Dylan Patel on Google's TPU going external: how it's already saving OpenAI 30% on Nvidia spend before deploying a single chip

and that's why uh organizations like semi analysis exist. And I believe we have Dylan Patel from Semi analysis in the re waiting room. Let's bring him in. Dylan, how are you doing?

I'm doing fantastic. How about yourself?

You know, I saw I saw the meme image that you guys uh put out there for me, so I had to wear the tank. Let's go,

dude. We need a bigger bigger uh screen for that bicep. We'll we'll work on it.

Let's go. Uh, where in the world are you?

I'm in Florida. I was spending Thanksgiving with my family here. Um, I'm trying to chill out a little bit. It's nice to have the family pamper me a little bit because I broke my foot a couple weeks ago.

I'm sorry. Do that.

How'd you break your

tripped over at TPU?

Family family reunion playing football in Texas. We're American as we can get.

There you go. There you go. Um, well, uh, we were just running through a little bit of the the TPU article. Can you uh can you actually set the table for me on like what do you think is new about it versus what has semi- analysis already been saying and this is more just like tying everything in a bow.

Yeah. Half of the article is just referencing recent

we've been saying this for two years. We've been saying this for one year

and even referencing con Google's own content about the TPU dating back

even even further. So

yeah, I would say I would say the majority of this piece was if you're if you're a client, it's already been pretty much all published. Um but it hasn't been tied together. It hasn't had a narrative around it, right? Cuz when we think about like what we put out as on the paid side versus what we put out on the newsletter, right? Um our clients sort of get uh you know what changed, what happened, uh here's the numbers. Um that's about it, right? We don't explain the technology that much because our clients are sophisticated, right? they're either in the industry or uh and and or they're finance pros who don't give a [ __ ] about the technical stuff. Um and so it's it's either of those two, right? Um and and so we we're just explaining here's what's happening. Here's the change. Here's the numbers, right? So for months we've been saying Google's selling TPUs. For months we've been saying, hey, here's TPUv7 versus Blackwell. Um we've even put out updates on here's what we think TPUv8 is versus what we think Ruben is. And so generally it was making it into a narrative and explaining the technology and the corporate I would say politics or uh dynamicism around it right so that's you know I think I think there has been bits and pieces put out by other other folks right I think the information has done great reporting on some of the stuff after we did but in the public space I think um you know for as an example right like um so so other people have put out bits and pieces surrounding this but they haven't put out the full picture um so so as far as like what's new it depends on where you sit in the stack. Uh but but you know Anthropic and and and Meta and folks like that have been talking to Google about buying TPUs for many months, right?

Um whereas whereas people externally are you know last week when Gemini 3 was launched or two weeks ago, people were just learning that TPUs are trained training Google's models, right? So it's it's where are you in the information spectrum, right?

Yeah, totally. So, uh, on that information spectrum, uh, the finance bros, they can probably just like if they if they read into this, oh, bullish Google or bearish Nvidia or whatever, like they can kind of trade in and out as as they please. But what on the on the more technical side, like are people using semi- analysis research to understand like, okay, I'm a NeoCloud. What do I want to rack for next year? Maybe I need to be putting in a TPU order. Is that is that how people interpret your research? like what happens on the technical side of the house.

Yeah. So, so as far as like some of the paid stuff we do, we have one model called the TCO model, right? Which is uh calculating the TCO of all these uh different hardware performance um building up the entire cluster cost, you know, breaking out into like a dozen plus different things whether it's storage or networking and breaking down the cost of everything. So there we put out research on TPUs because as soon as NeoCloud started getting offered, hey, you want to buy TPUs? Yeah.

We're like, okay, we need our own groundup model. So when you're negotiating a big contract, what you do is called a should cost, right? You you go and calculate what it costs for the company versus what it costs for me to deploy. Um and then you like think about like, oh, what's the margins they have? What is ridiculous to offer them? What is not, right? Because everyone always wants to know like, hey, what margin are they making off of me? Can I push that down a little bit? Um what is what is ridiculous to demand in a negotiation versus what's not? So we've already been working, you know, through this TCO model. We've put out four different updates on the TCO of uh TPUs V7 and V8 because there are NeoClouds out there um as well as labs who are purchasing TPUs that are using that to understand what's the cost now you know Enthropic I will say just already knew and figured it out because they've hired so many Google people but other labs are also looking at it right yeah

um and so so you know when you when you say hey on on the cost side of things on the technical side of things right there's a lot of network engineers now out there who have never deployed Google hardware that are now like okay I need to figure out how to do text, right? Like, you know, so there's there's people who have DM'd me that are like, "Oh, we've been, you know, as you know, we've been thinking about deploying Neil clouds, but your material on this is technically better and teaches me more than Google's own material, right?" So, it's like this is this is helpful to people on multiple factors.

