Meta Tokenmaxxing, Intel Joins Terafab, Frontier AI vs. China | Diet TBPN

Well, there is a whole bunch of news to run through. The first story is that Meta employees are apparently token maxing and competing on an internal leaderboard called Clawdonomics for status as a token legend. This is from The Information. Over a recent 30-day period, total usage on the dashboard topped 60 trillion tokens. And this sparked a huge debate over how much is Meta actually spending with Anthropic. Of course, the other big news is that Anthropic just past $30 billion in run rate revenue with one of the, probably the steepest revenue growth chart in human history. Absolutely legendary.

Yeah, this chasing status as a token legend reminds me of, maybe it was a year ago at this point, you were saying, will tokens ever become like eyeballs the way eyeballs were during the—

Yeah.

—during the dot-com era, right? Just optimize for eyeballs. Obviously, not every eyeball visit to a website is created equally, but people were optimizing for eyeballs. And now, the reaction to this, I think, has been generally, at least online, been, I guess, reassuring. A lot of people are saying— Gary Basin says, you, why? Marty says, Goodhart's Law, when a measure becomes a target, it ceases to be a good measure. And so who knows what's actually going on internally? But we do know Zuck is pushing the entire company to be as AI native as possible. And this guy loves spending money too, right?

I have a crazy bull case here that I will run through. Let's get through some of the story first. We got to pull up this comic from xkcd. In the comments here. When a metric becomes a target, it ceases to be a good metric. It's right under the leading post. There we go. And says— and the other counterparty says, sounds bad. Let's offer a bonus to anyone who identifies a metric that has become a target. It is good. I don't think that's going on here.

Lighter was texting a friend at Meta and sent the post we just discussed on token maxing. Yes. And said, true. And the person said, yes, it's pretty sad. But I mean, imagine, so Meta has been, there's been rumors of Meta layoffs for a while now. Sure. Unclear how many, if any, if any have happened. But if you're sitting there, the company, Zuck is saying like, we need to get AI native, Boz is saying we need to get AI native, and then suddenly there's a token leaderboard. Yeah. You do not wanna be at the bottom of the list. I will say that, right? Yeah. You know, you don't wanna be the, you don't wanna be the guy who's having to explain like, no, well, I'm actually getting the most out of each incremental token and the other guy's just like set up an agent that just counts one, one to a million over and over and over.

Yeah. Uh, yeah. I mean, you have to measure the actual output, the impact on the business. Uh, I mean, fortunately Meta has been a huge beneficiary and a huge winner of AI. Uh, the ads are getting better targeting. They're seeing, they're delivering more ads and, and the, and the quarterly earnings have been strong. The headline number here that, that, uh, sort of took everyone by surprise is that Meta staff used 60.2 trillion tokens over 30 days, which would pencil out to about a third of Anthropic's ARR was the number that was thrown out. But both of these claims are pretty questionable. And so Tyler did some back-of-the-envelope math to show that the one-third revenue estimate is way, way too high. And I don't know, do you want to take us through some of the reasoning there?

And then we can talk about the knock-on effects of all this.

Yeah. Okay. So $60.2 trillion tokens is the number. Like, we can just assume that's true. So basically, I'm going to assume all the employees are basically just using Opus 4.6. Yeah. So then there's basically 3 numbers you need to look for in the API cost. So there's input. There's cached input. And then there's output. Sure. So for Opus 4.6, it's $5 per million tokens on input. It's $0.50 per million tokens on input cached. And then it's $25 on output. Yeah.

So if you multiply that 60.2 trillion tokens at the highest possible rate, $25 per million tokens, then you do get like a billion dollars in a month, which is crazy. That's not what's happening.

It's a crazy number, but like you have to think about it like, you know, if you're using like Claude Code or any of these coding agents, you know, the vast, vast majority of the tokens used is input. Yeah. Because like, so imagine you're working on some, you know, coding file, right? Yeah. There's like 1,000 lines of code in the file. Maybe the model's only changing like 10 at most, right? So that's a very small percentage. So the output tokens are going to be a very small percentage of the total tokens going in, right? OpenRouter publishes like a lot of this data. So you can kind of use those ratios to figure out what is actually, like, what are the actual numbers of the input versus cache versus output? Yep.

