Limitless: An AI Podcast - This Week in AI: Nvidia "Acquires" Groq, Meta Buys Manus, Gemini is the Biggest App

Episode Date: December 31, 2025

Nvidia's $20 billion deal with Groq could totally change AI infrastructure. Meanwhile, Meta's purchase of Manus AI might not be so bullish. We celebrate Google’s Gemini app dethroning ChatGPT and discuss Alphabet's acquisition of Intersect for better energy control. Plus, insights from Andre highlight the future of generative AI in shaping user experiences.------🌌 LIMITLESS HQ: LISTEN & FOLLOW HERE ⬇️https://limitless.bankless.com/https://x.com/LimitlessFT------TIMESTAMPS0:00 Nvidia's $20 Billion Acquisition0:49 Understanding Grok's Inference Technology10:11 Meta's Latest Acquisition12:10 The Rise of Manus AI15:53 Google Takes the Lead with Gemini21:46 Google's Strategic Energy Acquisition22:05 Andre Carpathy's LLM Year in Review23:57 Wrapping Up the Year------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

Nvidia just spent $20 billion to acquire a company no one's ever heard of, and it's their biggest acquisition to date. GROC specializes in a very unique type of GPU, which will help Nvidia gain an even larger monopoly on the GPU stack. But they're not the only ones making acquisitions over the last week. Meta announced their 20th acquisition of 2025, Manus AI, which specializes in a very own unique AI agent. But I'm feeling pretty bearish about this. I think that this is one of many pitfalls that Meta is going to face over the course of 2025 and 26. And in other news, Google has finally dethroned OpenAI at the number one spot for AI being used. It is top of the charts for Apple. We're going to be getting into all of these and much

more in this episode. Yeah, so let's get into GROC, which is not to be confused with GROC, the other GROC. One is with the K, which is the one that we frequently talk about, which is made by XAI. this GROC that we're talking about today is GROQ, which is a totally different company that was just recently acquired by NVIDIA, but we need to make an important clarification about the acquisition. It wasn't actually an acquisition of the company. Invitya does not own GROC. What it was is a non-exclusive licensing deal for GROC's inference technology and a big aquire of GROC's top leadership and engineers, including the founder and CEO Jonathan Ross, who we're going to get into why he is so important in a little while. Basically, it's this non-exclusive license. for the technology, the inference technology. And I think that's what we're going to focus on a lot this episode is that Nvidia has the GPU, which we all know is the powerhouse of AI. But one of these weak points that we're missing is this inference time, this test time, this thing that needs to be quick and very cheap.

And what we've seen with Google and other companies is as they've started to introduce TPUs onto the scene, well, things get a little complicated for Nvidia's bulky GPUs. They're not as efficient. They're not as effective. they're not as cost effective, and it left them with this blind spot, which seemingly they filled with this GROC acquisition. Let's get into what that blind spot actually is, right?

So to your point, Josh, Nvidia makes these GPUs that are really good at helping AI models train, and it requires a different type of architecture to be able to do that. But when it comes to AI applications, what you really want to rely on is how quick the AI responds to you. I don't know if you've come across this, Josh, but when you write a prompt to your AI model, I sometimes get really impatient when it hasn't responded over like 20 seconds.

I'm like, come on, like, can you think a little quicker and just give me the answer now? And this comes down to like that test time compute that you mentioned. And the way that you enable a quicker response is if you reduce the latency. And there's many different architectural choices that you can make to do that. But the point is, Nvidia didn't specialize in that. But this company GROC has. And so, Nvidia had a few options to make about this, Josh. which is highlighted in this tweet here.

Option A, they just kind of continue using Nvidia GPUs. It costs a lot more and it's just inefficient. Or option B, they acquire or partner with GROC, which is the leader in these specialized chips that allow for lower latency. Or option C, they start creating their own specialized chips from scratch, but it'll cost them a lot of time, a lot of money, at which point GROC will already be the number one leader.

So I think this is a strategic position. They spent $20 billion, which is a lot of money for them, but they're able to acquire the number one leader in that case and own that monopoly for them and not be afraid of GROC kind of taking over Nvidia or providing an alternative to Nvidia in the future. Yeah, and if you'll remember, like Google recently, we made so many episodes about the TPU and how it's becoming a legitimate threat to Nvidia. Why? Well, it's because the TPU is good, really good,

at something that Nvidia is not, which is this very deterministic computation that happens local. and rapidly when it's actually delivering AI to the user. So this won't interfere with the core business of Nvidia, which is largely the GPUs and this pre-training part of the world. And we have a post that breaks this down really well, EGES, that you shared with me earlier that I kind of want to take a second to go through if you don't mind sharing.