Yeah. What what about uh the software side? Uh Google's built their own internal stack to compete with CUDA. Uh how much of that are they going to actually give to their customers who are buying TPU? because that feels like you it feels like potentially you could overrotate on oh well Gemini 3 is really good but why is it good is it just because of the hardware or is it also Google's incredible prowess multi- data center training all this fancy stuff that they have that they won't be giving you when they sell you the TPU

yeah so so that's the interesting thing is some of the stuff software will remain closed source but you can still use it right and then some of the software um they are trying to open source aggressively and then some of the software they're never going to get out there anywhere, right? So, it sits in three kind of buckets, right? Um the interesting I guess newer thing that we did in the piece was um we looked across all these different open source AI uh software, right? Whether it's PyTorch, whether it's VLM, whether you know, all these different um open source uh libraries and we calculated and counted up how many Google uh commits there were, right? And you can see there's a chart in the article where the number of commits that Google's doing on TPUs has exploded over the last handful of months, right? as they've decided to shift their strategy, sell GPUs externally. They also recognize software has to be open for this, right? Um, you know, only the gigab brains that like anthropic can figure out how to do everything themselves, right? It's it's those people outside of anthropic, you know, types that that need a bunch of open source software that builds on top of it, right? Um, and what's interesting is when you look at like, hey, Nvidia, you know, the biggest argument that Nvidia doesn't really make for GPUs, but they should, is that, you know, about 40% of the software that's open sourced is actually just from China, right, on on CUDA, and that's the CUDA mode, right? It's like 40% of the software is just like open source stuff, whether it's people committing to VLM or PyTorch or or all these other libraries, right? Byte dance open sourcing stuff, DeepSec open sourcing stuff. Um and and and and and Google, you know, they they don't have people open, you know, Anthropic is not going to open source software. So Google needs to catch up not just by, hey, here's all the software we have internally, let's open source it. They also need the ecosystem to build a ton of software on top of TPUs. And so that's the that's the real big uh challenge there. Um and and there's an element of software there that Nvidia's happy to open source. Um and customers of Nvidia are happy to open source that Google will never open source because it's it's you know, Google Cloud is selling the TPU. Gemini is the one actually using it and developing a lot of the software and these two groups are not always going to be aligned.

Yeah. Isn't that uh like I mean what are the other kind of uh just problems with uh Google becoming an actual like seller of TPU? It feels like uh there's obviously an opportunity because Nvidia has high margins, there's demand, it's a great chip, but culturally, structurally like Google tries a lot of different things. they have a lot of advantages, but occasionally like they fall flat on their face with just like they can't even get an RSS reader out or something like that. Uh, so like like are there other risks to the TPU not really finding its footing for reasons that aren't just the the laws of physics?

Yeah. So, so the biggest challenge I see with them is it's everything is non-standard, right? Google for years, they they developed liquid cooling first, right? Sure. uh for for AI computing. They deployed rack scale architectures, right? Everyone's talking about GB200 rack scale architecture. Google did it first with TPUs, right? But when they did all of this stuff, they didn't give a crap about, hey, you know, this has to go in 50, 100,000 different people's data centers, right?

Um this has to go in my data centers that I designed myself. So everything is super vertical. The entire liquid cooling supply chain is super vertical. Entire the the racks aren't even the standard uh width, right? So when I look at like a data center, it's like the door the loading bays because they're so much wider. The Google racks are like three times as wide. It's like maybe it might not even fit into the data center like physically like through the doors. Um so there's like all sorts of random like I wouldn't say random, it's Google from first principles design stuff.

Totally. Yeah. Yeah. Yeah. But but if you're Neo cloud and you're like the hot thing is going to be TPU next year or the year after and I want to be able to sell into that market. Uh it's not just flip a switch, drop in, replace with TPU. you have to maybe build a whole new building like it might be that significant.

