Google Battles Nvidia with TPUs and OpenAI With Pretraining, Automating Wall Street | Nov 25, 2025

0:00.

0:13Welcome, everyone, to the Informations TI TV. My name is Anita Ramaswamy. It is Tuesday, November 25th. We have a great show lined up for you today. First up, we have new reporting on Google's fresh efforts to compete directly with NVIDIA on AI chips. We'll discuss how it might be catching up there. And we'll break down one of the big concerns Sam Altman might have around Google's success with its Gemini 3 model. Plus, a closer look at the rising debt funding behind Oracle's data center build-outs. And a look into 2026. We bring on an expert to discuss what's coming in the IPO market. And we'll end today on a conversation with the CEO of Model ML, a fintech company that just announced a$75 million Series A round just about a year after launch.

0:59It's a busy show, so let's get right into it. The information published new reporting that shows Google is making strides in its efforts to compete directly with NVIDIA in the AI chip business. And it's doing so by talking to big companies like Meta about letting them use its chips in their data centers. Shares of NVIDIA fell in the aftermath of the story. Here to get into the details is one of the reporters on this story, Erin Wu. Hey, thanks so much for having me. Thanks for being here, Erin. So I want to start by having you give us a little bit of background. What exactly is Google's current strategy when it comes to AI chips?

1:35Yeah, so Google's been developing its own AI chips for a decade. They're called TPUs, Tensor Processing Units. And these are chips that are really made by Google, designed by Google for Google. And so they've been using them for development of their own AI models, including Gemini, which, as you discussed, has had a lot of success recently. And Google also has a cloud business. And so Google has also been providing TPUs for its customers through the cloud business. And they've signed a lot of high-profile deals, including Apple, Anthropic, Safe Superintelligence. But what we're reporting now is they're also starting to pitch customers on putting TPUs in their own data centers.

2:16So going beyond the cloud model that we're seeing right now and into putting chips on-premises. So can you unpack that aspect a little bit more for me, Erin? I mean, based on your reporting, what would you articulate as the change in Google AI strategy when it comes to chips? Yeah, so it's going from having chips in Google's own data centers to having chips in either the data centers of Neocloud, so companies like Corweave, or directly to the ultimate operating customer. So we talked about Google talking with Meta about a deal like this. Google's also pitching this to major financial institutions who maybe need to have data kept more safe and secure.

2:56And these talks, Erin, if they do succeed, how could they potentially close the gap with NVIDIA, who is far and away the market leader at this point in AI chips? Right. So NVIDIA is way further ahead than Google is right now. They have a lot of advantages because, in part, GPUs, they're made for customers to be able to use them versus TPUs, the chips that Google has developed. They're very specific from everything down to the design of the data center, to how power works, to how cooling works. So it's definitely an uphill battle for Google to climb in terms of getting customers comfortable using these chips.

3:32But the way Google is looking at it, it's like there's this massive market for all of these data centers. And the idea that Google could try to compete with that market is a really huge opportunity for them. And so when Google Cloud was thinking about doing this, they were asking, like, okay, like, what's the potential market? And the numbers were thrown around, like, okay, well, like, what if we could get to 10 % of what NVIDIA is doing, which is like a 10 plus billion dollar market? And so there's a really big, there's a really big opportunity here for them, even if they aren't able to dislodge NVIDIA.

4:07And I mean, it's definitely a big opportunity. To what extent do you think it's realistic that Google really gives NVIDIA a run for their money in this space? Look, I think it's hard. I mentioned that GPUs are already built essentially for customer use. With TPUs, this is a lot of what Google has to figure out is how to get people to use it. And so part of what we reported, they're not just giving people the chips. There's some software that goes on top of it, which is designed to make it easier for people to use. But I mean, right now, Google is in a place of real momentum and especially perceived momentum where they've been using these TPUs to train Gemini for a long time, but now with the success of it, there's, I think, maybe a moment for Google to come in and take some market share.

