You're surrounded by journalists. Hold your position.
Overnight success.
Double click.
Right.
That's misinformation.
Clearing order inbound.
Let's just roll.
We are surrounded by general. Hold your position.
Get up.
Trust the experts.
Five.
We are.
Founder.
Five.
I see multiple journalists on the horizon. Stand by.
UAV online.
Double blaze.
Triple blaze.
Double kill.
Cook is.
Team deathmatch.
We are experts.
Triple blades.
Let's just roll.
Right.
Hardy clearing order inbound.
We are surrounded by journalists. Hold your position.
Strike one.
Activate golden retriever mode.
Marky clearing order inbound.
Five pudding.
I see multiple journalists on the horizon. Stand by.
Founder.
You're watching TVPN. Today is Thursday, July 23rd, 2026. Jordy is wearing a bandana.
What corporate logo is on that bandana?
TBPN collab here with AMD Advanced Micro Devices. Uh yeah, beautiful. Thank you to the team for hosting us. Thank you to the AMD team for having us here at AMD Advancing AI. Uh we're on the road in SF today. Uh we have a bunch of great guests for you. Lisa Sue is joining in about 90 minutes. Uh but let's you take you through a little bit of the news of what AMD is announcing, what's happening today. So uh Lisa of course will be joining us at 1:30 p.m. We have Oliver Cameron from Odyssey, Muhammad from Ideogram, and Anne from AMP coming back to the show for round two. Uh round two or round three. We've probably talked to him a few times. It might be the third time. Uh, Advancing AI is AMD's biggest AI event of the year where it reveals new chips, software partnerships, and how it plans to compete with Nvidia. According to Jackson Force TV newsletter, we just found this out this morning. AMD is actually powering the new mod retro M64.
Very cool.
It's very Yeah.
Crossover Palmer Lisa.
Yeah. So, we're going to be spending all 30 minutes of our time with Lisa Sue on gaming specifically. How is How are they going to bring and suits? Yes. wear a fantastic suit collection.
Got a fantastic suit on today.
Uh so the three key stories that people are focused on out of AMD advancing AI are AMD unveil un unveils Helios the new rack scale AI system powered by the MI450 accelerator. Uh we'll talk to Lisa about that. Also a new partnership with a Frontier Lab. AMD and Anthropic are partnering to deploy up to two gigawatts of AMD Instinct Mi450s.
Sizegong.
Sizegang five billion dollar deal uh and a whole bunch of clo uh a whole bunch of uh collaboration across compute and the software stack. The big question around the software ecosystem around AMD is always a hot topic. I'm sure there's a ton of developments there. Only stage just a little bit ago, they were talking about how um Anthropic's own models are effectively speeding up their ability to ramp up new hardware. Yep.
Which is something that uh has been talked about for years now. Some questions around Yeah. Uh the CUDA mode, right? Uh if you assume that making great software will just get easier and easier, uh that that's been a big question. Uh and they're starting to see uh real movement there.
Yeah. But also a lot of that speed up requires actual deep partnership like this because there are some close source packages that you might need to uh you might need to work on. This is the old George hot interaction. Um talking about little bugs that he was frustrated with wound up working with AMD developer team who we've talked to on the team on the show. Um fantastic outcome there but interesting uh to see that flywheel just get tighter and tighter and tighter and I expect it will. Uh also AMD is expanding its partnership with Cerebras teaming up on AI inference. Uh we love the team over at Cerebras. Uh they introduced a powerful disagregated inference solution pairing the right engine to each phase of the inference pipeline. Uh this is what Agentic AI has been waiting for. The fastest production inference at massive scale. So uh excited to talk about that as well. But uh our first guest is here uh Oliver Cameron from Odyssey. I have a whole bunch of fun backstory. you can come on down. Um,
Here he is.
I first met Oliver, uh, almost 5 years ago. You were working at Cruz doing easy stuff, self-driving cars. Uh, and you took a ton of time to talk to me and walk me through exactly uh, how you were thinking about self-driving. You explained a lot of different stuff to me about what what's being done in machine learning, what's being done with C++. Very excited to have you back on the show. How are things going?
Things are going great. Thank you so much for having me. Great to have Uh, congratulations on the massive progress. Last round was what? 300 something billion. A million. A million raised. It's gonna be billion next round.
310 million.
310 million. Fantastic. And And where is the product right now? Like how are you describing uh how much is in the research phase? How much are you ready to deploy this with early adopters? How are you thinking about that?
