Tae Kim on Nvidia's culture, the TPU threat, and why Blackwell demand is 'bonkers'

and they're trusted by millions. Um, we have Take Kim, author of the Nvidia way and a Baron's senior writer joining the show. Take him the author of the NVIDIA way

for joining the show. And I'm sorry it [music] took us so long. Uh we've exchanged uh posts on on X many times and uh we wanted to have you on the show earlier, but I'm so glad you got to ask right away.

We did because it's the perfect time to talk to you.

Do you have roommates? [laughter]

Not.

No roommates.

Guys, longtime listener, first time caller. I'm so excited to be on this.

I'm so excited to have you here. Reminder, everyone, go buy the book. Seriously, uh Christmas is coming. I can't imagine a better gift for everyone.

For a four-year-old for even

I know what my son's getting. He's getting the Nvidia A Jensen one teachable copies.

Multiple copies. No, seriously, get get 10 copies. Give them that teenagers. A lot of teenagers have have read it. Like it's amazing and they reach out and uh it's an inspirational entrepreneurial book that a lot of parents are giving to their kids. So definitely

I love your uh your headset by the way because we're actually developing our own TVPN over here headset because this is just like this is the this is the ideal ideal setup.

I was telling my friend I don't care if I look like a dork. My hearing is going

so like I get locked in when I have the headset I can hear everything.

No. And it's wired. The worst the worst is AirPods. AirPods have a tiny lag. Like you're doing Zoom calls. It just it it ruins it. You feel like you feel like you're right here at the table with us.

It's not code red over there. He's Baja blasting. You know, he's Baja blasting. Um anyway, uh let's let's uh we we we I mean we have some time. Let let's run through uh I want [snorts] to know a little bit more about your perception of Jensen, your perception of Nvidia, and just set the table for us. We know how what is his management style? How does he he has all the direct reports? He reads everyone's like to-do lists every day. They have tons of employees. They never fire people. Like what makes Nvidia's culture unique? Set the table for us. So then we can go into the opportunities and challenges with that framework in mind.

So So the first thing I found out about their culture is it's very blunt.

Like I think in most companies and you guys have done startups, but I don't know if you work for large corporations.

Um bureaucracy builds up process it gets oified. Yeah.

Uh, Nvidia is the complete opposite. Like things are not going well, he'll chew you out in front of the whole company.

And that kind of blunt mentality, I think, you know, sparks better performance because you don't want to be embarrassed in front of Johnson in front of the whole company. Yeah. But [snorts] also it it just sparks an agility like uh when I talk to people at Intel or Google like the biggest problem they have is meeting paralysis and you need to get signoffs from like five different executives at Nvidia like you have a meeting Jensen makes a call he seeks out the right information and you move. So there's this agility at Nvidia. Uh the the other thing is just a meritocracy. Even from the beginning, like 30 years ago, Jensen's always asking who's the smartest person that you work with? Who should I try to recruit?

And uh from the beginning, like Dwight Durks, uh who's one of the top people right now, he he recruited him because he talked to this other guy and he said, "Oh, Dwight's really smart. Like I I really enjoyed working with him." And he just almost ruthless in a sense. He just goes after them and and recruits them, brings people aboard. So this meritocracy, agility, speed and just getting rid of the internal politics I think really separates Nvidia.

Uh how has most of the team uh becoming millionaires affected the culture

if at all?

Um I think a they had that thing in in the first 1015 years like people started getting sports cars and putting them in the parking lot but a lot of people

cars let's go that's the best news I've ever heard. That's fantastic. Sounds like it's had an incredible impact on the culture.

Job's finished. Okay. Yeah, you can end you can. You don't need to respond. You don't need to say anything.

I think [laughter] winning culture breeds winning and people want to stay with winners. Yeah. Right.

You want to win on the track then.

A lot of people I meet in terms of colleagues, they work for a company for five years and the chip doesn't work out. Yeah.

And you just wasted five, seven years of your life. So, you want to stay with a winning company and Nvidia has been winning for 30 years. So, it it it's kind of like winning begets winning. you get the talent and then then the talent stays like there's so many top executives at Nvidia that have stayed there for 25 30 years and

yeah there's also there's there's there's some benefit of people when if if if an somebody doesn't have a scarcity mindset right and they're just playing to win like they're just like they're they're they're they're no longer thinking like oh if we can just get to that next milestone and get this secondary sale and if I if I can participate and if I can vest my two years and sell into the next uh tender offer then I'll you know they're just like we're good. All that matters is just being as elite as we possibly can be and just uh and doing it uh for the love of the game basically.

