Claude Mythos’ Immense Power, Microsoft’s GitHub’s Growth & Outages, Is Nvidia Worth 400% More?

0:13Welcome, everyone, to The Information's TI-TV. My name is Akash Pasricha. It is Wednesday, April 8th. First up today on the show, Anthropic made news yesterday with its announcement of an AI cybersecurity project. We'll break it all down with Tomas Tunguz, general partner at Theory Ventures. And the pullback in crypto is pushing some companies to diversify their products. My colleague Yueqi Yang wrote about how Polygon is one of those companies. Plus, AI agents are having a big impact on GitHub's business. Our Microsoft reporter Aaron Holmes will join the show to unpack a column that he wrote this week about that.

0:48And finally, as Silicon Valley debates whether AI valuations have gotten ahead of themselves, Our senior finance editor, Ken Brown, has a new column out about how some valuation frameworks could suggest the opposite. It's going to be a fun show, so let's get right on into it. Anthropic made news yesterday with its announcement of Project Glasswing, which was a bit of an eye-opener in terms of how AI could impact the cybersecurity sector. I want to bring on Tomas Tunguz, general partner at Theory Ventures, to help us break it all down. Tomas, welcome to the show. It's great to have you here. It's great to be back, Akash.

1:21So what is Anthropix Project Glasswing?

1:27Tomasz Tunguz:So Project Glasswing is, well, it starts with Mythos. So Mythos is the largest model ever trained. It's about five to 10 times larger than any model trained ever. And it is incredibly powerful. To give you a sense, most model improvements when they become state-of-the-art measured on different benchmarks are somewhere between two to three percentage points of of improvement. Here we're seeing 10 to 20 percentage points of improvement overall. It's an enormous step function. And the capability of this model in agentic coding has another side of the coin, which is its ability to actually pull software apart and expose its vulnerabilities.

2:07Tomasz Tunguz:As a result of the capabilities of this model, Anthropic has decided, unlike all previous releases, to only share it with about 40 different organizations. And I imagine they're both companies and governmental entities. These are the organizations that build the major infrastructure for the internet and banking because we need to harden those core systems before anyone else has access to this model. Plus, as Anthropic has written, they need to develop the guardrails to make sure that this model behaves in a way that's aligned with our interests. So this is kind of interesting. I'm imagining these software companies making a phone call saying, hey, just give us a little while to figure out how we're going to deal with this.

2:53You know, I guess my question here is, so Glasswing is the effort to say, hey, we'll work with a few select partners to see how we can roll this out. I mean, what's your reaction to this? Like, is this conceivably very dangerous? I mean, does Anthropic, they're going to have to get this model out at some point, right?

3:13Tomasz Tunguz:I don't know. I suspect there's an era where you have to be approved to use certain models. This model has a certain designation, which means it can be a weapon system, and you don't deliver that to everybody. There are certain laws around which weapons can be sold to which countries and which kinds of companies. And so there may be a class of models that are so important and so powerful that only a handful of companies have them. And I think there's a big question here about, can these models be used only for defensive purposes rather than being built in products? Because, okay, the US government will have access to this model.

3:59Tomasz Tunguz:The networking providers will have access to this model. The security companies will have access to the model. They'll harden all those systems. But imagine now they start to commercialize the model and start to compete with proprietary access to a model that their competitors don't have. There's a king-making dynamic here that could exist and a company that can decide who wins or who has access and who does not. So I think the first is just like national security, and that's the priority. Then there's core infrastructure stabilization. And then there'll be a third question, which is around commercialization and access and I don't want to call it fairness, but accessibility within the marketplace.

4:37So in other words, you're predicting a very controlled rollout of this product. If you can even call it a rollout at all, it's more of like a certification. Like, you know, are you safe enough to use the product? Only then will we sell it to you, essentially.

4:58Tomasz Tunguz:Right, and so this exists in different levels, right? So you have SOC2, we can debate how safe that is, but then there's, for the federal government, there's FedRAMP, right? And there are different levels associated with that level of certification. And so, you know, I think we're all learning at the same time. And the capabilities of this model are incredibly impressive. And so making sure that country is safe and secure and the core systems are safe and secure, priority one, I think everybody has put aside commercial intent for that reason. But eventually we'll have to face that or answer that question.

5:31Now, I saw that cybersecurity stocks jumped yesterday. Yeah. And I imagine it's because everyone's like, well, the future is looking very bright for companies that can afford that that can offer any protection whatsoever. The question I have is, so you have that dynamic where you have agents potentially posing security threats. Then you have that other SaaSpocalypse dynamic, which all enterprise software companies have been struggling with. So is cybersecurity entirely immune to the SaaSpocalypse and the threat that AI can cannibalize their business?

