, it's an incredibly challenging problem. I think the model providers are working by proxy on the benchmarks that they can create, Right. And the big challenge there is, can you build benchmarks that are representative of the ways in which these models will be used in the real world? Right. And so the model providers will make progress. They'll hill climb against these benchmarks when they've seen meaningful improvements. They'll give them to, you know, model consumers like ourselves and ask us for feedback.

4:41Right. And so a big thing we're looking for are sort of these step changes in behavior, emergent behaviors. And it's sort of a feedback cycle. So you were overall impressed with the model. I wonder when you talk about feedback, what feedback did you give to OpenAI in saying, hey, this could use some work? Yeah, so I mean, the strongest piece of feedback we gave was certainly around the model guardrails, right? And this isn't necessarily negative feedback. We were actually impressed with this performance, but it was an emergent behavior that we hadn't seen from the GPD sort of predecessors. In terms of, you know, room for improvement, I think the things that we identified were, you know, the model is a bit rigid.

5:21in terms of styling, formatting. It has its own sort of preferences baked into its weights there, and it's harder to coax out these changes in prompting. And then with legal specific use cases, we did see the model perform notably better on transactional legal work than on litigation work. So these are the kind of feedback signals we're giving. And I just wanna go back to again, where you said 5.1 was highest on your benchmark. So based on your comparisons of 5.1 and 5.2 against Gemini 3, Gemini 3 is the one that made a lot of noise. You're putting 5.1 and 5.2 ahead of Gemini 3 right now. Am I understanding that correctly?

6:01Yeah. And what we're talking about here are decimal points of difference, right? So the current leaderboard is 5.1, 5.2, Sonic 4.5, and Gemini 3 Pro. Again, these are within marginal mean kind of differences. The thing that's more important for us is not the mean score, but where are the sharp edges for the subtasks of legal work that we're most interested in. Let me ask you a question here about this trend that there's this race to develop the best model and find which model is the strongest. There's also a discussion saying, hey, people might not need the best model all the time. We hear this discussion coming up with open source models that, hey, we just want to customize the model for what we need it for.

6:41where do you sort of land on this spectrum of we should be focused on the best model or we should be focused on making a custom model that is just good enough for what our company needs? Yeah. Yeah. I mean, look, the framing here is entire AI ecosystem is moving at warp speed, right? We had GBD 5.2 yesterday, Gemini 3 Pro and GBD 5.1 just a month ago and Sonic 4.5 about two months before that. So things are moving really, really fast. At the application layer, We're, of course, also familiar with urgency and sometimes what it takes to drive urgency. The things that we are prioritizing above everything else are model performance, production readiness, and client preferences, right?

7:23So model performance is measured by the benchmarks we just mentioned. Things like production readiness do come into play and are trade-offs with quality, right? So things like security and privacy, regional availability, and latency, reliability, all of these things do factor into the model we ultimately end up shipping to production. Great. Well, Nico, thanks for coming on the show. It's great to have your review, and we will talk to you again very soon. Thank you. Okay. Broadcom reported earnings last night. Revenue growth accelerated to 28%. Free cash flow jumped 36%. It wasn't enough to encourage investors.

8:00The stock initially moved up in after hours trading yesterday. It then erased those gains during the earnings call and swung negative and it opened in the red this morning. To break it all down, I want to bring on Tomas Tunguz, general partner at Theory Ventures. Tomas, welcome back to the show. It's great to have you here. Great to be here, Kosh. Thanks for having me on. So what did you think of the results? I think the AI ecosystem is on a knife edge right now where any perceived weakness, real or not, or any satisfaction of expectations leads to pretty significant drawdowns. I mean, you look at the overall growth of Broadcom relative to NVIDIA, it's not growing as fast, although it is much more of a services business than NVIDIA.

8:42You have the Google TPU business, particularly with Anthropic, that seems to be growing pretty quickly. Meta is a very significant customer spending a lot, massively increasing their CapEx spend at the expense of their virtual reality ecosystem here. And so you have a very, very healthy business, great margins, great cash flow from operations, approximately between 25 % to 27%. You have a pretty significant backlog. And I think the only thing that you could pick at the overall earnings statement was that some of that RPO might be a little bit, remaining performance obligations might be a little bit more delayed relative to where analyst expectations had thought they would come in.