So this is just to get sort of like market standard averages, like baseline benchmarks. Now Meta could be using these tools differently, but if we are to assume that the shape of their agent decoding efforts are similar to the average, this is what the numbers would look like.

So maybe that there is like some bad incentive where people are just saying, to the model, like count up to a billion and then do it again. So then it's like totally skewed. But if they're doing it relatively normally, so on OpenRadar, it's about 98.9% of all tokens are input. And that's including cached ones.

Yeah, because you're stuffing the context window with all your code base or a huge amount of context. It's going in and finding things.

And that's not changing every time, so you can cache it. So then that's like 1.1% is output. So basically, if you basically— get all the numbers, a million tokens is going to be $2.26. So that'll get you to something like $136 million a month for the 60 trillion tokens, right? So that's like way less than the 900. So that, that would be $1.6 billion a year, like run rate. Still huge. That's a lot, but that is still in the max. Assuming they're in the top. Yeah. That's assuming that OpenRouter, the kind of breakdown of how they're using the tokens is the same as OpenRouter, which I think it's not. If we assume that, that's like $4,500 per engineer if there are, I think, 30,000 engineers at Meta. Yeah, every month, $4,500 on tokens. $4,500.

That's actually in line with what I've heard a lot of other people spending in terms of their token budgets. Yeah, that's not like absurd.

That's not absurd if you're trying to incentivize people to use that.

Yeah, yeah. No, not at all.

But so you can actually see the breakdown on OpenRadar of how people are using tokens. So 17, The biggest plurality is OpenClaw, which is 17.6%. Yeah. And then Claude Code is 16.8%. Sure. So I think if you think about Claude Code, you would imagine that like in Claude Code, there's the kind of percentage of cached tokens is gonna be higher than in OpenClaw. Yeah. So I think Meta's usage is actually gonna be more heavily based on the cached tokens.

Sure.

So if you do it just based off like Claude Code usage, you'd actually see a higher percentage of the input tokens be of the total tokens. So it's only like 0.8% is the output. So then if you get all those numbers through again, it's only like $55 million a month, which would be $669 million a year. And each engineer would be like $1,800. Yeah, that's actually pretty low. Which is like, I think, very reasonable.

John Chu over at Coastless says, plenty of my Meta friends told me folks have been building bots that just run in a loop burning tokens as fast as they can due to this policy. It is an apt— it's an apt description. Absolutely stupid policy and is similar to how Meta uses lines of code to measure engineering output. Managers are supposed to use it as a proxy and dig in to understand where complexity, but plenty of managers are lazy and just don't. That was in response to Christina over at Linear saying ranking engineers by token spend is like me ranking my marketing team by who spent the most money. Yeah. We may not have hit our KPIs, but Joe spent $200,000 on a branded blimp that only flies over his own house. So he's getting promoted to VP.

I'm pro-branded blimps though. I like that idea. So my take on this was that, yeah, Yeah, it sort of ties to what Jensen Wong was talking about at GTC. He was saying that an engineer that's making $500,000 might soon command something on the order of $250,000 a year in token budget. Andrej Karpathy had a similar line. He said, "It's all about tokens." He said on a podcast last month, "What is your token throughput and what token throughput do you command?" And so Meta actually has two different harnesses internally. They have a version of OpenClaw called MyClaw, and then they also, of course, acquired Manus. But it appears that they're running Claude, maybe Opus under the hood to actually generate the tokens that come through those harnesses. The interesting thing is that at $250K AI budget per engineer, you're at like $20,000 a month. And so based on Tyler's math, this feels like Okay, there's going to be another maybe 4x to get to Jensen's prediction. I think it makes clearer the strategy with Meta Superintelligence Lab because if you're looking at, you know, it's clear that they're spending hundreds of millions of dollars on this just for internal code gen tooling, like running their business. They are going to spend an inordinate amount of money on frontier inference. And so training a model there, they will be able to amortize the training cost of the next model that they build, not just over can they get a product out that goes viral and becomes its own standalone chat app that people pay for, or maybe it's ad-supported. Like just on the internal usage, they could be running a, you know, multi-billion-dollar token bill that they would have to pay another lab. And so if they develop that internally, it's pure vertical integration. And then you also have everything that's happening on the actual ad targeting and content delivery side. And when you add up all of those, all of a sudden the big question has been like, does Meta, is Meta going to be able to launch an entirely new AI product like Vibes or something like that? And this is a data point that to me says they don't need to. Just from a pure vertical integration story, the investment in MSL can pencil out.