4:55And they've been, again, signing all of these big deals. Like there is just a deal with Anthropic, for example, for 1 million TPUs. Right. So you noticed, or you noted some of the challenges in this competition. And one roadblock that I've heard a lot about to Google competing with NVIDIA is NVIDIA's CUDA software. That's the software that AI developers use to program their chips. And from everything that I've heard, it's become somewhat of the standard amongst those developers. How exactly will Google compete in that realm? Yeah. What we've reported on in this most recent story, so Google's pitching companies this program called TPU at Premises.

5:32Like I mentioned, this is not something where they're just selling the chips. There's also some software on top. Like this could potentially be structured as like a monthly rental fee where like Google engineers are still managing some aspect of this. And so Google completely understands that this is something where they have to teach people a new framework. And it's like different than GPUs. But the way that they're trying to pitch it is this idea that this could be more cost effective. And again, there's this Gemini halo. The big company in your story and in your reporting, Erin, is Meta. but you also said in the story, I noticed that Google is in talks with big financial institutions about the use of TPUs.

6:11What more do we know about that? We know some more. There's some stuff that we couldn't put in the story. We do have some examples of who they're talking to. But this is being pitched as something that's good for companies that have very secure data needs and need to have things on premises because they don't want to send information to the cloud. And so that's part of it. There's also some interesting details where Google's trying to pitch this as something that's good for companies that are trying to do high-frequency trading. And so they're definitely making a play at this financial institution market.

6:47Well, I personally can't wait to see what details that you didn't get to put in the story that I'm sure will come out in your later reporting. Always a pleasure having you, Erin. Thanks for being here. Thanks so much. So sticking with Google here, the release of the Gemini 3 AI model has lit a fire under OpenAI and its CEO, Sam Altman. He told his staff to buckle down and prepare for, quote, rough vibes and temporary economic headwinds, as we reported in the information last week. My colleague who broke that story is now zeroing in on one specific concern that Altman might have, which is around how Google's model got so good in the first place.

7:23That's through a process known as pre-training. Stephanie Palazzolo joins me now to talk about it. Steph, good to have you back. Thanks, great being here. So just to get started, can you first remind us what exactly pre-training is? Totally. So pre-training is kind of the first step of training a model where you essentially show the model tons and tons of data from the internet, from books, from tons of sources to kind of teach the model connections between these different concepts. So this essentially kind of gives the model like a general understanding of the world and how the world works. Got it.

8:02And how is that different from the process known as post-training when it comes to AI models? Yeah. So post-training is a later stage in the kind of training process where, you know, once the model has learned this kind of general representation of the world, host training basically helps it improve either in specific areas. So for instance, maybe you want to teach it extra knowledge about math or coding or law, or it can also help the model kind of respond in a way that humans like more. So maybe you want to teach the model to be more polite or friendly or to have a certain personality. So why is pre-training something that Sam Altman is specifically worried about, especially when it comes to ChatGPT and its progress in pre-training?

8:48Yeah, so the reason why Sam is likely so worried about pre-training is because a lot of researchers think that pre-training is a much better way for models to generalize. And by generalize, I basically mean models being able to create content or make new discoveries outside of the data that it was trained on. So the reason why this is so important is that OpenAI and lots of other model makers are building these models because they're saying that they want to create AI that can cure cancer at some point or come up with new science discoveries or even do AI research to help improve AI. So obviously, all of these are brand new concepts and ideas that the models have never seen before in their training data.

9:34So this means that they have to kind of generalize beyond what they were originally trained on. So because pre-training is more helpful for that, you can kind of imagine that, you know, pre-training, again, gives you this general world knowledge. And so it helps the models kind of learn things beyond the training data. So this is why it's so important, because it does help with that, versus post-training can actually sometimes either, you know, it can actually have a negative effect on teaching models knowledge outside of the data that it was trained on. And so, for instance, if you want to post-train a model on how to be better at doing math proof, so you want it to be more logical and rational in its responses, I can actually hurt it in other areas, like being funny, for instance.