And a lot of these people that like they made it they could retire they could have retired 10 20 years ago but they stay and work still work 80 hours 100 hours a week because that's what's expected. Like even the marketing people inv 85 hours a week and that that kind of mentality I think is different at the companies. Mhm. Okay, let's uh let's shift into the competitive dynamic. I mean, Nvidia's been uh we I was revisiting the performance of the Mag 7 since the dawn of ChateBT. It's been three years exactly. Nvidia by far the winner, up 10x uh on market cap. Uh the next closest company, I think maybe 4xed uh by comparison. And so, uh, the clear AI winner in the public markets, the most obvious AI trade that just completely ripped. Now, you know, there's this whole narrative of like, uh, how strong is their moat? Uh, what is what is the TPU mean? Is the TPU going to be significantly competitive? Is there going to be margin compression? How have you been processing this new narrative that uh Nvidia might face serious competitive threats because they're so on top of the world that everyone owes them so much money [laughter] that people are saying I got to get a discount from somewhere and I'll go to Google maybe

I want to talk about this 10x move like it hasn't been like straight up to the right there's always been

every three six months there's always a reason to sell Nvidia like

the H100 problem the transition to Black Whale, uh, China, ASIC, Broadcom competition. So, this this stuff has been happening this entire 10x move up and the media loves to latch on to the latest thing to worry about, right? We had deepse earlier this year that the entire media establishment was it's over for the AI tree. A AI models have become so efficient when it was actually the opposite because the reasoning models and there was an exponential demand for computer. So I I find it amusing like the whole world kind of discovered that Google had a really good chip in the TPU

which they've been working on for a decade too.

Yeah, [laughter] they've had it for 10 years, right? They've offered it to uh clients 2018. This is nothing new. The Ironwood specs, you know, which I always take with a grain of salt with specs. Even even Nvidia, they talk about 25x improvement with Blackwell when it's more like

I'm really just focused on the name Ironwood is goes pretty hard. It's pretty good. They they got some good names over there in Google.

Ironwood came out all the specs came out in April. Like this is not all this stuff is isn't new. And uh another thing I want to say is TPUs their chips. Morgan Stanley estimates there was a huge decline in TPU shipments in 2025. And Google at Google Cloud Nvidia GPUs took more share than TPUs this year. It's like no one talks about this, right? And now everyone's going Ironwood is going to take over the world and Nvidia's in trouble.

Is that just because we're

I mean I think it was a Gemini 3. It was a No, it was a Gemini 3 thing. People are like Gemini 3 is the best model in the world.

Yeah.

On by the world for a whopping six days and chatbt is still number one on the app store. Let's

But I mean six days it was unseated by Claude which was also anthropic which is also TP potentially in the future.

Yes. and chat GPT um Microsoft the head of Microsoft AI cloud code refer that open eye is training um their next models on the GB300 and the L72 uh that what just went live in October actually earlier today the Nvidia CFO said it's going to take six months so I was a little disappointed in that

six months for the training run

yeah well she said the first models on Blackwell like on the superclusters are going to take six months so

it's a singularity

[ __ ] It's going to be another another

hop AI is going to get there. The the Claude and Gemini uh benchmark gains with uh pre-training. Yep.

That's the most bullish thing for the whole AI industry, right? AI AI adoption is the scaling laws are intact. Everything's going to work out and open AI is going to get there when they uh build their next training uh on the next model. So, so going back to TPUs, uh, thank goodness, a shout out to semi analysis. They do the best channel work in the industry. Everyone freaked out on Friday, right? They read that semi analysis. No. Oh no, total cost of ownership.

They're going to destroy everything. But like people that actually know the industry,

it was flaming bullish for Nvidia. Like it it just it was like so obvious in my face because uh Dylan and and the Sammy analysis they said, "Wait a minute. The next TPU V8 is not going to be that great. They lost a ton of people and the the set function up in performance is not going to be that great." So you know what's going to be great? Nvidia's Vera Rubin, which comes out at the end of next year.

So no company is

People are saying it's the Rick Rubin of chips. [laughter]

Are they related?

Yes. it and so anyway, so Bar Rubin is going to be dramatically better at the end of next year and even

you know the the Ironwood which just became generally available and they're ramping right now. Um it's it's it's no one's going to switch over for one it's a huge uh endeavor to put workloads uh from CUDA Nvidia GPUs and put them on there's always problems when you put put them on on a new chip.

Yes.

U and let's talk about TPU customers, right? Everyone freaked out that Meta might spend a few billion dollars in 2027.

That sounds like a lot, right?

That's less than 1% of Nvidia's uh expected revenue.

Sure, it nothing.

