I see a large IPO on the horizon.
You're surrounded by journalists. Hold your position.
Overnight success.
The place 3000 double right.
There's misinformation.
Courtney declaring order inbound. Let's just roll. We are surrounded by journalists. Hold your position. Come on. Get up. Trust the experts here. Five to our expert. Founder mode find.
I see multiple journalists on the horizon. Standby.
UAV online. Blaze. Double blaze. Triple blades. Double kill.
Cook is w team deathmatch. We are experts. Triple blade. Let's just roll right. Clearing order inbound.
We are surrounded by journalists. Hold your position. Strike one. Strike two. Activate golden retriever mode.
Mark clearing order inbound. 5. I see more journalists on the horizon. Founder.
You're watching TVP. Today is Wednesday, March 4th, 2026. New haircut. New haircut alert. We are live from the TV channel, the template of technology, the fortress, finance, the capital capital. Let me tell you about ramp.com. Time is money. Save both. Easy to use corporate cards, bill accounting, and a whole lot more all in one place. Uh we have a little bit of a shorter show. Let's pull up the linear lineup. Linear, of course, is the system for modern software development. 70% of enterprise workspaces on linear are using agents. We have Dean Ball coming on to give the other side of the Ball versus uh Thompson argument uh about anthropic. Then we have Scott Kupor and Jared Isaacman coming on to talk about space. We're going to the moon and they're going to take us there. And then we have Adam Bry from Skydio who I actually worked with through Andre Horowitz uh maybe 10 years ago. And then Matteo, CEO of Eight Sleep has some exciting news. And Dillon's coming on to break down the latest AI agent Hermes agent.
Well, anyway, uh there was some big news uh this week. Of course, everyone knew that Apple was launching new products. Uh but there were some surprises and uh there's some cool stuff starting with the MacBook Neo. Um my the the more downstream take is like how is Apple affected by the AI boom? But let's start with the actual products because uh this thing in the video it looks so small. It looks like an 8 inch laptop. I guess this is uh so this is the MacBook Neo sort of designed to compete with the Chromebook. Uh it comes at $599.
Well, not even just the Chromebook, just all PCs.
Uh yeah, but I mean you're not going to get this for like if you're in a professional work place, go to a desktop with multiple monitors, plug in, you're doing like advanced work, like this is for students. This is for the low end. And
this is a personal computer. I think I think it's designed for people that have an iPhone but don't have a Mac.
And now they can have an integrated suite.
It's Yeah. I mean it's a it's a crazy product. I I love the I love the launch video, too. Uh it's very weird when they start because it's just this like blank block and then they're adding features to it with CGI. Uh it does a really good job. It's sort of anticlimactic because every feature they add is just like what you'd expect. Yeah. I was kind of like, "Oh, is this going to be a touchscreen?" And then it's like, oh no, it's just a normal screen. It's like, oh, like, guess what? It has a keyboard. Boom. Mic drop. Oh, it has two USB-C ports. It is kind of cool that they put the uh the headphone jack a little bit further down. Thought that was cool. And then they're like, "Check it out. You're never going to guess this. A camera. We put it on the computer."
They're like, "Hey, we heard you liked having an iPad that you could attach a keyboard to." So, we basically took an iPad and put a keyboard. But it's very cool how how they like sort of expand the product like show you. Okay, there's the speaker. It's pretty big. It goes on the side. There's two of them. Shows you. But this does not look like a 13-inch computer to me. This looks like an 8 in computer to me. This looks like a small iPad. And 13 in is a very normal size.
Maybe the model is just like Shaq or something.
Yeah. Yeah. No, I I think they actually use someone with the big hands to make it like maybe look smaller or something.
But uh 13-inch display, 16 hour battery life. The price is the really crazy thing. So, $599 for students, it's $100 cheaper. So, $499. That is one of the cheapest Apple products they've ever made. I'm pretty sure the headphones cost more. The Apple Pro uh Apple AirPods Max cost like 500 bucks. Uh, remarkably cheap A18 Pro chip. Uh, 256 gigs or 512 of storage, four colors. They got it in ramp yellow. They call it citrus, but I think everyone's going to be
And it now cost the same amount as the entry-level iPhone. Yeah.
17e. Yeah. And it has Wi-Fi 6E, Bluetooth 6 connectivity, 8 gigs of RAM, uh 1080p front-facing camera. Like people are sort of like ah it's underpowered. But look at the price. This is this is crazy. The last time Apple had a product that was anywhere near this cheap was um well where where was it? Uh let's see. I I wrote about this. So um in 2014, so 12 years ago, it's long time ago. uh they sold the MacBook Air for $899 and now they have this down at $499 or $599 uh which is remarkable. Really really cheap. Uh cheap for 2026, cheap for an Apple product, even cheap when compared to other laptops. So uh this is sort of like their their their escape hatch, their pressure release valve. So they can actually increase the price on the other products more because if you're price sensitive, well, they have a product for you. And there was something similar that happened with the the new suite of iPhones where the iPhone Pro Max was getting so big and the camera notch was so in your face and the phone was bigger and heavier and the battery life was great. It was very powerful and a lot of people want that. But there were also a lot of people who were complaining, I just want a thinner phone. Well, then then you have the iPhone Air and there's going to be crazy trade-offs. You only get, I think, one camera. The battery life's not nearly as good.
It lags. But for some people, if they're going to be complaining about the all the features of the Pro Max, it's like, "Well, we got the Air for you. Stop complaining." And I think that's a lot of the strategy uh there. So, um the other products they launched, MacBook Air with a with an M5 chip, MacBook Pro with the M5 Pro and M5 Max chip.
Uh did you wind up going How much memory did you wind up going with?
I feel a little I feel a little soft.
Is that enough to pull up tweets and react to them on a live stream? What's up there?
Uh, apparently you you actually have been maxing out something like 32 gigs of RAM, which uh which took me by surprise. I have not had I've not been memory constrained, but I do I do actually love the idea of getting the top-end memory, which is the thing that constrains you on running LLMs locally. Is that correct, Tyler?
Yeah. I mean, you can always like uh if you have a really good model, you can like get around the memory thing, but then it gets like super slow because you're like loading it in and out.
Okay. You're loading it in and out for like every token that you're generating.
You need a lot of memory swap. Yeah.
But but so so if I go with the 128 gig, the top-of-the-line chip, I should be able to inference like a pretty smart model locally, pretty reliably. I mean, it's also like the you can always just quantize the models down and so like you can actually run like really frontier like source models on like pretty normal hardware. Now I'm I I'm just purely thinking for you know doomsday prepping the idea of you know how you have your library in case you're bored during the apocalypse. You have your DVD collection. You can watch all the movies from start. That's that's the big one for you because you're going to be able to watch all of American cinema from the beginning.
I I still I still don't think I'd actually watch
You wouldn't do that. What would you be doing? Assuming that the internet's down in this apocalyptic scenario, you can't scroll. There's plenty of other things to do.
Take a walk on the beach.
Uh, what if you can't go outside because there's nuclear fallout everywhere. You're stuck in your bunker. What are you doing?
Whittling. Did you get into whittling?
I actually used to love whittling as a kid.
Called it. He's not going to be watching movies. He's going to be whittling. I will be whittling while I watch movies. While I inference a local model that I can just ask questions to and and get all sorts of history. And there might be a few few uh a few hallucinations because it's not clearing the internet because of course in this hypothetical scenario the internet is gone. Um but I would still have access to you know all the pre-trained uh information and basically the world's best encyclopedia which would be fun. Um what else did they launch? Uh in a world uh in a world where computers keep going up in price it's kind of wild to see Apple drop a 599 laptop says Theo. I could not agree more. Fry says, "Oh man, they're going to sell five billion of these things." Yeah, this is uh this feels this feels like more of a no-brainer for kids than an iPad for some reason. I don't know if I'm just like old school and I'm like, if it has a keyboard, it's horrible.
iPad iPads with kids just have really bad aesthetics. It does. Kid with an iPad just zoning out, right?
You get them a keyboard, they're locked in, but you get them there.
I think there's something here.
You're like, are they are they day trading?
Yeah, they're doing something. Uh for sure.