What, are you laughing? I just want you to get to your schizo theory. What's the schizo theory? That this whole like token maxing thing distilling is like a barrage while they distill the model.

Oh, oh yeah. Yeah.

I mean, there is a world where if you're running, if you're generating trillions and trillions of tokens of a frontier model.

They're like, wow, Meta is really like burning through a lot of tokens.

And you haven't generated everything.

It's like, oh, we're just token maxing. Yeah.

I mean, there's another story about distilling we'll get to later in the show, but there is a question about if I have a, if I write an essay and then I have a model rewrite it, those tokens, They are from that model provider. They— I buy them, they become mine. Can I train on them? That's probably out of terms of service. So you would think no, but you sort of wind up in this Ship of Theseus world where if Meta pays Anthropic $100 million or $1 billion to go rewrite every line of code, every email, every Slack chat, every internal message, like basically map the entire organization, rebuild it. They wind up with an incredible training corpus that they can use for their next model. But I would imagine that they can't. And I imagine that the enterprise contracts go both ways. They can't, you know, the lab can't train on the corporate information. That's standard in all of the enterprise contracts. And I would imagine that the opposite is true as well. Although it is this fuzzy Ship of Theseus world where If you're using coding agents to upgrade your infrastructure and then you want to run and train some model on your infrastructure, do you have to pull out the tokens that were revised by the AI lab that you don't have the right to train on? It's all very interesting. Apparently startups that have gone out of business are able to sell their corporate histories for something like $1 million to data brokerage firms and AI labs now.

Have you heard about this? Yeah, heard about it. Skeptical? I'm skeptical. I mean, certainly there's a market for it, but basically all the code that a company built over a few years, maybe they raised— Code, but also usage within different enterprises. Yes, all sorts of different stuff. In other news, Intel is joining TeraFab. Yes, let's read through that. Intel is proud to join the TeraFab project with SpaceX, XAI, and Tesla to help refactor silicon fab technology. Intel says our ability to design, fabricate, and package ultra-high-performance chips at scale will help accelerate TerraFab's aim to produce 1 terawatt a year of compute to power future advances in AI and robotics. And throwing up a post of hanging with Mr. Musk himself.

Let's go through the Wall Street Journal's coverage of this. Elon Musk is partnering with Intel on his ambitious TerraFab project, which aims to build specifically designed chips for SpaceX and xAI as well as for Tesla. In an announcement Tuesday, Intel said it would work with the companies to design, fabricate, and package ultra-high-performance computing chips at scale. The company shared a photo of Chief Executive Lip-Bu Tan shaking hands with Musk, CEO of SpaceX and Tesla. The partnership is a win for Tesla, which has struggled in recent years— Intel, which has struggled in recent years, leading the company to cut production capacity when demand was surging for data center chips and when competitors like Nvidia and AMD have thrived. That was always a just such a tough pill to swallow when you would talk to the ASIC companies like Cerebras and you would say, hey, like you're doing something new, you're not doing Nvidia chips. Is there any way you could get off of TSMC? And they're like, no, like we still need to be in Taiwan. Obviously there's a huge geopolitical component here. We can get into all that. But last year the Trump administration reached a deal to acquire an equity stake in Intel for around $9 billion to help secure the American chipmaker's business. The US government held 8.4% of Intel shares outstanding as of March 20th, according to securities filings. The figure doesn't include warrants that could increase the government's equity stake in Intel. TerraFab represents a step change in how silicon logic, memory, and packaging will get built in the future. Tesla and SpaceX confirmed the partnership in posts on X. Musk unveiled the plans for a single facility in Austin, Texas, to make chips to be used by SpaceX and XAI, which merged in February as well. As by the publicly traded Tesla. He pitched the project as an opportunity to quickly experiment on chip design by designing and manufacturing the chips in one facility. The fab will make chips for use in Tesla's robotaxis, which they're already fabbing, I believe, at Samsung, although they do have Nvidia Dojo chips, I think, that are TSMC. Optimus will also need chips and they are planning to use Intel for that as well. So these are two areas of priority for the electrical vehicle maker as it shifts its focus to artificial intelligence-enabled products. It will also make chips optimized for use in space, where SpaceX is planning to deploy huge numbers of satellites capable of handling AI computing tasks.

