Cerebras IPO, Warshtime, VC Infighting, H100s in China, Robot Monkey Business

I see a large IPO on the horizon.

You're surrounded by journalists. Hold your position.

Right.

There's misinformation.

Clearing order inbound.

Let's just roll.

We are surrounded by journalists. Hold your position. Come on. Get up.

Trust the experts here.

We are experts found.

I see multiple journalists on the horizon. Stand by.

UAV online.

Blaze.

Double blaze.

Triple blaze. Double kill.

By coming is wins.

Team Deathmatch.

We are experts. Triple blade. Let's just roll. Right.

Hockey clearing order inbound. Come get up.

We are surrounded by journalists. Hold your position.

Strike one. Strike two. Activate golden retriever mode. Another one.

5.

You're watching TBPN and today is Thursday, May 14th, 2026. We are live from the TBPN Ultradome, the temple of technology, the fortress of finance, the capital of capital. Always. Hey, Ben. That is Ben. Indispensable. You got multiple Ben over here. We have a great show. It's Cerebras day on the show. Cerebras IPO. We'll talk about that. SemiAnalysis has a fantastic deep dive on the company. We'll go through that. Doug Laughlin from SemiAnalysis has joined the show. And then the founder and CEO of Cerebras is joining the show. But we have lots more folks joining. Amy Reinhardt from Netflix, the president of ads. Can you imagine? I didn't think, you know, you think about presidents, the president, you think president of the United States. I think president of ads.

That's right. Ben Hilac, Eric Visra, Steve Valo. We got a bunch of folks coming on the show today. It's going to be a fun one. So, there's a ton of news. Let's start with Cerebras. The IPO has gone spectacularly well. Cerebras doubled their valuation basically overnight. Brandon Gell had the good fortune of writing up some of the details of the Cerebras news in the newsletter today. TBPN.com. You can go sign up. Yeah, right now it's sitting at a $64 billion market cap. Yeah. And a lot of the prediction markets, they didn't even have a category above 50, right? A lot of people were just kind of trading or betting.

And when I wrote the newsletter Friday, Monday, I said a $50 billion IPO and was sort of being optimistic and it beat those expectations, which is great news. They deserve it. It's a true overnight success. We'll show you some charts of the valuation. Lots of troughs of disillusionment, but Andrew and the team powered through and wound up finding the perfect application for their technology at the perfect time during a mega cycle, which we'll go through. So, chip design company, Cerebras, if you're not familiar, they make a big, big chip, big chip company. Instead of taking the wafer, putting a bunch of chips on it, cutting it up into smaller chips, they use the whole wafer. It's a genius idea. It's one of those simple ideas taken deadly seriously in some ways. But it's trading at $350 a share on its first day of public trading which values the company much higher.

Three 300 now. 300. Yeah. 300. Okay. So it was okay. It was up at 350. It has since sold off slightly. And they raised around $10 billion. I think they were targeting six billion at one point. They've upsized that. I think there was a $3 billion raise initially, but they have a good amount of money in the bank now. The price on this IPO has been literally up only. On Monday, the price range was 150 to 160. Then they raised it, that was up from 115 to 125. And today we're seeing, you know, much higher prices.

Go back to that picture. Who's in the picture at the NASDAQ? Picture at the NASDAQ of the Cerebras team standing on stage. Someone should make a set in LA. You know, they have those like fake private jet sets. Imagine if entrepreneurs could have a set where they put their logo in the background like they're hitting it with a hammer and there's confetti going everywhere.

Yeah. But it's for your course. Yeah. Yeah. And you walk right from there to the Lambo. 1,000 students in your mastermind. Yeah. Yeah. No, I had this idea back in the day when, do you remember the ice cream museum? This whole thing. Oh, yeah. So, there was this trend. I mean, really bad news for the museum industry, but they're getting eaten alive. And so, some entrepreneurs, I think they did very well, started something called the ice cream museum, which was not really a museum in the sense of like a presidential library or, you know, the Norton Simon or the Getty or the, you know, natural history museum. It was more of like an experiential place to go and hang out. Good for first dates, good for, you know, taking kids maybe. And they would maybe give you some ice cream, but most of the museum was just like very Instagrammable things. So, there would be like a ball pit or a bunch of raining confetti and stuff and a huge fabricated statue of ice cream that was not a piece of art that would be sold.

Sprinkle ball pit. There you go. That sounds real. I don't know if that's real, but it sounds very believable. They had that. Okay. Yeah. So, and there were a number of other kind of copycats that were trying to jump on and do like, oh, we'll do like the waffle museum or something or the pancake museum, you know, because they just wanted to cash in. And my idea was just the museum of Instagrammable objects. And so it would have all of those. So there would be a private jet set and then there would be a Lamborghini set. And this one would fit right in. So it's just they have a big pink wall so you can go take the pink wall photo. And then there's a beach. And then there was a gym with fake weights so you could go and look like you're maxing out and benching 500 pounds. And so it just says, "Bring these clothes or we'll have them for you." And then you move from room to room taking the ideal dating profile photo reveal.

Yeah. Exactly. Exactly. Oh, you had kids here. You in the hospital. You can live an entire life through this fictional museum of Instagrammable objects. More of a meme than a real business idea. But no, John, the Museum of Ice Cream now has seven locations. Okay, so they're cooking. They're global. They're global. They're doing well. Well, let us know. Anyways, nice little tangent there. Miami is the capital of Gimmick Museums. Got to do a whirlwind tour. Anyway, let's go back to the serious stuff. Cerebras, it's a complicated company because they are so deep in the AI supply chain, but we'll break it all down for you. So, SemiAnalysis has a fantastic deep dive. It's a longer read, so we're not going to go through it all, but there are some very interesting tidbits in here that we can sort of summarize and contextualize for you. And then of course we'll be talking to Doug Laughlin from SemiAnalysis at 12:30. So, there's a bunch of interesting takeaways, some really solid positives. Cerebras chips work, which was something people were not expecting for a while. There was a lot of FUD around this company, just the idea of like, oh, that'll never work. Like what if the architecture changes? What if we go away from transformers or something? What if we need something quote completely different? Or maybe like the yields will never work because there was this idea that if you're using the entire wafer, typically as you're etching the chips onto the wafer, sometimes there's little defects and it's not a problem if you're going to break up a wafer into like 64 chips because you just throw away one. But if there's one defect on basically every wafer, well then your yield is going to be super low. We talked to Andrew about how he solved that by creating redundant cores and they don't actually activate all the cores and so they sort of built in that redundancy and got through that. But that was an early critique of the strategy.

You can use Cerebras chips today in Codex 5.3 Spark, and so they are very fast. And I think the most important thing that SemiAnalysis points out is that token consumers, customers, businesses have shown this revealed preference for and a willingness to pay for speed. And they sort of contextualize it and they quantify it based on their own usage and their experience with Anthropic's Opus models. So Opus 4.6 fast mode, famously, I like that they use famously because it's like famous to like 100,000 people, but famously charges six times the price for two and a half times the interactivity, although it's now under 2x faster. So effectively you're paying six times the price for two times the speed. That's disproportionately more money for what you're getting. You would think you'd pay six times the price for six times the speed potentially. But there were a lot of questions about would people really pay that much more for faster models, faster inference. And Andre Karpathy, Sam Altman was saying like, do you want faster models or smarter models? And he was like, I think in Sam's point was sort of like these models are very intelligent, but using them faster is sort of more of a magical superpower. And Sam was, I felt like Sam was sort of leading it towards like speed is really important as the next leg up on productivity. And Andre Karpathy was like, no, I just want smarter. I'll just let it run overnight. I don't mind that. But that's not what everyone is feeling. Some people, especially the SemiAnalysis team, leaned more towards interactivity or speed over raw intelligence power.

Well, yeah. And then there's the other aspect which is just capability, right? Capability, speed, and intelligence. Yeah. I think like that's the question I think people have had is like, okay, is there a 250 IQ model or is there just a much more capable model that uses tools more efficiently and is really quick. Yeah. And that's actually important to Cerebras because as we'll get into, the chips do face a hurdle with scaling in terms of longer context windows, all that stuff. But SemiAnalysis shared their breakdown. They were running $10 million on AI spend in April and they said that April was the peak, which is interesting because that was sort of, I would have expected a straight line continued growth, but they might have, you know, been really excited, tried it all and then eventually said, oh, okay, well for this we can probably use a cheaper model, we can probably optimize. We don't want, you know, 95% of our revenue going towards tokens. We want a little bit of margin to pay our team and obviously it's a business, they need to make profits. So SemiAnalysis was spending 80% of their AI spend on Opus 4.6 fast. And so they were willing to pay that 6x, like 80% of their spend disproportionately more, even though when their sort of expectation as they put it was that they would always want the smartest model. They would be very cost-conscious. They were in reality saying, I'm going to hammer fast mode. I want to spend on fast mode. And then I think the price was significant. And so there's probably sort of a renegotiation about when is the right time to use fast versus when do you want to leave something running overnight. But OpenAI is clearly very bullish on Cerebras. Cerebras has a big 750 megawatt deal with OpenAI and the chips are already serving GPT 5.3 in Codex under the name Spark as we mentioned. And I've used it. You should use it. It's a very interesting experience because I think a lot of people have interacted with LLMs and chatbots and they're sort of used to this token streaming in and it's sort of cute because the phone vibrates and it feels like you're talking to someone who's typing. But it's way better when you just land on a Wikipedia page, the full thing loads and you can just scroll however much you want. And that's the experience that I think people want and will demand across everything, especially if they're firing off a coding task. They just want the code immediately. And so you can also just go talk to the model like it's ChatGPT. You don't need to use Codex 5.3 Spark in a coding context. You can ask it whatever you want and it will just act like a normal LLM. And I personally think there will be huge demand for faster inference across all parts of the AI economy. There's this old late...

Yeah. Another another way to think about it is like if you have two employees with the same skill set, the same capability, but one is just five times faster, right? That person can create way more value in the organization. Yeah. Right. And for a lot of things, if they're if they're two times faster, they do command six times the price because over the course of the year, a sales rep that that that sells twice as much or someone who is twice as effective at their job might actually command a salary that's five times, six times um the the actual price. And so there's lots of other uh context across different business lines that uh you could draw to.

Uh there's also this old uh adage or saying about e-commerce that might may or may not be real, but it's probably been transposed so many times in think pieces. I don't know the real quote, but it goes something like every 100 milliseconds of latency costs Amazon 1% in sales. I don't know if that's the right way to think about it, but basically as Amazon was scaling, they realized that there were a bunch of things that they could do on the UI side, a bunch of things they could do on the layout side. Where does the buy button go? Where does certain information go, the price, the discount, all of this stuff, the images, they they were tweaking the front end. But as they did that, they added bloat, and the pages would slow down. And what they noticed was that the slower the page was, the lower the conversion rate because people were waiting for amazon.com to load, click on the page, it takes a second, they get distracted, they go somewhere else.

And I think that that's happening in LLM use cases all over the place. People fire off a query and they're like, "Ah, it's taking too long. I'll go scroll Instagram reels. There's always an Instagram reel." And they'll be like, "Ah, I I kind of forgot about what I was asking about. I didn't get my answer." And so, uh, and that's certainly true in business contexts as well. So, um, this is currently playing out in AI inference. Uh, companies are paying disproportionately more for faster inference, and this is good for Cerebras.

Uh, but SemiAnalysis does point out a number of potential headwinds and problems that the team at Cerebras will have to solve or contend with over the next few years. Uh mainly Cerebras chips are not currently as capable of holding larger models uh in the limited memory that they have or networking multiple chips together to serve larger models. We've heard about the NVL72 racks that wire a whole bunch of Nvidia chips together can serve these really large models. Uh that has potentially been a challenge. So uh SemiAnalysis says moreover the industry is trending towards larger context windows ad infinitum. 128k context will certainly not be acceptable for long especially with the prevalence of agentic workloads and it doesn't look like there's a simple solution of just scaling the wafer size larger because TSMC is set up with a standard wafer size uh and adding or adding more memory to the existing architecture because Cerebras' whole design depends on a lot of SRAM, static random access memory, directly on the on the wafer.

But SRAM is no longer shrinking as much with each new semiconductor node. So the last version of the Cerebras chip, they've done WSE 1, 2, and 3. They're on 3 now, but WSE 2 had 40 gigs of memory. WSE 3, you would expect, oh, we want a doubling, right? We wanted 10x or something. Uh, it got 44. So a 10% increase over one process node, one iteration. And uh, so analysis is asking the question of like, okay, is there an easy way to double this? Is there a question like how will this scale as the models get bigger um to add more SRAM? You might have to sacrifice compute area because everything is being done on one wafer if you want computation or memory. There's a direct trade-off because you only have so much space on the actual wafer. And so uh they they there might be much harder 3D wafer bonding approaches doing stack stuff like that. There are other potential ways but there's not uh like a linear oh yeah of course the next version's going to double again and so that is a potential problem that they need to work through.

Uh but in an agentic workflow I think it's entirely possible that you want like the biggest most powerful model like the vice president delegating things. You want the vice president junior, senior vice president, uh maybe just the president uh handling the critical work. Uh so future future models might not and that might not be on Cerebras, that might be on an NVL72 or TPUs or something. Uh but uh I imagine that we will quickly jump from the agentic age where you're firing the best smartest model at the full workload to the orchestration age and there will be hybrid approaches. So the biggest and best models will delegate certain tasks to smaller faster models just like they go and do database queries these days or they go and search the web these days and that's CPU bound.

There will be certain workloads that the larger smarter agent model like the boss model can sort of delegate to the Cerebras speed workers, the faster workers. So if you need to, I mean just yesterday we were I I was asking Ben about like uh he pulled all our guests together. It's like I'd love to know the geolocation of every single guest. And it's like, okay, run that same inference query of like look up this company, figure out where it is across a thousand or 2,000 uh individual rows and something like that's highly parallelizable if you want that to happen really fast. It might not need a GPT 5.5 class model, right? Uh it might be okay uh to run faster on a smaller model that works on Cerebras. Uh and so it's hard to predict the exact mix of chips that will power large networks of agents. Uh but uh these different designs to me they seem more complementary than directly competitive.