We are in the GPT2 phase of world models. Yeah. And what we're building is foundational world models. The idea being that this is a technology that can be useful and applicable to tons of different industries. Uh whether that's for example robotics or science or gaming or driverless cars.
And so we've been on this journey now for 3 years.
Is there still a market in self-driving cars? Because it feels like the world models, a lot of that research was done at Cruz and Whimo and Tesla and that was where I was first introduced to them. I was like, "Oh, they built a video game version of San Francisco and of course they're training it there." Uh is there a world where you wind up powering future self-driving initiatives and autonomy initiatives?
I think so. I think world models can power all sorts of virtual and physical systems. The virtual system systems might be gaming for example, uh might be education. Uh and then the physical systems will be driverless cars, robotics, drones, all that sort of stuff.
What I believe is that driverless cars today and lots of automated systems look very much like NLP systems in the 2010s. M they are these very complex very handtuned systems. They're intelligent but they're intelligent in sort of isolated ways and that those systems hold back what these technologies can do. And so I think you can replace these very handtuned driverless car systems with a single world model that has this very deep understanding of physics of cause and effect of human behaviors and just lightly tune that world model to the task of driving. Y what we're seeing early evidence of is you can indeed do that. You can take these very big Jammer Wolf models, tune them with just a few hours of experience and then the car drives itself. Pretty cool.
So what what are your interactions and relationships, partnerships look like with uh various types of robotics companies? So think of like if someone's building a humanoid or we we were talking with Travis Kalanick yesterday doing autonomy and mining. So like what what does an Odyssey relationship look like with those companies today?
Absolutely. So we very much believe that the robot companies are customers of ours. We provide that base intelligence. They then have their robot, their embodiment, their task and they can take that base intelligence and tune their model to make that task amazing. And what we see earlier evidence of is that using a world model versus a vision language action model, the number of examples you need to show the world model is dramatically less than a vision language action model, which means it's just much more adaptable to the world.
Than a vision language action model.
Makes sense.
Talk about data right now. Um, a lot of these world model systems, it feels like they were bootstrapped on Unreal Engine systems or just video game systems generally. Uh, I imagine that the amount of data that's getting poured in the pot and mixed together is huge at this point. But what are some of like the power law data sources generally that are useful in building world models?
Yes. So I I'll get to the point eventually, I promise.
So what I observed in driverless cars, yeah, and we've seen in robotics is that you typically train those systems on very narrow data distributions, right? So you're a driverless car company. You collect lots of data from the road, dash cams, basically dash cams, but also cameras around the car.
Exactly.
We've all seen the street view cars.
Exactly. And you're really teaching the car just from the perspective of a car, right? It's it's solely seeing what cars see. It's got the blinders on. And what those cars don't see in training is video games or is inside a conference or in an office or in a home uh or just the multitude of scenarios that exist in the world. And what we very much believe is that a world model shouldn't learn from a narrow distribution of the world like a series of uh driving examples or robot examples. It should be that you learn from every possible thing that could exist in the world and then you tune the model to the task of driving. So you have a base that's just very general and then you provide driverless car data at the very last step. And so really our belief is that that will produce a more robust more intelligent system than these very narrowly trained models.
I love that. Makes sense. Completely tracks with a GPT2 GPT3 progression and language models. Uh there was no point when people said oh well we really want to solve IMO level math. So, let's get rid of all the fiction. No, it was always include everything, right?
Exactly. You need Harry Potter to uh solve math problems.
Hilarious. Very odd scenario, but it is true. Absolutely.
Uh I'm interested in uh the more expensive pieces of data that I could imagine you'll be acquiring in 6 months, maybe a year, maybe 5 years, I don't know. But we've heard about data labelers making seven figures creating RL environments for very specific tasks, very niche use cases, just seven.
Yeah.
I mean, it's going to be eight next next month, next quarter. But uh but but truly like the idea that uh the data labeling that went into early driverless cars, but also early reinforcement learning uh human feedback on text LLMs was just is this acceptable? Is this readable? Does this have typos? And now it's does this actually track to case law or does this look like a financial model that I could turn in at an investment bank. uh what is the world model equivalent of that? Can you imagine what that will look like?
I can. So I think if you look at Alpha Star, which was Deep Mind's uh training of agents inside Starcraft to then beat Starcraft players, which was incredible.
Really what you had was agents learning inside this environment that doesn't update or improve. Starcraft is Starcraft. The agents then have a ceiling on their performance because it's always Starcraft that they're learning within.