10:25That's pretty wild that there's a direct trade-off sometimes when it comes to training the model. I mean, I'm just curious to get a little bit more context on how exactly Google got so good at pre-training the model. Yeah, so it's not really clear to us exactly what happened within Google to get them to be so good at pre-training. As my colleague Aaron Wu has kind of written about, there are specific researchers at Google like Noam Shazir, who was brought in during the$2.7 billion acquisition of Character AI, who sources have told us has been really important to the process of improving pre-training.

11:06But a lot of people kind of just say generally, like, it's basically boils down to having a lot of really great data, and then very good processes when it comes to training the models. So essentially, the researchers need to run like lots and lots of experiments. So it's kind of the process of like running experiments really well, testing a bunch of different ideas, and then moving forward with the ones that are the most promising. So it does sound like if we're going to really see a step change in how well these models work, it has to stem from a foundational improvement in pre-training. I mean, I'm wondering how you're thinking about the next stage of the story, because it sounds like OpenAI is a little bit behind.

11:45What exactly will they need to do to improve on pre-training just to make sure that they're staying competitive in this area? Yeah, so as we saw from the memo that we wrote, Sam Altman does seem pretty confident that the company is making bets that he thinks is going to pay off and is going to help OpenAI catch up to Google when it comes to pre-training and the kind of quality of their models. So I do think that, you know, whatever OpenAI is working on, they do seem to be pushing ahead pretty strongly on it. And I imagine that they will want to see results, you know, sometime in the next couple months or so.

12:21We already wrote, for instance, that they are working on a new model, which is codenamed Shallot Pete, which basically aims to kind of fix some of the pre-training bugs and issues that they had experienced in the past. So, you know, we might, you know, it might take a couple of months, but I imagine that we're going to see that shallot peak model from OpenAI pretty soon. And then I think from then going forward, we can see, you know, how much OpenAI has been able to catch up with pre-training. Have they put in significant efforts into post-training as well? Yeah, so, so far, OpenAI has been pretty focused on post-training.

13:00So this involves, you know, acquiring lots of really specific data from, you know, companies in healthcare or finance, for instance, to really improve the models along those specific areas. And then we've also written about how companies like OpenAI and Anthropic have, you know, spent a lot of money on data labeling, but also on a specific type of data labeling called reinforcement learning gems, which are basically kind of like fake copies of apps that the models can use to learn how to navigate these new apps. So definitely, OpenAI, I think right now is definitely also very focused on post-training.

13:40Got it. Well, thanks, Steph. I guess it remains to be seen and we'll have to watch exactly how OpenAI navigates these quote unquote rough vibes ahead. And I know that you're going to be on the front lines reporting on all of it. So hopefully we'll get you back on the show. Thanks for being here. Wall Street is growing increasingly worried about rising debt funding for companies that are building data centers for Oracle. My colleague, Myles Krupa, wrote a great piece on this story and he joins me here today to talk about it. Myles, how are you? Good. Thanks, Anita. How are you? I'm doing well. Talk about debt and data centers.

14:14So, you know, just to put the story in perspective for our viewers, how much borrowing exactly are we seeing around Oracle's data center projects? projects? Yeah, so we've seen data center developers building facilities for Oracle raise at least$65 billion this year in project finance, which basically helps them build the data centers and get them ready for when Oracle is ready to come in and put their chips in there. So can you tell me a little bit more about the big contract with OpenAI, Miles, that everyone seems to be talking about that represents a substantial chunk of this Oracle debt? Yeah, the thing that's really driving this is, of course, that Oracle signed this five-year, roughly$300 billion contract with OpenAI for cloud services.

15:04And basically, as a result of that, we've seen Oracle go around and sort of snap up any gigawatt-plus data center that's being built and this is all now being branded under the kind of Stargate branding that OpenAI laid out in the White House at the beginning of the year. So you know Oracle is a smaller cloud player than Microsoft, Google and Amazon and historically it's relied a lot more than those companies on outside data center developers and we're seeing it basically continue that playbook even as it's trying to really grow into this open AI contract. So I'm wondering, Miles, what exactly is the big fear or the big risk around all of this borrowing and all this debt that Oracle has been taking on through, you know, both directly and indirectly?