Uh it might be underpowered, but uh it'll get the job done. And they I I I think there's something about, you know, typing on a keyboard that does lend itself to more creation and less consumption. Like the iPad is a media consumption device. That's where where most people use it for. Throw on Netflix, play a game, do something there. There aren't that many people that are power users, but as soon as you get the keyboard, you're ready to type. You're ready to you're ready to work. Let me tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web apps, servers, databases, and more. Well, Railway automatically takes care of scaling, monitoring, and security. And let me also tell you about Octa. Octa helps you assign every AI agent a trusted identity. So, you get the power of AI without the risk. Secure every agent, secure any agent. So, uh, hilarious chart from Andre Horowitz. Apple on capex. Nah, we're good. Uh, they looked at the standard standardized quarterly capital expenditure for Amazon, Microsoft, Alphabet, Meta, and Apple. And all of the four companies that are working on AI, Amazon obviously through AWS and Anthropic, Microsoft through OpenAI, Alphabet through Gemini, and Meta through MSL, they are all absolutely off to the races, going to the moon on capex. And Apple is down 19% since uh 2015 on this standardized uh metric that A16Z put together. Very, very funny. And uh it's just interesting that Apple has been so unaffected by AI costs. Not just on the capex side, which is where every hyperscaler sort of had to say, "Yep, you know what? Our business is fundamentally different now. Uh we are not the high margin cash flow generating business. We're going to draw down on that cash. We might be taking on some debt." Like this is a big big different financial model. It's a very different financial model from years past. And buckle up. You're coming with us shareholders. And the stocks have done well because a lot of shareholders have believed in the goal, believed in the value of AI, believed in the value of building compute infrastructure broadly, and they've been rewarded for it in most cases. Uh, but Apple's just sort of sat that out and said, "Hey, we're not we're not going to train a foundation model. We're not going to build a whole bunch of compute infrastructure.
We're going to serve Gemini.
We're going to serve Gemini and we'll pay them what, a billion dollars a year, something like that." It's like,
yeah, I mean, this this year will be really interesting. It'll be fascinating to see what the new Siri actually looks like, what it can really do. We'll start to see that, I think, in Q2.
And then separately, we're expecting new devices from OpenAI, right? Teasing the puck, dime, whatever you want to call it. Uh it'll be interesting to see if they can kind of break the the uh AI hardware curse that I feel like is
well I I wanted to think about that because I I wonder if the story is over for some of those like cursed bad launches product like we saw friend apparently has found a little niche very controversial but Humane has sold to HP it seems like they're not working on that product anymore the pin the AI pin but the Rabbit R1, I think his name's Jesse. Uh, you know, that product, there's there's something there where with enough iterations, I feel like that might sell well. I was playing with the board the other day. Uh, that hardware, it's like a huge screen, touchscreen that you can play games on. Like, there's still these niche applications that I think you could potentially, if you have a hardware team, you have the ability to sell something. It's It wasn't polished. The AI wasn't good enough. Well, the AI got better. You still have the supply chain. Maybe there's a way to cut through, but it probably needs to be a little bit more niche. Like everyone's gonna have one. Tyler, what do you think about hardware in the age of AI?
I mean, I I think the Mac Mini is like a good example of like AI hardware right now.
Yeah. No, that's a great example. Yeah. I wonder I I I I wonder what uh like how exposed the Gemini APIs will be in iOS because if you if you fully expose like Gemini Kit like you know in iOS just like you can call like a weather API or maps API like Uber exists because there was like a there was like a GPS API that you could just call within an iOS application uh with a couple lines of code and you didn't have to like figure out where the user was because the iPhone just told you that as long as you clicked like, "Yeah, share my location with Uber." If you now get an app that has not just Apple's local LLMs that are sort of, you know, mediocre, but it can go and call Gemini and do a whole bunch of crazy stuff. you you start running into like will someone build an open claw app or something like how much of Apple's AI strategy will wind up being vended into the third party ecosystem which has served them so well for so long they've sort of turned their back on it they locked down Apple Vision Pro a little bit more there was a there was way less uptake there and and there's a ton of risk because if I develop like a free iOS app that and incurs like a ton of Gemini tokens and ton of charges. How do those get passed through? Is Apple just biting the bullet on it? But there's something interesting there where where turning app developers loose to to build software that takes advantage not just of local LLM capabilities, but Gemini under the hood uh and and sort of wrap it up in a neat bow that is, you know, available to the consumer could be could be very cool. Anyway, the big question that I had and I will get into the essay was uh so Apple's unaffected by capex costs, but they've also with this release been unaffected by memory costs and the RAM magdon. So I will I will take you through how I dug into the question of memory pricing at Apple. But first I will tell you about labelbox RL environments voice robotics evals and expert human data. Labelbox is the data factory behind the world's leading AI teams. And I will also tell you about console. Console builds AI agents that automate 70% of IT, HR, and finance support, giving employees resolution for access requests and password resets. That's a good sound effect. I like that one. I have one, too.
Can't believe you haven't used that yet.
I I know. No, no, because I was so burned by the last time I got access to the soundboard. I went too crazy. So, there was one one show where me and Jordy both had full soundboard access, all of the same sound cues, one single show. And it was truly a disaster.
What's the What's the metaphor from Greek mythology? It was the It was Pandora's box for me. Like, once I opened it, I could not resist and I became addicted and I was pushing the thing very much and it just didn't work. So, we were like, "No, let's have Jordy play the soundboard. I'll I'll do other stuff." the the issue for me as a as a soundboard artist is
when I'm not doing the show hang out
when I'm hanging out on the weekend I'm just thinking
I'm just I'm if somebody says something like we're at dinner
I'm thinking about the sound effect
I was I was hanging out with you and you said something crazy and I wanted to just be like
and so I opened up my phone I went to YouTube and I pulled up that sound and played it but then I had to turn on my volume and it took me way too long and the joke had passed. But we have a solution. Tyler Cosgrove over there has been working on a soundboard app that can go on your phone and you can have we should do a widget, too, so that it's just there on the home screen. Boop. Your favorite your favorite soundboard cues. So, uh, expect that soon. Anyway, let me tell you about AppLovin. Profitable advertising made easy with Axon.ai. Get access to over 1 billion daily active users and grow your business today. So the the the question the question like like I get I get the strategy of Apple saying look we're good with Google handling the AI inference stuff they have GCP they have the TPU they got Demis who's got a book written about him by Sebastian Mallaby he like like Deep Mind is great we're working with them. Um so capex unaffected I understand that. How are they not affected by memory prices? Ramageddon has come for us all. Um, and and and uh Ben Thompson of Stratechery and John Gruber of Daring Fireball actually had a bet about this. And I hope I didn't misspell John Gruber's name. Um, they had about they had a bet about this. They said whoever wins this bet gets a steak dinner. A steak dinner was at stake and Ben Thompson lost. John Gruber won because John Gruber said Apple always overcharges for RAM.
getting a if you went to build a computer or a phone or laptop or you know iPad equivalent and you just went to retail getting an extra 128 uh gigs of memory or or a slight memory upgrade uh might cost you 30 bucks in the Apple ecosystem that's like $200. So Apple has built in this uh this pricing buffer for a very very long time and and that buffer really saved them this year. But there's there's a lot more to it. So uh many companies are dealing with Ramageddon this year. Uh, and some stats that are interesting. DRAM prices rose 172% throughout 2025. DDR5 spot prices have quadrupled since September of 2025. Uh, Trend Force expects PCD RAM to roughly double in price just in Q1 of 2026 with LPDDR pricing seeing the steepest increases in history. So every wafer that's allocated to an HBM stack, a high bandwidth memory stack for an Nvidia GPU, is literally a wafer denied to the LPDDR5X module of a mid-range smartphone or consumer laptop SSD. So they are truly commodities. They there's only so many of them and they can only go to one place or the other. Apparently OpenAI's Stargate initiative alone could consume up to 40% of global DRAM output as it stands today. And this is the problem because there's only three major companies in the memory category. It's Samsung, SK Hynix, and Micron. Uh together they control 95% of the global DRAM production.
Remember Micron has been trying to build a new facility in New York and six people basically blocked it.
Oh, I I we talked about that, but
And the rumor The rumor is that it's like foreign funding.
Actually uh trying to uh hold Micron up, but
Yeah. Yeah. Uh I mean Micron has been very they've been a huge beneficiary of AI margins and higher prices. Uh and they so much so that they've actually reallocated all of their capacity to AI and they pulled out of their uh consumer brand which is called Crucial. So, if you ever tried to build a gaming PC, you probably found Crucial Ballistix, which was like this very aggressively RGB branded like PC Gamer uh DDR memory, which was uh you know,
like the monster energy of RAM.
Exactly. Couldn't have said it better myself. No, it really is like the the the branding on Crucial Ballistix is like insane. Uh anyway, gamers will be rising up soon with monster energies in hand to uh to take the the the RAM by by force if they need to. I already see memes about this all the time. Uh I see these like the worst AI slop video you've ever seen, just absolute junk. And then the caption will be like, "So this is why I can't afford RAM, you know, and or this is why RAM is $800 a stick or whatever." And then on the flip side, people will say, uh, you know, occasionally there'll be a good AI slop video. People be like, "Okay, it's worth paying $800 for RAM because this is so good." Right? Goes both ways.