And so I think what world models really promise is the ability to be a learning environment that's continuously improving and adapting to the agents intelligence. So for example, an agent should be able to learn within a world model that is cons consistently improving over time. It's getting more robust. It's getting more diverse. There's more scenarios. And so the agent's intelligence being trained inside that world model should also get more robust more.
So fascinating because yeah, training a model to win at a video game is one thing, but the real world is never fixed in place the way a video game is. Right.
Exactly. Yes. And a world model really can be thought of as a sort of infinite infinite simulation, right? It's continuously generating new types of environment that the RL agent can explore and adapt to and fight within and all these these cool things. And so I think actually one of the the key or the killer applications of world models will be a learning environment for AIS. language models, other types of AIS can all learn within world models. It gets very much um philosophical.
Exactly.
Yes. Staying on philosophy. Uh if there are there are many many labs that have made the bet that uh text is the universal interface scale and maybe coding software only singularity. There's a number of different buzzwords to sort of define this idea that like LLMs are on the path. Then you have the folks who are arguing that everything from LLMs are a dead end to deep learning is a deep end dead end blah blah blah. But uh what I'm interested in is your if you're looking at there's clearly progress going on in textbased codebased LLMs that feels like a path to something called AGI. Maybe that happened two years ago. Maybe it's happening two years from now. But it feels like there's a trajectory there. Are world models like the next step? Is it a separate curve? Will they come together and work together to produce whatever the next the next uh paradigm is? How do these two technologies and like paths converge or diverge on the tech tree?
One slight tangent is uh one of those critics for language models, Gary Marcus.
Yes, I wasn't going to call by name, but you're welcome to.
Me and him and many others in driverless cars have been bickering for.
Oh, yeah. Because he's driving he's an Uber guy. I forgot. Well, before language models were the debate, it was all about driverless cars. And when they would come to Mark, is it 2050? Is it 2040? It's like, no, this is going to happen much sooner. Conveniently, he's now moved on from that being the debate. But anyways,
So my my belief is that language models are there's some new there was some new cope yesterday as well that I was seeing that maybe I shouldn't bring up.
Cope. Okay. Anyway, let's continue.
Cope is evolving almost as basically as fast fast takeoff as the models themselves.
Okay. Yeah. Yeah. Makes sense.
That's the frustrating thing. The goalpost is just moving moving moving.
We have physical ghost that we move around the studio because there's always a new go new goal.
It's ridiculous.
But that's the that's what makes it fun. It would be boring if we reached the goal and we were done. We want to keep building.
That is true.
Uh and so what I very much believe is that language models are going to continue on the trajectory. They're exceptional technologies. They will lead to a form of super intelligence that's amazing. Changes the world of course. Um, what I believe they learn from though primarily is a sort of biased representation of the world.
And it's our writing of the world.
And it's also a remarkably, I think, inefficient representation of the world. If you were to describe what's going on here in text, imagine everything, every detail was somehow describable by text. How long would that representation be? It would be huge, right? I pull out my phone, I captured two, three seconds. I've captured so much detail of humans intermingling and everything. So I very much believe that they are almost distinct somewhat complimentary technologies.
There'll be cases like coding incredible for language models. I don't think world models will play in that market too much.
And creative writing and many other emails and things like that. Amazing for language models. World models I think will play this role of operating within either virtual or physical worlds and that those will be just distinct applications. it. A world model will prove to be a better driver of a car than a language model. A world model will prove to be a better operator of a robot than a language model, flyer of a drone, creator of a video game. Uh educational experiences even I think they'll prove to be distinct. I do think there is a convergence somewhat of the the learning environment thing I mentioned that world models will serve as this universe for language models to uh learn within.
Uh and that's maybe where it converges at some point.
Selfish question. uh on autonomous driving. How do you see the market evolving from here? I've been having a lot of conversations recently where people are um buying buying Teslas, not specifically because they want a Tesla, but because they want the uh autonomous functionality. And I feel like a lot of the manufacturers are in a lot more trouble than maybe they even realize because there's plenty of drivers today where if you have a maybe 10 20 minute commute daily, you're not really thinking about how do I solve this problem? But for people that are driving any longer than that, it's starting to be like top of mind. That's like the number one factor. Um how how do you see this evolving? Is it going to be like language models and that there's a bunch of different providers with like pretty good autonomous driving? Uh we're not seeing it quite yet. I think you can see, you know, the two wildly different companies that are both exceptional. Right. Right. I mean, right now you have Whimo and Tesla. Yeah. And but like wildly different tertiary players.