Yeah, so it's this post right here. And it describes something called LPU. So we've heard of GPU graphics processing unit. LPU stands for language processing unit. And as its name implies, it's specialized for LLM specifically. and efficiently processing tokens into whatever output you would like from your AI model, right? And there's a few things that make this specifically unique architecturally. One, it uses this thing called static random access memory.

So we've all heard of RAM. We've all heard of DRAM, dynamic random access memory. SRAM is an alternative to DRAM. So for context here, DRAM is like the most sought-after type of memory that people want to put into their AI chips to train their models. And Josh, you and I've been discussing this a lot. The price of DRAMD has like 5X over the last couple of months. Because we just don't have enough.

In fact, like the cost of a, what was this? The cost of like a PC memory card now, Josh, is the equivalent of like the latest MacBook. It is just insane out there, right? And so GROC has this alternative memory board called S-RAM, which they integrate into this LPU chip, which is what makes it so unique. Well, what's so unique about it? In a typical GPU, so take for example, in video's GPU, in order to process an LLM request, you need to take information from the memory and take it to the processing unit, the processing part of the chip. With S-RAM, you don't have to do that as much. You can just store the entire model weight on that part of the memory

component, and it's able to process way more efficiently. How much more efficiently? 10 times more So for a singular chip, you can store 10 times more data, so more memory rather, and it'll require one-tenth of the energy cost, which makes it an extremely efficient chip to get fast inference between your AI model and the person that's using the AI model. That's the unique unlock here, and it's just super cool to see. Yeah, so in conclusion, it's not that Nvidia bought GROC. It's closer to Nvidia bought the parts of GROC that actually matter the most. And a big part of that equation is the talent.

the people that were involved in GROC. And where did I find out about the people? Well, on Chimath's memo from 2016, nine years ago. So for people who don't know Chimath Polyhapitia, I mean, you probably do. He's a very, just very influential person in the world of technology. And he published his memo from 2016 very early on about this guy named Jonathan and about his decision to be the lead investor in the series A of this small little company called GROC.

And nine years later, fast-fellow. This tiny little company worth $10 million sold for, well, got a license acquisition for $20 billion. And it was fascinating to read this memo and kind of see what the worlds looked like in 2016. It's this funny snapshot. It's a very forward-looking and accurate representation of where we would have ended up at a time when this wasn't clear. So if you remember, in 2016, the transformer paper hadn't even come out yet. There was no such thing as the transformer. Josh, I was at university in 2016.

I had no idea that was happening either. No one did. And yet this memo signaled like, hey, this thing called a TPU is going to be very important towards machine learning, which is also going to be this very important thing. And what's really funny about the memo is the header. It includes a subsection fundamental human need, which says understanding and creation. Then it says, is this going to affect 25% of the population? Yes, but indirectly.

$100 billion by 2045, possibly. A special person, yes. And then as you read more through the memo, you start to understand that he invests in this company because the founder and CEO was, or the founder and CTO at the time, who later became the CEO, he was the founder and the creator of the TPU, which Google now has. This is the dude. The dude running rock is the dude who created the TPU, which is now the biggest thing in the world. So clearly they saw something early. Clearly it was the correct bet. And I mean, good for them. They deserve it. I mean, we certainly were not thinking about these things nine years ago when it really mattered. Isn't it crazy that he had the foresight to identify both the trend of machine learning becoming the most popular form of computation and coupling that with like we need a new chip? So at the time, GPUs were super popular. You and I were discussing this before we started recording, Josh. You know, we were using GPUs for gaming. We were then using it for like crypto mining around that time.