Who else do you think they need to get involved here? Because just the two of these, the two of these got, you know, Intel and Tesla coming together. It's good to have more involvement, but still, I think the entire project—

No, we've seen, we've seen a few of those, like, AI leader gatherings in DC where you see Tim Cook and Sundar and Sam Altman and Dario and all the, the whole, all the leaders are together. And I was always hoping that at one of those dinners they would say, okay, everyone's going to try and say the biggest number, but this time it's going to be how much you're committing to Intel and how much you'll, how much you'll buy from them if they come online with a competitive product. Because the demand side has always been a big problem for Intel that they have the capability, they have the plans to build the 2-nanometer, 3-nanometer plant, like a frontier plant, leading edge fab. But every other company has been so tied to TSMC. But I think everyone now acknowledges that TSMC is not investing super heavily in CapEx. They're not going, you know, they're not scaling up as much as as the industry would like them to. And so lots of folks have, have sort of signaled towards a chip bottleneck coming in the next few years. And Intel has the opportunity to communicate that this seems like the first step in that, in that chain. So companies including Tesla often design their own semiconductors but need a supplier to actually make the so-called— make them in a so-called chip fab. Musk's companies have sourced chips from a wide range of suppliers, including Nvidia, Samsung, Taiwan Semiconductor. Oh, I got it. Musk said that TerraFab is needed because his company's demand for chips is slated to far outstrip the supply it gets from partners. I was listening to Chuck Robbins from Cisco talk about data centers in space and the heating issue came up and he was like, yeah, I don't really have like a solid answer for that yet. But I do think that if you are bullish on data centers in space, you have to start with the fact that Starlink works in space currently because it is doing compute.

It's not doing—

You couldn't possibly put, let's be honest, John, we couldn't possibly put a computer up there.

Yeah, like there are computers with like, they can't inference frontier models. They can't, you know, it's not gigawatts in space yet, but there are, I believe across the entire Starlink cluster, megawatts of compute in space with solar panels. And they do heat up because you are running a chip that routes packets across the internet from one satellite to the next to get you your internet via Starlink. And so it's not that it's a solved problem, is that we are actually, we are on a path to deploy some level of compute in space, Tyler.

Yeah, I mean, we've seen like Philip Johnson, like there are chips in space right now, like there are GPUs, I think, Arthur, he said there were like 5 or 6 H100s, right? Yeah, yeah. So like they do work. It's like the— Yeah, I think most people's problem with space data centers is that it's like economically it doesn't make any sense.

Well, so yes, that is the correct angle, but a lot of people are getting hung up on—

Not that it's like physically—

No, no, there is a whole conversation about like, like it is impossible and you need to like move past that into the economic equation, which then gets you into timelines and actually thinking about what needs to happen to dissipate that heat. But clearly, yes, you can. I mean, you can put humans in space on the ISS and cool that. They're like, we, we have created ways to, to, to move heat around in space for decades. It's a, it's obviously a new challenge. But I think starting with, with the baseline of like there is compute happening in space right now, we're going to try and I mean, Elon wants to like 1,000x it, 100,000x it, 1,000,000x it. I don't even know how what the, what the scale is, but orders of magnitude. And so there's new engineering challenges.