6:05Tomasz Tunguz:It is the category that has suffered the least. The last time we analyzed it, it is security systems are insurance policies and the risks associated with AI compromise have increased enormously. It's almost unquantifiably. And so as a result, everyone needs more insurance. And those insurance policies are software systems. None of them can offer you a guarantee. but security systems are just layers of an onion or, you know, cushions, many, many more pillows around a very sensitive thing. And so I think the demand will explode because now all of a sudden, if you are facing an adversary, let's say, you know, a malevolent actor has access to a model like this, the defense that you need to have as an organization, a security posture that you need to have is incredible.

6:56Tomasz Tunguz:I mean, there was a tweet yesterday, I think it was Dan Romero tweeted, everything you assumed was private is now suddenly public, right? The security posture of many organizations when faced with the model of this capability is basically porous. And so now we need to get software to be built to be much, much more resilient. And the models are capable of doing this. There are new technologies, there's math that we can use in the form of formal verification, and then models like Mythos, but Mythos won't be the only one. And so we need to secure the perimeter. I mean, I'm sort of, it's a hard question, but I've got to ask it.

7:31I mean, in this world that where these tools do exist and where security becomes so much harder, I mean, I wonder how you think about how the structure of technology, products, software, how all of this could change. I mean, is it really just simple as cybersecurity products will get more popular? You know, we will employ more people at tech companies to build protection systems? I mean, is it as simple as that? Or I don't know, you know, do the, do the, do you get cybersecurity companies that, you know, they're the only ones that can make the next popular consumer app because they need to be safe?

8:08Like, I don't know. How do you think about the structure of the sector changing?

8:11Tomasz Tunguz:Well, I mean, I think you could see a clamp down on vibe coding in a very material way, because Vibe-coded applications tend to forego or at least delay the implementation of security. That's a nice way of putting it. I think the reality is, we've been shipping software very, very fast. And now all of a sudden, okay, the software has to be unbelievably durable, which means that we may lengthen or slow deployment cycles for hardening. And you can imagine enterprises spending two, three, four, maybe 10 times as much of token budgets on hardening as the original software implementation because that perimeter needs to be secure.

9:01Tomasz Tunguz:That's the big shift. I don't imagine that security companies will be developing the products just as they weren't in the past. We all have to level up. So this is kind of interesting. So what we could see, if I'm hearing you right, maybe we see these companies that are saying, token maxing, use Vibe Coding as much as you can. I'm imagining the opposite, where maybe you have these big breaches and the companies say, okay, hold on. We're canceling Vibe Coding for maybe a quarter while we short through this. Maybe there's actually a pullback. I think it's an and. So, I mean, Amazon had that. Not as dramatic, fine.

9:47Tomasz Tunguz:No, no, no. But maybe put it a different way, which is continue to token max and continue to iterate. Do not ship. Do not ship until the system is hardened. And then instead of spending the last 10 % of your maxed tokens on security, spend 50 to 60 % of your tokens on security. Got it. That I think is the more likely outcome. Okay. Okay, speaking of Anthropic, I want to talk about a column that you wrote yesterday. We got a look at Anthropic's latest revenue figures. And you asked the question in your column, when will Anthropic surpass NVIDIA? I want to know what the answer to that question was that you landed on.

10:28Tomasz Tunguz:Yeah, so it's somewhere between seven months to three years, the current course in speed. and you just you know nine billion of bookings for anthropic in february 10 billion of bookings in in march or april it's uh i have the months off there but it's uh it's march and april sorry no i had it right february and march anyway so 10 billion in bookings is all the revenue of databricks on a run rate basis plus all the revenue of palantir on a trailing basis in a single month there's a growth rate here we've never seen before in a software company and if they come anywhere close to sustaining or if you apply moderately conservative discount factors to that growth, you still see them exceeding NVIDIA's current market cap somewhere between seven months to three years from now.

11:13Tomasz Tunguz:And so current course and speed, they'll be among the top five most valuable businesses, no doubt. And so that's if they keep going at the rate at which they've been growing. My question is, do you think that it will happen? I think it will happen. I do. Historically, software companies, they trade at a higher multiple than hardware companies. They have more durable revenue streams. You look at NVIDIA's PE over its life, there are three significant surges there, roughly here, 50, 100, and 150, marked by three different waves. You had the crypto, sorry, you had the gaming wave, then the crypto wave, and then the AI wave.