It was becoming more popular. But he had the foresight to be like, this is probably going to be used for some kind of AI. ML inference in the future. And he saw it a decade ago and he waited it out. So if you look here, it says it would allow them to own around 28.57% of the company, which is just a massive bet. I mean, yeah, great. His 10 million is now worth a hell of a lot more. I think there's rumors that he made over a billion dollars on this is just insane to kind of see at that time. I just want to kind of put into context what this Nvidia acquisition looks like, right? So there are a, specific types of chips or kind of like hardware that you need to make a successful AI model

or an AI product in general. You need GPUs, which everyone's heard of. Now we have these LPUs, which now, I'm here on this tweet, Nvidia now also owns via GROC, the number one leader of this. And then there's CPUs. And we know that companies that are specializing CPUs include Intel and AMD. Earlier on in this year, Nvidia made very big investments in both of the leading companies, AMD and Intel. So I just see this kind of like this Thanos grip that tightens with every single acquisition that Jensen makes. And the craziest part about this, Josh, is that it's all pretty much free. When the news broke that Nvidia had acquired GROC, their market cap pumped $35 billion, which paid for the acquisition of more. That's so nuts.

It's just crazy. It's just crazy. So that's the big news on Nvidia this week, in an acquisition that feels very powerful and meaningful like it will actually move the needle. And now here's another acquisition where I'm not sure the same thing can be said. Just this week, Meta announced another acquisition. After the seemingly 100 that they've done in terms of personnel and companies this year

in a company named Manis. Now, Manus is a Singapore-based, Chinese-founded company, and the price hasn't been disclosed yet, although it's rumored to be somewhere between $2 and $5 billion. Manus is less than a year-old, Ejas, I believe. and it was valued at half a billion dollars just earlier this year, and they raised $75 million.

And then in nine months, they went from zero to $100 million. Oh, yeah, here's the post in eight months. They went from zero to $100 million annual recurring income, which is just this unbelievable hockey stick growth. So why would Meadow want to acquire them? Well, Manus is different than other companies. Manus is more of an agent than a chatbot. And if you'll remember, Medus had a tough time building a chat.

bot, but has a lot more use for an actual agent. So Ijaz, maybe you want to explain kind of what this acquisition was, why Meta made it, and what their hopes are in terms of what type of value this can add to a company that's kind of having a pretty tough time deploying AI in an effective and meaningful way. So I'm reminded of something specific that Zuckerberg said in his interview with Dwar Keshe earlier this year. Dworkesh was like, why are you investing tens of billions of dollars into AI. You guys are a consumer app product. And he goes, yeah, I don't want to build the best AI model. I want to build the best consumer AI app. Meta specializes in consumer apps. We think we're the best at it. And we think we can be pioneers in this new AI world. We're not trying

to get the best model. We just want to build the best consumer experience. I think this is another step towards that. So if we kind of like track back to like the age of the internet, which we currently kind of live in right now, right? We use websites. We download apps. It's a very kind of archaic thing when you consider that AI is probably going to automate a lot of this. How is that going to be automated? It's probably going to be in the form of agents. Agents that can automate a bunch of work for you, but also make your entertainment and lived experience way, way cooler, right?

So it can help you with research, but it could also help you kind of discover new things, new shopping experiences, stuff like that. And Manus AI is one of the top agent producers out there. Like you said, Josh, this hockey stick growth to 100 million ARR, even though I think that's a pretty vague metric. And I think a lot of that can be gameed. It's still very, very impressive. And I know a bunch of my friends that use Manus in many different ways. Can I just put my bear hat on for a second, Josh? I know we don't like the bears on the show, but my skin is tingling for this one particularly.

Why can't like a Claude code or a Claude in general just kind of recreate this entire product? I don't understand the unique capacity for here. And maybe it's because META didn't have the props or the talent to create this themselves. So they just saw this as an easy acquisition for a cheap $2 to $5 billion. And now they can kind of just put this a ready-made product out to their millions of billions of users. But that's also still bearish for me because they've spent $35 billion this year alone to acquire like a hundred different people and they've got nothing to show for it. I'm just honestly confused, dude. Yeah, it's aside from the actual strategic thing,

you could kind of look at what Manus does and then I guess back yourself into the conclusion from there. And you mentioned Claude Code. I would say Manus is kind of similar to ClaudeCode in the sense that it's kind of a harness for a base model. There are these base models that are interchangeable that Manus accepts because Manus does not actually have a base model. But it creates this harness around the model

that allows the AI to do more interesting things than it would if you were to just go to ChatGPT or Gemini. A few of those things include tool use and orchestration. So a big thing when you're using AI agents is tool use. What types of tools are they able to use in order to do things for you? And if you think of CloudCode as the agent for building code, you could think of Manus as the agent for controlling your computer, for doing deep research, for managing files.