Like a year or two ago when uh Daniel Gross wrote AGI Bets and was sort of like is Nvidia underpriced? I don't know if Nvidia he might have said that on on SemiAnalysis on Stratechery but uh like you know we we entered the AI age and everyone was like oh GPUs are the future. Nvidia is the company, but then it was like Nvidia GPUs are good and then also CPUs are good and and ARM is getting into it and Intel's doing very well and it it...

We're going to make big computers.

Big computers, big computers for sure. Uh and so uh I'm extremely optimistic obviously and very excited to talk to uh Andrew Feldman, Doug O'Flaherty, Eric Vishria, a bunch of folks who have been involved with the journey. Overnight success, Series A in uh almost a decade ago, 2016, uh and uh Benchmark are sitting on I guess as of this morning like many many billions of Cerebras.

It feels like they brought a huge team to the Nasdaq. Look at this photo, the second post we have. Um, uh, obviously Andrew is there. A lot of the team members, typically the the the the key banker, but, uh, you know, we we've been to some IPOs and, uh, some of them have have had smaller teams. This one feels extremely celebratory and, uh, feels like a very broad inclusive crew came together. Um uh what else?

Founders in control.

Uh Honam says, "This IPO illustrates the power of an individual partner over the brand name of the firm. Pierre Lamond was a partner at both Sequoia and Kleiner, but instead of those firms backing Cerebras, it was Eclipse, the firm he joined at the age of 84, that backed this little-known chip company multiple times in the early days. What a way to wrap up a career. He was born in 1930, the same year as Warren Buffett."

Wow, that is an awesome story. I love that. Uh uh Matthew Seagull is giving, the recovering CFA is giving some color on what happened to the order book. Uh one-third of the order book, the folks that said I want shares in the Cerebras IPO, one-third of the book got zero. And the top 25, I guess the top 25 investors took 60%. That's probably the big investment funds, the Fidelities, the State Streets, the BlackRocks.

Yeah, it was very much just like another day at the office for the team. Uh yeah, that's definitely what I was contrasting it to. Uh the Cerebras valuation every round: Series A in 2016, $100 million, Foundation, Benchmark and Eclipse co-led. The Series B in 2016, WndrCo led. The Series C in 2017, then 1.6 billion valuation in 2018, 2.4 in 2019. Uh 4 billion in 2021. Uh that that was like maybe a little bit of a slump. But then 2025, Tiger and Fidelity come in at 8 billion. Then Tiger comes in at $23 billion. Then in May of 2026, they IPO at 48.8 billion. I can sort of just do the work myself. I don't I don't...

I don't need any help. I don't need the soundboard to do it. Am I doing it for real or am I using the soundboard? You don't know.

Yeah.

Very, very good.

Very, very good. Well, our first guest is joining us in eight minutes. Uh let's run through uh the Kevin Warsh news because he has been confirmed as the Fed chair. Uh and we'll run through this and then we will come back to some of the other stories uh because we have a gap later in the show. So uh Kevin Warsh who uh is most famous for interviewing Alex Karp on CNBC while Alex Karp appeared to have popped a nicotine pouch and then spun a notebook on his finger. Did you ever find that clip, Tyler? Is that in the...

Yeah, we have the the video here. It really put Kevin Warsh on the map because uh this is what he's known for. Of course, he's recently...

He's really spinning it. I didn't realize he goes back to it like four times.

Okay. Wait, wait. But so, uh, I have so many questions. First, we have to get him to recreate that for sure. Um, second, uh, I I thought Tyler, I thought we were talking about that being on CNBC, but that looks like just a podcast. Like that doesn't have any chyron or...

Oh, okay. So, it was just like a random podcast and then and then uh when I've seen it on CNBC, they were playing the clip. Got it.

Probably a reaction stream over there. Um well, let's go through what happened because Kevin Warsh has been confirmed as the new Fed chair. The vote was 54-46 in the Senate. Uh the divided vote signals challenges ahead for Warsh who faces a Fed committee skeptical of rate cuts that Trump has demanded. And of course, we talked about the inflation news. Typically, you don't cut interest rates uh going into inflation and potentially economic stagnation. Uh you definitely don't cut rates in that's why stagflation is so difficult because if you have uh stagnation and low inflation you can cut rates very easily maybe the economy starts overheating a little bit you get a little bit of inflation but then you can pull back that's what we've done historically.

Vice versa if the economy is running hot you're seeing high GDP growth and high inflation well if you raise rates you're going to pull back on both of those but in stagflation you're seeing both inflation and economic stagnation harder to deal with as a Fed chairman, which is potentially the task he will be faced with. So, uh, the Senate confirmed Kevin Warsh as the Federal Reserve's 17th chair Wednesday in a largely party-line vote that reflected how tensions with the White House have dragged the Fed deeper into the political fray. I was looking back at the old Fed chairs. There's some absolute legends in there because some of them had really long runs. So very quickly you get back to the black and white portrait and the painting as you go back in time.

Volcker.

An absolute dog.

Um yeah, I don't know. Hard to pick. Hard to pick. Uh Warsh, who was known, who was nominated for the post by President Trump in January, won...

Confirmations, 54 to 45, earning support from all Senate Republicans, but just one Democrat, John Fetterman of Pennsylvania. Senator Kirsten Gillibrand of New York did not vote. No Fed chair has been confirmed by such a narrow margin since Senate approval became a requirement for the job in 1977.

Chair Jerome Powell, whose leadership tenure ends Friday, captured at least 80 votes in Senate confirmations for each of his two terms at the top of the Fed. Wow, Jerome Powell, just a fan favorite of both teams. 80 votes in the Senate, that's pretty significant.

Pretty cool.

Yeah. Is he going to be looked at as a future podcast throughout history? Get a maybe VC fund going. What do we think?

No, but when we look back, it seems like the last two, three years he's handled himself.

Yes.

I mean, he's had a really tough situation and he's...

When did he land the plane? When did Jerome Powell first become the Fed chair? When was he? He assumed office in 2012. So I think the... wait, oh...

2018.

2018. He was right after Janet Yellen. And so I'm putting him in the conversation, Jordi, but I'm not giving him the GOAT trophy, nearly because the challenges faced, he wasn't confronted with a great recession, a dot-com bubble bursting, a Black Friday. He took the economy from 2018 to today...

Global pandemic doesn't count? You don't count a shutdown of...

No, because I actually don't, because the economy was pretty strong in 2019 and it went into 2020 with pretty strong consumer balance sheets, low debt. There wasn't a shadow banking economy, there was no bomb in the US economy waiting to explode. And so although we saw high unemployment briefly and we did have to stimulate the economy, that's not his job. His job was to set rates. There was a little bit of like, I mean maybe you put the inflation, the end of the ZERP era and all of those gyrations on him, but the problems that were downstream of both the ZERP era and the end of the ZERP era were suffered mostly and benefited mostly on tech companies and Silicon Valley companies that had really long cash flow horizons. And so there was not a moment where it was a dire situation that the Fed had to intervene in a meaningful way and save the economy like in 2008. It's a big deal. He did a great job, but he wasn't faced with the same challenges of a Bernanke, for example. That's what I would say. Tyler, what do you think?

Yeah, I think that's reasonable. Also, if Powell was worse at his job and you saw some crazy crash because of COVID and then he brought it back, then it'd be like, oh yeah, he did face this massive thing. But because he did such a good job, maybe you didn't see any massive crash. So nothing super bad happening is evidence that he was really good as a Fed chairman, right?

Yeah, yeah, maybe he's a defensive back. If they don't score, there's no great plays because he's just shut down cornerback for the last couple years. Potentially, he's definitely in the top 17. I'll give him that. Anyway, Kevin Warsh, Chair Jerome Powell, whose leadership tenure ends Friday, captured at least 80 votes in the Senate. The previous chair, Janet Yellen, was a little bit more controversial, confirmed 56 to 26. Seems like not that many people showed up in 2014 to vote for Janet Yellen, with many senators absent because of bad weather.

Interesting. I wonder what would have happened. I mean, it feels like she would have cleared it no matter what. But a difficult economic backdrop and Trump's broadsides against the Fed's independence have set up the central bank for a thorny leadership transition. At the Senate committee confirmation hearing last month, Warsh faced intense questioning from Democrats over how he would maintain the Fed's independence from a president who places priority on personal loyalty. Warsh said he would preserve the central bank's monetary independence and that he had made Trump no promises about policy decisions.

Powell, citing concerns about political attacks on the institution, plans to remain on the Fed's board of governors. Defying Trump's insistence that he leave, he says, 'You're going to have to drag me out of here. I'm staying at the Fed,' says Jerome Powell. Warsh is 56. He's been immersed in monetary policy debates for decades, frequently as an outspoken critic of the Fed. Former Morgan Stanley investment banker, he became the youngest Fed governor in history at 35 when former President George W. Bush appointed him to the central bank's board in 2006. During the financial crisis that struck two years later, he played a key role in tying up rescue deals. He was the bridge between Wall Street and the Fed that Bernanke deployed, and so those are his laurels that he will not be resting on, but he will be drawing on from experience. So Warsh left the Fed in 2011. He had become a critic of its direction, concerned that as the economy recovered, the Fed's ongoing efforts to support financial markets went too far. So we will have to check in with the progress on Fed Chairman Kevin Warsh soon, but fortunately we have our first guest of the show, Amy Reinhardt from Netflix, in the waiting room. Let's bring her in to the TV show. Amy, how are you doing?

Fantastically. It is an honor to have you here. Our first ever guest from Netflix, I think. So thank you so much for taking the time. It only took like 2,000 interviews to get you guys on here, but we're excited to meet you.

Yes, yes. I mean, obviously big fans of both Netflix and advertising. But would love to start with a little bit of background on yourself, your experience, and just sort of your intro to how you found yourself as the president of ads at Netflix today.

If you take us back to the initial push into ads, what can you tell us about the trade-offs, the build versus buy debates that were going on at the time, maybe even the cultural changes? I think we are super, you know, we love ads. We think it's a fantastic business model. It's a way to deliver great value to customers at lower prices, and there's so many benefits, but culturally...

Yeah. What was the debate like?

Yeah. What was it like internally?

Into 15 more countries around the world in 2026. There you go.

Everybody's on board. Full speed ahead. How are you pitching the ad product today? Is this primarily brand marketing? Is there a timeline to get to more of a performance focus? What is your pitch to advertisers?

What are conversations like around how brands should, how much brands should want to associate with particular pieces of content? Because I think some brands might come in and say, well, I'm advertising strollers and I know that parents will be watching K-pop Demon Hunters with their kids, and so this is the most on-target and I want my brand linked to this particular piece of content. But we've seen time and time again that once the algorithms get good enough, once dynamic ad placement can actually flourish, every company tends to see better performance there. So where are the ad buyers today in terms of those trade-offs?

Yeah. I mean, it's interesting. Netflix has deep, deep experience in machine learning, AI recommendations, all parts of high-throughput data processing. But I'm interested in any learnings or surprises from building the proprietary ad delivery stack. Has it been as expected? Has it been that there's new skills that you need to bring? Because a lot of companies have been successful at scaling content and then struggle to figure out ad delivery. You obviously haven't. But then also there's this AI boom going on which can help with productivity, but also new algorithms and new ways to actually target content. And so I'm interested in where the build-out didn't match your expectations or surprised you.

than necessarily tech learnings because we're so used to that tech cycle of testing and learning and iterating. How are you thinking about the ad product feeding back into the content production? I've noticed that Netflix has been fantastic, in my opinion, of creating more engaging content. And I was watching The RIP with Matt Damon, and you click the play button and you see Matt Damon's face within like two seconds, and it clearly confirms that you're watching the right movie, and then the title card comes in, and that's a departure from 50 years ago. You watch The Shining and you know it's a helicopter shot of a car for five minutes, and they show you the full titles, and it is a different style of editing, and some people lament the old style. I particularly like the new style, and I'm wondering about — we went through a period of time when television, there was the famous like fade to black and then the ad break and then fade back in and you resume, and Netflix has never had to contend with that in media products, but is that going to come back? Is there a next generation pattern for creating content that can both have ads in it and not? Are you seeing glimmers of what the future — like the impact of ads might have on the editing structure and the timing and the pacing?

Yeah, a lot of it, to be honest, depends on our creators. So, you know, working with talent who — and some of that talent may be more tech-forward and are thinking through those types of things when they're writing shows. I'll give you an example. Shonda Rhimes was used to writing for broadcast and network for many, many years. So when she writes a lot of her content, she's already thinking about where those natural breaks are.

But not all writers do that, and that's okay. We can still find what are those natural breaks because we want to make sure, again, getting back to the member experience, that it's not intrusive or it doesn't come mid-sentence, right? And isn't cutting off any of the action. Yeah.

We're able to adapt to any way that our creators want to write the content and fit it into that member experience.

Yeah, for US markets. Is there any enterprise spending? I would imagine that a lot of enterprise buyers have been Netflix subscribers for a really long time, and maybe they're not getting served any ads at all, and so this is more of a consumer opportunity, or am I thinking about that the wrong way?

We've — most of our clients right now, the target segment that we're going after are enterprise top 400 clients, right? Because we think those are the ones who —

Oh, sorry, I meant B2B versus like B2C companies.

Oh, we think about this more as a B2C opportunity for the most part, and I think as we expand our learning and expand our offering, we may get into the B2B space, but for the most part B2C.

Yeah, yeah. I feel like all of that — like the higher-up market, more targeted — that's all unlocked with scale once more. There's more learnings on responses. I'm wondering what other signals you think might be valuable, because many times, you know, advertisements shown during a TV program are very passive, harder to track. But if someone's watching on their phone, there can actually be a call to action, a trackable link. Like, I imagine that the data is messy, but how important is it to sort of close that loop in an ad era where it's a little bit trickier, but there's a lot of things that you can do on the signal side anyway?