Let

right or or do or or like knock down some walls and then like you know I need to I need to go get liquid cooling not from Dell and Super Micro and and HPE who I've who service me already. I need to go get it from some random supplier who's only ever sold to Google. And usually they're sitting across the table from like some gigabrain engineer who has a team of 20 people working on liquid cooling instead of like you know my one guy who does liquid cooling procurement and negotiations and like also does procurement of like network stuff. Yep. Yep.

There there was a there was a tinfoil hat theory floating around that Meta leaked their TPU in Meta leaked their TPU interest uh to try to gain some ne uh sort of leverage over maybe uh some negotiations with Nvidia. I don't know if you uh see any possibility in that. But how do you think those conversations are going? uh Jensen doesn't want to discount uh and compress his margins, but at the same time, he can't do this kind of like equity rebate thing. If he if he took a big position in in Meta, he'd be very suspicious. I'd be very concerned.

I totally get the opening eye investment. That seems like it makes much much more sense than saying, "Hey, we're going long Meta, you know, is a $4 trillion company."

Yeah. At the end of the day, right, like TPUs have like a set of maybe 10 customers, right? Because you have to be super sophisticated. Yeah. Um, and so what really is challenging here is is, you know, Meta Meta looks at the numbers, you know, it's like, okay, open I'm getting 30% off because they're paying they're investing in me as a result. Obviously, they get equity, but they're investing in me and I get 30% off on these GPUs as a result, right? Meta, you can't do that. So, so Meta, they're I don't think that they're just negotiating, right? Like, you know, is is are they just negotiating with Nvidia when they buy AMD? No, they're any engineers, right? They're developing all the software. they're actually deploying uh Llama 405B was exclusively on AMD for a number of months, right, for inference, right? Um so so when you look across, hey, is Meta just like playing around trying to negotiate? It's like no, no, no. Like they're they're looking out for what is best, right? And Meta is power constrained and TPUs are currently way more power efficient. Meta is compute constraint. Um and TPUs are potentially higher performance per watt and higher performance per dollar, right? At least that's what we believe for TPUv7 that it is. So, they'd be dumb not to look at it, right? And they have the time, they have the people, they have the team. Now, Nvidia at the same time has to play the game of chicken, right? Yeah. Sure, they could discount the pricing somewhat. Um, and because what's funny is Nvidia is more vertically in integrated than Google is when selling hardware, right? Google has to pay Broadcom who pays TSMC, whereas Nvidia gets to pay TSMC directly, right? There's this vertical integration challenge where Nvidia could drop the price a little bit and they'll be fine, but they don't want to, right? You know, the whole point is you charge the highest price possible. And then the last thing is they've got this like um you know they've got this view about antitrust right you you don't want to cut deals for specific customers because that looks bad right instead you want you know right now Dell pays the same price for a GPU as Gigabyte as meta

now the networking hardware there's different pricing because there's a lot more competition and Nvidia can cut a lot more there but on the GPUs themselves Nvidia's pricing is very fair right fair in the sense that they're making a shitload of money off of everyone you Yeah.

Uh how talk about kind of Jensen's leverage uh that he has around Reuben allocations as some of these customers start to uh at least consider TPUs.

Yeah. So as far as like next year's TPU deployments, it's pretty set in stone for the vast majority of the volume, right? Anthropics got a bunch and then there's some sprinkled elsewhere. Uh but as we go into 2028 where Google can actually ramp um you know the flip side is Ruben is also ramping and at least based on our research looking throughout the supply chain um you know over a year ago when open started their trip team they poached like 15 Google people overnight right in one week like someone I knew I heard was like oh yeah I'm joining open I text like another three people I know and they're like oh yeah I'm also joining open like what the [ __ ] um so so Google's a lot of their best TPU engineers have left right they also have a ton left and so that what that's done is, you know, chip timelines are so long, that didn't affect TPV7, that's affecting TPV8. At the same time, Google's trying to diversify their supply chain, get from not just Broadcom, but also MediaTek. And so, Google's got a real challenge on TPUv8 in that uh it's good. It's an improvement, but then when you go look at what Nvidia is doing with Reuben, Reuben is so much better because Nvidia is just pedal to the floor, paranoid as [ __ ] Uh we we have to be the best and we have to be way way way better than everything because how much better I am than everyone else is my margin, right? And so Nvidia, Nvidia has the sort of like at least currently we think Nvidia is going to be so much better that they'll be fine and they'll be able to maintain margins. Right now things can happen. Reuben can delay or TPUs can delay and the position looks better or worse, right? Um there's a lot of unknowns to go through,