Yeah.
So I see it going something like this which is that firstly consumers word of mouth matters and as many people are now trying Teslas it's going to become almost um insurmountable for many of the car companies to compete with just how good a Tesla is. Yeah. All things not just driverless car driverless functionality. And so I I think we'll see the continued sort of um attrition in the smaller car brands converging into Tesla. That's one thing. Second thing is I think it's becoming almost um what's the right word here? Like uh um a a national crisis to not enforce driverless technology. Uh I have two kids and it's uh today still the leading cause of childhood death is car crashes and the idea that we have this technology now where I can literally walk out here and be driven fully driverlessly back to my house like 40 miles away.
But can you can you put on a a sympathetic hat and try and have some empathy for the ambulance chasers?
What are you I mean Yeah, you have to live to think that. No. Uh I didn't know where that was going.
Take uh coming back to uh diffusion and self-driving. I want to uh go back to that idea of like 2050 prediction. It sounds ridiculous because you can walk out on the street and see one driving around San Francisco. But if you rephrase the question on driverless cars to a billion driverless cars, full diffusion such that there are, you know, like they make themselves available everywhere over the world. It's the dominant vehicle pattern that actually does take time. It's more of an industrial process than a technology breakthrough. Um, how are you processing diffusion of AI tools, LLMs, and will world models follow the same trajectory as LLM diffusion, which is really fast, everyone uses them for everything, but they haven't taken all the jobs. And we're sort of in this interim thing where it's amazing, but it's additive and it was sort of unexpected to everyone. Both the people that were like, it's fake, and both the people that were like, it's God, you know, like we kind of got the middle case, right?
It's just like, okay, things are going well, keep building.
Exactly. I specific to world models, I think that there'll be a very similar sharp progression of adoption that we've seen with language models. There'll be some killer applications that become very clear very soon. The idea that you could type a prompt and get out a AAA level game doesn't seem crazy to me. Multiplayer games too. The idea that you could have generally capable robots in your house in offices and everything else enabled by these models becomes clear. And then yes, driverless uh technologies and many others. Um and so yeah, I think as what becomes the thing I always lent on in driverless cars, yeah, is even if you didn't believe in the technology, you just have to look at the volume of people and the smarts of the researchers behind it and how much dollars was going into investing in that category to assume it's going to get figured out at some point, right? It's not an impossibility that we'll have driverless cars and so you put enough smart people behind it's going to get done. And I I think the same with humanoids for example like there's just so much capital, so much talent flowing into that space. It's going to get solved, right? It's just inevitability. You just keep going. And I think uh this is why I have such faith that yeah, we should I mean there's no reason now we shouldn't have a billion driverless cars in the road.
But timelines by 2030 will I be able to walk in and get a Honda and have it have autonomous driving capabilities that's as good as Tesla's is today?
Okay. Not even not even not even what Tesla will be in 2030, but like a Tesla Ford, you know, these other manufacturers.
2030 is like two years away.
No, but but but I'm just saying like it's like I I I'm curious because I think every passing year all these manufacturers are just in more trouble.
Maybe. Yeah.
Yeah. I think they get in their own way, right? So I think a company will build a technology that they could sell to a Honda or other companies that is driverless ready on that scale in 2013. Absolutely. I mean, we've got literal super intelligence coming that can write code of crazy quantities and such. So, I I think so then it's really the car company's decision about do I want this? Will I adopt it or will I continue to burn money internally failing at building it myself?
Last question. Uh how important is at least developing some of the application muscle, the go to market like uh Sunno just went viral in our office. We're all addicted to Sunno now. Uh and and there's been these moments where like okay everyone's on midjourney it's really fun and like it's this is this platform and then it cruise a lot of value it can sometimes be baggage because you're building you're you're like okay now I'm a consumer company. Um how how do you think about uh value capture and how much you want to be the front end to just prompt get a get a game.
Exactly. So my answer would be that uh I'm actually not happy with my answer but sometimes I get asked what are we good at Odyssey right and so my answer to that is that we we need to be great at everything and the reason I say that is because if you're building a very foundational technology it really does need to be great at all of these different tasks um and we need to stay focused on that instead of getting really focused on a minute case and um so yeah long story short I very much believe that by building this foundational technology along the way you'll will discover by talking to companies, talking to researchers. Really cool stuff. Some of them might be partners, some of them might be in house. Uh well, I'm very excited for it. I'm so ready. I can feel that the moment's coming. Congratulations on all the progress and thank you for all the work you're doing. Thanks for coming on the show.