9:23I mean, look, let's call it what it is. I thought the numbers were good, honestly. You have accelerating revenue. You have free cash flow. They talked about the big customers they have using the XPUs, Apple, SSI, Anthropic, Google, a fifth customer that they didn't want to name on the call, but they obviously have been working with open air in some capacity, whether or not it's that fifth big customer. You know, the thing that I was looking at this morning was the valuation of Broadcom relative to NVIDIA. And based on the revenue estimates, Broadcom is actually trading higher, which I think might be what people were getting at.

9:58I think that's right. I mean, you have very elevated multiples. And I mean, we know from all the previous hardware cycles, whether it's telecom or also hard disks, that any miss here, you have a very significant drawdown. And because of the capex here, so intense. So if there's any weakness whatsoever or any perceived sign of potential weakness in the future, I mean, you saw it with Oracle earlier this week where the RPO was absolutely massive, but the timing of that RPO remains unclear. And will people actually end up using this? I think that's the major factor that people are trying to understand.

10:37Like, okay, OpenAI is committed to spending a particular amount. Google is committed to spending a particular amount. CapEx is growing, it'll probably be somewhere close to 3 % or more of US GDP next year in terms of data center build out. Will the inference be there? And all signs point to yes, but if there's any indication that there's a deviation in that plan, then the drawdown will be significant. And so I think that's why you have this sensitivity. The only other way I sort of interpret the market is the profit taking. The stock has ripped this year and there tends to be some profit taking before the end of the year.

11:10And so maybe that's another reason. Well, it certainly would be a nice Christmas present for sure. I want to ask you about Hawk Tan. We wrote this great profile about Hawk in our Weekend Magazine last week, and he's an iconic CEO in the world of chips. I wonder what the discussion is in Silicon Valley about Hawk Tan as a leader. You know, what do people make of his leadership and the way that he's running the company. Yeah, I mean, I think he's done a phenomenal job, right? You see the core technical advances. I mean, Broadcom is some of the best technology to what's called serialization and deserialization technology.

11:51And the speeds there on the GPUs, I mean, are 224 gigabit per second. Nobody else can come close. You know, there was a rumor that Google might be bringing some of this capacity in-house for the chip design. But I think given the relative performance of Broadcom compared to Marvell and Synopsys, it's really unlikely. So I think he's done a really nice job building a pretty phenomenal business. It's also diversified, right? Compared to NVIDIA, where 75 to 80 % of the revenue now is primarily within data center. Broadcom has an AI business that's growing about 75 % per year, but it represents something like 30 % of the overall business.

12:28And so if there were to be a drawdown, you have nice core intellectual property. It's very difficult, even for the hyperscalers to replicate you're talking about hiring thousands of people pretty significant uh engineering investment over the course of multiple years to come anywhere close uh and then a nicely diversified business so to your point like it's it's kind of um a head scratcher about the overall stock performance of this business well i i was texting with with some analysts last night during the call and and if you look at the the stock movement basically after this first question that analysts asked about Google bringing the capacity in-house, right?

13:08That was the answer that they were not very impressed with, investors I'm talking about. And analysts sort of said he didn't give as reassuring an answer as they would have hoped about their advantage there. I wonder how much of a threat you think that is to Broadcom's business, this idea of customer-owned tooling, I guess, that they could design their own custom chips on their own in-house. It seems like a big endeavor. It's a huge endeavor. These are massively complex systems, and there are many different dimensions to it. So I think two or three years ago, there was the rumor that Google was trying to build this in-house.

13:41Google pays something like 30 % to 50%, according to analyst estimates of the BOM 4 chip design. And so you can imagine that this will be an important thing for them to be able to do at some point. And it's incredibly technically difficult. No one has come close on some of these chip throughputs, particularly the serialization, de-serialization component, where you can imagine a GPU is processing a huge volume of data in parallel. It has to aggregate that to send it to the next step. And then that next step has to take it and de-aggregate it and parallelize it. That's the bottleneck for a lot of these GPUs systems, aside from in and around memory.

