Nvidia Invests $3.5B into MediaTek, Wafer Raises $40M Series A, VC Argues for Raising Less for AI

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Tuesday, September 1st. Today on the show, we have exclusive reporting on a new funding round for a founder looking to optimize open source models and also specifically looking to tailor its product to non-NVIDIA chips. We'll bring him on in just a minute. We'll then unpack NVIDIA's MediaTek Convertibles investment totaling$3.5 billion and what it means for NVIDIA's transition from a chip company into what it calls a broader infrastructure provider. And to close out the show, we will dive deep into one venture capitalist view that heavy capital requirements and extreme dilution in deep tech right now are driving the case for building software in the AI era.

1:00It's going to be a great show, so let's get right on into it. The information exclusively reported this morning that Wafer, a company that optimizes open source models, raised$40 million at a$200 million valuation. I want to bring on co-founder and CEO Emilio Andere for a conversation. Emilio, welcome to the show. It's great to have you here. Yeah, thank you so much for having me. Well, congrats on the funding round. I will encourage everyone to read Stephanie Palazzolo's article that came out this morning. It goes into deep detail about what you do, but explain to us, what is the proposal here from Wafer?

1:35Yeah, totally. So Wafer is an AI inference cloud, and the proposal is simply to use AI to optimize AI. I mean, to me, of course, it seems like a very natural idea. If you have a super intelligent AI, you should go ask it to optimize itself in a way, right? So we've been building from the beginning systems so that agents can do the menial work and very complicated work of optimizing GPUs so that LLMs run very, very efficiently on those GPUs. So just to confirm, you're not making wafers, literally speaking. We're not. We're not. We, the name sometimes confuses people, but we're just at the softer layer right now.

2:15It's not very good for SEO, I can tell, or, you know, for AEO, whatever we're using these days to find companies. Googling Wafer is, just makes you hungry. I know, I know. Yeah, you have to put Wafer AI and then. Wafer AI, there you go. So, okay, so optimizing models, and are you, you're focused only on optimizing open weight and open source models? We are fully focused on that. Okay. And does it matter what chip it's running on? Talk about that angle to your business. Yeah, so this is actually one of the coolest things. Obviously, NVIDIA is amazing, and they've done amazing work to make some of the best chips GPUs in the world.

2:54But there's so many other chips that are simply underutilized because the software is so hard to get right to run these LLMs. So one of the things that was very cool to see from our software was when we were able to run LLMs on AMD chips at comparable performance and cost of ownership than NVIDIA. So this was one of the first times that anyone has ever publicly showed that AMD, if you optimize the software correctly, can actually run at similar performance than NVIDIA. So I think we see a future where we're very chip agnostic. NVIDIA is amazing. We'll use a lot of NVIDIA. But I think we'll also use a lot of AMD.

3:39We'll use a lot of, hopefully, TPUs from Google, Tranium, Permatable US, SambaNova, Cerebris. There's so many chips coming along that we're very excited about. So you're saying that somebody can buy AMD chips for cheaper than, let's say, an NVIDIA, like even the Rubin chips, let's say. And using your software, you can get it to be as good as an NVIDIA chip? Exactly, exactly. And have you been able to show that yet? in your testing? Yeah, so we were able to show this with two open source models. The first one was GLM 5.2, where actually it was about 80 to 90 % of the performance as an NVIDIA chip, but it was half the cost, right?

4:23And then there's a second example with Kimi K3, where Kimi K3 was a very popular open source model that came out recently. And we were able to run it on AMD at similar performance than NVIDIA, actually better than B200s. and it was also for around half the cost. When you say almost as good, what unit are we measuring then when you say 80 to 90 % as good? Yeah, we're measuring tokens per second. So how much throughput the chip actually outputs per node of GPUs, which is eight GPUs. Right. Tell me, the way that you're doing this, the way I understand it, so you're actually using agents then to optimize the AI use?

5:07How does that work? Exactly, yeah. The simple way of putting this is AI to optimize AI. But if you dive a bit deeper, basically what we try to do is automate the super grueling and menial work of making LLMs run really fast on chips. So historically, there's been very few of what people call GPU kernel engineers. These are humans that are paid so much money and are paid so much money by the labs, by like Anthropic, OpenAI, NVIDIA, because they're so valuable. There's genuinely probably less than, I don't want to exaggerate, but probably less than like 500 people on earth that are extremely good at doing this type of GPU code writing and kernel writing.