Yeah, you're absolutely right. And this is an area where I talked about the testing and iterating. We're leaning in a lot on the testing. What does that screen experience look like? You know, again, how do you meet the customers where they're at without being sort of intrusive? A lot of testing going on in the space, but the biggest thing for us is, you know, privacy-safe. We want to make sure that we're leaning into, again, that member experience and taking care of our members' data.

Yeah.

But a lot more, I think, to come.

Have you been surprised by the return of the QR code in maybe podcast advertising? But I see it a lot because people are watching on — you know, they'll watch a YouTube video on a TV, and the creator will, you know, hardcode in a QR code to link out. And that was something I had completely written off QR codes, and then they made a major comeback.

No, I agree with you. From a member experience, it may not be the most simplistic thing. Understanding kind of the ad tech on the back end, I'm not surprised by it because it gets to be pretty complex pretty quickly.

Yeah, and there's some — we hit some sort of like inflection point where the — maybe it was in a certain iOS revision where the camera app became so easy to press a button and pull out, and then it detects the QR code so quickly that that flow — because you used to need to like have a QR app separately to scan it. Now it's all integrated, and so someone can just whip out their phone and run right to it. Something else?

Nothing, nothing super top of mind right now. I mean, last question I had is: do you ever expect Netflix to serve more short-form vertical video-style ads in something like the clips tab? I know the clips tab right now is focused on basically like content discovery, but I imagine in the future people will spend more time in a format like that, especially on mobile.

Absolutely. And that is one of the announcements we had at our upfront yesterday, is that as we roll out this vertical video content, that we are going to be offering that to advertisers along with our todom.com coverage in 2027. So yes, we think that is a big opportunity too.

Last question for me. I would love to know about the intersection between games and ads. That's been a huge, huge growth driver in other categories and other companies. But I'm wondering where that is in the roadmap, how you're thinking about that.

We haven't thought about that yet. Look, our roadmap — I could fill our roadmap for the next two to three years based on just some of the foundational things we want to do and a lot of the innovative areas we want to lean into.

But it's an area that we're keeping an eye on. And as we watch that games engagement increase, I would never say never. I've learned to never say never at Netflix, but it's not something that's on the near-term roadmap.

Okay, thank you. Well, thank you so much for joining. This was a really great to meet you, Amy. Thanks for breaking it down. Progress for us.

We'll talk to you soon.

Cheers. Have a good one.

Thank you.

Moving on from Kevin Warsh, who is the newest Fed chair, we have a debate. We have a debate going on in the timeline around General Catalyst's advertisement. Some are calling it an attack ad.

Particularly Andreessen Horowitz is calling it an attack ad. We touched on this yesterday. But if you did not tune in, General Catalyst, the large-scale — what do they call it? Gigafunds now? What do they call it? Platform fund? But there's something else. It's a huge, huge venture capital firm. They launched an advertisement, which we can play again to refresh everyone. Let's scroll down to the beginning audio. It's only Thursday.

Hi, I'm GC.

And I'm VC.

Who's your friend here, VC?

This is Woof AI, an AI-native companion platform that combines robotics and machine learning. You'll never want a real dog after this.

Well, I think people like dogs as they already are, though. VC, you don't need to walk it. You never need to tell the kids you sent Milo to the farm. We're leading the seed and could probably make room for you. Well, I'd love to hear more, but we actually have a really high bar around responsibility for these things.

Is Woof AI okay?

Of course, he's fine.

Oh, sorry, buddy.

Easy, easy.

Stay. Stay. No. No. Stop. Stop. Stop. Stop.

I'm sure you'll be fine.

Okay. Tons of thoughts, but you kick it off. What's your read on this? Take me through it.

So, the actual ad, the way it's shot, the timing, cinematography — it's fantastic. I do — I think the ending is funny, right? It makes me smile a little bit. The dog's going haywire. The robot dog's going haywire. My first thought is I actually think there's a huge opportunity for a robot dog that is 10 times better than existing robot dogs. There are robot dog toys out there. So I don't —

specifically in the toy market, because this is a Boston — this is probably, I mean, it seems like it's a Boston Dynamics robot dog, and those are not typically used as pets that I know of. I think that they are more deployed. Whenever I see a demo of a Boston Dynamics robot dog, it's like walking through a nuclear power plant that you don't want someone walking into. It's like an industrial product for the most part, and then obviously fodder for viral videos. But yeah, step through it, because it is a crazy back and forth from both firms. First, the stats. General Catalyst — oh, they're over a thousand likes now, but two million views. So a lot of discussion. Any Midhoff says it's a bit cringe, guys. Olivia Moore at Andreessen says if I was a consumer founder I would run for the hills watching this weird take from a fund that returns so much money from Airbnb and Snap. Because is the idea that like robot dog is a weird idea, but so was airbed and breakfast or Snapchat, right? The biggest B2C businesses always start out looking weird. Yeah, there's a way to do the same pitch for airbed and breakfast, which is, you know, that empty couch, you know, that empty room in your house? What if you were to monetize it? What if you were to allow strangers to sleep in the empty room in your house? And of course that was a lot of the criticism at the time was that it never was going to work, and it was strange, but then it created a lot of positive externalities. People built businesses, people get to travel and integrate with local communities. Yeah. So the main thing here is like it's —

I do find it fascinating considering that when you actually look at the portfolio overlap, it is insane. Almost all of their winners —

Yeah, yeah, almost all of their winners, especially in the modern era. Sure.

They are in both companies, right?

Of course.

And so they're backing a lot of the same companies. So it's super hard to counterposition against them. Now granted, A16Z has a very different media strategy. They're very loud. They're all saying the R-word a lot. And I think GC can kind of counterposition against that, but you're not going to counterposition against like what companies you're backing, right? GC — I had to look this up because I thought there's no way this is true. GC's in both Ky and Polymarket — big, great companies, right?

Those have been at the center of the debates.

But that is the center — those companies are at the center of the moral debate.

Yeah.

In tech, right? Certainly for the venture-backed controversial companies, there are some that are controversial that are not.

Yeah, and all the funds have backed various like betting, trading-related companies over the years. I think it was like GV did — what was it? Not DraftKings, but FanDuel, I think, back in the day.

So it's not to say — there's companies that start out as controversial and then they become completely normalized and everyone sort of comes around to like, oh, that's a good thing. Like Anduril is a good example. Both of these companies — both Andreessen and GC are in Anduril. When it launched, it was like, wait, you're building defense technology? That's insane. And now the whole Silicon Valley community has come around to the idea that it's really important to have a functional defense industry. And then on the other side, you have companies that come out as really controversial and people are like, this is the end of the world, and then they just sort of fizzle out. Like Cluey would be a good example of a company that has done okay, but it's not like, oh, no one's doing homework. It's completely upended education and it's so successful and it's bad. It's like no, it's like there was a lot of saber-rattling around that. Sora went through a similar thing where everyone was like, this is going to wirehead everyone, and then it was like just some funny memes and ultimately, you know, moved on from it. And so it's very hard to map like the controversiality to the ultimate outcome and where it lands. Like a lot of the prediction markets were not controversial when they were just predicting the election. That was not what was controversial. Tech started being controversial about it or saying it's controversial once it got into sports betting.

Yeah. So the ad is interesting from a couple of ways. One is trying to counterposition against GC being like the cool, hip fund that wouldn't back the robot dog company, even though when you look at the portfolios, right, there's enough overlap that you can safely say if a company is ripping or has a lot of potential, they're probably both interested in investing in it regardless of what category it's really in, right? And then the other thing is that historically, like A16Z is the firm that's trying to be like hip and cool and loud and do like new media, and GC is the one that I've always viewed as like more buttoned up, more behind the scenes, more traditional fin. And I love that. We love that.

I think Capital is a good example of another firm that's been like pretty straight-laced but still has like aura behind it, and it's like they have a private equity fund attached to the venture fund, and like Bain's done some fun stuff. I mean, Fogo de Chão, and they've leaned into things every once in a while, but they've never been like, oh, we're going to be the craziest brand strategy. It's been like, yeah, buy the book investors. We find great companies, back them, and you know, Sequoia's done that too. You know, their whole pitch is very austere, and that's worked for a long time. It doesn't need to pivot the brand, which is maybe what this is signaling towards. This is the first time — I mean, it's very rare for any VC to like take shots directly at another VC just because they're all syndicating deals to each other.

It's way more direct. I mean, a lot of people clocked it, but because the actor could look — I guess if you were far away enough, like —

Yeah, Steve Ballmer looks like Steve Ballmer.

Or Marc Andreessen.

But people are saying Marc Andreessen.

But I was looking — so hold on, I was looking at Marc Andreessen photos and I could not find a single photo of him in a vest. So I don't know about this, but clearly Marc Andreessen did take issue with it because he quote-tweeted it something like 45 times, which really amplifies it and creates more of a conversation about it. Strategically, you might have just wanted to mute this if you don't want this to become a thing. But maybe he does. Maybe he's like, yeah, actually, I do want to fight because like this is dumb and I'm going to fight back and I'm going to win. So, you know, that's the strategy. But it does seem like messy rolling around in the mud when you're wrestling with pigs, you're going to get muddy, right? Isn't that the saying? Yeah, it's just — again, like creatively it looks great and it's very fresh, and I think Reggie and his team did great work, but it's funny to just like take shots at a whole category of investing basically. That is kind of the bread and butter of GC's business.

Yeah.

Yeah, yeah. Drawing the line in the sand. What is the actual line? The really high bar for responsibility around these things, like because — first off, it's just odd because like robot dog feels very, very low on like the responsibility. Like if some VC was like, okay, we take responsibility really — yeah, so I, for example, the last robot — I would be like, don't fund gambling, don't fund cannabis, like don't upset Sager and Jetty basically. It's like what the way I would pitch someone if they were trying to be the responsible actor — fun and robot dog would be fine.

The last robot dog pitch deck I saw was a company that wanted to use robot dogs as a replacement for actual seeing-eye dogs.

Seeing-eye dogs are like incredibly expensive, you know, kind of out of reach for many people, and so the opportunity for a robot dog that can maybe, you know, travel with you like on a plane without, you know, barking and needing to eat, right? Like that's actually like a very world-positive kind of endeavor. I don't know if it'll, you know, work or what the business will look like, but in general, like there's a, you know, anyway, so —

I'm extremely bullish on robot dogs. I think it's very complementary. Think about — I mean, especially if it's a toy for kids. Think about how many vehicles kids have.

Gabe says service dogs can reach 60K.

Whoa.

And that doesn't count lifetime of all the other costs associated with dog ownership with a service dog.

Flex like, whoa, buddy. You to flex like that. No. The 60 — the — yeah. I mean, kids have like, you know, a bicycle, a tricycle, RC car that drives, and then another one. And like it's just like you throw the robot dog in the mix. I feel like that's going to be in addition to a dog. Robot dog rips is what you said when I said that. This is going — maybe a thousand robot dog startups will flourish. And the most ironic scenario is because the industry will become so deniable, General Catalyst and Andreessen both have to back it and they're like duking it out for allocation at every round, and the future robot dog founder out there is like, I just do nothing, win, or build a great company and win. Was there anything else going on with the robot dog debate? I —

You know, selfishly, you know, I think it's bad for the industry if you have two of our platform funds, you know, just making these somewhat petty videos that are basically ad hominems that, you know, basically throwing stones from glass castles, as one could put it. So I think it's generally bad, but for entertainment purposes, if they want to keep going at it and turn this into, you know, proper Drake versus Kendrick situation, this might be a good use of A16Z new media. Yeah.

Although much of the talent there is now over at —

I mean, I would love — yeah, I would love like a response video from Andreessen shot in the same way or something. Like there is an opportunity for like a rap battle here. That's a beef that's like very, very entertaining that I would absolutely love. The flip side though is that I agree with you, like if you're a VC, you're much better picking a villain that is something around like a lack of technological progress. Like Thiel did this very well with like stagnation is the boogeyman. Like if we don't get the robot dog, we don't get the seeing-eye dog, we don't cure cancer, we don't do the big thing, we don't visit Mars. Like that is — you can still have like a villain, but the villain needs to be something that the industry and America and all the constituents can align around instead of like picking fights where like these two firms are obviously aligned on like 99% of like where the future goes and what the goal is of building a business that delivers a value that consumers enjoy.

Yeah, which is so funny because when you look at cap tables, they're almost always probably touching. They're effectively holding hands on the cap table, right? Because like one of them might have more of one company, right? Slightly higher. One of them might have more of the other, right? But in general, you know, they're just hanging out.

Ollie did Anduril at GC and then went over to Andreessen. And so like there's more overlap than differences. Plenty of other things to take shots at. But I mean, to Andreessen's credit, they've done a good job of that. Focusing on geopolitical competition, focusing on China, focusing on stagnation and nimbyism, all sorts of different things that they've been aligned with like more of like an abundance view as opposed to punching down. Although, you know, there's memetics all over the place. It happens.

Yeah, just it seems like GC has an opportunity to be the buttoned-up —

Yeah.

Platform. I think so. You know, they don't need to say the R-word. They don't need to be super loud, right? They can just focus on the craft of investing. I wrote about it in the newsletter, but you know, run some ads. You know, you've been saying this forever. Run some ads in The Economist, Financial Times, Wall Street Journal. Just — there's a way to counterposition yourself without attacking.

You remember the last time General Catalyst went viral? Hamant, CEO, was on Harry Stebbings' show, 20VC, and he said like triple, triple, double, double, double is no longer good enough. We want to see 10x-ing every year because that's what's happening in the AI era. And it was a brash statement. It was bold. It's smart to debate looks — but look at the results. Like, you know, like there are companies that are doing this. Like we talk to them every day. And sure, he clarified it on our show. He clarified it on other shows. He was not saying that like you shouldn't build a business that only triples revenue every year or only doubles revenue every year. He was just saying that the reality of the market right now is that there are power-law companies that are growing exceptionally fast, unprecedentedly fast. And so you as a venture capitalist have to adjust your benchmarks and think about how you're allocating funds, what companies you're investing in, make sure you're in the best company in the category that's actually going to win. And it might be the company that's growing 10x a year, not 2x a year. And so that was something where it was like thought leadership from Hamant. It sparked a conversation. There was some debate around it, but it was —

So far that take was controversial but entirely correct when you look at the growth of a lot of the most exciting companies in the industry right now.