It goes both ways. But clearly people are feeling the the the the pinch. Uh,
but Apple fans will be sitting out the fight thanks to a few key moves from Apple that help them avoid disaster this year. So, uh, one, Apple has long-term supply agreements that allow them to secure memory 12 to 24 months in advance. They also benefit from vertical integration uh and custom silicon. So instead of buying that consumer-grade commodity DRAM modules that you see go into either gaming PCs or Nvidia GPUs for example uh they negotiate multi-year contracts directly with Samsung and SK Hynix for custom LPDDR packages that meet Apple's exact specifications. And so I
Yeah. And so this is like RAM's always been cyclical. Yep. And so it makes a lot of sense for Apple to say basically like lock up pricing. Samsung's locking up the demand. They they I I don't even think Samsung would have predicted this level of demand even 3 4 years ago. Otherwise, they may not have priced priced it in the way that they did.
Of course, uh the other thing that that uh maybe you get to is
and Samsung's still making their default margin. They're just like the opportunity cost is so high because we could we could reallocate. The other thing is is they they probably have so much margin in the upsell that they could have some margin compression and still be
so Apple at various times has had over a hundred billion in cash. They currently have over 66 billion in cash. Like they can take the hit there. They can also just take a hit to their margins.
Every AI lab looking at that cash being like Tim I could really put that to you right now.
Yeah. They're the only hyperscaler that hasn't really shelled out to the tune of 10, 20, 30 billion into one of the frontier frontier labs. Maybe that changes. I don't know. Somebody was throwing out like uh Apple should buy anthropic or or uh or Amazon should buy anthropic, but I think the price is too high at this point. It's just too it's just getting away. Um
but uh so I I'm not I'm not a 100% sure about the how the manufacturing supply chain for memory works, but I believe that because Apple's specific requirements, it's harder for Samsung to just instantly reallocate uh uh product that's been already manufactured um in a way where Crucial Ballistix, if it's going into gaming PC, it's easier to reallocate those wafers. So even if they didn't have those contracts, I think I think uh the the supply chain is a little bit stickier with Apple because they're vertically integrated. Um and then yeah, of course they have the money to overpay. Uh we talked about how that the the MacBook Neo is this pressure release valve. So even if the memory does increase, they have this they have the cash buffer. They also have the they're charging 200 bucks for something that costs 30. If it goes to 60, they can just take a hit on margins buffer. Then they have the contract buffer. Then they also have this Neo where if the MacBook Pro goes up in price, the MacBook Pro buyer is probably less price sensitive. They're like, "Yeah, you know, this year it was $4,000 instead of $3,000. I'm buying it for business purposes. It's not that big of a squeeze." And then the price sensitive buyer says, "Yeah, you know what? I'm not going to go for the $2,000 laptop. I'm going to go for the $600 laptop this year." And they're still in the Apple ecosystem. Um, but my takeaway is like this is an incredible testament to the operational prowess of Apple uh to pull this off at this scale. The Sony PlayStation 6 is allegedly going to be delayed until 2028 or 2029 because of rising memory costs. They're going to wait it out because there are three players. It's not just TSMC in this case. Um, it does feel like there'll be more pricing competition and and there's more of a race to build capacity. Whereas a TSMC, if they don't build the next FAB, they can just continually raise prices and and increase margins uh once the next contracts go up for renewal. Whereas if if Samsung has a bunch of capacity and Micron doesn't, like Samsung wins. And so there's more of this game theoretic like they got to they have to build. And so I think a lot of people and Tyler was talking about this with me earlier like we we all expect this memory shortage to be uh a few years not more not like five or 10 years. So if you push out the Sony PS6 to
you not bullish 28. He's bullish.
No I'm very bullish but I I think yeah because because there's more players like I I think the the shortage will be uh
say you're not AGI pilled. What do you mean?
It I mean there is a question about like if you truly go exponential forever and it just like okay yeah there's like you know trillions and trillions of dollars of of data centers being built. It's like can even Samsung Micron SK Hynix keep up forever like can
Yeah. No. So it might we might be living in like a permanently higher price just because like it's a more valuable product like we get more economic time might be different
Um also the Nintendo Switch 2 might go up in price next year. It's really bad for for gamers. Gamer gaming graphics cards are going up in price. Nvidia is pulling out of gaming graphics cards or they're skipping a year. So, it's mayhem for gamers. Uh but long term, the competition between the memory manufacturers should stabilize prices. Uh and uh but you know, just watching all the big companies that are not Apple get caught flatfooted is just crazy to watch because it's not like it's not like the folks at Nintendo or the folks at Sony are like, "Oh, yeah, first time manufacturing something." Like if you told me that like the team behind the Rabbit R1 was like suffering from higher memory costs, I'd be like, "Yeah, like it's a new company, new product. They haven't like Sony has been making PlayStations for decades and they still got caught flatfooted on this." And a lot of that is a testament to the brand power that Apple has, the network effects, the the the the scale they built that they are in this position and still not getting squeezed even when other hyperscalers are getting squeezed on all sorts of stuff.
Tim Cook still underrated. He cooked. He has cooked.
And uh and yeah, the operational side of Apple is always it's just habitually underrated even in the face of like slight misses on features and products and uh and and marketing and oh this this ad didn't land. Well, the business is still thriving and dealing with just crisis after crisis like the tariff should have been a major crisis, they got through it. This memory ramageddon should have been a crisis, they got through it. Uh, it's all been very, very good news for Apple. Let me tell you about 11 Labs. Build intelligent real-time conversational agents. Reimagine human technology interaction with 11 Labs and get that chuch-ching ready because I'm going to tell you about phantom cash. Fund your wallet without exchanges or middle end and spend with the Phantom card. Um, Apple is exercising the Coast conjecture. They're waiting for open source models to get better. Applying pressure to Frontier Lab pricing, eg rejecting anthropic overtures. Anthropic wanted more money from Apple. They went with Google and Gemini. This is from Saurin Larson who is an economist who I really respect. Uh it may rather be that im it may rather be that impatient buyers have adverse selection revealing actually poor business positioning. So uh he is uh he's he's he's identifying uh Apple's leverage in the system. They're you know this consumer aggregator and they've been uh they've been doing very well. Um Kyle Harrison says that quote we ran the math. OpenAI is buying three to four times more memory than it could possibly need in the short term. The most charitable explanation is aggressive forward positioning. The less charitable one is that they're cornering the supply to kill on-device AI before it starts. Well, if the on-device AI happens on Apple products, that doesn't seem like that real of a scenario, but it is uh it is an interesting thesis. I think that uh across the frontier labs, everyone is just very AGI pilled. They're seeing the the the increase in capabilities, but also the the these sort of like stacked exponentials. Ben Thompson was explaining this today that when you went from LLMs to reasoning, the number of tokens, the number of compute required just to get a single answer went like, you know, 10x all of a sudden. And and then we saw the same thing with coding agents where these long running tasks are are sort of like squaring the amount of compute and so you just keep stacking these things on and you just get more and more compute demand but you get more uh more economic output. So it winds up being all worth it. Uh when will the when will Apple release the iPhone? The only other thing only other thing on the on OpenAI uh memory side is
they plan to sell a lot of consumer devices.
They they do. Oh, so they might be buying it for that, right?
Yeah. So So we we don't really know, but uh but yeah, I think that uh obviously opening eyes kind of rolled back some of like that that $1.2 trillion number. Uh but seeing seeing the growth in demand just in this first couple months of the year starts to look like hey maybe maybe maybe he wasn't so crazy after all.
Yeah. Yeah. Yeah. Totally. Like the the the models are justifying the compute. And I think that we talked about this yesterday a little bit with the orchestrators just this idea that uh you know where will the next 10x come from? Well, it's probably from just running 10 instances of cloud code or codeex and and just firing off 10 times as much inference and and people are ready to do that. They're you see it all the time on Twitter, people talking about how much inference they're they're doing. Um, when will Apple release the iPhone 18? Uh, Khi has before October at 35% before 2027 at uh at 34%. What is the chance that they don't release it this year? That would be probably zero. And I think it is it is zero. So, uh, stay tuned. Um, I'm sure Mark Gurman will have good coverage on um on the next iPhone.
Let's get the Gurminator back on.
Yeah. Well, well, I I I I wanted to invite John Gruber to talk about these, but uh Mark Gurman, of course, is welcome at any time. Uh, but we've we have yet to talk to uh John Gruber, and he has been covering Apple fantastically throughout his entire career. So he would be an interesting gentleman to talk to. Let me tell you about Gusto first. Gusto is the unified platform for payroll benefits and HR built to evolve with modern small and medium-size businesses. And let me tell you about reream. One live stream 30 plus destinations. If you want to multiream go to reream.com
submitted a new semiconductor fab construction plan for the southern Taiwan science park last week. M the site is located on Tynan's Anding District adjacent to the F18 FAB and covers 15.46 hectares. They're doing more 2nm production, the leading edge with eight hectares allocated for the production area and one hector for administrative facilities. The project is expected for completion and production start in 2028 says semi analysis. So, uh, we'll see what what happens in the actual capex guide from TSMC. Ben Thompson has been very disappointed with where, uh, TSMC has been guiding on capex, but maybe this is a shift. Uh, I wonder how much faster it will be to build another TSMC fab over there. If they ramp production of this new facility before Arizona, if they if they eclipse and and and there's a flipping and and this gets to scale before Arizona, um it will reveal some very serious problems with America's uh building and permitting system, I would imagine, or maybe maybe talent pool. Uh I would I would certainly hope that both uh that that Taiwan, I mean, this is their crown jewel. They should they should get this up and running fairly quickly without many delays since they've done it before and they are heavily invested in TSMC. Of course,
more importantly, Wendy's US president post new video amid burger battle with McDonald's and Burger King.