It's more of the consumer-facing version of CloudCode, where it's really great at tool use. It's really great at running virtual computers. It's really good at analyzing real world's competitiveness through deep research of many, many different variables. It's really good at UX and design. And it has this unique skill set that doesn't require a core base model in order to use. And I think that's important because meta does not have this core base model yet. In fact, I suspect if they're going to continue using manners, they're going to have to use some third-party application until they could deploy their new supposed model named avocado, which is their first close-source model that will hopefully release some.

time soon. The way you describe that, like, you know, it doesn't use a base model. It kind of sounds like a rapper, right? And, you know, there are a bunch of critics against rappers. I actually think rappers are great and fine. But isn't that just cursor? I guess it's the same as what cursor is for coding. Manus is for agents in a way, and it doesn't matter what model it uses. I don't know if that convinces me enough behind this acquisition. And I'm purely coming at that from the fact that they have acquired so much supposedly good talent and teams over the last six months, and they've had nothing to show for it. And I'm kind of confused at meta strategy in general. They say that they want to go into consumer AI, but their vibes app, which is basically AI TikTok that they released,

got completely swept under the rug. No one cares about it. No one uses their AI chatbot. In fact, they're kind of gaming metrics by allowing any Instagram search is now an AI model search. I don't know if you've seen that, Josh, but like it comes up with this really annoying. kind of summary when you're on Instagram. I'm just like, why are you doing this? I don't really care about this. So I'm not convinced, but let's see in Q1 whether they release this new avocado model and whether it's actually something that people want to use. I'm going to remain kind of like uncertain on this for now, I think. Yeah, I guess in terms of the agentic use case, it does make sense because like Meta has, they have billions of people generating real world data

all the time to improve this. And they have the glasses and the Quest headsets that they want agents for. They have, like you mentioned, Instagram, WhatsApp. the app, the messenger as places to use. They have the distribution. So if you can apply this agent, this agentic force on top of it, you can generate a lot of value for the user. So maybe, I mean, again, we'll see. And then one last point is there's this funny little chart that I like to reference, which is Yahoo. It shows their acquisitions by year up until 2014, because, well, if you remember after that, they didn't do so well and then had to get acquired. But there's an interesting chart that follows this. And this is kind of reflective of meta, which shows their

market cap over time. And you could see that their market cap exploded and was at the highest it's ever, or just about the highest it's ever been, as they were making all these acquisitions, they were desperately trying to acquire the talent needed in order to be competitive in this new world. Unfortunately, that didn't work. And just two years later, they were acquired by Verizon for 90% less. That's crazy. They acquired everybody. They made so many acquisitions two years and one year prior to them going under. And what we're seeing with meta, not to draw a direct high, but there's this kind of almost irrational acquisition spree

where they're spending tens of billions of dollars on individual people. They're spending billions of dollars on companies. And they're just trying to get talent in-house, but they still have nothing to show for it. They're not able to generate revenue. They're not happy to make the products better. So I don't want to directly compare it to it because I don't see them going down 90% like Yahoo did.

But it's just a testament to show that normally when people start a, or companies start acquiring at this rate, it's just something to be aware of is all. But we should probably move on to the next bit of news for this week because there's a lot going on. And that is the Google Gemini app back in the news, right? Josh, my favorite company is officially number one. And it's a massive really being able to say that.

Good for that. So what are we talking about? So Google Gemini is obviously Google's leading AI model, but it's also the name of their AI app, which is a dedicated app like the chat GPT app, You can go on and you can speak to Gemini 3 Pro and all the latest bottles. And for a while now, it's been creeping up the rankings. And it was in the top 10 for a while.

And then Gemini 3 Pro released. And they shot to number two. And it's been at number two for a while. They are finally number one, Josh. And this comes after learning a few weeks ago that more people are, or rather, people are spending more time using Gemini models than the average chat cheap. user, which is just crazy. And the number of daily, weekly, and monthly active users have been creeping up to the point where it's almost, not quite yet, but almost at parity with OpenAI's

chat GPT users as well. This is a major milestone, in my opinion, because for a while, I mean, you and I know this very well, Google's just been on a rollercoaster of a ride through AI. They created the transformer paper, so they were supposedly in the lead, but then they didn't do anything with that. Chat GPT and Open AI ate their lunch, and then they made this massive. they've 180 degree turns, like your Britain came back, and they've made this crazy comeback. And it's just kind of cool to see them

take the crown of this position, both at the consumer level and at the chip layer with their TPUs. And now everyone's using Google search in AI mode. It's just crazy to see this type of comeback from Google.