Thank you.
Have a great time here. Uh while we bring in our next guest, let me tell you about RAMP. Time is money. Save both. These use corporate cards, bill pay, accounting, and a whole lot more all in one place. And we are we have a soundboard here. We are joined by the founder and CEO of Ideogram, Muhammad. Welcome to the show. How are you doing? Thanks for having me. Looking good.
Thanks for looking good, too.
Yeah, it's uh it's the right place for us. Um uh why don't you uh introduce the company a little bit? I I I I know a fair amount, but for the viewers who don't, the the shape of the company and uh the product and uh sort of the the target market at this particular juncture.
Yeah. So, if you think about it, we already have content creation in front of us. If you think you know we had the early days of image and if you look through your social media feed you see a lot of video but we think there's an inflection point that design marketing can also take advantage of yeah generative AI and that's what we're building at the AI foundation for enterprise adoption of the next uh phase of content and design. So.
Do you think that something like image generation can be fully solved within a shorter period of time than let's say intellectual intelligence?
It's kind of interesting because uh when you think of image generation you can have uh a very detailed document and diagram and really technical design as part of an image you know. Oh yeah. It's not at some point if you're saying generate a blackboard with a solution to a IMO level math problem on it. The model needs to understand that level of math.
Yeah. Like think of the circuit design or architecture design and and that's the same level of complexity as a lot of the reasoning problems we're trying to solve now. So.
One of the first how you look at it.
Yeah. Yeah. I was asking from the lens of of I think everyone on Earth now has now viewed an AI image, not known it was AI and just assumed like okay looks like a a normal image to me. And I think when you're talking about like if Pepsi is working on a campaign, they don't need like as soon as it looks photoreal, they don't need something that's necessarily more real. And so I just wonder what that means for the shape of Idoggram and where at at some point like what the what the core uh competency will need to be for Idoggram to continue to evolve as a business if that makes sense.
Yeah. A few a few things. One is we still have issues with consistency. It's not at a level that a brand can use for their uh marketing for their design. And we focus a lot on image generation and not as much on design generation which is okay I have certain typography I want to translate across languages have my logo it has to be 100% accurate you know uh with product photography it has to be exactly correct I can't buy a piece of clothing and then it doesn't it doesn't match the picture uh so when it comes to consistency uh we're still lacking but then part of the dream is to help with the design of future products. You know, the creative part of uh design requires a lot of understanding of the context uh and then the brand DNA and putting them together to design the next generation of cars or design the next generation of shoes. And that's where we are focusing on and uh we are working with brands to help supercharge their design and uh production.
How are to me that is I will just say the most my most addictive AI experiences is designing products that don't exist yet. just anything I could possibly imagine in my head and then the process of like prompting and and you know getting getting close and you know getting to maybe 95% and then trying to get it again you know the last 5% ends up taking you know 10 times more time than than getting to the 95.
Yeah. Um, so related to that, what are enterprises looking for specifically from you like this month? Because we just went through the token maxing up and down where enterprises were just, you know, spraying tokens all over. You can use whatever model for anything. Check the weather with, you know, the best advanced model. Um, but I imagine that your customers are concerned about cost, but also speed, quality, consistency. Uh I know you're partnering with AMD obviously, but uh walk me through a little bit of like the problems that you're hearing from enterprises and what you're focused on solving in the short term.
Yeah, one thing is data sovereignty and obviously when it comes to design, IP is extremely important.
Lots of competitive industries that okay, I can't let my competitors see the design of my future car and imagine sending that over to the cloud and then uh that that creates a lot of potential. Does that mean they want to run your models on prem or does that mean that they're just want to know that you're hosting it in a certain certain secure way with a certain SLA around? We're not going to train on this.
Both. And uh with the newest model that we released, it's actually a relatively compact model that's at the frontier and it's open vein.
And that creates a new set of opportunities for us in terms of licensing and partnership where companies can host these models on prem. Yeah. that brings down their their cost but also data sovereignty is really important. So on the design side that's one one thing we're hearing a lot. Yeah. And then on the brand side this is still uh consistency is an issue but we think we're going to solve that in the next few months.
Yeah. There's a big there's a big debate over uh openw weightight models, geopolitical discussions and one of the themes that's coming up is like o American openweight companies might be at a disadvantage if uh other countries openweight models have less respect for intellectual property. And so I would think that you would be fighting with one arm time tied behind your back. But based on your business clients, they might not want a a model that infringes on IP because that could wind up getting them in trouble. So, is it less of an issue for you or do you see uh like sort of the move fast and break things that may or may not be happening internationally actually being a like a headwind to your business?