Let's see. Let's watch. Let's rate it.
Okay. I like the way that he's holding it. You know, he's
That's a burger. This is good.
I like talking talking about food in his mouth. Okay. Yeah. Going going back for a second.
He's saying burger. He's not saying product.
Yeah. No hesitation. No hesitation.
Wait. What did he dip in the frosty? Was that bait? What did he What did he do at the end there?
He dips a French fry in the frosty. Whoa. Okay.
Okay. Doesn't call it a product. No hesitation talking with the food in his mouth.
Yeah. Yeah. That's bold. That's bold.
I'm giving this an 8 out of 10.
If you scroll down, I like uh Colonel Sanders after seeing this just knowing.
Also also the way he kind of puts the fry in and quickly, you know, kind of pulls pulls the hand back.
I I I It's got style points.
Dude, here's a good job. If you're looking for a gig that will bring home the bacon, uh, apply for the job of CEO at McDonald's. Apparently, the guy makes 18 million a year.
Wait, do they have an open job application?
No, but after this, it's entirely possible that that if you if you put on a masterful performance, chomping down, uh, you could make you could make a run. I don't know. It is funny that there's just like one after another of CEOs. Uh, this is probably good for all of them. I don't know. Uh, this is this is fun. People are I want no jump cuts. Somebody should have seen swall Somebody should be seen swallowing the food. Oh, okay. People are people are analyzing.
That's somewhat somewhat fair.
Analyzing the details. Let me tell you about Gemini 3.1 Pro. Gemini 3.1 Pro is here with a more capable baseline. It's great for super complex tasks like visualizing difficult concepts, synthesizing data into a single view, or bringing creative projects to life. Uh Berber Jin came in and shared Alman defended open AAI's defense department contract in all hands yesterday calling the backlash quote really painful. This was an example of a complex but the right decision with extremely difficult brand consequences and very negative PR for us in the short term. he told staff uh yeah I I still I still stand by that uh given given the events that have transpired in the world yeah
like being there to like meet meet the government
feels like uh being on uh on the right path
it was just yeah it was just odd that uh the that like Google strategy was to not say anything do nothing win basically I I would think or or do nothing don't go up or down like uh Gemini just
but but I don't think Gemini is involved in this contract. It's it's Grok.
Grok was the other one that that uh came in with the same term.
So Gemini doesn't have any relationship with the Department of War. I think they do. I I I think I I thought that there was a there was
I might be I might be they might I just don't think they were involved in this contract negotiation
because I've seen Yeah, I've seen Palantir like show that they have like three or four models that they sort of like mix together to get responses. I don't know. Uh OpenAI announced its Defense Department deal on Friday, hours after the defense secretary Pete Hegseth designated rival Anthropic a supply chain risk. There's still a question as to whether or not that's going to go through. So, uh, Khi has it at a 42.8% chance and it hasn't really moved. It's unclear if that's just like staying in tweet land and is like bluster or will there be something there or will will there I mean I have to imagine that at this point uh both the department of war and anthropic are talking to each other. Do you have more information?
Uh yeah, Gemini does have deals with the DoD. Yeah. So there was the 200 million one that was uh in July. That was Anthropic was also in that one. And then they did there was the recent like genai.mil. Do you remember was tweeting about this?
Um but they're like involved in that too.
Yeah. But I I am what I'm so shocked by is how does AWS have an advantage in 2026 on Fed Ramp and like federal classified networks? Like I get that it's like hard. You have to do a lot of cyber security. you have to really harden the servers. But you would think that Microsoft and Google would have figured out how to fully support the deepest depths of the US government and get full approval before like like may maybe Amazon would have gone first, but I would have thought that Microsoft and Google would be like a year or two behind. And so you would see you would see Gemini and OpenAI like checking those boxes really fast. Um but but I I don't I guess I guess AWS does have like a a material advantage there still. Um
Alman said that while he didn't regret signing a deal with the defense department, he wished he hadn't announced the decision so quickly, telling staff that it looked opportunistic and not united with the field.
And I think that's I I I think that's a good point. Um I agree with him on that. His remarks echoed a memo he shared with staff and on X Monday in which he said the deal looked opportunistic and sloppy. A ground swell of AI researchers and the at the company and across Silicon Valley spent the past few days criticizing Altman and the company for what they saw as a capitulation to the Pentagon by essentially agreeing to a deal that allowed AI to be used in all lawful cases. Um, and we will see where where where the laws go. I' I've been digging a little bit more into like the rules around surveillance. There's there's actually like a ton of them and it's a very nuanced thing and so it feels like it's hard to b to to bake it down into a single line but a a single line is what hits on social media for better or for worse. Yeah. Um and so you can see a single line and be like that's amazing. Uh it's exactly what I what I think it means and that and that could either mean good or bad. So very very tricky. Uh Scott Bessent went on CNBC this morning and said, "No private company will ever dictate the terms of our national security anthropics attempts to push uh to push use clauses into their contracts with the United States government are unacceptable and their product will no longer be utilized by the US Treasury or any other government agency." So doubling down there. Uh there was also a story in in Semaphore this morning from Reed. He writes, "Anthropics investors don't have its back in its fight with the Pentagon."
This is very interesting. Anthropic may be standing its ground against the Pentagon, but the AI powerhouse is doing so with a noticeably quiet quarter. Its own high-profile backers. Despite the company's defiance, Silicon Valley's biggest players have remained silent. In a recent meeting between Amazon CEO Andy Jassy and Pete Hegseth, the issue of anthropic came up, according to two people who were briefed on the encounter. Amazon has invested billions in the startup, a crucial part of Amazon's custom chip strategy as the largest consumer of the company's Trainium AI chips. and Hegseth has been threatening to designate anthropic a supply chain risk which would make it impossible for many of the military suppliers to do business with the company. Jassy demurred declining to take anthropic's side, the people said.
Uh he wasn't alone. Despite anthropic CEO Dario's very vocal opposition to the Pentagon's demands and rival OpenAI Sam Alman's running commentary on the matter, a number of anthropic investors have remained silent. One investor told Semaphore that speaking up might further inflame things with the admin and that they were still holding out hope the issue could be resolved. Another said that Anthropic requested they say nothing. Anthropic and Amazon declined to comment. The Pentagon didn't immediately respond to a request for comment as well.
The most important thing here is that we get Pete Hegseth on Twitch. Amazon Andy Jassy is sitting down with Pete Hegseth. Hegseth likes Twitter. He likes X. Get them live streaming those press conferences. Put them on Twitch. That's the goal here.
Yeah, we we'll we'll we'll talk direct Ball about the back and forth.
Uh Noah Smith summed it up honestly in the Ben Thompson versus Dean Ball debate. We have Dean Ball on the show at noon. I think Ben is right. Noah agrees with Ben. There was just no way or in any nation state in there was no way America or any nation state was ever going to let private companies remain in total control of the most powerful weapon ever invented. Uh this is a big question about like where are we like uh um we like like are are we there yet? Like is this the most powerful weapon ever invented? that feels like this feels very much like uh like a training run for something that's coming in the future and it's an important conversation but we're not necessarily there yet. Um so we'll see. Dean Ball replied he said also Dean versus Ben is wild and an honor for me as a guy who's been reading Ben since the days of Android will not commodify iOS and use
Ben as a core inspiration for my own project today, even though I disagree with Ben here. You will hear much more from me on this soon on a certain podcast. I hope he's referring to TBPN. But the thing is, Ben is anti-regulation and does not own the consequences of state seizure of AI.
There was another post by Dean Ball. I want to get him to unpack a little bit more. He said, 'It is so clear that the important fissure in AI politics is not now, right now, is not liberal versus conservative, Democrat versus Republican, EAC versus EA, or safety versus anti-safety, but instead takes advanced AI seriously as a concept versus does not take advanced AI seriously.' And why this is so interesting to me is because it feels like both Dean and Ben are in the 'takes advanced AI seriously as a concept' camp. So we'll see if he agrees with my understanding of both of their positions, because it seems like they both are in the same camp of takes advanced AI seriously, but then they disagree about the level of pressure that the government should place on a private corporation. And so, very, very interested in digging into that.
Let me tell you about Figma. No matter where your idea starts—Figma Make, Figma Code, or Sketch—the Figma canvas is where ideas connect and products take shape. Build in the right direction with Figma. Let's move on to Will Manidis. He has also been chiming in on Dario's statements.