but as far as like what is Jensen's leverages, look, I'm going to make the best hardware and plus my software advantages and I'll be able to continue to be dominant and dominate the market, right? Um there's there's curveballs I could go which is like oh Google software they could open source enough software that actually their software ecosystem is not far behind Nvidia maybe they don't want to right or hey they could um you know they could execute everything and Nvidia has a three sixmonth delay now all of a sudden they're a lot more competitive right um and and so all these things are still open questions but it's it's Nvidia can play the allocation game as well of course right um hey I'm going to give all of the GPUs initially to companies that probably could buy uh TPUs but that ends up all the AI labs and hyperscalers, right? Um at least, you know, like Meta, right? And bite dance, people that would actually be willing to buy TPUs. Um and then you end up with this like weird situation where okay, well that's like 75% of the GPU market anyways when I look at the the AI labs through the NeoClouds, right? Um when there's, you know, Nebus and Iris Energy and all these other, you know, Coreweave and all these folks are buy are deploying for OpenAI anyways, right? Um, you know, this this sort of ends up being like, well, sure, I could stiff like some people on the allocation, but at the end of the day, everyone who was a potential customer for TPUs uh is sophisticated enough to be where they were going to be on the beginning of the allocation anyways,

right?

What what uh how are you framing ClusterMax these days? Is it is it for customers who want to buy services from Neoclouds? Is that the primary goal of cluster max? because I feel like some people look at it and they're like, "This is a buy rating. This is a sell rating on the stock."

So, so the funniest thing is like Cluster Max V1, the title of it was Cluster Max, how to rent a GPU, right? Because we discussed all of that and then and then in Cluster Max V1, I believe we put Irish energy and underperform, right? At the same time, the research side of the business, uh, we explicitly were like, "Dude, they've got these data centers. It doesn't matter if they suck at running GPUs. They've got these data centers. They've got this power. if you just value them on a watts per you know how much money they could make it's it's a long so like at the same time as like um Jordan right Jordan he's running cluster max is like Iris kind of sucks uh and it was other people in the technical team before him you know it's like it's like Jeremy who's running the data center side and I think he's been on TVPN is like dude Iris energy is a stock right so it's like it's it's kind of like you know it's like what what what the technical side of the house does versus what the you know research side of the house does yes they talk to each other right Jeremy did ask the team like, "Hey, what do you think of Iris Energy? I think it's a log." And and the team working on Cluster Max is like, "I don't know." Like, you know, it's it's it's a bad cloud. And it's like that doesn't matter. So, ClusterMax has nothing to do with the stock, right? Um, now obviously there's going to be some correlation with how good is a stock versus, you know, who's going to want to rent from them. Yeah. Um but at the end of the day right like cluster max is the the goal purpose sole purpose and what we explicitly say in there is it's for people renting anywhere from like you know hundreds of GPUs to you know right below the AI lab scale right the AI lab scale there's different considerations um but in that range tens of thousands of GPUs all the way down to hundreds of GPUs that's who we're targeting plus we're saying we're giving a bunch of feedback for people to make the cloud ecosystem better

the unsung hero between cluster max v1 and v2 is that we move the bar up Right? You know what it required to be in gold

like was was much more. What it required to be in silver was much more because everyone improved so much, right? And as as we continue to like increase the requirements, make it harder and harder to