5:50So our thesis from the beginning was, well, agents are actually getting really good at coding. So what if we just specialize them for this domain and then have them be really good at optimizing LLMs for GPUs, writing that code so that it runs really fast on GPUs. And then that has proven to be just a really great tool to accelerate LLMs to make them run as fast as possible on GPUs. So you've gotten acquisition offers, Stephanie reported? Who is approached to you? I, yeah, no comment on that. Would you sell? No, not really. I mean, yeah. For the right price? Everyone has a price. I mean, I guess the thing that we really care about is like what we call maximize intelligence per watt.

6:37I think we're very excited about a future of intelligence too cheap to meter. There's probably some future where we see some alignment with a company that believes the same. But to be honest, like one, we are having a lot of fun doing this and we're just like, yeah, things are going very well. And I think we have a very promising technology in front of us that we've sort of focused on. and yeah i mean you're you're talking you're talking to a bunch of customers i imagine who are considering a very heterogeneous uh suite of chips in their back end i'm curious what are you hearing from customers in terms of how they uh look nvidia is top of their game uh nobody's disputing that but people want different options i mean we've been talking on the show about cerebrus about somanova about these inference focus chips uh tpus are are certainly getting traction so how do you sort of size up the chip market right now and from your view yeah so nvidia is definitely still dominant to be clear like they're they're amazing at what they do and they've been the market leaders for many years now but you do see this sort of gap getting um shorter and shorter into, as you're saying, as to what the other chip providers can do.

7:55So I would say the second biggest sort of closest contender to what NVIDIA can give you today is AMD. They've also invested a ton of resources into making their chip good, but there's still some gap between what they can deliver out of the box and what NVIDIA can deliver out of the box. I think there's also very exciting offerings that are sort of taking a completely different angle than what NVIDIA is doing. So you can see Cerebras where they do the wafer scale chip. You can do Samba Nova where they do the reconfigurable data flow. There's all these different ways to attack the market for saying, hey, maybe I'm not as good as a GPU in all these general case scenarios, but I can be very specialized in these particular directions.

8:41The great example of Cerebras is like ultra low latency. So I'm particularly very excited about combining chips in different ways that are interesting. You see a lot of people combining AMD plus Cerebris, NVIDIA plus Cerebris. You saw NVIDIA by Grok and combining it in a single system to run different parts of LLMs. I'm very, very excited about the heterogeneous world that will become in the next couple of years of GPUs. And also just of people having more choice to choose their chips. Let me ask you one more question about the agent landscape broadly, since that is where your technology is focused.

9:19We saw this week that OpenClaw released OpenClaw 2.0, and I was texting a couple people about what their thoughts were. Is this a big deal anymore? We've seen companies introduce their own versions of OpenClaw to help people create agents. It feels like there certainly wasn't as much buzz around 2.0 as there was around the initial OpenClaw. have you uh have you put any thought at all into 2.0 and whether or not open claw has kind of fallen out of favor with people have you tried it at all what's your view there yeah so um not that much into open claw uh to be honest but i do use a lot of open claw like agents in my daily life like for example a shout out to instinct who's obviously amazing um been using them what's the what is the whole what's the give us a story here on instinct i keep we haven't talked about on the show yet and and that's probably that's on me we should be talking about it more yeah you guys should be they're awesome so they're basically an open claw thing that you just text on your phone and it can connect to all your sources now i'm just now i'm just promoting instinct um i have no affiliation to instinct um but you can use them you can use them on your phone you can text them like you're texting a friend or a co-worker and they connect to all your sources and they can just do things for you.

10:38So you can say, hey, find me a reservation every Saturday for me and this person. You can do all these complicated agentic workflows that OpenClaw supports, but it's more geared towards people that don't want to set that up and don't have the time or maybe even technical ability to set OpenClaw up. It's still pretty hard to set up OpenClaw for most people in the world, right? So yeah, pretty excited about the form factor of like instinct-like tools where people don't have to do much to set it up. and it's a pretty intuitive experience. And do you think that Instinct will, the same way that we have sort of stopped talking about OpenClaw and OpenClaw 2.0 has basically gone unnoticed, I would say largely on the timelines if you look at it.

11:21I mean, do you have any more confidence that Instinct will be any more sustainable than OpenClaw as a tool? Yeah, I think OpenClaw was interesting and very different from Instinct because OpenClaw was open source And it was also like the creator was bought by OpenAI to do internal projects right there. Right. So I think instinct, I don't, I don't know them very well. Again, no affiliation, but my understanding is that they are a, like they're a startup whose sole purpose is to make this happen for the world. So I'm way more excited about the durability of something like that. I, yeah, hopefully they, they do stay.