Totally, totally. And so that sort of more narrow, staying-in-the-lane view — it got a lot of attention. It did break through. It caused a conversation, but it still didn't sacrifice it. It didn't feel like, oh, he's taking shots at someone specific, right? It was just like a market analysis from someone in a position to give that exact analysis. So anyway, Jensen Huang is over in China. Jason Calacanis has a photo that looks extremely real. Zero AI detected, but he's bringing two huge boxes of GeForce RTX 5090s, which are —

This is a picture from when he was in Alaska, too.

Jason says, never stop selling. I agree. There is some news which we will cover later in the week around the dynamics around H100 sales and Blackwells, what's actually happening. It's all in flux as the Trump-China summit plays out on the front page of The Wall Street Journal every day this week because it is headline news. High-stakes US-China summit kicks off. There was another drama in the tech world yesterday, but we'll come back to it after our next guest, Ben Hilac from Raindrop, joins. I believe he's in the waiting room, so we'll let him come in. He's the co-founder and CTO. We've had him on the show before. Welcome back, Ben. How you doing?

Well, man. How are you?

Fantastic.

Great to see you.

What's long time in your world? Reintroduce the company quickly and then tell us the news.

Sure. Raindrop — we make observability for agents. So the main thing we do is self-healing agents. So what it means is that when Raindrop hits a problem in production, we detect it, we fix it.

How do you do it?

That's a good question. So at the end of the day, like, we consider ourselves like the intelligence for your intelligence. What that means is that we are the best and fastest way to essentially look at anomalies. So what that means is that like, let's say you make a change, right? We're able to very, very quickly find out that like, oh, users all started complaining about something, or the trajectory, the traces are kind of starting to evolve into a different pattern. And so it's kind of a combination of agents but also more like classic ML techniques, a lot of like custom-trained models for every customer.

Walk me through the shape of the agent market right now. Like the way you're talking about it, you know, sort of illustrates the broad diffusion of agents and custom agents. I think that a lot of people think Claude Code and Codex, and I don't know if you're doing enterprise deals with those firms or that's the goal, but I imagine that every startup, many legacy companies have built some sort of agent, some sort of harness. And I'd love to know the shape of how broadly diffusing like custom agents are in companies versus — is it the domain purely of startups that create an agent for legal or an agent for sales and then they vend that into a company?

Yeah. So I would say that there's two kind of categories of customers. We started with super high-growth startups at the time — startups. So those are companies like Clay, for example, Framer, Speak.com, some of the fastest-growing companies in the world. And those were some of our earliest customers, and we're lucky they grew a ton. So, you know, that has helped our growth.

Always helps.

It always helps. Yeah. Someone once mentioned that like, you know, this kind of business is a lot like early-stage seed investing, actually. It's kind of interesting. Like, you know, you have to be pretty picky not to work with companies that are going to die. Because if like especially analytics, like all these sort of things, like you succeed as a company when your customers succeed. Like if all of your customers are terrible, it's like everyone's like, well, why do I — why are your —

I had a portfolio company that was working on like agent infrastructure roughly two years ago and pivoted because he was like, okay, this is clearly going to be a big thing someday, but right now he's looking at all the underlying companies and he's like, I don't believe that any of these like agents in their current iteration are going to work. No, maybe there's —

I think it was very counterintuitive at the time, but I think we chose to find companies like Clay.com, right, which were clearly on an insane trajectory, but at the time weren't necessarily as large. And so I think a lot of our customers now are pretty large, but at the time weren't necessarily as large. And then in the last few months, we've been moving into Fortune 50s, Fortune 100s, and like a lot of amazing things happening there. And again, it's kind of like two shapes of a product. Like one is like our bread and butter is like, you know, companies that are redefining the way people, you know, interact, you know, in different verticals. But then yeah, there are like Fortune 50s, Fortune 100s that are also deploying agents internally. I think the shape of that looks very interesting and like it's something that like being on the forefront of like understanding how these companies are deploying things — like there's not that much I can talk about right now, but yeah, always very interesting.

What do you think is a generally underhyped agent category right now? I'm sure you're seeing the future a little bit.

It's a really good question. I think that — I mean, you know, I — this is a tough question. I — what I want to do actually is pivot the question a little bit because I want to talk about our launch today, if that's okay with you guys.

Yeah.

Okay.

I'll tell you my questions. You tell me your answers.

Okay. Okay. Sounds good. Does that mean that you want me to not answer this?

No. No. No. I'm just messing around. Go for it.

Okay. Cool.

The joke is — yeah, I botched it. The joke is what questions do you have for my answers? And some CEOs show up and they act like that where it's like you could ask me anything and they're just going to redirect to a talking point. But it's fine. I want to hear about the launch today. So just tell us about it. Let's talk about it.

Okay. So guys, there's been this crazy thing that has been missing for a very, very long time. That's why I want to talk about it. So like people have been building agents.

Mhm.

You're building them locally. Like you're using some sort of SDK. It could be open, it could be for sale, whatever SDK it is. And —

What do you mean locally? Like it actually has to run on your laptop or —

Right, like before you push to production? Sure. Right.

Sure.

It's on your laptop.

Yeah. Yeah. Yeah.

There's no way to see what it's doing. Like no standard way, nothing. Like so people will send those traces out to like a server. Like Raindrop is like one of those, you know, and there's a bunch of others.

Yeah. But they might also just like drop the logs in like a non-relational database.

They'll just print it to, you know, console.log, like, oh, here's what was happening. It's like that bad.

And the other problem there is like so you can't see like a nice trace, or you're sending it to some server and it takes like seconds to see everything and like whatever, it looks terrible.

But then also your coding agent can't like see the traces either. So then when you hit a problem and you're like, hey, you know, this response was wrong. Claude Code will just make stuff up. Like it'll just be like, oh, I think that like maybe this tool was wrong or I think maybe like this happened. Because it doesn't have any of that data. It doesn't actually know what the coding agent did. Yeah.

So I think that like as someone building agents, as like our company building agents, we're like it's actually kind of embarrassing how long it took us to solve this problem. No one else solved it either, but yeah, that's what we launched today. Free, local, open-source tool. Raindrop.ai/workshop. And it's completely free. Like it's just open source.

Why open source?

That's a really good question. I mean, I think the genuine answer, and I think part of like why competitors haven't done it — I mean, there's probably other reasons for that as well — but I think it's that it can be, right? Like someone else can do it. You know what I mean? Like I think that it running locally is the best experience for people. Mhm.

And to be clear, like there's still things that it enables, you know, if you connect it to your production Raindrop, which is like you can pull in a remote trace and replay it. And then Claude Code or Codex can just keep doing that loop until it works. So there's still benefits for us. But also the truth is that we want people to hack it. We want people to meld it into whatever works for them. So I mean, we use a lot of open-source things here, right? So it makes sense to contribute back as well.

Yeah, that's great.

Um, yeah, I'm wondering about other just like predictions about the next breakout category of AI agents, what you're seeing. Feels like we're so close to being able to book a flight, but maybe no one wants that. I don't know.

I mean, I'm not sure if you guys saw I had a little bit of a thing with Ryan Chesky earlier about Airbnb. Oh, yeah. Talk about that.

We talk about that. No. I think like, you know, I use Airbnb a lot. I love Airbnb. I think if I had to guess, I would say Brian Chesky knows a lot more about Airbnb than I do and probably a lot more about being a founder than I do as well. And so I think there's probably a lot that I'm not considering. That being said, I think it's like if Airbnb had an API, I would use it and I would book Airbnbs with it like through Claude Code, right? So it's like I know I would do it. I find Airbnb very, very hard to search. And I think that there's a lot of — I think the tough part and like what I see industry-wide right now, everyone trying to figure out is you see companies almost reducing themselves into an API with like absolutely no moat. Like you look at like Photoshop, Illustrator, etc. They're like, oh, we have a Claude Code integration now. MCP. At the point where people are just using Photoshop, Illustrator, etc. as like an MCP, they've sort of lost the game, right? Like if no one's actually touching the UI anymore. I think that right now companies have to do that increasingly because they have no other choice. I think that there will be a point where the incentives don't make sense anymore. Like I can give an anecdote from when I was at Apple. You know, do you guys remember App Clips?

Yeah.

Yeah.

Here's my app. Very few.

Where did those go? I only see them go with like parking meters sometimes, I think.

So like one of the hero ideas there was like, oh, you know, like imagine you're in line at Starbucks, you don't have the Starbucks app downloaded. Like, well, why not just, you know, scan something, have the app go order your drink? And it's like turns out Starbucks doesn't want that, right? Like that's the last thing in the world Starbucks wants. Starbucks wants you to download the app. They want you to have stars. They have an entire — like there's a reason why DoorDash and Uber Eats and like whatever, you know, god knows other apps exist. It's not because they need to, but because each have companies and money and goals and like so why would they reduce themselves into an easily interchangeable API? It doesn't actually make sense.

Yeah. But I think it's important to be careful around using like a tool like Photoshop interchangeably with like a retail store like Starbucks or like a marketplace like Airbnb or DoorDash because I really think that these marketplaces provide, you know, an exceptional amount — all the value is not in the UI, right?

Agree. And where — and like the value of Starbucks is not that it's a pretty app, it's because they have specific drinks that they can like pretty much anywhere, you know, someone would be.

Yeah, I think of that company by the drink company. They got started during the direct-to-consumer boom. Obviously, they would have some beautiful Shopify website. They didn't. They just went direct to retail and they had Amazon. You could order it online, and if you went to their website, it would just say go to Amazon. And they did fine. Billion-dollar company. And because like the value is not in the e-commerce experience. They didn't play like the stars game. Of course, Starbucks is maybe sacrificing a piece of that business model, but it's not giving away the whole cow. I don't know what the —

There are going to be ways to monetize this, right? Like there are going to be successful business models built on top of this sort of layer. And to be honest, as Raindrop goes into the future, like that's the future we're building towards. That's the future we want. I mean, we're going to be announcing a partnership with a really large — one of like the large coding companies as far as like integrating with them more, where it's like I don't see Raindrop as a company that's going to submit PRs and production to people's code bases. Like someone else is going to be doing that. We're going to be the layer that's really good at finding those issues, diagnosing them, and tracking them. So I just think it's going to be interesting into the future how much companies are willing to sort of just like be the API without all those hooks, without the, you know, knowing everyone's email, like having the mailing list, like all that sort of stuff. So that's a very interesting trend I'm looking at.

I feel like you're generally on the frontier and cutting edge of like adopting all these tools. You mentioned your Claude Code use, and I'm wondering about — give me a reality check, a health check on your experience with computer use because you're lamenting the fact that Airbnb doesn't have an API, and I imagine you could create a scraper or download the HTML and interact with it, like treat the front end as the API effectively, you know, puppeteer the computer through computer use. Like where are you on like the AGI moment in computer use from what you've seen? Like where does it work, where doesn't it work, where would you recommend people get started if they want to play around with it?

Yeah, that's a really good question. There's definitely places where it works. Like I think that Codex has done a very good job of implementing like browser use, actually. Both for like debugging applications that you're working on and in general, like this is something that Claude just like doesn't do, like creates a really — again, that kind of like I think the next couple months the thing you're going to keep hearing from me but also everyone in the world is like self-healing loops, loops, loops, loops, right? Like how do you create loops where it's — how do you close the loop? How do you have like Claude Code make a UI change, see that it sucks, and then just keep going, right? And a lot of like do we have AGI or not is how many loops in a row can you do before — right, before things just end catastrophically, right? Because there is sort of this like it gets worse, right, in many cases. So yeah, I think like there's a lot of like ways to answer this. Like I'm a fairly like security-conscious person. I —

I think that, like, you know, I'm not like an open claw guy. I'm not going to give all of my like cookies to some, you know, agent, etc., etc. But yeah, I think...

DVD. Yeah. Yeah. Cool. Well, yeah. New challenge. Book an Airbnb with an agent. Can it be done? Is that where the goalposts need to be set? Let's figure it out. Anyway, thank you so much for coming on the show.

Great to see you, Ben. Congrats to the team. Congrats on course. Last thing, if you want a hat, you can. We have a new CLI. You can run raindrop drip. You can get a hat, an umbrella, a couple other things.

That's a fun way to give out merch. I like that. That's very creative. Thanks for coming on the show. Great stuff. We'll talk to you soon.

Okay, back to the debate around Figure. We've had Brett Adcock on the show before and he had a live stream. We talked about it a little bit. Watch a team of humanoid robots running a full eight-hour shift at human performance levels. And Brett Adcock said, 'This is fully autonomous running Helix 2.'

All right. Pull up this post from Pete. Yes. And the stream did fantastically. It was 24 hours. It got 3.4 million views. But at a certain point during the stream, there were some questions about whether or not the humanoid robot was in fact...

Back to the beginning. Back to the beginning. Okay, let's play this. All right, so it's cooking. I mean, the speed is actually...

And we were extremely impressed by this. This was remarkable. Remarkable. Even if it's teleoperated, it's extremely impressive. Yeah. Yeah. Yeah. Like the robot's clearly working. This is very... But they're saying that it's not teleop.

Okay. So then the robot starts missing things, being a little bit like an inch off, and then reaches up and touches the robot's head. The robot, which is something that wouldn't normally be necessary. It doesn't have like a logical explanation or conclusion. And so a lot of people are asking...

It does have a semi-logical conclusion, which is that Brett is claiming when it reaches across its body to go to the right that it puts its hand up here to get the hand out of the way. That's what I was thinking was that if the hand is halfway up...

It might be blocking the sensor, the camera sensor, and so even though the robot might reach the hand up further to move out of the view, so then the robot can look at the next package. So that's one possible explanation, but a lot of people are asking even harder questions saying that potentially was there a human in the loop? Was this teleoperated? Which is something Brett has said it's fully autonomous. I feel like that means no humans in the loop. But Teor Taxes has an artist's representation of Helix 2, Figure's in-house neural network running entirely on board, and it of course is a human in a VR headset. Very debatable. We'll see where you stand. But there is a third option which I have shared, which is potentially no humans involved. I don't know if you'd call it autonomous, but you would call it no humans in the loop.