On the image side, actually seems like American companies are doing really well and we are at the frontier. So, so it's not as big of an issue. side as long as we give them the protection they need. But one other set of customers that we are seeing is actually generating data to train really sophisticated models for manufacturing problems for really unique uh defense problems where where you don't have a ton of the training data. And these ones are getting to the quality that they can generate images of defects. Yeah. uh images of really rare scenarios that you can't find in the real world. Yeah. And that's another example where again having it on prem is important because uh they don't want to Yeah. share that out share that data with everyone else. Yeah.
Interesting. Interesting. Um what what has it been like working with AMD? We've talked about the software ecosystem around AMD. you're obviously g give me some like benchmarks on what you've actually seen from your performance and then I want to know about the process to actually optimize one of your models for an AMD stack.
Right. So it started with us testing some of the early models. It was MI300. Uh and then we got to know about the road map. It's a really impressive road map with uh MI450 and and the following ones uh 455. So we started testing it. we got good results and uh with AI agents it's become so easy to you know give it the the architecture and then get it to be optimized for uh for the chip and uh the AMD team has been really accommodating in terms of helping us out whenever some issues uh arise and then we went live a few months ago with our four 4.0 model.
How much does the road map at AMD and other semiconductor companies uh affect how you're designing your product? Do you work backwards from this model, I know it's going to be able to do X, Y, and Z, hold this sort of memory, this flops like, and then okay, let's design the best model for that rack or that chip.
Yeah, exactly. So we we often customize the details of an architecture based on price per performance latency and then work backwards and train the model when we uh get to the uh details of the architecture.
Uh there are certain changes that you can make uh that wouldn't have an impact on the quality of the model but would have an impact on the downstream performance and latency. So I think everybody does that in the frontier AI that you kind of work backwards. But the problem is you also don't know too much about the future chips. Um.
Yeah, maybe that's why you got to come here, meet the people, and whisper. Oh, really? That's what it's going to be? Okay, we'll work backwards from that. Or maybe you just have a proper, you know, deal and everything's above.
We didn't get to this, but uh before you leave, what were you doing before this?
Uh oh, I was at Google before this and before that I was a computer programmer, then turned AI researcher. Wow. And then uh started idoggram. At Google I started image generation. this brand called imagine and I thought I can have a bigger impact uh you know.
Uh how much of the team is researchers at Ideogram now.
Yeah we have a relatively small team it's primarily engineering.
And half of it focusing on the product and half of it is focusing on the model.
Sounds like a dream well congratulations on the progress thank you so much to you.
Thanks for coming on we'll talk to you soon.
Uh let me tell you about MongoDB what's the only faster than the AI market your business on MongoDB. Don't just build AI, own the data platform that powers it. And I'll also tell you about Codeex. Codex is a powerful workspace for getting work done with AI agents. Whether you're writing code, analyzing data, creating content, or automating business with workflows, Codeex helps you move projects forward from start to finish.
Welcome back. How you doing? Great to see you. How you doing? I love it.
Just saying hi.
Amazing.
Thank you, man.
What's happening?
Live.
What's up, boys? Good to see you again.
Good to see you.
Good to see you again. I think this third time.
Oh, man. That jacket is so sharp.
The green.
The TVPN green.
I That's the TVPN green. Okay.
We'll We'll coordinate. We'll connect you to our tail.
Do Do you have a Do you have a brand color yet for AMP?
We have We have like the the cream and the for right there. We actually have a Forest Screen coming.
Okay. There you go. There's a a new video model that just got announced this morning called Flux 3 that's training on the AMP grid and uh it's from Black Forest Labs.
Very cool.
And they're up in the Black Forest like near Fryberg in Germany.
Wait, say Amp Grid again.
The Amp Grid.
That's a good sound effect for the AMP Grid.
I like it. It's a good branding.
We'll get you We'll give you a sound effect.
We'll have our tailor talk to your tailor. We'll figure it all out.
Yeah.
Careful what you wish for, man. I would love the recommendation on that. Black Force Labs, their color is forest green. Okay.
Yeah. Because etc. Um and I find it's a very soothing color.
Yeah.
There's some like I think there's some like neurosychology around like green immediately calms you down.
Yeah. I I.
You don't want caution yellow or you know high viz orange that's a little bit more aggressive.
Well, there's some companies that make orange work.