Speaking of space, looks like Elon is going to use SPCX as a ticker for the SpaceX IPO, which he had to acquire from Matt Tuttle, hence the ETF's ticker change shown below. Eric from Bloomberg says, we predicted this could happen in a December note. Nice catch by Will, who famously gave the Meta ticker to Zuck. I did not know that Will Hershey had the Meta ticker. Previously. So, uh, we know, we know somebody that squats on—

who, who had the Meta ticker?

Uh, a guy named Will Hershey. Oh, interesting. There's a company called Roundhill. But we know somebody who's— I thought it was Matt Ball. We had somebody here. Yes, yes. Outside of show hours and say that they were squatting on a bunch of tickers. And the, and the idea seems so— I, I think, I think what, what, what might be the reality is that you actually— that it needs to be it needs to be further along than just reserved. I don't know. I think so. Having it, you can go if you're a startup today, you can go reserve your ticker today. But I'm not sure that that actually gives you enough leverage to when, when Elon comes knocking ready for an IPO, you actually have priority. All right. We got to talk about a corporate retreat that went badly wrong. Okay. Technology company Plex took its 120 employees to Honduras for a week-long bonding experience. It was a disaster from the moment they arrived. Senior executives at the tech company Plex were eager to treat their 120 fully remote staffers to a week-long corporate getaway in a tropical paradise. Pop quiz, Tyler, do you know what Plex is?

I don't know about Plex. No.

Have we seen Plex before?

I don't know either. So we all failed, but now it's your job to figure it out. I will continue. The plan for the Honduras trip was simple. Company meetings and team building by powdery soft beaches during the day and island fun at night, at a cost of roughly half a million to the company. They'd build the trip around a Survivor theme with teams and challenges, but it'd be fun, not too physically grueling. The CEO of Plex, a free streaming platform, would play a role similar to that of Survivor host Jeff. Perhaps the executive should have taken it as a sign that just as the first bus of staffers pulled up to the resort, the chief executive was already in his hotel bathroom experiencing the initial waves of violent stomach infection. What followed was a comedy of errors, including military drills that outpaced anything this group of office workers had in mind, a rogue porcupine, stranded airplanes, and one syringe to the butt of an employee. Corporate retreats are generally assumed to be torture, or at least a semi-stressful chore, what with their forced fun activities and hybrid work-play environments that leave workers confused about boundaries. Is that— is that like the industry standard. That seems wild.

I don't know if I've ever been on a corporate retreat.

I've been on some like Founders Fund events, but those aren't really retreats. Those are more just like conferences. But I don't know, corporate retreat seems, I don't know, never unexplored territory for me.

It's no wonder the new season of Jury Duty, a comedy series that tricks an unsuspecting non-actor into believing his off-the-wall fictional circumstances are actually happening, is set at a corporate offsite. But in real life, PlexCon 2017 beats anything on TV. Here's the story of an all-staff company getaway told by 6 people who were there. A trip where most everything that could go wrong did go wrong. Nearly a decade later, they're still working together and still talking. So it's crazy that they—

it was bonding experience.

Yeah. Well, yeah, it's crazy that this is now, now coming out. So Sean Hoff, 42, founder of Moniker Partners, an independent corporate retreat agency that planned the trip. About 3 weeks before we arrived in Honduras, we got an email from the hotel's general manager that said, I will be departing. I wish you the best with your retreat. I knew something was off. 3 days later, another email. The head chef was no longer going to be at the hotel. Uh, Scott, 52, chief product officer and Plex co-founder. We get there. We've got to take the bus from the airport, dirt roads. You start getting closer and there are guard towers around the property, people with machine guns and stuff. A lot of people were like, where are we going? Keith, CEO of Plex, 54. We usually go a day early and we set up if there's any little thing we have to get it right just so the employees have the best experience possible. Keith woke up the day that people were coming in Sunday morning and he is sick as a dog. Everyone there is fried. Basically, people are telling me, don't eat the vegetables, don't eat the vegetables. That's like the same. No, no, no, because they clean it, they wash it in water. It's usually not filtered water, right? Because it would just be kind of crazy to—

yeah, yeah, here it is.