11:54I think the tool is great. Okay. And last question for you. Yeah. OpenAI's jalapeno chip. I saw that you had some thoughts on that. What's your view? Yeah, I mean, it's interesting because it goes kind of against what I just said of chips being separated for different parts of the LLM inference. They basically just decided to put it all in the same chip, similar to how GPUs used to do it. And they're actually showing amazing, like really good performance graphs compared to NVIDIA's latest flagships and AMD's latest flagships. So first of all, like the quick things are, first of all, insane performance by the OpenAI team and the fact that like they were able to stand up the hardware, I think it was in like nine months, some absurdly fast number.

12:40And then two, it seems like they actually got a performance chip in one of their first iterations of the chip, which is just all just, yeah, the overall impression is very impressive effort by the by the OpenAI team and very excited by it. Great. Well, Emilio, I want to thank you for coming on. That is Emilio Andere, the co-founder and CEO of Wafer here on TITV. NVIDIA invested$3.5 billion in convertible bonds issued by Taiwanese chipmaker MediaTek. For more on this, I want to bring on Mustafa Nimuchwala, partner at New Enterprise Associates. Mustafa, welcome to the show. It's great to have you back.

13:18Gosh, thank you for having me. So the way I understand it, MediaTek competes with Broadcom on making custom ASICs. Is that right? In the context of Google's TPU, the answer is yes. MediaTek has been doing the inference version of the TPU, the last iteration, and Broadcom is doing the training version. But MediaTek, as you know, the semi-industry is very intricately tied together. So MediaTek, when it comes to ARM CPUs, which is what we're going to talk about here, can compete with Intel, can compete with Qualcomm when it comes to autos and other systems on chip opportunities. And so quite intricately tied.

13:59But yes, I think in the context of a lot of the accelerators that we talk about, MediaTek gets in that design category where they can license their IP for folks in new domains. Okay, so you said ARM CPUs. So is CPUs the reason why NVIDIA is so interested in media tech, or what's the calculus here on their end? Yeah, there's a couple of things. So first, it is ARM-based PC CPUs, right? So NVIDIA announced their own Vira CPU, which is their server CPU, and they announced a tie-up. $5 billion they put into Intel last year, and that was for x86, which is Intel's architecture-based CPUs, but also for laptops and PCs.

14:39And now they're also doing the ARM version, Again, for the laptop and kind of that domain specifically with MediaTek. And so it's kind of an all of the above approach. NVIDIA wants to be everywhere that compute happens. And then there's also a bidirectional thing here where earlier, you know, NVIDIA had announced their NVLink Fusion partners and MediaTek was one of them. Which is NVLink Fusion. What is that? Yeah, I think the simple way to think about it is interconnect across the rack. And maybe the kind of thing to think about in this context is MediaTek is doing TPU v8 for inference, and other folks are going to build other, you know, we call XPUs or these custom accelerators.

15:17And NVIDIA wants to have NVIDIA interconnect across the rack and across a lot of these data center designs, no matter whose accelerators are being used. Is that the same as networking chips? Yeah, precisely. Interconnect is one of the domains of networking. So this is the context of rack, let's call it almost rack level. Right, right. Because IP as well on the chip to chip side, et cetera. Because correct me if I'm wrong, I mean, NVIDIA has been getting deeper into the networking chip business. And I mean, you know, rival number one there is Broadcom, which is a leader in networking chip. And so I sort of see this as a way to maybe build out competition against Broadcom.

16:03MediaTek, of course, is sort of a, they're coming in as a lower cost competitor to Broadcom on a separate side of the business. But I sort of see this as a way of getting at Broadcom. Do you agree with that? I would say everybody is kind of getting at each other. But NVIDIA's mindset will continue to be that they're going to own everything in the system. And so that includes certainly the networking. That includes certainly Certis, which is where a lot of these original design manufacturing companies have IP. For example, MediaTek has Surtease IP, Marvell, Credo, of course, Broadcom. And so candidly, I think in some sense, you could almost argue that Broadcom already has a lot of other folks competing against it in that business.

16:42NVIDIA is trying to almost think about its own future and how can it prevent Broadcom from competing against it, if that makes sense. Because NVIDIA is the lion in the ecosystem. Even Broadcom, even the great Broadcom is a cheetah, maybe a tiger, depending on the context. Okay. Okay. Tell me, is this another data point to the circular financing risks that everyone has talked about? I think in this case, it's harder to make that argument because it's, you know, I think to me, the clearest circular financing arguments are when one person is another person's customer and vendor, and you're paying, you know, the vendor is paying the customer to buy its products.