16:01Yeah, it's still early days, but basically lenders who are active in this project finance market are starting to get worried that basically Oracle is a huge component, a huge portion of the market and of what is sort of like coming to market, the kind of flow that they're seeing month to month. And, you know, these lenders, even though they are sort of protected by some of the contractual terms around the lease that Oracle signs with these developers, they're still leery about getting overexposed to any one company, and in particular Oracle, because it's so reliant on OpenAI, which, as we all know, isn't projecting that it'll be profitable until 2030.

16:52Within this particular market, what else is included besides Oracle's debt? I mean, what other sorts of projects are being financed? Yeah, I mean, we're seeing other tech companies use third-party data center developers more and more, and then those developers will go out and raise debt in a similar fashion to the way they're doing it for Oracle. Corby was also a pretty big player here. One of their data center partners recently raised more than$2.3 billion of debt this month. So it's a very active market, more so in 2025 than ever before. And I think we're just going to continue to see increased volumes going into 2026.

17:36Having read your reporting, Miles, I think you've spent a lot of time sort of thinking about questions about systemic risk. And I was wondering what your perspective was on sort of the broader risks to the market. As we see this data center build out progress really quickly, we see companies trying to build data centers and scramble to get them out as quickly as possible. Yeah, I mean, I think the thing to watch for is the banks that are taking on the risk in these deals. are they able to get the debt syndicated to a wide array of banks and other asset managers just because these deals are getting so huge, you know,$38 billion.

18:14That's a lot for just a few banks to swallow. And then in that sort of broader syndication, you know, what kind of role are insurance funds playing? You know, we've reported on how life insurance funds owned by Apollo, KKR, how these are starting to play a bigger role in data center financing. I think that's definitely an area to watch if things go sideways. You know, how are these insurance policyholders affected? So those are a few sort of points I would follow. And if you're an Oracle investor specifically, are there any specific concerns that you should be watching for? Yeah, I mean, because Oracle is so reliant on these third-party developers to go out and raise debt, you know, if the market starts clogging up and Oracle has to bear more of the burden of raising the debt to build data centers, you know, we already know that its balance sheet is pretty stretched.

19:15It's on a credit rating watch from S &P and Moody's. It has a sort of limited ability to continue borrowing even more on its own. And so if this market were to clog up in a really material way, you know, that could potentially be problematic for Oracle. Sounds like there's a lot of risks that investors need to be apprised of. Really appreciate you coming on to talk about them. Thanks so much, Miles. Thanks, Anita. As we head into year end, we're taking a look at the IPO market and what's ahead in 2026. Joining me now is Ron Benzer, partner and co-chair of Capital Markets and Public Companies at Fenwick.

19:55Ron, glad that you're here to join us because now that the government shutdown has ended, you must be extra busy helping companies prep for IPOs, right? Thanks for having me, Anita. Yeah, it's been really interesting. So we obviously had a ton of momentum heading into the fall this year in terms of the IPO markets. That was obviously slowed down over the last month and a half or so with the government shutdown. The government's back in the office now. They are very clear that they have a long backlog of filings to still go through. So we are seeing significant delays for a number of companies.

20:36The good news is that the government is prioritizing companies that were very close to going public before the shutdown. And so we are seeing them move to clear those companies and get them out on the road. And we're hoping that they are able to clear the backlog here over the next few weeks so that the markets are up and running in January. Ron, what exactly does that backlog look like? I mean, how many filings are we talking? What sorts of companies are in that pipeline? Yeah, I mean, so they've been on record to say that they had about 900 registration statements that were filed during the shutdown.

21:19So they do have a significant backlog to clear through. A number of those, obviously, are technology companies, and we're working on a number of those IPOs. But, you know, it's everything from from, you know, the registration statements that you would see for a traditional IPO process to registration statements that that existing public companies have have had to file and that they they are required to review as part of their mandate. Can you name any of the ones that you're working on at this point? uh and for unfortunately i can't because a few of them are uh publicly filed and uh are in the the quiet period but uh we've had a fantastic year at fenwick in terms of uh in terms of ipo work uh this year we worked on um on the core wave ipo and and on figma which were two of the largest venture-backed uh ipos this year got it so i want to look ahead to 2026 you know everyone talks about their IPO predictions around this time of year.