Because you have... Well, it is an autonomous system, right? It just sort of runs. Yeah, I would consider this autonomous. It's the image that I shared in the production chat. It's not of a human and it's not quite robotic, but there's no human in the loop. And so this could explain it's the system is running with no humans in the loop. If you make that claim and you follow this...

I think this qualifies as no humans in the loop. If you have a giant orangutan in a VR headset puppeteering the robot via teleoperation, you could say that this system does not have a human in the loop. And you could make that... and I could make the argument that it's autonomous. Yes, the chimpanzee is running its own. It has somewhat of a neural network... like a neural network. Yes. Yes. No, no, no. No one knows.

Says, 'I think there was a human physically inside.' Ooh, physically. That is another option. Yeah. I mean, the thing that I'm... So I want to talk with somebody at a place like Amazon who I imagine does this kind of thing all day long. Yeah. And are they asking for a humanoid to do this process?

Yeah. Yeah. Like this seems like something that e-commerce fulfillment and logistics companies have been doing for many, many, many, many years. Yes. Is there not a purpose-built robot that sits right there and makes sure that yes, the packages are in the right orientation? Does it have to, you know?

Yeah. If you watch an episode of How It's Made, you will see every variety of custom-made machine for flipping around, sorting packages, that type of activity. There are custom-built machines that run at scale. They might cost like $10,000, but they last 50 years. And anytime you see a Diet Coke factory or gum manufacturing line, all these things like the gum that you have there comes off and the gum rattles down and is sorted into the pockets of the packaging and then the sleeve is wrapped around and glued and all of that is done autonomously but just with a bunch of machinery that was built probably like a hundred years ago. Honestly, if it works, don't fix it. But you can clearly see how this type of task, package sorting, would be like on the curve to a more economically valuable humanoid robot. And like if I was going to buy a humanoid robot to do my dishes and you showed me this video and it was in fact fully autonomous, that would be an encouraging demo to me. That would be something that I would look at and say, 'Oh, well, like if it can do this successfully for hours and hours and hours, I'd probably trust it to put some laundry in the washing machine. That doesn't seem well beyond the scope of capabilities.'

It's so interesting how quick it is when it's just sorting packages there and then it does the bite and walk on the way off. Like if you rewind for a second. Yeah. Yeah. Yeah. The walk is not... walk. I only use that terminology because that's the terminology that is used. Yeah. Like look, why does like... if you were able to shuffle like this so fast, you'd think that you'd be able to hustle a little bit. But maybe that's V2. Maybe that's less relevant for this particular task. You know, there's a lot of different options, but we will dig into it. Brett launched day two. I mean, regardless, putting up views, sorted 32,000 packages. Day two is live. And he shared more details on what's going on. The original goal was an 8-hour run. After zero failures yesterday, we decided to keep going. We're now over 24 hours of continuous autonomous operation without failure. This is uncharted territory. The task is small package sorting. F03 detects the barcode, picks up the package, and reorients it. Barcode face down onto the conveyor. Humans average around three seconds per package. F03 is now around human parity. The robots are reasoning directly from camera pixels. The robots are fully autonomous using Helix 2, our in-house neural network running entirely on board F03. There's no teleoperation. Every action comes directly from Helix 2.

Okay. Well, I feel like that rules out the monkey business. I think teleoperation would fall if you had a monkey puppeteering this thing, I think it would count as teleoperation. So he is denying that allegation from the timeline. Yeah. But the timeline seems convinced. YouTuber commenters started naming the robot Bob, Frank, Gary yesterday. So they added name tags to each robot. And if the robot gets stuck or the AI policy goes out of distribution, Helix triggers an automatic reset. You'll occasionally see this happening during the live stream. If a robot has a software or hardware issue, it autonomously leaves for maintenance and another robot takes over. We run our labs and figure this way to maximize uptime. If we haven't had a failure yet, we haven't had a failure yet, but statistically we probably will at some point. So, very, very fun going back and forth. Who else is chiming in? People are... I'm the LA Dar says I'm the last person I'd expect to rush to Figure's defense and I'm looking forward to hearing Brett's take here and in any and all cases I stand with PBD King but IMO this demo seems authentically autonomous and could see this being learned behavior from teleoperators that collected the data for this model with their VR headsets. And PBD sucks who broke the story or went viral first time said he actually has a pretty reasonable sounding excuse but doesn't give me tons of confidence on the model's brittleness for crossbody reach. The policy lifts its arm to avoid hitting the metal chute. Nice try.

I wasn't sure if he was going to reply to this and sort of engage or just sort of let the timeline run wild with it, but the metal plate does seem like a piece of what's going on. But people are still hungry for teleoperation bombshells. It sort of cuts both ways. I remember Jason Carman did a video maybe with 1X and everywhere in the video they said this is teleoperation. We're doing teleoperation. We're bullish on teleoperation. Put it at the bottom in the text in the description like so told everyone. And still people were quote tweeting and being like this is teleoperation. And so people are sort of grappling with like what is real, what is fake constantly. Well, is there anything else on the Figure story that you'd like to dig through?

Uh, no. Switches hands after working more than four hours straight. Huh. Well, we can... there is some... we've in just a few minutes. A newly released OGE form, Office of Government Ethics 278T, discloses that President Trump filed 3,642 trades involving stocks of public companies between January 1st and March 31st. Transactions include hundreds of stocks and ETFs such as Nvidia, Microsoft, Broadcom, Amazon, Apple, Alphabet, Meta, Goldman Sachs, AMD, Airbnb, Palantir, Netflix, Costco, Walmart, JP Morgan, DoorDash, and others. Individual purchases of Nvidia, Microsoft, Broadcom, Amazon... individual. So he's averaging around roughly 40 trades a day. 40 trades a day. Check my math there. Um... that is in Q1. So a lot of trading activity. We talked about this like should you just give Jane Street access to the federal government? Should they just be able to change the laws to optimize for max GDP growth? And feels like we're one step closer to the economic singularity of the hedge fund running the country. Anyway, we have... Oh, yeah. What else?

I'm trying to find the history of presidential day trading. I don't know if there is one. Jimmy Carter famously divested from his peanut farm because he was worried about conflicts of interest. But we are in a new era. Anyway, we'll have to figure out if Trump is long or short the Cerebras IPO. He's probably watching right now to hear Doug O'Laughlin's take on it to understand what's George W. Bush with Cerebras says not a day trader but had a famous controversial stock sale. He sold 200,000 Harkin Energy shares in 1990 before bad news came out. Okay, interesting. And there's no other evidence that we're finding of presidential stock traders. Well, we'll dig into it, but we have Doug O'Laughlin from SemiAnalysis in the waiting room. Doug, how are you doing? Welcome to the show.

Good, good, man. You know, pretty busy day, another day, dude. Honestly, every day is a busy day. Every day is a busy day. Take us through it. How do you think the market reacted to the Cerebras IPO, to the SemiAnalysis deep dive on the company? What is the overarching story here?

So, I think the market was obviously positive. I don't think we're quite as positive as the market, but it's a bull market, baby. I think the takeaway is that Cerebras got to IPO, which at one point in time we didn't think that would happen at the SemiAnalysis world. We've historically been very bearish on SRAM. Okay. But I think there's a path forward for them to be a disaggregated pre-filled chip or maybe even AFT chip, meaning attention feed-forward disaggregation. Okay. So yeah. Yeah. Unpack sort of the competitive dynamic, like what the fear around Cerebras, as far as I could tell years ago it was like will this ever be useful? Will they ever actually be able to make it? Will it have defects? Then it became certain applications, demand-side customer concentration. But where do you think they are now? How has that journey evolved?

So first and foremost, Cerebras is about SRAM. SRAM is like the fastest possible memory and it's kind of done on a logic process. But the problem is SRAM scaling is dead, meaning that you can't make smaller and smaller SRAM scales. So pretty much they kind of committed to this dead-end process by having the biggest scale-up world as a wafer size, but then the models got much bigger than just a single wafer. And so they have really, really fast inference, but only at a certain size. And I think the real capability problem is can they imprint models larger than a trillion parameters? And I think the answer as we think right now, it's pretty unlikely in the near term.

Yes. So I understand all that. I'm just wondering about the world where should I view it more like a CPU? Because when the AI boom, the ChatGPT moment happened, the obvious buy was Nvidia because we're going to need a lot of GPUs. No one was really expecting a chip shortage in CPUs, but then agents wound up using CPUs for a bunch of stuff. You have to keep the GPUs filled. And so CPUs are now in demand. And I'm wondering if there's this world where there's this yes, we're going to move past the trillion-parameter models, but we're going to keep using them forever, just like we use relational databases forever, even in an agentic AI world. Or you have a scenario where you have a big model that is giving sort of orders, workload, delegation, something delegating to a smaller model.

Yeah. I think in a perfect world where there's no silicon constraints that might be true but obviously there's silicon constraints and I think Cerebras is really well optimized for a certain problem and we think they do a great job at answering that which is fast inference at a certain size of model. Maybe that market's going to be large enough and I mean honestly I don't think I was ever bullish the entire time but now that we're here like non-ironically 1% of a very large market works and I think they got like 1% of a very large market. When it first started, I was like, 'Oh yeah, what are you going to do? 1% of a very large market. That's going to be a few hundred million.' And that's like the classic seed pitch too. You know, 10 years ago, they're... Yeah. Is there any... for a long time there was a lot of fear around ASIC companies around architecture changes. We're going to move past the transformer and they're all going to be locked in the past. Is there any optimism around there's an architecture change that actually is to the benefit of Cerebras and makes them more relevant in the future?

Do you think... you know, that might be my pay grade, that's too gigabrain for me right where I'm at and the understanding. There is a narrow path for them I think and I think they're going to be able to inference maybe one trillion parameters at very small context window sizes or smaller models at very, very fast speeds. Yeah. But I don't know, man. Maybe I mean like you know the true gigabrain take is Mythos is so good or whatever that it makes compute efficiency super easy and boom, you know, your model is inefficient and AGI understands. Yeah. Yeah. Yeah. Yeah. Distill yourself so you can run on a Cerebras chip just as effectively. Okay, now we're talking, that's the gigabrain thesis. But I think... I just think that there's demands, right? Like clearly we're in a shortage and ironically in a shortage it's not the best company who wins. I mean you can look at Nvidia's stock chart and that tells you it's the second, third, fourth best companies where the demand overflows right and so we're seeing all of that today and I think the reality is the market's big enough for a lot of demand and is in that space. So they've done a really good job and I mean it's a cool engineering problem but we think it's kind of a solution looking for a problem because the world of LLM blew up at a much faster scale than anyone could have ever thought of. The size I think is really the difference.

Yeah. Yeah. Give me a little primer on Groq, how Groq fits into the SRAM machine market. What the view is because it felt like Nvidia's move there with the license acquire as you put it was defensive against Cerebras. Is that the correct framing? Like how does Groq fit in on this?

Okay, so let's talk about exactly where Groq fits into the architecture. So in the transformer architecture you have like the multi-heads of attention and then there's a feed-forward network that's a portion of essentially the entire transformer block and what's become really hot in the last few years or not even few years like probably a few months man is you've been disaggregating all the different parts of inferencing into subsequent specialization. So we're talking about GPUs and ASICs being a specialization over CPUs, but now we're actually starting to break the essentially the constraints of inferencing into different compute and memory-bound pockets. And so for example, we're finding prefill ends up being... prefill being essentially loading all the weights ends up being compute-constrained. So you don't really need a lot of memory bandwidth. So why don't you just use a very flops-heavy portion and you disaggregate the memory onto the decode portion which is like extremely memory-bandwidth limited. And so this is Groq where this fits in the strategic thought process here is in the GB200 rack what you can do is you can pass the activations over to the SRAM in the Groq LPU rack and that is an extreme speed-up and so it's like that's like a perfect example of another break-apart of the transformer architecture. Pretty technical, but that's like the thought process here is that the memory is so fast the memory bandwidth or the speed of the IO doesn't really matter and you don't need a huge scale-up world size because you're just streaming the activations.

That problem wouldn't work with the previous chip because you're kind of... it's an island, right? You think of it as an island of compute. It's really, really good at everything in the middle, but moving anything off the island is really hard versus moving something off the island onto a Groq chip because there's a plug at the end of it is a lot easier. And that's kind of the calculus, I guess. Yeah. So Cerebras lower memory bandwidth, lower interconnect speed off the chip, but on the chip, it's as fast as hell. Yeah. Okay. So what does that mean for the Groq Nvidia ecosystem? Because is this something where the default configuration is going to be a Blackwell and a Groq chip like in you know 50% of racks, 80% of racks or is this like still some sort of niche application where Groq is going to be deployed you know sort of sparingly sprinkled into specific use cases? Do you have an idea? Yeah, I think I don't have an idea with high precision. I think you'll find that a lot of these things there's a lot of different ways to split up and serve your model. So expert parallelism, pipeline parallelism, tensor parallelism, right? And so the correct optimization per hardware rack is going to kind of depend on the shape and architecture of the model. And we don't really know with high precision what is what. And there's been kind of like different roadmaps along the way in terms of what they wanted to do for speeding up inference. A perfect example of this is the CPX rack which was mostly built for extra parallelism. It kind of remains to be seen if this is like if the Groq GB200 speed-up is going to be like the way forward, but it's definitely a technology tree that I think Jensen is excited about. So, I mean, we'll see.

What about Lisa Su at AMD? Is she excited about this technology tree? Can you give me an update on how AMD fits into all of this? So AMD is mostly just trying to get the last thing to work, which is the rack scale-up. And I think they're going to do a good job of 450. I think what's going to happen is that like you know it's a compute shortage, right? So you're talking about overflow demand. I think Lisa is going to figure it out. But on the inference serving side, I think there's definitely some demand or desire to probably match the Nvidia roadmap. And I wouldn't be surprised to see if there's some kind of fast SRAM offload FFN chip in the next 12 months. But the thing is the number of candidates there is actually like pretty low. I think Intel's really... Intel's going for Sanova which is a little clever. There's like HPM2. There's a few other players out there too that pursued SRAM scaling, but I think that in this specific case Lisa's mostly just focused on the last thing and I think AMD is definitely good enough right now.