There's a strange—there's a set of strange little fictions that tech has chosen to believe as some sort of basis set for the perceived morality of their work. Chief amongst these is the idea that the legal system doesn't exist, and when it does, it is—it's a precise mechanistic process. Tech is a positive sum in a community, which means you pretend that the standard weapons that exist to the modern corporation—coercive litigation, IP blocking and trolling, standard white shoe law firm stuff,
Really taking shots at white shoe lawyers.
is negative sum or something bad we don't do here. This leads to a weirdly formalist view of the legal process that we must avoid it at all cost. But if we must engage in it, then it's a fair and repeatable process that must necessarily yield, if not the truth, then at least the same outcome repeatedly and provably. It's this instinct that gives you Anthropic's odd contract formalism. This is how you end up thinking your best move is to show your cards at the forefront and thinking you've won the hand. There's incredible potential for outperformance available to you for betraying any of these silly norms that tech uses to feel good about itself. The modern corporation must avail itself of all possible edge, and holy wars require holy weapons. So get busy.
The—the—the idea that sticks with me here is that if we engage in it, then it's fair and a repeatable process. I talked to someone who was—who was—who was doing something with the legal system, and he said that like on this particular issue, it was a true like left-right issue. Like a liberal judge would always side on one side of the case, and a conservative judge would always side on the other side of the case. So no matter how well he argued, it came down to did his case get a liberal judge or a conservative judge. And I was like, 'Oh, well, like that seems simple. Like, like get a judge that's aligned with your company and your position.' And he was like, 'Yeah, bad news. The way they decide which judge I get is they literally spin a physical wheel.' And I was like, 'And so there's nothing you can do?' He's like, 'Yeah, no, like it—it is secure. It's like a real process.' And like that's by design. Like some things are better left in the democratic process to inject a little random randomness so that you can't put your finger on the wheel and pick the right judge and get exactly what you want. Like that's not possible. But it—it is—it is potentially like the least bad system, but still very stochastic and still very random. And so you wind up in this weird situation where things are not necessarily repeatable. Repeatable is the interesting thing that you—that—that you can—you can go into an engagement in a—in a legal context and wind up with a different scenario just somewhat randomly, and—and that is almost by design.
Let me tell you about Turbo Puffer—serverless vector and full-text search built from first principles on object storage. Fast, 10x cheaper, and extremely scalable. Dario was at the mic at the Morgan Stanley Tech, Media and Technology Conference today and was dropping bombs. On defense, he said, 'We believe in defending America. Anthropic has been working with the national security community for two years. We are the most lean forward.' On AI acceleration, he says, 'We do not see hitting a wall. This year we'll have a radical acceleration that surprises everyone. Exponentials catch people off guard. We are at the precipice of something incredible. We need to manage it the right way.' I'm very interested to hear like this idea that he thinks we're at the end of the exponential. I feel like most technological processes—Carpathy was saying this, maybe he was saying like—like AI is a continuation of—you know—previous technological revolutions, right?
That was why he said we're not going to see like massive growth in GDP. I think that was his reasoning. Yeah, it's just kind of the same process.
You saw computers, then you saw internet, then AI. It's the same curve.
Yeah, it's very interesting. I mean, there—there are people that like he's saying there, like we don't see hitting a wall this year. But it begs the question, like do you think that there's going to be a plateau? Because at a certain point, like you—you—you plot out the AI revenues and you get to something that's larger than GDP, and the GDP has to move in order to grow. Like—like what happens at the end state? What does the end state look like? I feel like we got to talk about this. Me and—me and Tyler were—were talking about the—just the general like AI safety discourse on both sides always ends with people just saying like we got to talk about this.
No one has any answer, you know. We got to figure this out, guys. We all got—we got to talk.
We got to talk to each other.
Yeah. Yeah. No one's like—no one's like, 'I have a clear plan. I think it'll work. You might disagree with me, but this is my plan and this is what we're going to do, and here are the steps. And like I think this gets us the good—the good outcome.' Everyone's—everyone's just like, 'I don't know.' It's very, very weird. Just from a comms perspective, it's very weird.
I mean, I feel like that's how decisions get made in companies. Like if—if a decision is really obvious, you can handle it over text. If you—if it's not, it's like let's—let's discuss live.
Yeah. But I—I mean, you talk to a lot of companies about—about their plans, and—and I mean the good companies can think in decades and—and—and think like, okay, we're going to do this and then we're going to make a—I mean the original Tesla plan: we're going to make a sports car, and then we're going to make a sedan, and then we're going to make a cheaper car. And like that's the plan. And—and there was no like, 'Oh, we got to talk about this.' It was like—it's like, yeah, we're going to make a bunch of cars and then we're going to make more factories to make more cars, and like this is the plan. We're like—we're just not—we're just not really seeing that from—from—from lab leaders at the moment.
But still plenty of optimism. The—the—the square packing—circle packing discourse has hit the timeline. I absolutely love this. First, let me tell you about Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents. So John Bidwell in 1997 found a more optimal way to fit 17 squares into a larger square. The smallest square that can contain all 17 looks like an absolutely deranged solution. Everyone would expect that it's just a nice even grid. You take—you add one. But if you pack them in this way, you get 17 squares of absolute chaos. On the next—on the next slide. And—and someone made a waffle maker. Someone made a waffle maker that optimally packs 17 squares into—into the squares of the waffle.
You will get mocked at breakfast.
You will get mocked by anyone except a mathematician, I suppose. Kevin says, 'For years, society was limited to only 16 syrup squares per waffle, but with recent combinatorial optimization breakthroughs, our research department has achieved previously unheard-of densities of waffle syrup.' It's great, great news. Very, very funny that this came through. And then someone in the previous post was posting the—the universe is not real, dude. This is ridiculous. I'm going to fight God because it looks very crazy to pack an arbitrary number of squares within—within another square. But that's—
Someone should do this for cubicles in an office. Can you imagine if you're in the one cubicle that's slightly off-kilter and everyone's like, 'Look, dude, like you get—you have the same amount of space as everybody else,' and you're like, 'Yeah, but I'm at a 37-degree angle to the wall and it's driving me insane.' That would be very, very odd.
Did you see researchers at the University of Southern Denmark developed a limbless soft robot that crawls using air-filled chambers and Kirigami-inspired flaps? You're going to love this.
I am the Indiana Jones of tech podcasting. I hate snakes. I hate—I absolutely hate snakes.
Very, very weird. 'It's over, boys,' says Autism Capital. Very weird, but you know, exciting that people are working on robotics, and I'm sure there's some—some odd application.
Yeah, I'm sure there's no like ultra-nefarious use cases for something like this.
We're going subterranean. And already working on it. What do you got, Tyler?
These look like—these would be good for like—
Medical. You make it really small. Yeah, maybe that's also like—
Yeah. Would you volunteer to as a trial subject? I think you should go first if you think—if you think they're so good.
I think you should go first.
Someone launched siggies.app, the most comprehensive archive of Chinese cigarettes to exist on the Western internet. Discover new packs and their history, track favorites and packs you already tried, share your collection. The Chinese cigarette industry is absolutely fascinating. It—I think it makes like hundreds of billions of dollars in profit for the government. The—the—the main cigarette company is nationalized.
And so they make—they make an absolute ton of money. They fund their military with it. It's fascinating.
Yeah. Really lovely website. We—we really got to go into the history of nationalization because the nationalization and privatization is interesting all over the point—all over the place. I looked into nuclear weapons and how those were nationalized. Fascinating sort of like hybrid public-private partnerships with a lot of consultants in the private sector, and then obviously the—the defense primes making the—the missiles, the ICBMs, but not the warheads. But in the tobacco industry, there are some fascinating examples of countries that start cigarette manufacturing businesses, develop large sort of deeper supply chain operations, and then privatize those. And then that when they sell that asset to a Western tobacco firm, there's sort of like a cash windfall for the government that can be very beneficial. They can go build roads with the money, or I don't know, cut people's taxes. But—but there's the—the other side of, okay, a company is—is getting large and then it gets nationalized. But plenty. Are there any other industries that stuck out to you, Tyler, as you were—
I mean, there's like—I'm still like kind of looking. There's so many like weird like ways to—to view nationalization though, because like so at least in the US, like historically it's been mostly basically during wartime or if like a really important company is—is facing like huge crisis, right? So like 2008, like I think GM at one point government—
Yeah. The Treasury owned like 60% of—of GM. Yeah. So—so there's all like interesting—
Fannie Mae and Freddie Mac.
But even like if you think about tobacco in the US, like almost like having super strong taxes is like almost like kind of a form of nationalization, right? Where you're basically taking like a ton of the profits away.
Yeah. Yeah. Yeah. I mean, that—that was always the pushback of—of, you know, if—if there's runaway profits by the AI companies, just, you know, instead of nationalizing them in that—in that scenario, you can just tax them and then redistribute the—
profit. There's like this real spectrum, right? Because like if you think of like Intel—
So the government now owns like—it's like 10% of Intel, right?