And Ben Thompson was very smart and astute. He's like, who's going to buy the TPU? Who are the biggest buyers of AI chips?

Yep.

They're the hyperscalers, right? So, so Met maybe Meta will put a portion of their workloads 1% Nvidia's revenue.

Is Amazon going to buy TPUs? Well, John just asked the CEO of AWS if he was going to buy TPUs. He dodged that question.

He didn't say yes.

I mean, there's no way in hell they're going to buy TPUs. They have their own tranium. They're not going to support their number one like one of the number one

arch rival.

Yeah, they're not doing that. So, is Microsoft going to buy?

You should have been like, "Yes, I'm going to buy one so I can like study it." [laughter]

Microsoft's not going to buy TPUs. They're the number two player in cloud computing and they're not. Are the neoclouds gonna buy TPUs? Now you're gonna say yes, Google got has some neoclouds. You know what happened with those Neoclouds? They're financially backstopping those Neoclouds. So Google is financially giving money and and backs stopping the debt for those Neocloud. So So there's a handful of small NeoClouds, but is Core Weave going to buy TPUs? Probably not, right?

Who are the other customers of AI chips?

Enterprises, companies, sovereign AI. Yeah. Anybody that wants to run like a fine tuned model, some small model, something like that.

Yeah.

90. [laughter] Um,

so like if you just go down on a first principles basis and look at the customers of AI chips, like they're going to stick with Nvidia. The millions of developers know CUDA. So you don't have you really need like Dylan talked about this is you really need like top-notch like software sophisticated engineers that can like work with TPUs and learn learn jacks and all that stuff. So most people aren't don't have those cracker jackack engineers, right? So they're going to stick with Nvidia because everyone's used to Nvidia. Nvidia is backwards compatible and forwards compatible. So like 20 30 years of this stuff.

Mhm. And if you buy it, Nvidia Collat app the CFO talked about this morning. It's um you can use it for training and you can use it for inference. It's all on the same architecture and it's going to work. Like I I I talked to an AI startup CEO a few months ago. He tried AWS training. Oh, it looks a lot cheaper.

Total cost of ownership, but then it crashed. There were bugs, the reliability, they they couldn't figure out what happened. And there's like they just threw up their hands. I give up. Like no one is going to like if you have reliability problems, bugs, crashes, the best thing about Nvidia is all that stuff has been ironed out over the last 1015 years. If you have a problem, you can figure it out because

yeah, it's like giving giving an F1 giving an F1 driver like a car that that is unreliable and saying like, "Hey, go race, go race, have have fun out there." And then it's like, you know,

and the specs DNF immediately, right?

Specs specs look awesome. It look seems great, but then when actually build your business on it, you you put the future of your business onto something. You the number one thing, it's not price. It's like it better [ __ ] better work. [laughter] It better work.

And it works.

Yeah. But what what uh react to the this idea that Dylan Patel was uh joking about as uh TPU is a stocking horse. So this idea that Sam Alman is already saving 30% on his Nvidia purchases effectively because just the threat of going to TPU is enough to get Nvidia to make an investment or slightly discount in one way or another. I

I don't think that's reality and I don't think that math actually works because I think he's confusing the AMD deal where AMD gave, you know, free warrants to Open AI,

right? First of all, the deal is not done.

Sure. It's a letter of intent has

none of these deals are done. They're they're not done.

AMD AMD's AMD is done. They they they signed an agreement where they're giving away a percentage of their company through these free warrants. The Nvidia deal hasn't been signed yet. And they actually language in the Tank Q that it might not happen. Doesn't OpenAI have to buy AMD chips in order to get the warrants?

Yes. Yes. Yes.

And so it's still it still could be that they don't actually end up going through with the purchase and then they wouldn't get they would point.

Yeah. Yeah. So, so the whole like let's do a side step here with the circularity and all that stuff.

All this stuff it's like one gigat at a time. There's a milestones on open AI there's milestones on AMD technical milestones they have to achieve certain targets. So all this talk about, you know, everyone loves the the big number that adds up five years of capex a lot like that it it's it could get the leverage could be up or down depending on how things happen every each year of the way. So you know it might not be that big number if open AAI doesn't come out with an amazing model or AMD isn't able to hit the milestones they said for their next 450MI chip, right? So like it, you know, don't worry about five years. Like take it one year at a time. Right now demand is off the charts. Now going back to the Nvidia 30% discount, like that's not how equity investments work. If Nvidia does invest 10 billion, 10 billion up to 100 billion. Say that say that