17:22This is, I think, more of an ecosystem play where I think this is a partner, financing a partner. and also I think NVIDIA, I think we can clearly see this in the ecosystem today. NVIDIA is becoming the primary financial investor in this ecosystem as well. And so I think a lot of these deals they're doing, cannibally, have a financial angle to them. Like you look at this deal, structure, it's a convertible note. Downside protected, and then the strike normally in a convertible note would be above where the stock price is today. So they get the upside if there's option value as MediaTek does well.

17:52Maybe even for that matter from Google's TPU inference product. I mean, that's crazy if you think about it, because now NVIDIA may make money from its own competitors launching their inference product into the market. So I personally... Explain that a little bit slower, I guess, for me and for everyone else. Yeah, happy to. So the convertible node structure is essentially a form of debt, right? It's a note first. And that note gives them downside protection. So let's say MediaTek, for some reason, were to go bankrupt. NVIDIA would have a debt-like instrument that it can capture value from. And then on top of that, convertible notes, the convertible portion of it tends to be like, you can think of it like a warrant or an option that's tied above.

18:31And the strike normally is above, the strike meaning the price that convertible note converts into equity is usually above where the current stock price is. So effectively, the upside happens if the stock probably meaningfully goes up. And right now, we're looking at a world where, you know, what they call implied volatility or the upside stock appreciation and media tech has been gigantic this year, you know, up 200 % this year. And so in that sense, point number one, NVIDIA is doing these structured deals in a very smart way. If you were really talking about it purely like, you know, quote unquote, dumb circular financing where you're just paying somebody to buy back from you, you would care much less about the structure.

19:08and the customer would be the other counterparty would be much less willing to take debt on in that kind of equivalent context. So that's point number one. Point number two is MediaTek is a part of the broader semiconductor compute ecosystem, including like we discussed within the Google TPU ecosystem, which TPU and certainly Tranium, but those two are probably the largest, you know, alternative accelerators to NVIDIA that are already going to market. And TPU obviously has the largest, broadest team behind it. So now MediaTek, which is taking on the inference share and there's rumors around what happens with the next iteration and rumors around meta, becoming a customer, et cetera, as well.

19:44NVIDIA can now benefit from that financially. But also you have to assume NVIDIA gets a little bit more say in the decisions that are made by being more intricately tied in. And then, of course, having NVLink Fusion, this interconnect, baked into more rack designs that MediaTek thinks about. And then, of course, there's the whole CPU thing we talked about. Probably will play out to cars and autos over time, where NVIDIA can get a room there as well. So it's just NVIDIA getting its tentacles at EPR, but also getting an opportunity to get money no matter what happens. Right. Well, and so I hear your point that this is NVIDIA in some ways benefiting from the success of TPUs should MediaTek make more chips with Google in that way.

20:25I mean, this sounds a little bit conflict of interest. I feel like NVIDIA won't have this. They're not going to have a say into whether or not MediaTek does business with Google, but they can certainly benefit from it. I certainly see that. Tell me, in this ecosystem play then, my understanding, though, was that the NVLink Fusion, so is NVIDIA not paying MediaTek anything for technology to help build the NVLink Fusion chips? Is that not the way it's working? NVIDIA will pay for sure. NVIDIA has been paying MediaTek for this ARM CPU design. So the way it works, right, is NVIDIA is going to bring...

21:05Right, but it's not MediaTek buying from NVIDIA. It's not circular in that way. Exactly. It'll be the end customer, you know, the laptop assemblers or wherever the CPU is going, who'll be paying for it. And these guys are essentially partners. I think that's a very unique part of this. Same only with Intel. With Intel, there's a bit of a more of a customer relationship of NVIDIA fabs on Intel's Foundry business, which is their TSMC competitor is coming up. But in this context, we're talking about the Fabulous slash. Also, this gives NVIDIA more space at TSMC. Not that it needs that. It already has a lot of capacity at TSMC, but MediaTek has its own capacity as well.

21:40Yeah, but that's the big scarcity right now is how we produce more. Precisely, right? TSMC advanced packaging. I also think the other analog I would drive a little bit on this is kind of like what happened with Anthropic where Google and Amazon, who are theoretically competitors, both invested. I think similarly we're seeing here. MediaTek, obviously, big Google partner and has been, but now monetarily a big NVIDIA partner. And I think you just expect to see more of these tie-ups in the polyamorous AI compute semiconductor industry that we live in. It's almost like that should be taken for default, that if you're not polygamous, then somebody else will enter into that polygamy relationship.