22:21Do you think 2026 is going to be a big year for tech IPOs? And if so, why? Yeah, I mean, you know, I think we're certainly seeing a number of companies approach us and start to think about IPO readiness. I think there's a lot of backlog of companies that had obviously raised significant capital, even kind of going back to 2020, 2021. and they've spent the last few years kind of retooling their business models. And then obviously with what we're seeing in terms of developments in AI and incorporating AI into their business models, I think companies are increasingly confident around their ability to tell a really nice AI story.

23:07And I think that bodes well for the IPO market heading into 2020, into 2026. I also think, unfortunately, again, with the government shutdown, you will see a number of companies that were just kind of delayed as a result of this backlog that we were talking about that will go out early next year. But that should at least kind of set us up with some momentum heading into 2026 and throughout the year. Sort of on the flip side to your point, Ron, I mean, we've also been seeing the trend for a while now that many companies, especially in tech are choosing to stay private for longer. We had a scoop at the information last week that Databricks, for example, which has already raised about$15 billion in the private markets, is in talks to raise another round.

23:51And I wanted to know what you thought was actually driving this trend. Yeah, we represent Databricks, so I can't speak to them specifically. Sure, just an example. It is a trend that we've certainly seen companies are wanting to stay private longer. You know, one of the really interesting things with the new administration over at the SEC, Paul Atkins has been very vocal around making IPOs great again. And part of that is streamlining a number of the rules and regulations that apply to public companies to make the public markets ultimately more attractive. I don't think that we will see a complete reversal and move back towards IPO-ing early in a company's life cycle.

24:34I think companies are staying private longer for a number of different reasons. But, you know, as someone who represents very innovative technology companies that are kind of working on groundbreaking technologies and business models, oftentimes it's a benefit to be able to do it outside of the public limelight, at least for as long as you possibly can. And so we are seeing companies take advantage of the ability to raise capital and significant amounts of capital in the private markets and then kind of delay ultimately their entrance into the public markets. I'd love to know if there are any specific administrative or policy shifts that you think have taken place in this administration that really will encourage the IPO pipeline to be loosened up.

25:25Um, yeah, I mean, I think we're seeing a number of things that they're talking about. So for example, emerging growth company status, which is something that came into effect back in 2012, with the Jobs Act, was actually, I think, a significant tailwind in terms of encouraging companies to go public, it streamlined a lot of the disclosures that they were required to include in their filings, which ultimately I think encouraged companies to decide to list. They're talking about now ways that they can extend the benefits of emerging growth company status so that companies can stay in a scaled disclosure regime for longer periods of time.

26:10So I think that is one of the things that they're doing. They're also kind of looking at a variety of different proposals next year in terms of streaming filing statuses, shareholder proposals has been a big topic of discussion recently at the SEC and figuring out kind of ways for companies to be able to reduce some of the costs that are associated with being public are all things that are sort of on the table for this administration in terms of how they're trying to kind of spur companies to go public. As you're looking ahead to 2026, what kinds of tech companies specifically do you think are going to be going first?

26:57Yeah, I think, you know, one thing that we're seeing across the board is you have to tell a really good AI story in order to be a successful public technology company. You know, we have already started to see some of the infrastructure companies go public. So the core weaves of the world. It's going to be interesting to see, obviously, the LLMs and what they're going to ultimately do and whether they are near term IPOs or longer term. But, you know, I think really the interesting thing to me is, you know, back in 2020, 2021 and before that, the bread and butter IPOs really like the software companies that we were we were taking public.

27:37And, you know, I think a number of those companies have still tried to figure out their AI story. I think Figma is a great example of a company that was able to both start to kind of incorporate AI technology into their platform and then be able to tell a story of why AI is going to be a tailwind for their business and not a headwind, given all of the innovations that they've been able to show to their platform over time. And I think that is a really good model, ultimately, for other enterprise software companies that want to go public, being able to tell an AI story and then kind of explain how that AI story ultimately impacts the business model and the cost structure.