So like, but what does that actually mean for like running Intel? Like do they have a say? Like right now—
They're not supposed to. Maybe that changes. It's like this—this interesting spectrum.
Anthropic is on track to generate annual revenue of almost 20 billion, a projection based on current performance, more than doubling its run rate. This is from Bloomberg. And Zero Hedge says, 'Unprecedented damage control, leak after leak. Just how bad is it?' And Matt Slotnik says, 'The leak is bad, sir. Really bad. Now everyone knows we're the fastest-growing software company of all time. We must put a stop to this.' I think what Zero Hedge was getting at was that the reason you leak, you know, strong economic performances is to, you know, allay concerns around, you know, pressure from the government, supply chain risk, because the revenue would not be growing if like companies were pulling out because they're so worried about the supply chain risk that they got to—they got to move on so fast. But still funny.
Clearly adoption is outpacing any kind of pullback from—from, you know, the Department of Treasury and—various other—various other kind of organizations in the government.
Yeah. Ara Karassian over at Ramp has some extra data. He said most people in tech know Anthropic commands the majority of API spend by US businesses. As of January, Anthropic took 50% of spend on enterprise AI subscriptions too. OpenAI still leads on business count, but the biggest spenders go to Anthropic, all from Ramp data. And Ben Thompson had some extra context here about the data, but I was—I was actually DMing with Ara last night, and it is fascinating what's happening. The market is so—so big that everything is growing, and you see these charts and you're like, 'It's over for this company, it's over for Cursor,' and then Cursor puts up two billion, and there's just more and more growth that's happening because of the diffusion problem and—and—and the fact that companies onboard and just more and more companies are onboarding to AI tools every day.
Ben Thompson said there are some caveats to the Ramp data. It only includes OpenAI, Anthropic, not Azure Copilot or Google Cloud Gemini. Ramp is also probably tilted more towards startups and smaller companies than larger enterprises. That noted, it certainly matches the vibes. Anthropic established that coding foothold and has been on an absolute tear over the last six months in particular in terms of not just its model capabilities, but also Claude Code and the overall sense that it is in the lead. And enterprise CTOs are a smaller and more easily reachable set of customers than consumers at large. Meanwhile, an AI subscription business model has always been more compelling for the enterprise than for the consumer space for the same reason that productivity apps always end up as enterprise products, not consumer ones. There's no consumer CRM, for example. Enterprises actually value productivity and are willing to pay for it. Consumers don't and won't, which is why the business model that makes sense for the consumer space, at least at scale, is advertising. So he's once again ad-pilled. What else is in the—in the news?
Eddie, endowment Eddie put this into context. Even in Q4 of 2021, public SaaS quarterly net new ARR peaked at 2.6 billion.
Anthropic has added nearly 11 billion in early 2026. Yeah, this is—
Of course, is like run rate, not actual—
ARR necessarily, but still absolutely shocking.
Tyler said for those kind of tracking at home, Anthropic just added a Palantir in revenue in the month of February alone.
Yeah, pretty unprecedented.
Howard Lindzon said, 'Anthropic pulled off the rare quadruple Lindy: great product, people pay for it, people love it and promote it even though they pay, and the people like Dario's handling of government. Next up, the people tear Dario down.' There is this interesting like tall poppy syndrome in Silicon Valley a little bit where like whoever's at the top gets like, you know, triple-glazed and then it needs to be torn down a little bit. So we—we—we'll see where all this goes. But—
Putting—putting AI growth in the context of the shoe market. Yes. Darren Rovell, former guest of the show, shares that New Balance is up almost 20% year-over-year at 19%, Adidas 13, Lululemon nine.
And then every other brand here in the negative: Puma at -8%, Under Armour -9.4, Nike -10%, and Jordan Brand—
Down 16%. Really insane to see such a huge line. If he likes the sneaker era, the era—the Jordan era is over for now.
Is this more of like changing in taste that New Balance and Adidas are growing and Nike is not? Because if—if they were all—if all of these like major household name brands were trading down, I would say like, 'Oh, this is like the Temu effect. This is the white label effect. This is the—the Aldi, the D—like just more competition from newer brands, challenger brands.' But it seems like there's more of just like a reallocation from Nike to New Balance and Adidas. Like Adidas might just be on a—on a brand tear doing some good advertising. But New Balance has been killing it with their positioning. Interesting to see total revenue as well. Nike is huge, 46 billion. Adidas is down at 26. Lululemon's at 10. Lululemon has done—has done really, really well. New Balance—they're bigger than New Balance.
Anyway, 'my portfolio every day at the opening bell looks like D-Day,' says Human. This is an absolute disaster. It really is so crazy. I was trying to explain yesterday to a friend like how the market—I was like, 'You're probably getting destroyed, right?' And he was like, 'Yeah, it's terrible.' And I was like, 'Well—'
Is this your retail investor friend?
Yes. It might shock you to learn that the market overall is like half of the year. It's—it's an absolute tail. Korean stocks are crashing. The Kospi index, the Korean stock exchange, is down 10%. 'Korean stocks are—are crashing,' says Joe Weisenthal. The South Korean stock market fell over 8% and triggered a circuit breaker. Jim Cramer said, 'Beware of South Korean spill-over to our markets. WDC, STX, SNDK, MU, all still vulnerable.' So there's lots of companies that are at risk. We will dig into the South Korean stock market collapse soon, but we have our next guest in the Riverside waiting room. First, let me tell you about Sentry. Sentry shows developers what's broken and helps them fix it fast. That's why 150,000 organizations use it to keep their apps working. And without further ado, let's bring in Dean Ball from the Riverside waiting room. Dean, how are you doing?
We're fantastic. I've been enjoying your coverage. We had Ben Thompson on the show Monday. It's very funny that this debate is phrased as the Ball versus Thompson debate because it feels like you agree on way more than you disagree on. But I would love for you to sort of set the table for us and just kind of give us your run of like what actually happened because there are a lot of things that have been said but not enacted. There's been bunches back and forth. So what's the—what's the—what are the rules of the road? What are the—what are the—the table of facts that you think like are the most reliable here?
First of all, there are very few people I respect more than Ben Thompson and who have inspired me more than Ben Thompson. So, yeah, I think we agree on the vast majority of tech policy issues and other tech stuff. Yeah, I mean, I think basically what's going on here is that, you know, there's a dispute about a contract, right? And that's fine. And grown-ups can have disputes about—contracts, and they cannot come to terms of business and then they don't sign the contract or they cancel the contract, and that's fine. The problem here is the punishment. The problem here is that the government is saying if you don't do business with us, we'll destroy your company or we'll threaten to, and that just can't be the way that America works.
Yeah. And I don't really like—you know, there are people who say, including I think, you know, this was an argument that Ben made, that well, what should Anthropic have expected? You know, what should they have expected? That the way they talk about AI being so powerful, how could they not expect the government to—to bully them and quasi-nationalize them in these ways?
And my view is like, well, no, what we should do is have like modest technocratic regulation so that we can avoid these sort of extreme outcomes of either like, yeah, total decentralization, total anarchy, or like bull—bullshitism, right? Like we can avoid that. It's called capitalism. It's called like the system that we have now. And I've spent the last two years being confused as to why we don't try to do that.
Okay. Okay. So I guess so far Anthropic has not been labeled a supply chain risk.
Yeah. Officially, is your belief that the Department of War is just really busy right now for obvious reasons and they haven't gotten around to it yet, or do you think that it's coming and—and that's why, you know, like where—where do we actually go from here?
I truly don't know. I operate under the assumption—so War Secretary Pete Hegseth said last Friday that they're going to do it, and they're going to do an extremely expansive version of it. So I operate under the assumption that he's not lying about that. I basically take him at his word. It'll happen eventually, but I don't know.
Do you think he—I mean, I—I—I know that there was some debate over the shape of supply chain risk, whether that means you have a partnership with Amazon, Amazon works with the government, you can't work with Amazon at all, or Amazon can't use your software specifically, your model specifically on government contracts. Do you have any idea of where the clarity might be found there?
Well, so in terms of what the supply chain risk designation actually allows, my reading of the law is that it would be only in the fulfillment of the—of the government contract and not more broadly. But there's a lot of things. First of all, you could just write it more broadly and try to fight it in court and see what happens, right? And who knows, right? You could also jawbone, right? You could—you could have the president call Amazon and say, 'Hey, we don't think you should be working with these guys.' Right. Very hard to sue about that because that's not a policy action. And there's also a very, very large range of other things that the federal government could do if it wanted to, if it wanted to really harass, which, you know, the irony of this is that we all know this because of what happened to Elon's companies during the Biden administration. I objected to that then and I object to this now.
Yeah, unpack a little bit of that history. Take us through it.