14:18And so as long as Broadcom has head and shoulders advantage there, then I think they'll still have a phenomenal business. But if you're Google and you're trying to compete with them or trying to capture more margins on the TPUs and meta and others, will you be trying? I think you will be trying, but I can't imagine this changes in the next two or three years. Great. Well, Tomas, it's always a pleasure to talk with you. Thank you for coming on the show. We really appreciate it. My pleasure was mine. Thank you, Vikas. Talk to you soon. Okay. Wealthfront is going public today in an IPO that values the company at$2.6 billion with all the stock options included.

14:53The company has built its business on robo-advisory and services. It also has cash management products and is moving into the lending business as well. Joining me now is David Fordinato, CEO of Wealthfront. David, welcome to the show. It's great to have you here. Thanks so much. I'm excited to be here. Well, congrats on a big day for the whole Wealthfront team. I want to talk a little bit about the IPO window, which is something we've talked a lot about on our show. Look, we talk about it opening and closing. Why did you decide ultimately that this was the right moment to go public? Well, so there's a long answer to that and sort of a short timing thing.

15:30The short timing thing is we were planning to go public in October, and obviously we're impacted by the government shutdown. We got together with our advisors as the shutdown lifted, talked to investors, and the feedback that we got was we were in good shape to go. If you think about the longer sort of trajectory of Wealthfront's history, we've been a profitable company for years, have been growing consistently, have done 26 % year-over-year revenue growth, and have continued to make progress towards being a mature company. We felt like this was the next step for us. Now, you guys have built a business on robo-advisory, which has been around for a long time.

16:06Now we're in this new era of AI, which seems like a natural application for the technology. I wonder if there are scenarios here where you think AI is not well-suited for wealth management and what parts you are not applying it to in your business. So we don't apply it to our core investment strategy. If you think about the things that AI is amazing at, you know, its use of natural language and being able to apply complex judgment. In the case of kind of best of breed passive investing, these are things that have been set out by academics for long periods of time. We apply those time-tested academic strategies.

16:42We make them very low cost to implement and we share the savings with our clients so that they can get the best outcomes at extremely low cost. And so what are you using it for then? You know, we use it across our client support, engineering teams. There are very interesting applications in financial planning that we'll work on over time. We're very excited about the possibilities. But for investing, we want to stay focused on time-tested, academically proven strategies. But I just want to go back to it. So with respect to passive investing, and you talked about the academics having studied this stuff for decades.

17:16So you don't think there are any applications or any suitable sort of scenarios where AI should be able to make the decision for you based on all the data that it's collected? I mean, the core of passive investing is basically looking at an optimization problem based on expected return and risk. It's a relatively straightforward mathematical equation to solve, you know, defined by Harry Markowitz in the early 1960s. There's been research that's gone into it with Black Litterman. There are obviously views that you could express with AI tools. And some people have tried to do that. It's still very early to do that.

17:52One of the things that's, I think, different about investing is that it's a competitive market. Everything that you do, you get reactions to. And so if you look at the big trading firms that do or are starting to use AI, they are being very careful with it because everyone else is competing with everything that they do and updating on a real-time basis. It's not something that you can just implement an AI strategy and then let it run forever. Let me ask you about another flashy technology that I trust you're probably staying away from, at least in the near term, which is prediction markets have taken the world by storm.

18:27And stick with me for a minute here. Just as an outsider looking at this flashy sort of quick moving trading ecosystem that is popping up around prediction markets, What are your thoughts on it? Well, so there's been a bunch of amazing research that's come out of UPenn. Philip Tetlock as a professor talking about the importance of both teams of experts and his judgment project, the Good Judgment Project, and prediction markets and how they work. I think it's a fascinating technology. It's amazing for learning about things. It obviously is used for speculation. It's very difficult to see a long-term investing angle to prediction markets.

19:09because it's a negative sum game. The fees to run the exchanges end up taking out, you know, expected return from prediction markets. And so if you were just to invest randomly in prediction markets, you would expect to lose money over time. And what about these other tokenized offerings for private companies and, you know, being able to buy some kind of a proxy for a share for OpenAI and stuff like that? There's so much out there now. What do you make of it? Yeah, I mean, I think it's a tough market to be successful in. Obviously, companies are valued extremely highly. There's a lot of risk in those companies.