Yeah. Well, that's not a Anthropic does not have a high viz orange. Although a high viz orange is maybe the one uh white space in in branding. A16Z is orange.
But that's not high visive orange. You know what I'm talking about? You see it on the road. That is an aggressive. That's hard to pull off as a brand.
How are you thinking about the brand?
You do you have a whole brand book yet?
Uh we're we're going with the vibes and the aura, man. It's like
I like it. As I've learned just the vibes generally the vibes like are in the space of Frontier which you guys have been an extraordinary voice on. Stuff changes every day.
Yeah. And so nobody's got any time to look at the first principles, the details, what's actually happening, what's like which model's coming out, what's the difference between Fable 5 and like CEX blah blah blah like and so it's all aura based decision-making right now because the brain just needs some easy heuristic to.
I'm surprised we haven't seen a chief aura officer.
That's that's the next like chief chief brand officer. But when somebody does that sell everything. So, it's it's probably too much once there's a new.
But you make a good you make a good point which is that if if you if you have if you have motion doesn't matter what your site looks like or what your logo looks like people just like attach their own feel what they feel when they're interacting or to to whatever you're putting out.
Motion is underrated. There's a there should be an AI company research in motion. That would be good.
Yeah, anyway. I'm not that. So, but but I want to answer your question which is what is the brand? The brand is what you know you stand for and we stand for output maxing.
There's just too much like what are we living through right now? We're living through the AI scaling era.
Bitter lesson holds that scale scale scale.
But the brain is naturally anchored to large numbers.
Yeah. But nobody like if if you actually deconstruct what's going on this space there's so much wastage because nobody's doing the output maxing efficiency right which is what's the output divided by the unit of input.
And if you me start measuring businesses or model capabilities or whatever through efficiency the story is completely different than than what just the headlines say.
A lot of people were token maxing and I think we will look back on that era as really valuable exploration. Actually totally worth it. Uh, but I like output maxing as the correct coinage as opposed to token maxing. I was trying to say token minmaxing, token optimiz. It's way better output maxing. I love it.
It's like what are we all here for? It's to try to grow GDP. Not at all expenses. It'd be grow GDP in the compute optimal way. Yeah.
Right. You cannot beat China at pure scaling.
It just not going to happen. Industrial man like industrial scale coordination where they go, you know what? Ban the H100s, H200s. all of you use Huawei and like get together and figure out how to scale with even if you don't we don't have the leading edge chips put out Kim K3 I don't care I'm Xiinping like I just want for us to be at the frontier and make you know 100 new nuclear plants whatever build build build baby like you know build scale whatever is required you guys the permitting who needs permitting just like you know they will just clear all the bottlenecks that they have been at a scale we don't understand right they will bring at least 100 gawatts of new energy capacity online in the next decade without even blink blinking.
Meanwhile, they were already they weren't even AI pill and they were planning on doing it anyway. They're like, "Yeah."
Uh so what do we have to do? We have to be smarter. We have to do output maxing. We got to be more efficient. Only 15% of every single tenant data center today in the United States is being utilized. Net flops utilization is less than 20% in the United States.
Is is that just downtime from chips not being used at the right time like misconfigurations like the water supplies out or something? What's going on?
Basically, for every dollar of long-term lease that an AI lab takes today, you lose about 30 40% flops in just bad scheduling. The nodes are just straight up not allocated. Y and then it's there, but it's just not getting work at the right time.
Then within the chip, you only have like 15% of the chip being utilized. That's called MFU model flop utilization cuz the chip is waiting around for some other process to complete. storage, memory, networking or some other chip to hand it off. So as a result, if you compound those two things, of every dollar, only 15% of flops are actually being utilized. That 85% that's a national security crisis.
Because that's what's keeping us from staying at the frontier.
That's why we all keep complaining about, oh, communicate 3 came out. Sure, you could talk about distillation. Distillation is definitely happening. However, everybody who feels computed in the United States should be asking, am I doing both? Am I getting more capabilities? Am I am I building up more supply and thank god Lisa is is on it?
But also are we utilizing the existing capacity to get the maximum output possible and both those things are what's going to keep us at the frontier. Yeah.
Yeah. It's sort of I mean it's natural cycle. Any company that that has grown really quickly or any industry that's grown really quickly, there's just ends up being all this waste and misutilized resources and all this stuff. And then but but the key and I think what you're pushing for is like even if you have this exponential growth you can still try to be more it's still important to be more efficient now so that we don't give up the advantages that we do have.