Yeah, so, you know, basically Elon being kind of a political enemy of the—of the Democrats, the Biden administration brought on a very large number of—of investigations, regulatory actions, FCC, DOJ, a lot of different other things. I think there—there's a chart somewhere and it's like—it's like 12 to 20 things, you know, it's a huge number of different things that they brought, and each one of those is going to entail millions of dollars in compliance costs and legal costs for these companies. So it's non-trivial amounts of harassment that's going on, and I again, I think we shouldn't do that kind of thing.
Sure. Is—is this broadly under the umbrella of like lawfare?
Okay. Walk me through this—this idea of the fissure in AI politics right now is not liberal or conservative, Republican versus Democrat, EAC versus EA, safety versus anti-safety, but instead takes AI—takes advanced AI seriously as a concept versus does not take advanced AI seriously as a concept. It feels like you and Ben Thompson both are taking AI seriously as a concept. Does that map with your reading on his arguments, your arguments? It feels like there's just a different downstream understanding of how this plays out in a world where advanced AI is a serious concept to grapple with, different than making pencils.
Yeah. Right. Like, I mean, I think the—the other way to put this would be like, do you think that Dario Amodei and Anthropic are directionally—and the other frontier labs, it's worth noting, are directionally right about where this is going, or do you think that—you think they're not? And—and also not just do you think they are, but do you really take that seriously? Do you own the consequences of believing that these guys could really be correct? And I think a—a consistent theme of my writing over the last year especially has been kind of being confused that there's people who just do not take—they might say that like, oh, AI is progressing very quickly, but they don't actually—like their downstream assumptions and like conclusions about what we should do don't seem to—to—to sort of reflect that. It's very hard—it's a very hard thing to do, but yeah, that would be like my broad take.
So this is—this is where I—I—I start to get confused because if I take advanced AI seriously—I do—then it feels like nuclear weapon-level technology in a few years. Maybe it's five, maybe it's two. But that leads me more into the Ben Thompson camp where I say, well, of course there's going to be a—a dramatic showdown between the US government and the private company that has invented something akin to nuclear weapons. And so walk me through how you are thinking of, you know, if we're developing new—new, you know, powerful technology, what—how should the government actually interact with that aside from like the lawfare in the short term? Like what's the long-term solution?
Yeah, I mean, the long-term solution, I think, is probably going to be a hybrid system where there just is regulation of these frontier labs, and there is government oversight of them, but it's just like—it's like, you know, JP Morgan is a private entity that lobbies against government policy all the time, right? They're a totally independent actor of the government and yet they're also deeply intertwined with the government. I'm not saying that we should regulate AI companies like banks—I hope not—but I'm saying that like, you know, technocratic regulation exists in, you know, capitalist republics and it's not communism, and we didn't like nationalize the banks, right? So it would be—my proposal would be to sort of take step-by-step approaches to solving actual problems, not just doing regulation for its own sake, but really trying to shape this so that it isn't some sort of violent—or—not literally violent—but some sort of conflict, and instead is, you know, a gradual development of like, yeah, this is how the government and the public ultimately achieve oversight of these labs, because I would agree that if we are going to this place of like near-term superintelligence or something, that isn't going to just exist in an unregulated fashion. That's kind of been my point for, you know, for a year and a half now.
Yeah. So what's the gap between regulation and nationalization? Because at a certain point, if we're in the nuke analogy, it feels like the government takes that and then they get to decide where they—they point those nukes. And—and as I was reflecting on the—the creation of the—the atomic bomb, I was sort of left like optimistic about the American project and optimistic about the way that played out. I don't know if you have a different side, but I feel like the government did get control of nuclear weapons. There hasn't been a nuclear war. And the whole concept of like the person that you're voting for will have their—they'll have the nuclear football, they'll have their finger on the button. Do you trust them? That is an animating force in our electoral process. Whether you like the current administration or the previous administration, whether you trust them or not, but like Americans vote on that question every four years. And I think that—that's a good system. I—I—I like that.
Mhm. 100%. I think that the nuclear bomb analogy starts to break down. Okay. But one thing I'd say about—about nuclear weapons is that it's true that nuclear weapons, you know, we—we got that and that went I think basically pretty okay as far as these things go, but you know what we didn't really get is nuclear energy to the nearly—to the extent that we could have.
Completely agree with you on this. Yes. Continue.
Right. And so that created this single regulatory sort of point of failure. Single points of failure are often problematic in complex system engineering. So, you know, what do you—what do you do? Well, you know, maybe—maybe we shouldn't have done it. You know, maybe we should—maybe the government effect on the sort of—the equivalent of the consumer side, the economic benefits, we didn't really get those as much as we could have. So that would be one thing. But the other thing is that even nuclear energy, you know, it's not directly useful to you. It is not—nuclear energy is not like—you are not personally going to express your liberty, exercise your liberty through nuclear energy or nuclear bombs, but you will with artificial superintelligence, and you certainly will with today's coding agents, right? And so the question is like, should the government be able to own that? And my view is that without really substantial guardrails, that is just almost certain to devolve into a profound act of tyranny. And so I think that having the companies separate and apart from the government but still overseen by the public through regulation, that is kind of like my view as to what is ideal. But the problem is that saying that view—a, it's really hard to get that right. Regulation is tough to get right. You want to get it exactly right. And it's possible that all my ideas are terrible and they're, you know, they're all going to go to shit, right? That's totally possible. But what you want to do is you want to—yeah, you—you want to—you want to at least try—try to get that right. The problem is that—that has been—that view has been characterized very often as sort of supporting regulatory capture.
So it's like, you know, what did Anthropic expect the way they talked about these models? Well, maybe they expected for like California's SB 53 to pass and for there to be like modest technocratic light-touch regulation that only affects the frontier labs. And like, I don't think that's lobbying for regulatory capture. I think that's actually just the prudent thing to do. Even if I don't agree with every regulatory impulse that Anthropic has, I think that in broad strokes that's the play.
Yeah. How do you—Oh, sorry, Jordi, you go.
Yeah. I wanted to ask like how—how much do you think information asymmetry played into the dynamic of the negotiation and the results? I mean, you have the Department of War who presumably everyone on that side has some idea that we're about to enter a conflict, and Friday comes around. The Department of War knows they're hours away from a strike. Dario, according to Emil, like isn't kind of engaging with him, and out of that maybe the admin says, hey, this is not a partner that we can rely on, and we don't want other groups in the government to be relying on it as well, and it just kind of blows up.
Yeah. I mean, so I think some of this is about personality—some of this seems like it's about personality clashes. Some of it's about politics undoubtedly. I think some of it is also about principle though, right? Like—and I think the DoD is not crazy at all. Like my—my problem here has never been with the DoD's policy principles here. They're saying Anthropic can't put usage restrictions that look like public policy because we set public—Dario Amodei doesn't decide when autonomous lethal weapons are ready for prime time. We do. And I think they're 100% right about that. But the solution to that problem is you—you don't try to destroy the business of Anthropic. You tell Dario no thanks to the business, and you move on and you find somebody else who will—who will do business.
Which is what Dario said in the CBS interview. He said there's other providers, we can agree not to work with each other.
Right. That's fine. That's fine. I mean, frankly, the tokens are probably higher margin for Anthropic, you know, sold to private entities, not to the government because they're sub—Anthropic is subsidizing them for the government. So it's a bad business as we're seeing.
What's a—what—what's a—what's a good outcome here?
I think we cancel the contract. And—and look, the other thing is like the DoD says, well, we have all these other contractors, there's Palantirs of the world, and Palantir might rely on Anthropic, and so we can't have our—our contractors can't be relying upon Anthropic because if that's true then we're still relying upon Anthropic. So it's fine for them to say no one who has usage restrictions like this can do business with any of our contractors. I think that's fine for—in the fulfillment of DoD contracts. Like I think that is perfectly fine.
Isn't that basically supply chain risk by any other name?
Well, no, because supply chain—
Supply chain risk would be specific—would be done at the company level as opposed to like at the—
At the contract level. Yeah. Yeah. Yeah. Yeah. Yeah. I—I—I agree with you. Also, can you—can you explain a little bit more about the—the logic of supply chain risk? Because I was shocked when like DJI is not on there, DeepSeek doesn't seem to be on there, Qwen doesn't seem to be on there. Like, this feels like very bad if I'm just trying to think of like my short list for things that I might want to supply chain risk. Like, yeah, it's good Huawei is on there, but I got 12 other—
Yeah. Yeah. Unitree, I don't think, is on there. Yeah. Like—like when should the government apply this? Does it matter if it's international? Is there—is the national versus international designation important?
Well, one thing I'll say is we don't know everyone on the supply chain risk list because sometimes those designations can occur in classified contexts. But—but it is the case I think that—that the supply chain risk designation, if you go and look at the statutory history, it's really intended for foreign adversary companies. This—this is about—this is about—it's really about China, frankly. It's really about China. And so it is kind of, you know, wild to be in a—whatever you think of the chip export policy of this administration of, you know, wanting to sell more Nvidia GPUs to Chinese companies, it is kind of wild to have that policy and then also to say that we're going to treat American companies that we don't like like Chinese companies, like—like enemies of the state. I don't think we should be doing that.