Nvidia gets ownership of the company. It's not like a freebie, right? You're giving away uh ownership of your company. So it's not really a discount. you're getting you're getting uh uh ownership of the company. So I I don't really believe in this 30% discount thing because um Nvidia Jensen will say they're they're investing to accelerate open AI and uh they would they they they're looking forward to you know open AI going

I mean it's definitely creative. It's definitely a new structure. I'm just trying to I would I would steal man in that like if I'm an entrepreneur and somebody comes to me and they're like I'm going to invest $und00 million in your company over a series of milestones and you're also going to buy something from me. I'm like yeah I'm taking some dilution but realistically like this is a way less of a headache. Like where else was I going to get hundred billion dollars from if in in O OpenAI's case like it's a great it's a great source of funding that yes it will be diluted but the whole structure is all diluted all the time because of all the different ownerships

except for this kind of sentiment thing that we had the last few weeks. So open AI hasn't had any problem in raising money for venture capital.

Yeah. Yeah. It's true. It's not just it's not like Nvidia is the only source of funding for open AI like everyone wants in the revenue run rate it's like 5 billion to 20 billion at the end of this year.

Um

so what was your take on the code red what do you think about the code red?

So uh I saw that you you you showed that Ashley Vance uh

interview really interesting Chen was talking about how they kind of focused a little too much on reasoning and their pre-training muscle wasn't there. Um I reasoning we could talk about this later. Reasoning is like the biggest kind of accelerant of AI demand in the past year. So I I think it's actually really good and supposedly Ilia was you know doing the research for reasoning. Reasoning is awesome, right? But they they kind of focus on reasoning the the past year with 01 and 03.

Um and now they're like okay we have to go back to pre-training. So, o OpenAI knows that pre-training still works because Gemini 3 had great pre-training results and and Cloud Opus uh 4.5 did. So, now they're going to do the pre-training.

So, they had their focus on one thing and now they're going to do the other thing and make their their model much better. I I do agree that Open AI has been a little too maybe

diluted like they're doing apps.

Sure.

They're doing hardware. They want to compete in AI infrastructure against Microsoft and Oracle. They want to compete in AI chips against Nvidia. Like I I I thought it was really interesting like Satia repeatedly said um he wouldn't name who he's talking to but it's like

I think it's important that we realize this is not a zero sum game and this could be a win partnership. That was during the anthropic Nvidia uh deal with Microsoft. Right. Sure. He said that a couple times and I think the the person that he's talking to is Sam

Alman,

right?

Let's let's Nvidia, Microsoft um made Open AI as successful. They were they were the partners like why are you competing with your the partners that brought you to the dance, right? Let's let's go back

focus on making the best AI model in the world and uh don't compete with Nvidia and Microsoft. maybe maybe 5 years from now, but like it seems a little aggressive to compete with them uh right now.

Let's talk about China. Uh there's been a ton of debate over Nvidia selling chips to China, legacy chips, older chips. Um we've gone back and forth on it so many times. Uh what's your current thinking about the the the the best policy for Nvidia exporting chips to China? Generally, I

I think the best thing is to keep it one or two generations behind

uh the current state of the art.

Yeah.

Like this is a really nuanced policy that people, you know, everyone's either hawkish or dubbish.

Totally.

Whatever. The best policy is to keep I don't want to use the word that Howard Lutnik used that got China very upset and forced [laughter]

forced China to like tell his companies not to buy H2.

Um but the best policy is to get China still on the Nvidia stack. So Nvidia gets $50 billion of revenue per year that can help R&D and fund R&D and make the chips even better. Like

Nvidia and the US already won, right?

They have 95% market share. like why are we going to give $50 billion of oxygen to Huawei and all these other Chinese AI chips uh companies that now Chinese uh companies that need to buy AI chips are going to buy Chinese AI chips like why not keep China on the Nvidia tech stack one or two generations behind don't give them the best stuff but maybe one generation behind I think that'll be the best compromise um

for for for both sides but I don't know

what do you think what do you I think what do you where do you place a likelihood that uh that the Chinese market has opened up again at at some point in the next 12 months?

Maybe 50 5050. I mean that sounds like a copout. Like I was much more positive 6 months ago, but

um you know this has been just so crazy. First the Trump administration banned the H20.

Yeah.

Then they didn't ban it. They said it was fine. But then China was like, "No, you hurt our feelings. Um we're not going to let companies buy the H20."

Yep. And then they ban it.

And then

Yeah. And then maybe Trump is going to let Nvidia sell the H200

or a Chinese specific version of Blackwell that's kind of like hobbled a little bit.