I don't think we have a choice because the supply chain is backed up for like two years. So we're just not like we at AMP the AMP grid we're procuring energy and new capacity for 2030.
And we and there's not enough sites.
Um we we're going to need nuclear. Where's are you spending time in nuclear companies?
Yes, actually there's Seth right over there. Hi, Seth.
Seth is a founder of a company called Maver that we we've just invested in. I don't think.
He was pop on the show. Pop on the show after.
Yeah. After he gets off. Well, yeah. We'll talk to you.
Go over there. Talk to me. A good time.
Yeah. Yeah. Talk to the team. Seth was at the DOE.
Yeah. um and was one of the people responsible for a bunch of permitting that just finally got like approved for new nuclear projects in the United States. Cool.
And he's Thank you for your service, sir.
Okay, we'll we'll give you a proper mic in a couple minutes. We got one more question. We I know you.
Yeah. So So walk us through the shape of your business right now. You have the fund, you're making new investments. uh you got a big allocation in I think obviously every you know anthropic ground to date uh but otherwise like what is the shape of the business you're building out like you're you're to me it's like two different businesses they work they're very interlin but how are you spending your time.
Yeah um so we've got AMP foundry which is our venture capital firm and capital business and we raise venture capital funds in.
You don't make a single chip there.
We do not make any chips yet.
Not yet who knows cuz I I wouldn't put it past you.
Lisa's a great partner. Uh on the but our our infrastructure arm, which uh we're actually um spinning out pretty soon as an independent entity.
Uh and we're not ready to announce the name yet, but you guys are going to love it.
So, that's for a followup. Um is building out about 2 GW of capacity in the United States.
And uh we're not making our own chips yet, but we're building out new uh there's new sites. Y.
Yeah. You're you're marshalling the capital, getting the right partners in place so everything can come together and the compute can come online.
Uh speaking of two gigawatts, uh what can you tell us about this anthropic AMD deal? I think it's 2 gawatts. $5 billion investment.
Oh, six gigawatts. Is that what it is?
Um but I'm interested in first, it just seems like another positive sign for the industry for anthropic and AMD. Uh and it feels like we're also I don't know. I haven't checked all the reactions, but it feels like are we post uh post circular deal FUD? Uh this seems like a very logical deal. You know, everyone's seen the ARR charts. Everything's growing. Businesses are using the tools. Like the full it's a full ecosystem. The economy is growing and so this part this type of partnership makes sense. But what do you tell to what how are you telling an earlier stage founder if they're trying to do a strategic deal?
Um well well past the circular part I mean these were all concerns maybe two years ago right agree um and so if you look at the value creation when you have like entropic go from zero literally when we started the business like five years ago to well north of 40 billion in run rate right in then and and most of that revenue is coming from everyday customers using cloud to code that's not circular that's net just new value creation right now the question is how How do you keep a healthy independent ecosystem of all kinds of frontier capabilities churning?
And the problem is that the the core problem one of many core bottlenecks is that private uh credit markets do not see startups as investment grade counterparties. So when they're trying to procure capacity to compute capacity like 3 four years out, they just can't get those contracts are not financable. And so what we need in the United States is a financing program that says, "Okay, we're going to be able to um turn startups, I mean, Antropic was a startup not too long ago that was not considered investment grade. Um at this stage, they actually likely are still not investment grade. Neither's OpenAI, but they're starting to become more underwritable by the financial markets, right?" And I think what we've got to do in the United States is do a little bit of innovation to go, how do you get the ecosystem to understand the quality?"
Of of these startups. Um they're the engine of innovation. There's where you know Anthropic is 5,000ish people today. Google is uh last time I checked like 60,000 people. Accenture is 75,000. And neither Accenture nor Google are responsible for frontier model innovation today. Like Google still hasn't put out a anthropic grade coding model.
And they're trying their best. But it's a focus thing, right? It's a focus culture like talent dance teams that are that are focused get more done.
Speaking of that, are large founding teams underrated?
Are large founding teams underrated?
Because there's a world where anthropics like the exception of the rule, it's working for them. Crazy. I mean, young company, but huge company. All the founders still there.
And is another great example. And as five uh and you know, strong team. Uh there was a while where you know the the the Silicon Valley wisdom was like two maybe three, right? Um yeah I most startup advice as you guys know is designed for the media fun for podcast.
Yeah. Yeah. But the outlier businesses they they have large founding teams. They have family founding teams like the stripe brothers. The conventional wisdom is don't start you know like business. Yeah. and Dario and Danielle are brother and sister. The Collison's are brother and sister.