Okay. On Intel, what is the latest there? It feels like the round table has been assembled and sort of everyone has held hands and decided to maybe jump across the transom at the same time. Take the leap of faith. But it also feels like you know lithography machines are majorly backlogged like there's a whole supply chain that they have to answer to that's backlogged and so really high expectations but also what are they... what is the next milestone for them after they actually get these deals with Apple and Elon Musk and Elon Musk. Yeah. And the gigafab sort of like once they get those signed like what does the next couple years look like? I think it's about execution. It's kind of crazy to me that I think the stock price is ahead of the technical turnaround and I think that... I think Liquitan clearly has like righted the ship and gotten the right people onto the party if that makes sense. And I really do think the government Intel deal was a stroke of genius because Pat Gelsinger spent you know three years trying to build a bottom-up demand to essentially come to the fab and Trump's like, 'Yeah, none of this. I'm going to sign the deal from the top and what's going to happen is you're going to come play because we're in the United States government or else.' And so I think people are there. I think the customers are there. I think the process is good enough. I think 148 will be also good enough given how much of a shortage N3 at TSMC is and it's all execution risk from here but the historical Intel has quite a bit of execution problem so we'll see.

Okay. Before we move on to TSMC which I want to go to next, are there any other interesting ASIC projects on the horizon? We've talked to a few of these companies, but I'm interested in like the shape of the differentiation. Like you explained a little bit of the divergence and strategies between Groq and Cerebras, but there's Etched and a bunch of other companies that are working on new chip designs, and I'm wondering if any of them stick out to you as particularly differentiated. You know, I'm not going to go too into the details because I feel like some of them are even like still figuring out their roadmap. I think Maddox is kind of interesting the way that they're kind of trying to pursue the memory problem. I think Etched I'm excited about the kind of yolo bet if it makes sense. Just make a big systolic array, but I think there might be like niche cases. I think the problem is like at the end of the day, Nvidia's big bus is still really good for the majority of cases and you're going to have to like start to make really opinionated bets on the ASIC to find what niche market ends up being all like a diverter of demand into their ASIC. And so the ASIC specialization from here, I feel like you have to make some pretty big brain bets in order to make your bets come pay off. And I think most of the bets that I would have guessed when you originally did them wouldn't have paid off. And the ones I didn't expect did. Like it's kind of crazy.

Yeah, it is. It is a very weird market dynamic where a couple years ago we saw ASIC and new chip companies, new silicon companies raising hundreds of millions of dollars or $500 million. And it was like, well, for that you're going to need this massive market. Are you really going to flip Nvidia or something? And then the market grew so much that the 1% of a huge market sort of potentially maths out for some of these companies now. It's a fascinating development. Jordan, do you have something? China trip. Yes. Oh, yeah. What are you tracking on the H100? Oh, so honestly, do you guys see the parade? You know, Trump loves a parade. Oh, yeah. They're winning. Good parade. I was like, dude, I was like, man, if I'm not much of a parade guy, but I was like, dude, if they showed up in that parade was for me, I'll be like, these guys could be friends. Yeah.

My impression is that the executive branch really wants a deal and I think you saw the H200 list, the verified H200 list. I expect probably more lightening up on the executive branch. Something that's really interesting is if you look on the legislative branch, there's actually more export control bills going through the House than like ever in history time. So there's kind of this tension, but I do think Trump's a businessman. He loves a deal. I expect a deal. So, yeah. Somewhat related, TSMC, Ben Thompson was writing that potentially they weren't ramping capex fast enough. What are you tracking on TSMC being a potential bottleneck for the AI buildout? Just as more and more Cerebras is now trying to get allocation, it feels like a particularly sharp place to do business.

Yeah. So I think at the end of the day TSMC is kind of a kingmaker in terms of supply and there's no reason for them to really let the market go out over its skis and I think they're happy with the pace of what they're expanding out because like hey they're growing their capex like whatever 40% but in absolute dollars these are big numbers. We're going to run out of TSMC engineers in the island of Taiwan pretty soon here. So I think this is all kind of good on the margin for overflow demand which is actually Intel's definitely reflecting some of that but I think the shortages specifically at TSMC is driven by clean room. It's a long lead time item. It takes 3 to 5 years or let's just say 3 years to bring a clean room up. And so in order for them to have like figured out and like perfectly match demand two years ago, they would have to been like we have a 10,000 foot house and we need to buy a 50,000 foot house with conviction, right? It wasn't that clear 2 years ago. And so I'm going to expect supply to kind of lag over and over and over, but demand signals will continue to essentially command premiums, move up wafer pricing, move up orders, and that's what's going to make TSMC invest more next year and the year after. But they're going to do it in an incremental not a revolutionary way, but like an evolutionary way. They are very methodical and do steps one at a time.

Okay. Clean room fungibility. When you say it takes five years to build a clean room, I immediately go to SpaceX. I imagine that Elon can build big things quickly. Is there some world where that partnership accelerates Intel regardless of your timeline for the mass driver, fab on the moon, all the crazy long-term stuff, but just having Elon around the table to say, 'Oh, we need to build something big and it needs to be capable of operating as a fab.' Like is there something where he brings more to the table than just dollars potentially? So, I definitely think Elon is the man to do it. I forgot who said this, but like Elon makes the impossible late. I don't expect it to be on time. You know, talking about the cigar in the Terapab, I'm really kind of doubtful. It's, you know, I guess from first principles, it's easier to just clean the entire room than to make like really hyper concentrated pockets. And that's what I would guess the bet is. But I still think by the time Elon figures it out the supply response will have reacted already. We're still two three years out and there is some clean room fungibility but and you've already seen this actually. Micron bought an old power fab. I think this is the PSMC deal. People are buying display fabs. Essentially every bit of clean room that is not accounted for in the world is being snatched up and retrofitted to kind of meet the supply demands.

Interesting. Yeah. I mean that's happening all over. Didn't Ford just announce some sort of AI play today? The stock's up on something. It's all over the place. I am interested in terms of like... 6%... getting powered shells. Ford is worth more than Figure now because last year around a year ago I remember Figure Robotics was worth more than the Ford Motor Company. Rough time, but now they're both AI companies, I guess. But what are you tracking on the American data center buildout? Domestically or terrestrially before we move on to space capabilities. Basically, how... Oh, go for it. No, no, no. Just I'm just curious about... I mean we're starting to see glimmers of pushback at the municipal level, different data center bans, and I'm wondering about what are the big levers that need to get pulled to actually continue to bring capacity online in America.

Yeah, I think that's a good question and you're already seeing the first level of this is the delays. Yeah. My favorite clickbait is 50% of all data centers in America are delayed or cancelled. Implying 50% is cancelled when it's really just everything is delayed. That's like my favorite clickbait. I got to steal that in the future. But I think that... I think it's going to be local municipal and people have to really believe and demand and desire the jobs. And I think one of the ways that we're seeing this is like, you know, capitalism works and effectively the dollar per megawatt has been going up. It's like a one-way trade in the same way that like, you know, the power per rack has been going up. The cost of making these data centers have gone up and one of the ways that happens is it leaks into labor, right? So essentially, you're super against it, but all of a sudden it offers 3,000 new jobs to your home and you're like, 'Well, maybe I'll take it.' And I think that with enough economics often times money finds a way and that's kind of how I would guess but it's going to be like a county-by-county fight right and some places are just going to say hell no.

Yeah. On that note we were debating this earlier today. There's been a couple of examples in like viral photos and articles about like they... I bought a beautiful house in the countryside and then they built a data center right next to it. And you know, no matter how pro AI you are, it sounds annoying to have a huge building that's an eyesore and maybe noisy, maybe smoky next to you, but have you been tracking like how feasible is it just to throw the data center like truly in the middle of nowhere? It feels like America has a lot of land, but what goes into selecting these data center sites these days? Do you have something else? Yeah. So I think pretty much two fiber pairs is the big desire. Essentially like you're more than willing to go to where the power is because you have to go to what the biggest actual bottleneck is and power is the biggest bottleneck. So you can just... in the past you're talking about like hey having these imprints or rather like let's say point of presence near local cities right but power was never constraint in that world it was just the biggest constraint was getting this video from TikTok to your phone as soon as possible. If the biggest constraint and the largest part of the cost is going to be power why not move the data center to power and then essentially hook it up with fiber and so I think that we're going to put them in the middle of nowhere that's just how it's going to work. To a certain extent there's going to be more densification in some of the inference near the population, but I still think the ROI makes the most sense to kick it out of the middle of nowhere. Yeah.

Has the political backlash pushback updated your thinking at all around the viability of space data centers? I remember you know we talked as this idea has like gained popularity you guys have like consistently said yeah technically you can do that but like maybe it won't be long. There are space data center players now that are kind of loving the pushback against terrestrial data centers because they're like this the more pushback there is the more it could make sense for us to put these up in space but what's your view? I still think economics is going to win out. You know, something a pound on Earth is probably 10 times more expensive in space. And it's really hard for us to go to essentially beat that out with a new completely specialized supply chain for what's going to be a smaller market in the near term. It's a real adversary against the adoption in the short run. In the very long run because I'm sure you saw the Anthropic Colossus thing where it's like also interested in space, right? Like the biggest maxi vision of this is like AGI we have you know 30 ter... you know we have a thousand terawatts of GPUs on earth and we're like we got to put a terawatt in space right so like in that world I think space data centers work where a small percentage actually... so big... it's 1% of the market again it's just like... one and it's a trillion dollar... trillion... VC is vindicated... yeah yeah VCs are vindicated... TAM pitch deck slides vindicated... yep yep yep Yeah, it's literally as big as a galaxy, bro. Just there's no end to it. Actually, think about how big the TAM is. It's huge. So on like I think what is more likely is if it continues to be painful to do it from a zoning perspective in America, it will essentially slip into other geographies probably in the Western Hemisphere. There's a lot of power in space in Brazil and I think that that's probably good enough, right? There's definitely ways to make this work. I definitely think the only way you do it is by paying more and finding someone who's like, you know what, I'll hit the bit. And so that's the important part. But you know... and finds a way.

Is that sort of the bull case for sovereign AI initiatives? I was always super skeptical because like Europe didn't get like France's Google like they just use Google and for a lot of consumer aggregator type consumer internet companies it's like Spotify is from Sweden but it could be from America and it wouldn't matter. YouTube is from America and they use that over there. And you didn't need a national champion in every consumer category or there were certainly returns to scale and a lot of the American companies just won. But so I never really bought the whole idea that like oh the French need like a locally trained LLM and the Germans also need a locally fine-tuned something or other. But if every country has some sort of excess supply of energy or space or regulatory capacity for data centers bringing that online and just operating like a neocloud could just be economically valuable for that country regardless of whether or not they're vertically integrated to the point of the consumer or the business that's running an AI agent. I think that's probably the case where at the end of the day, economics is going to kind of push it through and there is FOMO and Europe did do a lot of investment in the internet like really late if we're going to use 1999 as an example. I think the thing I keep thinking about is that this AI thing is going to be a big deal. I continuously am shocked and surprised by the magnitude and scale. That's a narrative. I don't think it is right now. I feel like we are in a particular moment where... No, there's just the people calling the top and the bubbles like they're awfully quiet right now. And that makes me even more scared. That is okay. So, to be clear, you know, the true top there's no... everyone's bullish, right? Everyone's like, 'Dude, it's actually going to be bigger next year. It's actually just going to be a bigger bubble. Shut up.' Yeah. So... yeah, I was not concerned about... I was not concerned about a bubble when everyone was saying... it's a bubble. It's a bubble. Yeah, exactly. I am... I mean I'm a little concerned it's a bubble, but at this point in time I think if you look at the big... I've been reading... honestly here's my view. Here's my view. It's not a bubble until you guys are spending 120% of revenue on tokens. Yeah. Our gross margin goes negative. Yeah, you're just like, we're raising a major fund and we're not going to be investing in it. We're going to be burning it. It's actually not a bubble until SemiAnalysis goes public and trades up 600%. There we go. I'd like that. That's the real top. No, I think there's a few things that have to happen. I think OpenAI or Anthropic...

Someone has to go public, and it's going to be this year. Like, we have to hit that keystone before it's all over. But I also keep thinking about this: dude, this is a big technological revolution. I think it's bigger than the internet, and I firmly believe this. I don't think I believed it would be bigger than the internet maybe even two years ago, but I'm pretty convinced this can be bigger than the internet. And if you look at the past, these big technological changes are often sometimes bigger than everything else. It reshapes the entire world.

Yeah, for example, on the sovereign AI thing, maybe you're like, 'Yeah, you don't need to fine-tune LLM.' But what happens when AI becomes such an important, fundamental, almost like society-level institution that a government can't control it? That becomes really uncomfortable and weird, where it's like, hey, an Anthropic can just put 5% of the compute of Mythos and run a really effective government whenever you wanted. And you're like, whoa, whoa, what does that mean for us? And so this wave is so big that I think people are going to, out of fear and concern that they're going to be left behind, and that the institutions that AI will bring are going to be bigger than the original thing that we're doing. I think that's the problem, right? Like the industrial revolution changed everything.

Yeah. The other thing that we were joking about in Q4 of last year is like, John was like, 'Great, the bubble popped.' Like, the bubble inflated and then it popped, but then we got agents, and then you have this sort of re-acceleration of every metric across the board. And so the other thing that we're like, we're trying to comp the AI boom to the internet, but the problem with the internet boom is that we didn't have the internet. So everything just took like, or the internet was coming online and people were getting access to it, and so the entire buildout and all the capabilities and all the companies took a lot longer to sort of grow, right? And now you have that core infrastructure, and so when you're layering on more infrastructure, that accelerates all the underlying trends.