And they forced the $20 billion acquisition for NVIDIA and GROC. I mean, I assume if GPUs didn't become big, that would not have been a pressing issue. But this is not the only time that Google was in the news this week. There's already more news coming out of Google, which is, as it relates to energy and infrastructure, EGES, what's going on with this? I didn't get a chance to read it.

So the headline here is Alphabets acquired this company called Intersect. I think they acquired it for $5 billion. So what does this company do? Intersect is an energy supplier and producer. And so you might be like, well, why is this important? Well, Google's building out all these crazy AI data centers to help house their TPUs and train an info. they're Gemini models. But there's one big issue, Josh. There's not enough power to power these GPUs and TPUs up in the first place. There's never enough power. There's not enough power.

America's electric grid is constrained at the moment. And it is a big, big worry going into 2026, because everyone's spending tens of billions of dollars, hundreds of billions of dollars to create these GPUs. But if we don't have any power to supply them in the first place, then these become kind of defunct pieces of metal that just collect dust. And Satya Nadella earlier this year actually spoke about that saying he has a warehouse that has hundreds of millions of dollars worth of Nvidia GPUs and he can't do anything with that because he can't get the electrical supply to it. So I see this as a very strategic move. What sounds boring on the surface is actually a really important move. And Google obviously sees meaning here and it's kind of making its next steps

towards solving that crisis. Yeah, part of being an AI behemoth now requires owning the entire stack from electron generation through token output. And it's becoming increasingly clear that if you want to be a serious competitor, you must compete across every facet of the spectrum, including actual energy generation, which is what we're seeing from Google. And it's a signal that what we predicted is kind of coming true

where you can't really compete unless you are one of these giant behemists. And the race is now getting shrunken down tighter and tighter and tighter by the people who can actually afford to compete at this level. Google has won, Open AI is one, Microsoft XAI, but there's not that many left. So it seems to me that if you really want to be a big player, you need the money. Google's doing it and they are making big strides in this. And the final topic of this episode, Josh, Daddy has officially spoken. Andre Carpathy has given us

his 2025 LLM year in review. We also did a year in review. That episode is out. You should have to go check it out. But I was really keen to see what Andre would talk about because his history has been across all the major AI model providers. He's worked at Google before. He was actually one of the founding engineers there when it comes to the AI world. He was involved in the research side of things and Open AI as well, right? And so I was curious, like, you know, where does his allegiance lie? And some of his takeaways was super fun. One of my favorite takeaways, Josh, I think you would agree with this. I'm scrolling right to the bottom here, one second, is nano banana. He loved Google's texting.

image and image editing generation model because he sees it as the future of graphic user interface. So if you have to navigate to a website or look at a screen to interact with the internet today, the future of AI will be generative. And he believes you don't have to go to a website. The website will just be custom made and created for you. And he thinks Nano Banana is the first step to. Do you agree with that, Josh? I know you're super excited with this.

Oh, dude, I love Nanobanana. It's one of my favorite model. I think it won my favorite model of the year in 2025. It is a miracle that that thing works. And I love Andre's framing of how he thinks it is most impactful when applied to building user interfaces. And that seems really cool because as we go to this dynamic world where everything is generated in real time, being able to generate really powerful interfaces or just visual

elements in general is so important. And Google, again, is leading the charge. So it's cool to see Andre on board with that and agreeing in a way that is very thoughtful. This is a really good post. I would encourage everyone to go and read the full thing if you get a chance because I really enjoyed it. And that is it, folks. A crazy week for a week where supposedly not much is meant to be happening. You know, we've got a quiet Christmas week. New Year's is around the corner and we have a casual $20 billion acquisition or licensing acquisition if we want to be technical about it before the year's end. I'm excited for 2026. We have an episode releasing soon, which is going to be our predictions for 20206.

Josh and I have thought long and hard about that to the extent where Josh brought out his glasses. so please stay tuned for that episode. It's going to be an important one. And we have some really, really good takes. But aside from that, if you're not subscribed, if you haven't turned on notifications, if you haven't subscribed to our newsletter where we drop Alpha twice a week,

please do all of those things, and we will see you in the next episode. Yeah, it's the, guys, it's the holiday season. If you do anything for us for Christmas, just hit the subscribe button. That'd be pretty cool. We got a great episode that came out early this week. We have the predictions coming out on Friday.

Subscribe, share it with a friend. That's it. We'll see you guys in the next one. Thank you.