got keep moving the goalpost,

right? People keep improving the ecosystem and actually, you know, this is this is the funny thing. It's like cluster max is evil. It's like when I when we look at the quotes and we've got hundreds of quotes on clustermax.ai, all these companies are like, "Dude, I love this. this one specific bug that this Neocloud had, they fixed it as soon as you wrote about it. Yeah. Right. Or like, hey, help me understand the reliability, help me understand this or that. People are like, "Love clusterbacks." Um, and and you know, altruistically, like I think we're generating billions of dollars in value just from hey, like all these clouds are more efficient and there's less failures and it's easier to get your workload running on any random GPU cloud and the market is more efficient. Um, now I'm not making any money off of that. How am I making money off of cluster max? I'll be very clear is people who hire us to do due diligence, right? So, people who want to acquire a NeoCloud, people who want to sign a massive massive deal that's not just like thousands of GPUs, but tens of thousands of GPUs, and then lastly, it's people who want to um you know, invest in NeoCloud. Those are the three areas where we're making money off of quote unquote cluster max, but not really. We're not selling ratings. We're not, you know, you know, we're in fact like a customer will do a consulting project with us or want to want to buy some research from us and I'll explicitly put in our Slack share Slack or I'll send an email to the CEO like, "Dude, just so you know, the people working on this are not the people who are doing cluster max rating, right? Um, you know, the people who are buying, you know, the research on like these data centers are there and this is the power ramp or here's the accelerators or here's the TCO. That's not the people doing cluster max, right? And I don't care about, you know, whether you buy it or not. I you know at the end of the day Google and Amazon and Microsoft are way bigger customer than you know Flipstack and like you know those kind of companies right and yet one some of those are ranked in silver and some of those are ranked in platinum and gold and that's because technically what's what matters not hey you know obviously when we talk about who buys our research the biggest companies in the world are going to pay me more than the mid-size companies in the world. Okay, question from the chat.

And the price has discriminated based on that.

Uh, would you change the rating of a Neocloud if Shalto promised to do the dishes for two weeks straight?

You know, there was an argument. I saw someone was like, "Who does the chores?" And it's like, "Brother, we we live together by choice, you know, we we pay someone to come once a week. If you cook something, you do your own dishes." But like, you know, um, frankly, we're we're working so much and I think like,

you know, I think I think Dores Ces has ordered pizza from the same spot three nights in a row before, right? Like it's it's it's Is being an adult man with roommates underrated.

Um, so I'm I haven't lived with people in years and then when I moved to SF.

So you came back.

This is crazy. I moved with I moved to SF uh this year, you know, I'm like, "Oh, you know, I should live with friends just so it's more fun." Um, and the first house kind of fell apart, so I moved into this house with these guys and we've been talking about it for months. Um, I love it, right? It's like, look, we we we we have, you know, if you think about, oh, what if we all rented our own places that were good and then we pulled that budget together? We have a nice place. Yeah. Right. And then in that place, we have plenty of space for ourselves. We we pay for someone to come and clean once a week, right? So, at the end of the day, what is what is the negative here? Is like,

well, we're living with our friends, but we have enough space to where like

and the beauty is if you if you do bunk beds, you have more room for activities.

Exactly.

Anyway, no, no, sorry, sorry. Actual question from the chat. Uh, when is TPU going on inference max? We got to know.

So, we're working on it, right? We we're we're working with Google um technical folks. Um, you know, funnily enough, actually, um, we triggered a security warning for this Google engineer. Uh, Kimbo went to a a Jax conference, right? Jax is is the opposite. It's like PyTorch for but for TPUs. It's the most simple. It's it's Google's own internal thing, right? Um, that people do use externally. Um, he went to this PyTorch or this Jax conference. A Google engineer presented something. He's like, can I get the slides? They send it to him and then Google security alert like locks him out of his computer because he sent us like a some technical like information. said like for 3 days the guy can't work and he's freaking the [ __ ] out and I'm like I I emailed Jeff Dean. I'm like bro this is like do not fire this guy. He sent me stuff that you presented at a public conference and he's like oh okay yeah yeah I'll get that fixed but anyways like um we're working with we're trying to you know implement it. Uh we have access to some TPUs.

The software stack is different right? Yeah.

You know just

so you have to so you basically have to rewrite or reimplement inference max like like the code that actually

I won't say I won't say it's that much work like as much as like completely redoing inference max but there's a ton of work. Right. So we're moving as fast as we can internal target is this year. Um you know then the obvious question is is like I I feel like inference max is my north star for TCO relative in AMD versus Nvidia land. Uh there was a bar chart of TCO for GPU versus Nvidia. It looked like it looked like TPU was doing really well on that chart. The bars were very low. Where did those numbers come from? Do you have confidence in those numbers or do you think the numbers will change once you actually get TPU on Inference Max?