Yeah. Yeah. I mean, the lab revenue multiples are like an order of magnitude or two off of dot-com peak multiples. And in the public markets, Google, Amazon, Apple, all the hyperscalers are at like pretty reasonable price-to-earnings multiples still, even with all the capex and stuff. And so, your pushback would be it's on free cash flow, because you can make earnings look good instead of free cash flow. But like, I think the revenue continues to be real, the demand continues to be real, and until you just see demand evaporate, like, it's hard for me to sit here and be like, GPU prices are up a ton, Claude Code is really valuable to me, I still think I'm an early adopter, and you know, this is all going to end tomorrow. I envision myself using it every single day more for the rest of my life, which is kind of crazy, and I think I'm an early adopter, and so I just think it's hard for me to envision this not being a ginormous deal.

And it's kind of like we just got the, like, I really, I wrote this whole thing like Engels' Pause or whatever, like it's going to change everything. Like the amount of net output that's going to increase is going to just blow our minds. It might be bad for GDP, ironically, because GDP will be unmeasured. Like we're going to, GDP might be broken as a concept. GDP got invented in the 1930s to measure how much output you could make to not screw over the domestic economy for World War II. Like it was a way to essentially organize the American economy, and it's a statistic, it's an estimate. Like I think we're going to attack, and a lot of institutions and ways that we're doing things and ways we measure are going to be attacked by this because it's such a big change. We have to rewrite the playbook over again.

And it's funny, I think wasn't Ben Thompson talking about this in a recent interview of like people are comping this like, okay, Silicon Valley brought crypto online, and then it wasn't maybe as big as some people had pitched it to be, even though it's been powerful. And then even the way you're talking, you're like, you know, we're still early, you know, which is like a classic crypto. Or you're saying like, you know, in crypto is like, well, like a community could have a DAO, and that DAO could be worth a billion, that community could be worth a billion dollars, but there's just no way to measure that. But now we have tokens, and you're saying GDP. But anyways, I'm trying to like unlearn I think some lessons from that cycle because there are a number of things that are quite different.

It's also what about the reflexivity that people do have a little bit of an immune system to just running away with everything, because you could believe this and then bid Nvidia to 10,000 times earnings or something, and like at a certain point you have to start grappling with the reality.

What about robotics? Has Figure had a major breakthrough?

I mean, one, I have not been following the feed as close as I should be. I just think robotics feels a little further out than the hype would let you believe. I feel like robotics is much more akin to the self-driving car paradigm, where it's like, oh yeah, it's definitely going to come and automate everyone's jobs, and then it takes a lot longer and it's a lot unsexier. I think the scary or positive thing about AI is since it's information work and it's already been distributed and it has the perfect network to run on, which is the internet, it can disperse very quickly, and that's what we're seeing right now. And so yeah, I'm just not anywhere near as bullish on robotics as I am on the fundamental.

I'm bullish on the next SemiAnalysis. I don't know, what are ClusterMax and InferenceMax? What are those called? Dashboards or analyses or rankings?

Dashboards. Dashboards. Now everything's a dashboard.

Dashboard. Well, you need to make a new dashboard. GTP, gross token production, is what we're measuring now. This will be output of the United States, gross token production, GTP.

We need to, I mean, I think more on this soon actually. This is like a place we're doing some research on, but I think, you know, the real bubble metric is if we're like, you know, how many tokens, what's the token replacement cost? That would be some really good bubble math where it's like, yeah, a software company has a really low token replacement cost per market cap, but like a hardware company has an extremely high token replacement cost. And then it's like, oh no, no, it's just enterprise value divided by token replacement cost.

Well, the real bubble one will be to go to the full Mary Meeker eyeballs metric, eyeballs multiples. So you will value companies purely on token consumption. You'll say, 'Oh, well, they're consuming 10 trillion tokens, so they must be worth a billion dollars.' And then you'll get really weird gyrations.

That'd be great for SemiAnalysis. That'd be really good for Semi. We are consuming a lot of tokens.

Well, you also put in a lot of good stuff. I really enjoyed the... Would you guys ever make a sort of political-style attack ad against another research firm for having AI psychosis?

Is that a reference to the G?

Sorry, it's a reference to General Catalyst attacking Mark Andreessen. You know, life's pretty long.

I actually think SemiAnalysis is just peerless. I don't think there's like a neck and neck with someone else. Like it's just you guys.

I was going to say I don't really know who our competitors are. I don't really think about it. Mark Andreessen or another research firm like that. Maybe one day we will go through AI psychosis.

Honestly, you guys need an arch nemesis. You need an op. Moody's?

I guess it would be Gartner. I had to say they're like... but this is not a good option.

You need to sell SemiAnalysis hype cycle and it's up only.

No, no, no trough of disillusionment. Straight line.

It's a straight line on a log graph. That's what it is.

The SemiAnalysis hype cycle. I love it. Gartner doesn't stand a chance. But thank you so much for coming on the show. This is fantastic. Always great.

Full analysis. Yeah. No more analysis.

Those guys would kick our ass, man. If they had full analysis, they would kick our ass. It'd be so over.

So anyways, take care guys. Great. Have a great day. We'll talk to you soon. Cheers.

Goodbye. Next we have Andrew Feldman from Cerebras joining in 20 minutes. We'll go back to the timeline because the OpenAI Elon Musk trial is in its final day. The trial is ending. People expected four weeks of trial. We only got three. They're cutting it short. What are the prediction markets saying about who's going to win? I want to know that. And I want to go to Mike Isaac, the Rat King, because he has a breakdown of what's going on. He says, 'Good morning. Closing arguments of Musk versus OpenAI with special guest Microsoft are happening today, Thursday, May 14th.' Again, Mike Isaac, of course, he kicks it off with what his lunch is. He's got an Epic bar. He's got the bison snacks. He's got a La Colombe latte. He's got a couple other good things. He looks like he's prepared. He's got a bunch of snacks. I feel like he's in a better position today. Learned his lesson. Three weeks of sort of like recursive self-improvement. I think that's what's going on here.

So the Kalshi sheet, will Elon win his case against OpenAI? It peaked at a 58% chance. Where is it now?

April 28th. It's now sitting at a 30% chance.

30% chance. Okay. So right now the judge is instructing the jury on the criteria by which they should be judging the outcome of the case. Important because if the jury listens and carries this out, it is a very, very specific lens through which they view all the evidence. Ostensibly, it's where theater ends. Listening to this and being read out in court for the last 23, 30 minutes is very helpful because it's clarifying on how high the bar is for the plaintiff's side proving some of these claims. Sort of feel bad for the AV guy during this trial. There's been feedback. There've been mic drops, but not in the good way. The mics have been dropping out. Vicky video feeds. They need to revamp this place, says Mike Isaac. LMAO, the first joke of the tweet storm. He says Musk's counsel is going after OpenAI execs Altman and Brockman and has the mugshot-style photo of Altman on the screen again. Battle of Photoshops of executives in this trial has been entertaining to watch. You want to depict your opponent in the worst possible light. Musk's counsel going back and forth hammering the point they've made over and over. The argument essentially painting a picture: Sam Altman, liar. Chipping away at witness credibility has been a core strategy for the plaintiff's side. And we're back to everyone hates Google again. Molo is using Larry Page, who they claim doesn't care about humanity, as a foil to the noble Musk, whose only care with respect to AI is the future of humanity. Musk's counsel is painting the 'don't trust Sam' picture in a bit more detail to the jury. Also, Musk's side has a picture of Elon and Altman on the screen now. Sam's looks like he's about to be processed by a US marshal. Musk's looked like he's getting ready for the Met Gala. Lol. Lots of Musk closing side arguments. Semi-populist track of pointing at OpenAI and saying these billionaires are making gobs of cash while writing a charity for the supposed good of the world. I'm curious if the jury can register this argument even if it comes from Elon Musk, the world's richest man. Ouch. OpenAI's counsel begins closing argument with a broadside against Musk. Even the people who work for him. Even the mother of his children can't back his story. Oh yeah. Back to the war of the Photoshops. OpenAI closing remarks now in the digital displays on the monitors for exhibits. All the OpenAI executives look like Olan Mills photo shoots. Do you know who Olan is? He says it's complimentary. I need to get up to speed on my photographers. Olan Mills is a portrait, offers portrait photography. Ooh, it does look very nice. You pull up the Google images on Olan Mills. Anyway, short summary of the closing: Musk camp, all these OpenAI executives are rich as hell and lying all the time. OpenAI camp, all of that is a sideshow. And literally all the claims Musk is bringing cannot be stood up by actual law. The Microsoft camp disappears into bushes. Dota got mentioned again. They love mentioning Defense of the Ancients. Incredible Photoshop from the OpenAI camp of a calendar of events complete with little characters and a timeline of events. I wonder if they're using image gen 2 or if they're doing it the old-fashioned way. I can't wait until it's entered into evidence this afternoon so he can show us. Sort of want to buy this meme guitar, but I also have two tele. Is that just completely side note. Gamer has entered the blog. The Dota moment has been mentioned nearly every single day during this three-week trial. AI researcher. We got to have Mike back on the show. It's so good. Saw as a true breakthrough in the technology. So Mike Isaac says, I played... What is the timeline for the jury to meet? Are they... Is this something they're doing today?

They're getting a 30-minute recess. Most they've had in a month. I might actually be able to go outside and get real food. There's a Popeyes across the street. Is it a bad idea to get a bucket of red beans and rice? That's what he's thinking about doing. So not much news on when this will close. It is 1:10 Pacific time. I imagine that they will wrap up by, what did he say, 3 p.m., 4 p.m.? So 30-minute break that happened 40 minutes ago. So I imagine that... And but they've been taking Fridays off is kind of what I'm getting at. Oh yeah, because this could... So maybe this happens to Monday. This is just closing arguments. It's not necessarily the end of the trial. Might get the results or the jury might make a quick call. Well, there was an update 11 minutes ago. A lawyer for OpenAI on Thursday defended the company's chief executive Sam Altman from withering character attacks by Elon Musk's legal team as both sides delivered their closing arguments in a trial with potentially seismic implications. The stakes are high. Mr. Musk, who was not in the courtroom on Thursday because he was in China with President Trump, is asking for more than $150 billion in damages. He is also asking the court to remove Mr. Altman from the startup's board and to stop a shift the company made last year to operate as a for-profit company. They push back. Sarah Eddie, member of OpenAI's legal team, tried in her closing argument to dull the attacks on Altman's credibility and to argue that there was never a firm agreement among the founders that could have been breached. Not one in this case other than Elon Musk has testified to any commitments or promises that Sam Altman or Greg Brockman or OpenAI made to Mr. Musk is what she's saying. And there is a new update that just dropped in after the recess. William Savit, OpenAI's lead counsel, told the jury that Musk does not have a claim against the startup unless there was a specific agreement between Musk and OpenAI describing how his donations to the nonprofit should be spent. That agreement does not exist, Savit says. So that's where I guess OpenAI is leaving it for now. We will continue to cover the story as it evolves. Is the jury allowed to use... could be done in one and a half hours. There's other tech problems going on. Max Zeff over at Wired has been covering the story as well and says Musk's lawyer brought a big monitor, maybe 36 inches, into the courtroom. OpenAI's lawyers asked to use it. Musk's lawyers said no. The judge told Musk's lawyers that they have to let OpenAI use it. Then OpenAI said it might not be possible to connect our laptops to it. AGI is here, but we'll still need a dongle. I suppose a dongle has entered the courtroom. Says actually there's about 15 lawyers standing in the middle of the room right now talking about how to use this big monitor. This is wild. They should have talked to OpenAI about sharing their monitor. What I always do, I always tell you when you come in here, talk to the other side. We don't have the technology available right now, so we don't want to use the TV. We think we should just get rid of it, says the OpenAI lawyer. Sam Altman just walked into the room, by the way. So that happened four hours ago. One of Musk's lawyers carried the big monitor out of the room upside down, wire dragging behind him. Defeated lion retreats. That is a very, very funny story.

In other news, break it down. Tim Draper says, 'I think I broke a record. I took 52 pitches in 52 minutes at below 40 degrees. Welcome to my office. #DraperUniversity #SurvivalTraining.' What do we think about going in the ice tank?

How cold are ice baths typically? You've done ice baths. I feel like I did one and it wasn't as insanely difficult as people said, but then I checked the temperature and I don't think it was 40. I think it was closer to 50.

Yeah, you can totally get closer to... I... because there's a couple companies that sell... personally, when you're going surfing and the water is below 45 degrees, it can just be very painful. So even in a wetsuit, anywhere that's not covered, a lot of people are putting gloves on. Booty, what do you think?

So apparently Joe Rogan's at like 34.

34?

Yeah.

Wow.

So that's like the cold plunge. He's the top of the mountain when it comes to ice baths.

He's the final boss.

Yeah, this is just a crazy picture. I did think it was AI. But turns out it's real. It's just funny because it looks like, what is this setup?

Yeah. What are all the trash bags there? And the wall is like sort of decrepit and there's piping.

It looks like kind of like a prison ice bath.

Yeah. This is not what you'd expect from... I mean, isn't he a billionaire investor? You'd expect some sort of palatial... You see the properties that Mark Zuckerberg's acquiring, that big investors are acquiring, you would expect something that would be much more regal. But he's doing it the old-fashioned way. Whipped this up himself, bought some trash bags and took some pitches. Yeah, and you know, who knows? Maybe the next founder of Cursor, Figma, Ramp is right now.

52 pitches in 52 minutes is crazy.

A minute is crazy fast for a pitch. I mean, we do 10-minute interviews, 15-minute interviews. Barely get to the meat of the interview.

And this one, you got four minutes.

Four of the founders.

Four founders jumping in one minute. That is remarkable. I am not a stranger to controversy though.

Yeah. Joe Londale says, 'I am not a humble man, but this is legitimately beyond my capabilities.' Absolutely wild. Well, Vercel, friend of the show, is apparently running an ad campaign on Lyft by buying custom license plates and deploying them through Lyft drivers. Is that what's going on here? No. You think it's random? The guy Peter, the driver, was like, he must love Vercel or worked there or something. I don't know.

If he was the eighth employee at Vercel, I don't think he'd be driving.

Hopefully not.

Unless he just loves it. Truly just loves the game. Loves driving.