Can you talk to me about what your 2025 was like, what your goals were, what of those goals have been achieved? Like give me sort of a scorecard on like AI policy in DC.
Well, you know what I—what I wanted to do is—I mean, first of all, you know, obviously getting the action plan done, that was a, you know, 18-, 20-hour-a-day kind of thing for four months in my life. So that was—that was an intense sprint, and getting it done is obviously a career highlight for me. I didn't get it done unilaterally. Of course, a lot of people worked on it, but, you know, that was a big part of it. Sort of course-correcting from what I thought were some of the excesses of the—of the Biden policy, a lot of which was, by the way, this national security-inflected control over the labs. So it does disappoint me to see us sort of regressing back into Biden-era mentality, which is what this is. So there's—that is a little bit concerning to me, but yeah, like that. And then more broadly, you know, I don't—I just—I want to AGI-pill people, you know. I want people to like feel the—feel the—like the—the importance of what is going on, and—and hopefully, you know, try to think more strategically about this issue because the thing is is that it can't just be one or two people doing this. It has to be a community of people.
So—so walk me through what AGI-pilling means in March of 2026. More Mac minis than people on Capitol Hill. Yeah, I—I—I mean like—I mean like there's terms like software-only singularity, that robotics will be delayed, that there might be—Dario talks about like the end of the exponential. He doesn't seem necessarily superintelligence-pilled. He seems like we will get someone who's 150 IQ and you will be able to run it in a data center. Like the—what does he say—a million geniuses in a data center. It's not—it's not one god—country of geniuses in a data center. It's not one god that's like—so like time traveling and teleportation—like it's not entirely all this crazy sci-fi. It's a little more narrow. So in terms of timelines, expectations, like what are you telling to elected officials these days?
Well, I think the very challenging part is that there's the trajectory of the technology and then there's the effect of the technology in the near term, right? So the technology that we have right now is already legitimately science fiction in my mind. But we don't live in a science—we don't—the world is not obviously more science-fiction-esque than it was three years ago. We don't have flying cars. We don't live on the moon, right? We're not doing it, and—and we don't have Dyson spheres assembled around the sun or whatever. And the reason for that is that it takes a very long time to transform institutions and to—to transform the way that organizations are structured. And so again, like that's the important thing is that to me it's—it's—it's about the actual winner of the AI competition is going to be the—the country, the civilization that is like most imaginative at like—we went from artisanal types of, you know, manufacturing to factories to eventually assembly lines. Right. We—we—we're going to do that, but for the next generation of things and inventing that. What is—what is all that? What does it look like? And, you know, that's the—that's the—that's going to be where the real economic benefits come and the real things that change the world come from. But those take time and they're really conceptually difficult.
How do you think about competition with China with the backdrop of nationalization? I heard this—I heard this quote that was like if you nationalized SpaceX a decade ago, like you just get NASA. Like you don't get SpaceX if it's owned by the government. It needs to be completely independent. And so, with that backdrop, like what does a healthy American AI lab ecosystem look like that can actually go and win on a geopolitical scale?
Well, I think first of all, yes, it's—it's got to be—it's got to be private companies because they can move faster and be more innovative. Also, very importantly, it's about trust, right? Like selling—if we're going to—if we're going to sell these services internationally, we need other governments and foreign companies to trust our companies, to trust that they're not just ultimately linked to the US military, right? Because that's the problem is that doing business with China, you sort of know, well, at the end of the day, every company in China is ultimately a military asset. And that diminishes trust in Chinese companies. We're—that's something really profound that we're eroding with this action right here. And that worries me quite a bit.
So I've been thinking about that in the context of like sovereign AI throughout Europe, and I was just like, look, France didn't need its own Google. They could just use Google. Like it—it was fine. But when you actually take AI seriously, then you start to wonder about, okay, well, does their strategy, as behind as they might be, does it actually make sense?
Yeah. One of—one of the things I hope we invent is—open-source models I think are—are useful and good, but I actually think what's—what's more important right now is like open-source infrastructure such that more people can sort of like really put their—put their own imprint on models, and more people can feel that sovereign experience without having to like build, you know, multi-gigawatt data centers and whatnot. I hope we can do that. That's a really—
Like a national compute reserve or something like that that—like that—like academic institutions could access.
It could be like that, but also even more than just the compute itself. Also like all the infrastructure that connects the compute and all the like—I—I think there's something really interesting there, and I'm excited to see more people do things like that.
Yeah. What do you think about the concept of like an FDA for new models? You're releasing a new model, like it has to pass a battery of tests or benchmarks that are maintained by a—like a federal institution in the same way that when you—when you release a new cancer drug, you have to test it on mice first, something like that.
Yeah. I think the problem with that is that it's—that the—there's too much to test for because the models can do too much. So we don't know what to test for. So we have very clear endpoints when we're testing a drug in a person, right? We have very clear things that we want to test for. It's much harder to do that in the context of a generalist AI model. So that's one of the things that worries me about licensing. In addition to—licensing can slow things down, and it can—regulatory capture is very real. It's—it's a very real thing. I—I've always thought of AI as being a little bit more like—like financial services where the regulation is not actually targeted at the products per se. The regulation is more generally targeted at the entities themselves. So we regulate banks. We don't—we do regulate loans, but we do that by way of regulating banks. We don't like have a—there's not a government agency that like approves every single loan that a bank makes. We—what we do is we look at the entity of the bank and we say this—this is a sound institution. This is a—
soundly governed institution. That's the kind of thing that I think will be... I don't know, again, I don't think we should regulate AI like banks per se, but I think that is a useful structural intuition for thinking about AI regulatory policy.
Yeah. Jordi, please. How much attention is DC paying to the news out of Square? The pushback from Block, the pushback from Silicon Valley was, hey, you're blaming this on AI, but it doesn't feel like maybe that's a small part of the reason to do this riff, but certainly not the entire story. But then a bunch of groups are going to use this as, you know, example number one as to changes in the labor force due to AI.
Well, it's a bad combination because it's all these companies, all these tech companies that overhired in the sort of immediate post-COVID. And then now they're also the most exposed to AI, right? They're the ones that are adopting AI the most aggressively. They do the things that AI is best at. There's lots of software engineering. And so like it's both things happening at once, I would say. And that creates an exaggerated sort of funhouse-like effect of the actual impact of AI on the labor market. I'm very skeptical that like we're going to see things like that at like wide, you know, firm level anytime soon, you know, in the broader economy. I think we could see it more in the software engineering discipline because there are a lot of big software engineering enterprises that hire too many people, right? So like, you know, I think that's a very live possibility, but DC is DC. People are going to interpret the news for whatever is convenient for them, you know. And when you play the long game, you have to try to just build credibility over the long term with people. But yeah, I mean, everyone's going to say this is evidence for whatever I'm on about.
Yeah. Hypothetically, how do you think something would play out if like there was a battle on the moon and NASA says, hey, we want to buy a bunch of rockets from Blue Origin or SpaceX and we want to go to the moon to fight this war on the moon, and the private company CEOs say like, no, we're not interested in that. We want to just launch more satellites. Like it feels like another analogy for like the nationalization conversation. And maybe it's just a weird time because I feel like in previous eras, a lot of the, you know, you go back to like the dollar-a-year men, a lot of American industrialists sort of like stepped up to the charge and were like fans of the government and now there's like much more, it feels like we're more divided than ever. But like what is the correct way to marshal a cornered resource as a government?
Well, I would think in practice that the companies, the SpaceXs and Blue Origins of the world would actually eagerly help the government there. But let's take your hypothetical and say that they didn't. There's an authority called the Defense Production Act, Title One, which is a... there's something called priorities authority and the government can quite literally just put itself at the front of the line for any launch of any rocket and that's what I would suggest they do in that situation. Rather than commandeer the rockets themselves. That's kind of what I would suggest.
That makes sense. So Defense Production Act, but it requires extraordinary circumstances.
The president can do it basically whenever he wants. The president can make a unilateral finding. The other thing I'll say that's funny about all this is that like, you know, one of the things I was enthusiastic about putting in the action plan was a little notice provision of the action plan is basically saying in the event of a national emergency, DoD needs to know with the hyperscalers how it's going to commandeer the data centers for like if it needs like tons of compute in some sort of national emergency, that's in the action plan. So like I'm not here saying like total libertarian, you know, whatever, like no government involvement. That's not what I'm saying. I'm just saying that like Anthropic has a right to exist. If they don't do business with the government, we shouldn't kill them. That's all I'm saying.
Yeah. Yeah. I think the Wall Street Journal called it like a fight about vibes. And I think most people are not at the point where they say, yes, AI is actually nukes. And yes, like one company should like, you know, like they should be destroyed if they don't give it all up. Like that's a very middle-of-the-road position that I think a lot of people agree on.