11:58Tomasz Tunguz:And software companies hold on to their multiples for much longer. So I think there's a tremendous amount of upside. And if the company can go, you know, I mean, add$9 and$10 billion, and we think about the overall penetration of AI within the enterprise, I think it's still less than 5 % in terms of overall terminal demand. There are a lot of tokens to be processed. And we should be clear, are you an investor in Anthropic in anywhere? No, I hold no shares. What about OpenAI? I mean, is it yesterday's news? No, I think there's a pendulum swing that happens. Who has the hot hand? And the pendulum continues to swing.

12:41Tomasz Tunguz:There are brilliant people in both of those businesses. As we just talked about, the market is absolutely enormous. OpenAI is predominantly a consumer business that's trying to build its enterprise business and successfully. And Anthropic is a company that's focused almost exclusively on the enterprise business. I think for OpenAI, figuring out the ads model will be essential in order to commercialize that ads business. And there's plenty of market cap there, right? So I think it's about$500 billion ad market today. What fraction of that will go to AI? It'll be pretty significant. And so I think there's plenty of room for both of them.

13:17Tomasz Tunguz:and the narrative will keep going back and forth and back and forth. Now, one more question for you, because I know that you like to dig into the numbers. So one point that our co-executive editor, Martin Pierce, pointed out in his column last night was how you measure revenue and revenue run rate between these two companies, between OpenAI and Anthropic, isn't necessarily always the same. You know, in some cases, the revenue share agreements that these companies have with the cloud providers may or may not be baked into the revenue projections. And look, they're not public companies, so we can only go based on either what they say or what my colleagues report at the information.

13:55But have you done any kind of research or analysis into the extent to which if you incorporate revenue shares, how that would affect the top lines for these businesses?

14:08Tomasz Tunguz:No, I haven't. I've read Martin's article, which is a great piece. I think the idea is OpenAI revenue is net of partner costs, let's say, or channel costs. And then Anthropics is gross, is my understanding. I haven't done any work to substantiate that. But it could very well be true. And we'll see how the market values it on a comparable basis when both of these businesses are public. And we see the S1s. I'm really keen to see what's under the hood. Last question. Who do you think goes first, Anthropics or OpenAI?

14:43i would bet it's anthropic yeah i think so too i would bet it's anthropic i think i think spacex

14:50Tomasz Tunguz:going first is a bit of an advantage because everyone will sort of figure out their allocation these are just enormous right these three ipos are greater than the sum total of all dollars raised in ipos over the last decade and so where does that money come from what's happening to the indexes? Do those probably need to be sold down in order to make room for these businesses? And is there cyclical rotation out of materials and industrials in order to find the liquidity to fund these IPOs? Those are very real questions. It's amazing. I mean, I think AI as a whole is breaking a lot of legacy systems all the way to the financial markets.

15:30All right. Well, Tomas, I want to thank you for coming on. That is Tomas Teguz, General Partner at Theory Ventures here on TI-TV. Amid the crypto downturn, Polygon, a blockchain company, is looking for ways to diversify out of the stalled market. The information's crypto reporter, Yueqi Yang, joins me now to share more about what she learned in her reporting. Yueqi, welcome back to the show. It's great to have you here. Hey, Akash. Well, tell me about what we learned about Polygon. So Polygon Labs, which is most known as the developer of the Polygon blockchain, is right now in early talks with investors to raise as much as$100 million to launch a new stablecoin payments business, according to sources.

16:16Okay, now this is the same polygon that we heard about years ago, right? I mean, I feel like I haven't heard about Sandeep, who was the, I think he was the founder, and there was a foundation, and this was a layer one blockchain, am I right? This is one of the earliest blockchain that enabled fast and cheap transactions that's compatible with the Ethereum blockchain. And they kind of rose to prominence back in the days, especially around 2021, as one of the first developers that was able to achieve this. and they are known for the Polygon blockchain, which is right now used by crypto platforms and also prediction market, including Polymarket.

17:00Right, and maybe I'm forgetting, maybe it wasn't layer one, maybe it was layer two because I remember it was compatible with Ethereum. Anyway, it doesn't really matter because there's probably five more layers that we haven't talked about right now. What is Polygon Labs pivoting to now? What is their new focus here with this$100 million funding round? So according to sources, Polygon Labs wants to launch a stablecoin payments business. And this is the kind of business that Stripe is providing. There are also startups such as BB &K, ZeroHash and others that are also providing. And it is one of the red hot area in the crypto industry right now, which broadly speaking is facing a downturn.

17:43So it is one of the only few Brightspot and Polygon Labs, which is traditionally known more as a blockchain developer, is also trying to get into this regulated stablecoin payment space. And I will say that it is a rare move for a blockchain developer to get into payments business, but this is the direction of travel we're seeing broadly in the crypto industry now where companies are trying to pivot and trying to reposition themselves for the next phase of the crypto industry. So what exactly would be the product or the infrastructure that it provides? Because we know stablecoins, we know USDC, for example.