Or he's just super illiquid. He's just like, pay me zero. Actually, I'll drive Lyft. I want all equity. I'm super bullish on Vercel.

That's a possibility. Well, Alex Conrad says, 'Is your startup even sponsoring Lyft license plates yet?' It's an outside-the-box strategy. Someone should pick it up. Someone should do it. Get a bunch of license plates for cars, rent them out to Lyft drivers, get those impressions. Wix is down a bunch. This seems like a very logical company to suffer in the age of vibe coding. People are vibe coding websites all the time and Wix is a supplier, you know, service to build websites based on templates. But Wix was buying 30% of its shares at $92 six weeks ago, but the stock is now down another 45%. And so I was wondering about this. I almost asked Max Levchin about this yesterday, but when you're going through this world, like it seemed like he was very confident about the SaaS apocalypse and did not feel the need to respond or take any dramatic actions, just sort of wait and let the metrics do the talking. But I was wondering about, you know, are you tempted as a CEO when your stock trades down on a narrative that you know does not apply to you but you're just sort of collateral damage, are you tempted to do a quick buyback and just sort of get a good deal on your stock even if it's just three months down then right back up? Imagine being a public company CEO and buying back your stock and then getting a return on it has to be one of the most euphoric experiences.

Totally. Not actually getting a return, but obviously decreasing or increasing everyone's... Wix is a 2.9 billion company now.

Yeah. They acquired this company Base44. Remember this was like a one-person company and they were growing. I think Base44 has been growing revenue quite quickly. It seems like pretty much any of these vibe coding tools. Yeah. Just the experience is so magical for people that a lot of them have grown revenue.

Stock chart if you zoom all the way out. So during COVID, 2021, ZERP era, stock was at $300 a share. It's at 52 today by the way. It traded down after ZERP era ended all the way to $50 a share, $60 a share, and then post-ChatGPT moment, 2024, fantastic for the stock, it gets back up all the way to $250, $200 a share. But then since 2025, as AI has gotten better at coding, vibe coding websites, doing front-end design, there has been a significant selloff that continues today. And so rough go.

I was looking to get a comp. I looked up Squarespace. Squarespace is no longer publicly traded. It was traded on the NYSE, but it was delisted after being taken private at 7.2 billion.

That is tough timing. Taken private in October 17th, 2024.

Oh, interesting.

And at the time there was not a SaaS apocalypse narrative. You couldn't one-shot a beautiful website with a single prompt. It's going to be so hard for this firm to make money on this deal.

Yeah, it feels like a new customer problem just because it's not the hot new technology that you're hearing about. Like the podcast ad conversion has to be a lot worse. But I would be very interested to know what is retention like because I know some people that have built businesses basically.

Yeah. I know some people that have built these website generator companies and then they just keep growing and growing and just sticking around forever because once someone has the magical experience of building a little website for their company or their personal brand and then they just let it run forever and they're like, ah, 10 bucks a month, I'll just let it keep going. Well, yeah, Squarespace had done around 1 billion of revenue in 2023. I'm assuming they grew into 2024. We don't have the full year numbers because it was taken private in Q4, but it's pretty reasonable revenue multiple. But if they lose out on a lot of those new customers because there's every single company in the world, every single company in the world it seems like is trying to make a box that will make you a website.

Yeah. Yeah. Everyone. Anyway, you know what? Very few companies are making a nightstand that turns into a bat and a shield for defense. I like this. It looks so unassuming as a nightstand. Very believable. No one would guess, but then something happens. You grab your bat and shield and you're ready to rock.

Would you pick one of these up? It has a little bit of a hotel vibe to it. It doesn't... And also, I like a nightstand that has... All I would say is don't bring a nightstand to a gunfight.

Okay. Yeah. Well, people are having fun with the AI-generated videos showing that. Yes. In fact, if it's not bulletproof, it has some trouble. If you disable... Ben Thompson says, 'If you disable open at login for the Gemini app launcher that the Gemini app installs in the background without asking, Gemini app launcher will immediately re-enable open at login. I will now, needless to say, delete the Gemini app and don't intend to install it ever again.' And so this is very, very odd. Gemini login. Oh, so it automatically logs in no matter what. He says, 'I'm actually struggling to remember a bigger middle finger to a user from an app ever. It's bad enough to install a helper app, but to immediately undo the user's explicit setting change, incredible.' And Josh Woodward from Google chimed in and said, 'This is a bug. It will be fixed in the next release, aiming for right after Google I/O, more if you're interested.' So that's good. They did receive the feedback. Well, we should talk about Nikita Bier. I was reminded of this because he screenshotted and posted it, the greatest growth hack of his career for one of his projects. This happened, was this a year ago? Gas or Explode app? This was about a year and a half ago, pre-joining X and working with Elon Musk over at X. He launched a company called Explode or an app called Explode. And he had a very interesting growth hack where he incorporated the company as TapGet Inc. And so in the iPhone App Store description under the name of the app Explode, it would say TapGet and then right below it would say Get because it's a free app.

And it doubles down on the call to action.

He made the entity a call to action.

It's genius. These little things really add up and you've seen them all over X. And he's done a good job of creating re-engaging areas and I just feel like the UI of X has been improving significantly. I'm really enjoying the latest UI feature where if you're watching a video and you want to speed it up, you can hold on the right side of the screen, which is fairly common in video apps these days. Doesn't work in the iOS native video player. I don't even know if it works on YouTube, maybe. But what's really cool is that if you press and hold it, you will temporarily be in 2x speed mode. But now in X, if you press hold and you're in 2x speed mode and you drag down, it fills a little circle and keeps you locked in that 2x speed mode and it actually changes the speed of the video permanently until you change it back. And so that little delightful touch is something that I'm seeing more and more of from the X team and I'm a big fan of. Well, without further ado, we have Andrew Feldman from Cerebras in the waiting room. Let's bring him in to the TV room. Andrew, great to see you again.

Looking sharp.

Feeling sharp. How you guys doing? Feeling... Congratulations. How has the day been? I would love to get just your reactions from the day. It seemed like there were a lot of people there. Take us through your emotions today.

Well, you know, this was better than we'd hoped for. All right. I think a chance to celebrate. We did bring a lot of people from the company and we brought families and to share with the team. We brought everybody who'd been at the company for longer than nine years and their families. We, you know, when you do a startup, the family is a meaningful part. It takes patience from them and a great deal of it. And so they came and we celebrated. It was really an extraordinary day. We opened up, you know, we priced at 185, we opened up 350 and we settled at about 320. What an extraordinary thing. We're just so proud.

Yeah. Take us through some of the history of Cerebras. Has it been a straight shot? Has it been an overnight success? How do you characterize it? What were the darkest moments? What were the highlights? What are the good old days to you? What does that mean?

Well, look, I think in the hardware business, if anybody tells you it's a straight shot, you can call BS. I just don't think that's the way our business works. I think the first time you build a chip with a new architecture, it's a little more than a prototype, a little more than a proof of concept. The second chip, you iron out your challenges and you begin to show it to customers in mass. The third one often that really takes off. And so it's a long, long road in innovative hardware designs. And so, you know, we were founded in 2016. We're more than 10 years old. We sought to solve problems that other people thought were impossible. As we showed you last time, you know, we tried to build a chip that was the size of a dinner plate. And everybody told us it was impossible. And the truth is, for a while it was. And you know, we didn't solve it until August of 2019. We built this extraordinary chip. We were faster than everybody. And absolutely nobody cared. Nobody. And AI wasn't ready and it was still sort of a novelty and nobody cares about how fast you are when it's a novelty. But starting with GPT and in 2025, the models got so darn smart they became useful and suddenly everybody wanted to use AI and you use it with inference and business was rolling.

Yeah. What were those early rounds like? Become thinking the benchmark round, CO2, bunch, you know, Eclipse, a bunch of others.

You know, we had the advantage of the founding team had been together at our last company that had paid pretty well for the venture capitalists and the team. And so we had some wind in our sails when we went out and raised money. It's not like today where we're four guys and the word lab and you're raising a billion pre for your A. That's not us. But we went out, we made eight calls, we got eight term sheets, we chose Benchmark and Foundation and Eclipse and we got going, you know, less than a year later.

I was expecting you to say like, yeah, I mean, it was a slog, you know, we were so... Other rounds were a slog. Other rounds were a slog at the beginning. Not so. You know, Thomas Lefant at Coatue came in shortly thereafter and we did a round with them. I think the truth is between about 2020 and 2023, it was much harder. AI was sort of in this situation where everybody was saying, 'Oh, that's cool. Look what this model can do. Look how big it is.' But it wasn't being used anywhere. Right. Nobody was using it. They were pointing at it. They were saying, 'Wouldn't this be nice?' and they went back to whatever they were doing before. And it wasn't until really sort of 2025 when the models got good and you just saw this tidal wave of people using AI and demand for AI compute and that's been exceptional. It's just been an amazing thing to ride.

Yeah. You mentioned like if you have four guys and your company name ends with lab, you can raise a billion dollars. There's a little bit of that going on in the market with just like chips, semiconductors, AI. There's not that much that needs to be explained, but what were the key ideas or thesis that you needed to explain in the roadshow to investors that wanted to go a layer deeper than just AI chips?

Yeah, I think there were... First, the market size and dynamic. And I think Jensen said some time ago on Brad Gerstner's podcast that the demand for inference will grow by a million x and nobody believed him. And you know, at the same time you saw Sam Altman, you know, displaying real vision and going out and trying to lock up huge amounts of compute and memory and data center and power because he saw it too. And I think trying to share what that means, what an exponential demand means, and that we're still so early and yet the demand for AI compute is overwhelming. I think sharing that was interesting and I think helpful in educating the financial community. The other thing is that there are lots of ways to do this. The GPU isn't the only way. You've got a TPU, you've got Trainium, you've got us. There are lots of different ways to build a solution here. And finally, maybe the notion that CUDA is sort of this grand lock-in is overplayed and that, you know, the Gemini 3, which is an excellent model, was trained on TPUs with no CUDA, that Anthropic's models were trained on Trainium with no CUDA. I mean, that lo and behold, some of the best models, some of the most interesting things are being done without CUDA and that lock-in might be overplayed. And I think these three factors were really important in educating the financial community.

Going forward, how do you and the team think about sort of calling your shot and sort of trying to predict where and how inference demand will look in, you know, 2030 and beyond versus like working closely with the labs that now have, you know, product lines with billions of dollars of revenue and their own roadmaps that you can work with.

Yeah. You know, like the Babe, I'm going to point out to left field and just say, wait, this is where it's going, baby. No, I don't think that's the way it works. Look, I think we're calling our shots every day by making big investments in data center capacity and collaborating with the leading visionaries in the field in working not just with OpenAI to service sort of the cutting edge and deliver their extraordinary models, but also with AWS to make sure that we can get access to the largest enterprise customers. And instead of having to work with these enterprise customers' procurement organizations who provide master purchase agreements that are the size of a Bible, you know, you can say, 'Look, why don't you buy us through AWS and it'll count against your annual commitment.' And so I think those are really important ideas and ways we get access to the market. And then we're taking huge amounts of data center capacity and so that's the other bet we're making.

Yeah. Makes a lot of sense. How do you think the year will play out in terms of just broader consumer awareness of what fast inference feels like? I had a really magical moment using Cerebras in GPT 5.3 Spark in Claude Code and even outside of coding tasks, just talking to the model and having it respond instantly was sort of... it felt like a new breakthrough or new paradigm. And I feel like this hasn't fully diffused, but it also feels like when it does, there will be potentially like entirely new ways of working, entirely new paradigms that might emerge. How are you thinking about actually diffusing the technology?

We think that's exactly right. And we think that the experience of engaging with a real-time AI will encourage people to do more things, to stay longer, to work on harder problems and to invent new things. I mean, if you remember, you know, when Netflix started, they delivered DVDs in envelopes. Right. And when the internet got fast, they became a movie studio. And they didn't get better at DVD delivery. They became something completely different, something that had never been in existence before, a movie studio that delivered directly to your home. I think that's exactly what's going to happen. And you can just sit back and you can ask yourself, I mean, how big is the market for slow search? Zero. How big is the market for dial-up internet? I mean, how much would I have to pay you to swap out broadband at home and bring in dial-up? Right. 1,000 a month, 1,500 a month, 2,000 a month. I mean, no way. I mean, it just wouldn't be worth it. And so the community is going to engage with inference in the same way. And that fast inference is going to be all of the market.

Yeah. So you make the chips. I believe you also make cooling infrastructure as well, cooling units. Are there other products on the roadmap that you think will be required to roll out and scale Cerebras over the next couple years?

No, I don't think so. I think right now we build the chip and the system and the system includes... it's about the size of a dorm room fridge there. You put two of them in a standard data center rack and the cooling infrastructure is built into the system. And I think that's where we want to focus. We want to be measured on our ability to build AI computers that are faster than anybody else.

Yeah. How are you thinking about scaling on-chip memory? It feels like there's some concern about, well, what if the models go to 10 trillion parameters? What if it gets too big? How are you thinking about that challenge or maybe it's an opportunity?

It is an opportunity. I think a 10 trillion parameter model is hard for everybody. It's actually easier for us. Okay. Right. There's a reason we're not at 10 trillion. It's because it's really hard and expensive to serve for everybody. I think one of the things that we've been able to do for the larger models is to tie together a bunch of...

Mhm. And that way we can train and do inference on trillion, multi-trillion parameter models in ways that I think are much more intuitive than on GPUs that have much smaller compute. They have off-chip memory, but their problem is the compute. They don't have enough compute per chip. And then how are you talking to customers about potentially bringing Cerebras in not as a full replacement to their entire semiconductor supply chain or stack, but as a complement to everything else that they're running? Because I have this vision of like the next generation of AI agents. You get this genius model, but it needs to use a small model over here, an open source model over there, a super fast model for a certain thing. If it's looping through some...

Yeah, delegation. How are you thinking about that?

What do you think the company would look like today if you guys had access to today's frontier models when you started the company? Like, are you feeling like how, and how do you think about just like the speed up in, you know, at the company today due to how good the models have gotten?