Yeah. So, Inference Max shows performance TCO, right? You know, it's great. Great. Like, you know, like guess what? Like, um, you know, TCO of like a Raspberry Pi is incredible. It's like five bucks, right? Um, you know, versus versus a GPU is $50,000. Performance divided by TCO is what matters. So that bar chart is saying look TPUs are cheaper and at least on quoted specs you know now now let's make some assumptions around utilization and in the in the article we explicitly said look we don't know what the utilization is it's going to change customer to customer um here's a range worst case it's like a little bit worse than GPUs best case it's way better than GPUs right um and and so inference max will tell us what the actual performance is in inference um because we don't know yet right um currently the open source software for TPUs is not good enough for good enough for us to just take the open source software and say that's the performance, right? Because that's obviously like not real, right? Anyone who like is actually buying TPUs is going to spend engineering hours to work on it. And so we're trying to work with Google to get a real performance number that is achievable by people. Um, you know, and and will be upstreamed into the open source software because this is an in progress thing, right? No one cares what TPV7 can do today. It's about what it does in six months. Um, and so, you know, obviously we don't want to be, you know, today TPUs, if you're using VLM are worse performance TCO than GPUs without a doubt, but the target is moving very fast. And, you know, there's a ton of like lowhanging fruit for us to implement before we actually put a number out there, right? Um, and so where does Google sit there? We'll see. Um, I I personally believe the TCO side of things, the total cost of ownership is based on what we know on supply chain, right? How much do uh how much do the chips cost? How much do the racks cost? How much do the liquid cooling cost? how much does the memory cost, how much do the cables cost, etc., etc., etc., right? That's based on our estimates up and down. So, I think the TCO side of things we're pretty confident. Um, it's the it's the performance side of things where we don't know, right? Um, there's a wide range and that's what we sort of tried to state in the article, right? Performance is a wide range.

Uh, can you can you uh explain more about Google and Broadcom's relationship? Max Hodak from from Neurolink and Science was was asking on the timeline last week why why have Broadcom as a middleman. Couldn't couldn't uh Google do the design and and place the orders from TSMC themselves, but but what's your read on on that relationship and how how durable it is?

Yeah. So, when you think about chip design, there's a few different stages, right? There's defining the architecture and then there's actually like implementing that architecture onto a process technology. There's laying out that architecture into gates in the on the chip and then there's like the whole supply chain side of things, right? Negotiating contracts, getting allocations, etc.

That takes like 18 months, right?

Isn't that like an 18month process basically?

Yeah, 18 months or more, right? Um

I would say I would say actually like Nvidia is is faster side and Google's on the slower side just because you know Nvidia's been doing it for longer. They have a bigger team, right? Um and they they you know but at the same time Intel has the biggest chip design team and they move even slower than that right they take like four years at least that's what they did a year or two ago we'll see what the new CEO can get into the you know you know reorgate right um but as far as like Google you know when they first started the TPU it was a very few people and they relied heavily heavily heavily on Broadcom to do everything right they just defined the top level architecture and Broadcom did everything I said below right negotiating with supply chain figuring out proc uh figuring out how to lay out the gates everything right As time has moved forward, Google has taken on more and more of this. Right now, they use, you know, they've talked a lot about Alpha Chip where they use AI to help floor plan the chip, right? Once you have the architecture, how do I physically lay it out onto the chip, right? Um, they've done more and more and more there. They haven't taken over everything yet, but that's that's sort of the point. But Google Broadcom has this like super big advantage, right? Nvidia, they they acquired Melanox, you know, call it five, six, seven years ago. Huge acquisition. Who's the biggest networking company in the world? Broadcom right Broadcom is the biggest networking company in the world and you know when you talk about AI it's it's the AI it's the architecture of the actual processing elements it's memory um which you're buying from you know you know the memory companies right and Samsung and Micron and then it's networking right when you try and boil it down to the most simple things and software right the networking side of things is so important and the let's say technical competence of everyone around the world besides Broadcom and Nvidia in networking is so Oh, or rather it's just not as good as them. They're actually good, but it's like Broadcom and Nvidia are just so good and Broadcom is better than Nvidia in many ways at networking that you know when you think about what is Google doing, yes, they're defining how the network topology is, but when you're talking about the physical networks, you know, how how packets get transferred, all these different things. Um, Broadcom has heavy heavy influence there. So, to this day, right, Broadcom is still charging margins like they did three, four years ago, even though Google has taken up more and more of the work. Um, but at the same time, Google can't leave until they figure out how to do the networking and supply chain themselves or with a partner. And so what are they doing on TPUv8 that is potentially a distraction that's slowing down their execution is they're working with MediaTek, right? MediaTek at times has helped Cisco with their network chips. MediaTek has uh a lot of work on some of this networking stuff. They're nowhere close to Broadcom, right? On revenue, right, for you that's that's that's one metric on technical competence, you know, that's another metric. I I think Medatech is good, right? But like they're just nowhere close to Broadcom. So now Google is having to work with, you know, I don't want to say subpar vendors, but uh inferior vendors to Broadcom,

and that's just to increase their margin on TPU8.