I, I've got to have a salad, just one salad. So I got E. coli, which may be the worst thing you could get possibly ever. Just as people were arriving on the buses, I was like— I had lost 8 or 10 pounds. They had a doctor come to me, which apparently is pretty standard. They nailed an IV bag to the bedpost. It's just— people are arriving for a party that night. The next day is Survivor theme kickoff. There's not one person on the planet more excited about Survivor Keith and his wife, they have watched every single episode. My wife and I met Jeff, the host of Survivor. What I wanted is when everybody shows up, I do a Jeff, welcome to the island, here's the theme for the week, but Scott got to do it. The opening Survivor thing was a contest where people on their different teams open up a platter. You have to eat what's on the platter. Sean, Sean, who's the Plex head of business development. Are you gonna call him? Yeah, somebody is, somebody, is cold texting me, pitching me their startup, and they've called me a bunch of times today.

Wait, is it actually them or is it their AI agent?

I wish I could pick up. It's just like a little bit. Yeah, it's a little too much to pick up. But yeah, cold texting somebody, like getting their number, I don't think that's the new meta. No, it's, it's, it's bold.

We heard from an executive in, in tech that they are getting dozens of emails every single day trying to recruit them. And every email comes from a new Gmail account that's unregistered, brand new, but it's all like, you know, LLM written, very different, like doesn't really do all the research but has a few keywords in there. And it's clear that someone is building sort of like a next-gen recruiting agency that's basically just a lot of spam. Feels like the end result will be like a return to relationship building and not like broad top of funnel.

I should read the cold text from this morning. And I have nothing against cold email and just being bold, but I did read this out loud to you, John, so I'll read it to everyone. So I got a text from an unknown number today at 7 AM. All right, Jordy, good news or bad news first? This is blank, and I'll leave the name out. And then I just get a PDF of a deck and then a text. All right, Jordy, the bad news is this was an unplanned introduction and on the surface, probably lukewarm outreach. The good news is that there's zero doubt you're now in touch with the founder with the most grit of anyone you've interacted with the past 12 months and likely anyone you'll interact with over the next 12 months. 50,000 seed round passes over the past 10 months. Here to make 50,001. So, so, you know, you should be coming in being like, I've been passed on 50,000 times.

Yeah, I'm hoping this is the one that gets through. That gets through. That seems like a rough estimate, though.

Every month of feedback and iterations have made it better. So you're seeing more quality, more quality presentation than rejection. 10,000. Looking forward to your message.

The chat wants the builder to pitch.

They want you to hear this out. Everyone's in favor of this. The chat wants you to get on the phone with them. Do it live. I mean, they wanted to do live. I don't know if you should do live, but you should take the call.

I will take the call. I will take the call.

But let's go back to the corporate retreat.

Okay. So they hire— they hire a former Navy SEAL. Okay. To basically haze the team on the beach. And you can pull up a picture, an image here.

The quote is, "This is not a super fit group in general. One of our biggest mistakes was hiring a former Navy SEAL to pump the team up. As I'm in my room dying, I could hear them out there doing all the drills and yelling, and so I'm in here thinking, 'This is terrible,' but it sounds terrible out there too. We're doing army crawling on the beach. It was 100 degrees. I bailed out partway through. I went into the ocean just to cool off. I went in probably on all fours because I was tired. It's not a fit group." not a super fit group in general. The ex-Navy SEAL is like, we can tone it down, no problem. We get up there and it's hot and humid and people are passing out. I don't think he'd ever seen quite such an unfit group. We ended on, I guess, what's probably a golf course. On command, everyone had to hit the grass. Everyone's silent. We're pretending we're Navy SEALs, but I happen to land in the wrong spot. I'm just like, oh God, what is happening? I was sitting on a fire ant hill. I was wearing shorts. I jumped and had my— I had hives and bumps from the bites.