19:48I think if you look at the venture returns as an asset class and compare that to the NASDAQ 100 or other tech-focused indexes, really a diversified portfolio that invests in many things is likely to do better for individual investors. What's the average age of the people on your platform? The median age of a client on Wealthfront today is 35, but new client cohorts are coming in with a median age of around 24 years old. And you started the company a while ago, and I wonder if you can speak a little bit to how you've seen investing change among young people. You talked about the new joiners, right?

20:25I mean, you talk to people in their early 20s all the time. How are they talking about it differently compared to when you started the company or when the company started, I should say? Well, I think what happens is there's a group of investors in every generation that's interested in speculation and taking advantage of the opportunities that they think that they can bring to the market. And there's a group of folks that are more focused on their families, their careers, and building wealth over the long term. We don't attract people that are interested in the speculative opportunities that exist in the market.

20:53We attract people that are focused on their families, focused on their careers, and building wealth gradually by saving every month and putting their money to work. They want to delegate that to someone, and we want to be the folks they delegate that responsibility to. Great. Well, David, congrats on the big day. It's a big step. And like you said, it was months and years in the making. We'll let you get back to the festivities. That is David Fortinato, CEO at Wealthfront. Thank you. Okay. Shares of data management company Rubric have more than doubled since its IPO last year. The company recently reported its third quarter results, 48 % revenue growth from the prior year.

21:31Free cash flow was more than 4x what it was this time last year. Joining me now to discuss the company's playbook is Beepul Sinha, co-founder and CEO. Beepul, welcome to the show. It's great to have you here. Thanks, Akas. Excited to be here. So, look, I want to talk to you a little bit about the growth you guys have been seeing. Where is 48 % revenue growth coming from? We are delivering cyber resilience to every business and government around the world. And cyber resilience is the number one cybersecurity category because everybody has spent so much money in prevention and detection of attack.

22:06But attacks are still happening. You can't prevent the unpreventable. And our strategy is that be ready for attack that is inevitable and ensure that your businesses are up and running even when they are confronted with an attack. Now, we talk about attacks. I do want to talk a little bit about the broader landscape here of AI and what it means for cybersecurity. But sticking with the growth here, is the growth largely coming from existing customers spending more or is it new customers? Rubrik has three core pillars for our business. Our core pillar is data protection, which is the business that is at a scale and we are growing it quickly.

22:46Then the next business is identity resilience, which is a new business that we created about three quarters ago and we hit 20 million ARR in just three quarters. We doubled the number of customers just in Q3 and it's very fast ramping business, early days, but very exciting. And the third part of our business is our agentic operation, which is a new product that we just launched. So we are delivering resilience everywhere across data, identity, and AI. And if you look at the combination of all three, this is what is driving the demand and what is driving the growth. And if you think about these customers that you're serving, one of the things that we've and talking a lot on the show is how much M &A companies need to undergo to be able to offer these new innovative products, given that they can't seem to build things fast enough in this fast moving landscape.

23:44Look, your stock is trading higher than other software companies, at least on a multiples basis. Are you looking to do any M &A? Rubrik is a unique company in terms of our culture. And we obviously have both organic and inorganic strategy, But if you think about identity resilience, a new business that we created completely organically from scratch and able to launch it. So that shows that our engineering, our product, even at this scale, is very, very innovative. And then we are layering on top of it team products technology to accelerate our roadmap. So we are always open. We are looking at great teams, great technology.

24:22We just acquired PrettyBase, which is an AI model fine tuning and serving platform. we want to serve our customers whether we can build it or we can acquire. Do you think AI is net good or net bad for cybersecurity? Because we've been hearing a lot on this show about how AI makes it a lot easier for attackers to come at you in different ways, shapes, and forms. I'm sure it has ways to protect you as well, but it seems like it's making the game a lot scarier. I'm in the overall IT industry for about 30 years and doing technology for a long time. Look, any technology that comes in, it has both positive and negative impacts.

25:08If you look at AI, it's no different. AI has given 100x more opportunity to everyone, but also it creates 100x more rest. And what we need to do is we need to give organization a framework to govern AI, to have confidence in scaling AI, while ensuring that they can deploy agentic work to actually accelerate their own business and be competitively more viable in the marketplace. And that's why we launched Rubrik Agent Cloud that delivers agentic operations while managing risk. In fact, we say unleash agents, not risk. But will it be enough? I mean, do you anticipate we're going to have more of these large-scale attacks because of AI?