Um I would I would even say their angle when they started this project was never we're going to sell TPUs externally. It was dude, we're paying, you know, a 3x markup to Broadcom. Um and half the cost of this chip is memory. Like what the [ __ ] are we doing? Right? um you know at the same time it's like well sure physically the cost for the networking is not that much but what value does the networking bring is you know sort of Broadcom and then Broadcom's also doing the like game theory not science of like well you can't really leave us so we're going to charge you what we think is fair or what we think we can charge and Google's like oh no we're stuck to you right so media is taking way way way less margin they're not passing the memory through them right and so you know this this ends up being like hey that's a huge advantage for them flip side is like well they've They've they've got to engineer all this work that Broadcom was doing. Instead of working on a way better architecture, they've got to work with a worse vendor, right? Objectively worse, although MediaTek, like I said, is very good. Uh to try and implement TPUs uh more directly with TSMC with less Broadcom sort of in the middle.

Very helpful.

And and Google, you know, because it's risky is going down both paths, right? They're continuing to work with Broadcom on TPV8 and then separate TPV8 project they're working with MediaTek, right? cuz they can't risk, you know, fine whatever 30 points of margin, 40 points of margin, 50 points of margin. I can't risk the TPU being late because ads runs on that. Gemini runs on that.

Yeah. Uh can you can you give any takes on the Nvidia's $2 billion investment in Synopsis that that got announced this morning? I don't know if you you saw it. I'm assuming you did.

Yeah. So, in a time where, you know, let's say the two biggest chip makers, Broadcom and Nvidia are making more money than ever and everyone else in the supply chain and all the hyperscalers are trying to design more and more chips. Everyone's everyone's sort of working on that. You've got you've got the the EDA vendors are at the lowest possible valuations or lowest valuations that they've had. They're still very expensive, but lowest valuations they've had on a earnings multiple basis for a long time there. And and and this is on the eve of hey like objectively are there going to be more chip designs or less chip designs in five years? A lot lot more. Right now the flip side is AI chip design is coming. There's 20 plus companies doing AI chip design. And it's a we've got a really long article coming on that soon uh that will sort of explain the landscape. But AI chip design is going to shake up everything. And so the question is like

AI this is AI AI chip design correct

AI helping chip design whether it's for AI chips or for like power chips. Okay.

Yeah.

Got it. Um and so so the question is like you know Nvidia Nvidia has a lot of tools internally right the the dirty the thing about EDA is that there's three companies that own 95% of the revenue but at the same time Google and Nvidia and Broadcom and all these guys also design a lot of their tooling internally although they are massive customers of all three vendors right so it's kind of like an oligopy where the customers also contribute a lot um and so Nvidia's whole goal here is like how do I get every EDA flow working on GPUs because today a lot of it is running on on FPGAs a lot of it's running on CPUs. Um, and AI AI chip design is going to get a lot more AI influenced. How do I get everything working on GPUs? Um, in terms of like the operation of it, even if it's helping people design, not GPUs, right? Um, and and I don't have enough engineers to work on all the software. They've open sourced a lot of software, right? Like KU litho, it's it's software for lithography, right? And they've got all this software up and down the chain all the way from lithography to laying out chips and all this other things. They just want to make it all run on GPUs. Um, and and so that's that's what their goal here is, right? And now they've given Synopsis a huge huge they're buying Synopsis at the lowest valuations that Synopsis has ever had with all this cash that they were going to give away in dividends or buybacks anyways. Um, and and they're getting Synopsis to now make GPUs first class, right? And so I think this is a win-win um for Synopsis and Nvidia.

Well, we could go way longer, but I know what's on your calendar. You got to hit the gym. Thank you so much for coming by and chatting with us. This is really really helpful. Have a great rest of your day. Enjoy the holidays with your family. Uh great catching up. We'll talk to you soon.

Cheers.

Goodbye.

See you guys.

See you.

Uh let me tell you about public.com investing for those who take it