This is Ridiculous.

Someone saw an alligator on the golf course.

Sounds like a ridiculous—

There was a porcupine that fell through one of the ceilings.

This is like a Fyre Festival for corporate retreats.

The Fyre Festival of corporate retreats.

Anthropic is taking steps to arm some of the world's biggest technology companies with tools to find and patch bugs in their hardware and software. The company is making a preview of its new AI model called Mythos available to about 50 companies and organizations that maintain critical infrastructure, including Amazon, Microsoft, Apple, Alphabet-owned Google, and the Linux Foundation. Cybersecurity researchers and software makers worry that artificial intelligence is becoming so good at exploiting vulnerabilities that it could cause widespread online disruption. Security experts have predicted that AI models will discover an avalanche of software bugs, and the effort is set to help companies stay one step ahead of cybercriminals and other threats. This feels like a very good rollout strategy. Generally, both because we've seen a huge amount of cyber attacks and hacks and accidental releases. Like, even if it's not, you know, there's been— we had a member of the security team from CrowdStrike on the show last week talking about the rise in cyber attacks broadly, getting the most frontier models in the hands of big companies early. Great from that perspective and then also just great as a product demo, which will get the entire organization excited about deploying the technology broadly. So very good as like a, as a B2B go-to-market motion. This makes a ton of sense. In some other more positive news, OpenAI, Anthropic, Google are uniting to combat model copying in China. This is a bigger discussion around AI safety. We've talked about this.

Would you look at that? Some, some faith in human users.

Who knew that you could get along?

Yes. I mean, like, I mean, I'm sure people in the chat have seen the New Yorker article where there's just tons and tons of quotes from various AI leaders, all, you know, upset with Sam Altman. And the inter-AI drama has been bubbling up since the dawn of OpenAI. Like OpenAI was started as a reaction to Google and then Anthropic leaves and teams up with Google, and then Elon doesn't like Anthropic, and then Ilya Sutskever and Mira leave, but they don't join Anthropic. And so there's been so many personalities and so many disputes. I feel like the takeaway is that this is all extremely high stakes. There's a technological transition happening, a huge amount of money on the table, a huge amount of influence on the table.

And so everyone is sort of clamoring for their share and it's creating a lot of friction.

So rivals OpenAI, Anthropic, PBC, and Alphabet Inc's Google have begun working together to try and clamp down on Chinese competitors, extracting results from cutting-edge US artificial intelligence models to gain an edge in the global AI race. The firms are sharing information through the Frontier Model Forum, an industry nonprofit that the three tech companies founded with Microsoft in 2023 to detect so-called adversarial distillation attempts that violate their terms of service, according to people familiar with the matter. The rare collaboration underscores the severity of a concern raised by US AI companies that some users, especially in China, are creating imitation versions of their products that could undercut them on price and siphon away customers while posing a national security risk. And so I was trying to square this question of distillation and model commoditization with the news that Anthropic has reached $30 billion in run rate and has an agreement with Google and Broadcom for multiple gigawatts of TPU capacity. Like clearly there is insatiable demand for frontier tokens, frontier models. They're incredibly expensive to train. We saw in the Wall Street Journal that these expected training costs from— yeah, it was training and inference, but it was hundreds of billions of dollars. And so The hope is that you, is that you're able to amortize that over at least a couple of years, you know, a long time ideally. The shelf life of a model after you train it is pretty limited if you're being commoditized and copied. If you're being distilled, it's even faster. At the same time, just staying on the frontier clearly leads to an incredible ramp in revenue. So is commoditization a real problem? It feels like it's almost just more of a problem from an AI safety perspective because you can't have the geopolitical conversation like what Bernie Sanders is proposing around different labs working together, potentially pausing or slowing down or just even adding more constraints and reviews before models get released. It's harder to do that if you have a different country that's racing ahead and moving much faster and trying to close that gap.

Leave us 5 stars on Apple Podcasts and Spotify. Sign up for our newsletter at tbpn.com, and we will see you tomorrow.

Goodbye.