22:14And you may not benefit financially either. And of course, Google and Amazon have both benefited massively from their anthropic investments. Right. Great. All right. Well, Mustafa, I want to thank you for coming on. That is Mustafa Nimuchwala, partner at New Enterprise Associates here on TI TV. Funding rounds are getting bigger and bigger by the week, it seems, as venture capitalists pour not just into deep tech startups, but also into Neo Labs and Neo Clouds with astronomical valuations. One investor suggests plain old software is one of the only investable asset classes right now. I want to bring on Finn Barnes, co-founder of the General Partnership to walk us through his thoughts.

22:53Finn, welcome back to the show. It's great to have you here. Yeah, thanks. Thanks for having me. Well, last time you were on the show, you got Corey, but this time you get me. It's an upgrade. Corey is, you know, he's, Corey is, I think he's running, what, 10 years at the information now? I mean, he's been here since the start almost, so I would hardly call it an upgrade. But I want to talk about a piece that you wrote. So the title of the piece is Raise Less and Build Software. Why did you decide to write this piece? I feel like there's never been a better time to build software. I think we've gone through a tremendous period of innovation around intelligence.

23:35And we've seen the scaling of the models. There continue to be bottlenecks in the supply chain, as you were just talking about with the previous guest. And when the hardware and the supply chain to support software is constrained, I think it's a great time to take advantage of building software that is more efficient, that leverages intelligence in new ways, and makes it available to the broadest audience possible so we can see the full impact of this innovation across the economy. So are you only investing in application layer companies, in software companies only? What's your playbook here? Yeah, so currently that is my focus.

24:16I think as a firm, we've always looked for capital efficient businesses, people who understand how to maximize the learning per dollar spent at a startup. So they have a thesis, they have a belief, they want to test that thesis and we can fund them to do that. I think sometimes those are larger seed rounds. I think there's nothing wrong with a large round as long as you're maximizing sort of the efficiency of that spend and learning per dollar spent. But I think today, the application layer represents a tremendous opportunity. You know, everyone is talking about agents and harnesses, and these are all just words for software.

24:50Right. Tell me, the NeoCloud business model then, what's your view on it? And I'm, of course, pointing to a post from Nikesh Arora and Palo Alto Networks. He got a lot of buzz around him basically saying that once the supply demand economics sort of neutralize out, you're going to see much lower valuations for neoclouds. Is that directly attached to what you're saying here? What do you think of the neocloud business model? Yeah, I mean, I think it's a very different piece of the market, obviously, than the software piece. Neoclouds having value depends on people having demand for what they do.

25:28I think that demand will be rising, whether it's met through hyperscalers and you see some consolidation or whether you continue to see neoclouds, specialized clouds focused on certain workload deliveries. I don't know and maybe over dinner I could argue with the cash about this, but I think the point around neoclouds is there is a bottleneck today. They're taking advantage of that. Whether there's lock-in in those workloads, I don't know. What I do know is when you build software to help people get value from intelligence and you understand their workflows, their data, and you can leverage that in a way that's unique, making their business better, then I think regardless of where the bottleneck is in the ecosystem, you'll have a valuable business.

26:12Right. Can you think of a time in the, you know, since you started doing venture capital, Can you think of any historical parallels here or moments when investing in these capitally intensive businesses or in deep tech was as popular as it was now? Because we sort of came from the SaaS honeymoon period where everyone is investing in software companies. We've since migrated to chips and infrastructure, stuff like that. Can we learn anything from history here on this? I think there's been ways. I think you could look at the initial clean tech bubble, I think probably the initial semi-buildout as the internet was starting to happen.

26:56And you could probably see some things that rhyme with what we're seeing today, both good and bad, risky and certain to build tremendous enterprise value over a very long period of time. But I think when I look at what's happening today, the biggest thing I see in terms of a shift that's different than historically is the venture capital firms themselves, I think, have a very different model in terms of the scale of the firms, the fact that many of the mega firms are effectively index funds across categories. And so the decision that folks are making in this sort of new asset management approach to venture capital, they have quotas in terms of capital deployed.

27:35And so when a new category or a new company gets funded, the decision inside every other partner meeting is, is it easier to mark that company up and follow quickly, or is it easier to find a competitor? And so I think that shift actually is in many ways sort of tail wagging the dog when you look at kind of the expansion of these rounds and the ability for people to aggregate tremendous amounts of capital to pursue ideas sort of up and down the stack. And I think that's a big driver there. And we do see already, I think Sarah Rao just had a tweet the other day about sort of the pivots that she's seeing in the AI space.