18:24We obviously know Tether is out there as well. But I mean, when I think about payments infrastructure, what is it that hasn't been built? What is the gap that they think they can fill? Yes. So in order to process stable payments, you need a few components. And Polygon Labs has been making acquisitions this year to assemble the components. So first, you need a wallet provider, and they agreed to acquire one in January. And then more importantly, you need an on-ramp, off-ramp provider. So Polygon Labs agreed to acquire CoinMe earlier this year, which provides the ability for users to convert their US dollar into stable coins and vice versa.

19:08And then Polygon Labs obviously also develops the Polygon blockchain, which is also widely used for stablecoin transactions. So they're in the process of building out different components. And their hope is that when they go to a pitch meeting with a client, they're able to say that we have all the possible pieces you need to provide stablecoin payments. And you can just sign with us as one counterparty in order to offer stablecoin processing ability. How much competition is there in this market for stablecoin payments? That's a great question. And there is a lot. And there's going to be more.

19:46I will say the biggest one is Stripe. They're pretty dominant. The stablecoin payments business is a huge push for the organization. Obviously, they made the acquisition of Bridge and Privy last year, which gave them the component to offer stablecoin payments. but also other companies like Coinbase are also pushing hard into stablecoin payments. And each firm has different advantages but also different weaknesses. For a company like Coinbase, for example, they're a crypto company and they don't have an existing network of merchants that could sign on to their payments product unlike Stripe. And Polygon Labs is obviously in a different position as well.

20:31So it's an uphill battle. do you think that they can have any success based on your years of covering this sector? We'll see. And I will also add right now, the fundraising environment is very challenging in the crypto industry. We've made that show that fundraising amounts dropped by 69 % in the first quarter from the prior quarter. Broadly speaking, it is a bear market right now in crypto. So it is hard for companies to raise money although payments really related deals are still pretty active and a few acquisitions especially the bigger ones that we've seen are all related to payments companies right in other words you're saying that there's no guarantee that they actually will be able to raise all 100 million dollars it's still in the the talks phase of fundraising yeah this is a still ongoing talks and we look forward to see what's going to happen Right.

21:26Let me ask you one more question. So going back to the old Polygon, I guess, the Polygon of 2022 that I remember covering when I was a crypto reporter, there was the Polygon token. What's been the fate of the Polygon token? Is that still used for anything? Is the underlying blockchain widely used? I mean, it was, you know, as I recall it, people were saying, oh, this could be an alternative, in addition to Ethereum. People were looking for ways to diversify out of Ethereum. What happened to that whole business? So the prices of the Polygon token has not performed well. They have fallen by about 90 % in the past two years.

22:10I will say that broadly speaking, most of the altcoins, which means crypto tokens that are not Bitcoin, not Ethereum, are down in the past two years or so. And Polygon tokens are part of it. But with this new payments business, Polygon Labs is hoping to be able to drive more volume, especially stablecoin transaction volumes, to the Polygon blockchain, which will hopefully be beneficial for the Polygon token holders as well. And remind me, is it still called the MATIC token or is it called Polygon now? I think it's called Polygon now. Polygon, okay, yeah, because there was that whole thing. It was like Polygon, I don't know what the origin of the name MATIC was, but one more question for you, Echi, before I let you go.

22:57Is Polygon the only crypto company that is trying to diversify away from just having an underlying blockchain? I imagine that there are other crypto companies that are trying to make similar pivots. Yeah, there are a lot of pivots that we're observing right now in the market. Given the downturn, I think a lot of companies are rethinking their strategy. The bear market is entering six months now with really no end inside. So a lot of the companies are in survival mode right now. And if they don't pivot, they could potentially be washed out. Great. Well, Yueqi, I want to thank you for coming on.

23:33That is Huechi Yang, our crypto reporter here at The Information. AI agents are having a big impact on GitHub's business, not just in terms of the products that it is working on, but also traffic. I want to bring on Microsoft reporter Aaron Holmes to share more about a column that he wrote this week on that topic. Aaron, welcome back to the show. It's great to have you here.

23:55Ken Brown:Happy to be here. So what's going on at GitHub these days? Yeah, so basically GitHub has this platform that developers use to store and edit code. And as anyone who has been talking to software developers recently knows, basically every developer out there is now using AI agents to write more code than would be humanly possible. And as a result, you know, GitHub is essentially getting flooded with traffic from these agents, according to their COO, who told me that the amount of commits, which is like the number of times that an agent has saved new code to GitHub, has surged to about 14 times the amount from last year.