Who knows? Like Nvidia needs to convince the Trump administration and then China to to buy the chips. The the the worst part of it is China was willing to buy the H20 and it's just all the kind of geopolitics and hurt feelings. Um you know that that ship has sailed. So, I don't know what's going to happen. Um, but I do think it the ideal situation is Nvidia could uh sell one generation behind, make $50 billion a year and and keep u the competition from ch Chinese AI startups uh out of the way.

Yeah.

And even understanding that at some point in the future, China's buying effectively zero chips from Nvidia, but it would be 5 10 years in the future. Um

it's like you have to you have to assume you have to assume that like

right now

go for it.

Uh the the amazing thing is [laughter] sorry uh it's 0% right and Nvidia's revenue accelerated for the first time in two years. Nvidia's revenue accelerated in this latest quarter and this is like not talked about enough. This is the first quarter that the NVL72 AI server has been available in volume and then revenue just skyrocketed without China which is incredible right and that's why I'm so bullish over the next few years because next few quarters let's say that

because this product cycle is going to last at least three four quarters um the the key tell is the revenue acceleration first time in two years and and the other key tell is that the networking segment for Nvidia was up 162% year-over-year a year and typically a lot of these data centers um and these neoclouds buy the networking stuff 6 months ahead of time. So the next six months from now like the the the GPU numbers for for Blackwell and the NVL72 server is going to be it's just going to be bonkers. It's going to be off the charts. And people don't talk about these NVL72 AI servers. They're 3 to4 million, right? There's 72 GPUs, 144 dies, um, uh, one and a half tons, 5,000 cables, and the, and the prior version was 8 GPUs. So, so these these AI servers are I I call it the iPhone 3G moment. Do you guys remember the iPhone 3G? Like,

this is big for the Christmas shoppers out there. Uh, if you if you want a gift, it's a step up from the Nvidia way, I recommend. Or you want to bundle something,

picking up an NVL72,

it's only $4 million.

Yeah. But for the right 12-year-old in your life or or the potentially the intern, I think Tyler,

it won't fit under the tree, but it could be fun. Keep it in the garage.

It's like it's like a Lexus. You put it in the in the in the driveway with the bow on top. That's the way it does.

With the fork, you drive the forklift. Come drop off.

Exactly. I got you an NVL72. Enjoy. [laughter] Um and and and then we have the reasoning model thing where exponential compute and like companies are actually seeing like huge cursor 40% productivity gains.

Yeah.

Uh CH Robinson 40% shipments.

Um

rocket mortgage 80% reduction in paperwork cost processing. Um this is like the next year because of AI reasoning because of the MDLMD2. That's why uh Amazon and Microsoft said every quarter this year they raised their capex and everyone's like they're going to cut their capex. They're no every single quarter they raise their capex. That's because they're seeing the demand and that's why Amazon and Microsoft are going to double their data center capacity over the next two years. I mean that that that's crazy, right?

Yeah.

In September quarter more leasing there are more data centers leased than entirety of 2024. This is like exponential step function up and and people aren't talking about it. They they want to talk about TPUs like destroying Nvidia.

What about talking about

what what about some of the kind of demand guarantees that have been happening? Is that a is that a concern at all? Do you think about it much? Is it

I not really I mean demand guarantees like you're talking about Nvidia and coreweed it it's like when that happens analysts every every quarter or on the conference call like did you use that you know demand guarantee like it's not happening yet like coreweave's 5-year-old GPUs that everyone says are useless are 100% us utilized right H100 a massive cluster before uh it expired probably a three-year contract they got like 95% % of the pricing. This is like unheard of. And and the reason is there's overwhelming AI demand and there's not enough capacity. Overwhelming AI demand, not enough capacity. And and and people just are are just they don't care about what's happening in the real market. This is real life facts, evidence, numbers. Nvidia going from 56% revenue growth to 62% revenue growth on 57 billion with zero China re I mean these are bonkers numbers [laughter] we we talk about the stock price being up 10x their revenue is up 10x in like two years this is like beyond history the last 30 years of following technology

I love how you're the you're the only person without Nvidia fatigue you're just like you're not bullish

David Gogggins of the

[laughter]

I I mean they can't keep running. They can't keep running. This is not me just like making stuff up. This is like the numbers are there right in front of me.

You make you make good points. You make good points. I I like it a lot. Uh we'll have to have you back on the show soon. This is a lot of fun.

Yeah, let's do it. Let's let's make this a regular thing. Super fun to have you on finally.

Thanks so much for all your time. Uh the book is The Nvidia Way. Get it at wherever books are sold. Get 10 copies. Give it to everyone in your life. Also, give the gift of Finn.AI, the number one AI agent for customer service. Automate your most complex customer service queries on every channel. Um,