Anthropic Confidentially Files for IPO, Nvidia’s New Chip for PCs, Ex-Meta CTO on their AI Playbook

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Basritcha. It is Monday, June 1st. We have breaking news to start the show today. Anthropic today filed confidentially for an IPO. It was just last week that the company closed its latest private funding round at a$900 billion valuation before the investment. And that figure is bigger than OpenAI based on its latest funding round. Now, there have been a number of reports that OpenAI has been looking to expedite its IPO filing process. So we'll be closely watching to see if and when we see OpenAI's filing. We're going to have more coverage for this to come to stay tuned.

0:54Today on the show, we will unpack NVIDIA's latest PC-focused chip that it unveiled alongside Microsoft. We'll also preview Microsoft's Build Conference, which is kicking off this week. We're then going to unpack our profile of Denise Dresser, OpenAI's new Chief Revenue Officer, who was previously the CEO of Slack at Salesforce. We'll also talk about forward-deployed engineering and the boom across that category of roles across the AI ecosystem. And to close out the show, we've got Meta's former CTO, Mike Schrepper, coming on the show to talk about his fund, GigaScale Capital's big new$250 million in fundraising.

1:34It's going to be a fun show, so let's get right on into it. Microsoft and NVIDIA unveiled a new chip for PCs, putting NVIDIA squarely in the game of personal computing. I want to bring on our Microsoft reporter, Aaron Holmes, to break down that announcement. Aaron, welcome back to the show. It's great to have you here. Happy to be here. Okay, so what did Microsoft and NVIDIA unveil today? So the biggest part of this news is NVIDIA's new chip called the N1X. And specifically, you know, NVIDIA has for a while dominated the market for chips in data centers. But with this chip, they're actually, you know, releasing their first core processor for PCs.

2:15So they are now entering into the PC market, which has for a long time been dominated by Intel and Qualcomm and some other competitors. And on top of that, we saw Microsoft as well as some other PC makers preview new machines that will use this chip. So in Microsoft's case, that is the newest version of their Surface laptop. They haven't really said many details, including what it'll actually cost. But the idea is that this laptop will have, in a lot of ways, higher performance and be able to run more AI features locally than previous versions of the Surface. And to what extent did NVIDIA have PC chips before this?

2:54Because I know NVIDIA certainly made its name in gaming. And I know, I mean, there are tons of PCs out there that are good for gaming. So was it already in this market? So NVIDIA had mostly been making graphics processing units, which are, you know, essentially what handle graphics on PCs as well as, you know, GPUs in data centers. They hadn't done as much central processing units, which is, you know, the core processors for what your PC runs on. So this is definitely an expansion for them. And, you know, we also have seen that CPUs are in short supply right now. So I think that they are well positioned to capitalize on that supply crunch.

3:35So to what extent do you think then NVIDIA could actually then pose a threat in the CPU market to the other CPU giants? I'm thinking about ARM. I'm thinking about Intel. The companies that Microsoft and the likes of them have traditionally relied on for these chips. I mean, is it just a scenario where there's such a shortage right now that any supply is good? Well, you know, what's interesting is that this chip is going to sit inside of NVIDIA's what it's called RTX Spark, which essentially combines the standard computer processor as well as, you know, a GPU and a memory in one system. And I think the idea there is that if NVIDIA can offer all of that for essentially what is a more cheaper or more bang for your buck, then that might be appealing to PC makers, but also to the end customers.

4:32Again, we don't really know how much this is actually going to cost, so it's hard to say how appealing that will be and whether this is going to be something that any PC buyer would want versus someone who wants extremely high performance. but either way, it definitely seems like an opportunity for NVIDIA to seize on that market a bit. And what about for Microsoft? What do you think this means for Microsoft's business and its PC line? Microsoft has been trying for a long time to get their Surface devices to stand out from the competition and we've seen Surface undergo a number of overhauls in recent years.

5:09They, for a while, were marketing co-pilot plus PCs, which were essentially, you know, PCs that were designed to run more AI locally using Snapdragon processors. It's not clear if that really caught off with consumers. But I think that, you know, with this NVIDIA partnership, they have essentially a new angle to promote the Surface. And in a lot of ways, I think they're trying to encroach more on the MacBook market, especially for developers who want to be able to use their personal device to develop and run more advanced applications. So whether that pays off, we have to wait and see. Can I just ask a quick question here?

5:51You said Snapdragon processor. What is a Snapdragon processor? That's just a previous brand of processor that they used. Well, I guess, I mean, the names, as we know, in this category are sometimes very creative. So it's better than 5.4, 4.7, whatever the other naming convention is. I'll tell you that much. Okay, I want to talk about a story that you published today, Anthropic. You went deep on how companies are using Anthropic's mythos model. And just to refresh, folks, I mean, this was the killer model that came out. Everyone was scared because its capabilities were so strong. Anthropic said, we're only going to give access to a couple of companies here to use the model.

6:37You went out into the world of cybersecurity. You talked to some of the companies using it. What did you find? Yeah, so essentially, we already knew that Mythos was going to be more expensive than Anthropic's earlier models, which are already on the pricier end of AI models. But I spoke to people who were testing it and got a sense of what that actually means in practice. And I learned, you know, a lot of testers saw that they ran through$1 million worth of tokens testing the model in just a couple of weeks. You know, Palo Alto Networks has been using it to test their internal source code. And within three weeks, they had used over$1 million worth of tokens.

7:15But what's also interesting is that... And that's, just to put this in context, that's a lot, right? Like, you know, because these are companies, hundreds of billions of dollars of companies. This is still a lot, right, based on what they usually spend? Yeah, for just a couple of weeks of usage, that is pretty expensive. And what's interesting is that companies still are planning to budget for this just because the amount that they're spending they think is worth it because they've found so many vulnerabilities in their code that they wouldn't have found otherwise, or at least it would have taken them a lot longer without this model.

7:47Just to clarify here, this is a company like CrowdStrike using the model to find vulnerabilities within its own code, which is itself the cybersecurity product that it offers to sell to customers, right? Yes. And it's not just cybersecurity companies doing this. You know, a lot of companies like Microsoft and others that have widely used software are using the Mythos model to scan essentially all of their source code and just find vulnerabilities and quickly patch them. And the idea is that they want to do that before a model like this falls into the hands of hackers who could start doing the same thing and exploiting vulnerabilities before they're patched.

8:30And on the budget side of all this, so they're finding it's more expensive than they may have thought. Are executives telling you that they are planning to expand their budget to make room for models like Mythos? Or where do they stand there? Yeah, I mean, I also spoke to Zscaler, which is another company that is testing this. And, you know, they told me that while it is essentially a step change more expensive than their previous code scanning tools, they are just budgeting for it in the year ahead. And, you know, in some cases that might have looked like a difficult conversation with the CFO saying, look, this bill looks crazy, but trust me, it's going to be worth it.

9:09And from what I've heard, I think, you know, CFOs, CEOs, and even boards are starting to understand the gravity of this, especially with just the amount of, I think, fear that's surrounding, you know, mythos and comparable models. So from what I've heard, you know, companies are willing to put up that cost. But at the same time, they're trying to be very judicious about how they use these expensive models to make sure they're not spending more than they need to. Right. So, OK, so the cybersecurity companies are saying, hey, we're going to make room in our budget for this. What does this mean then for spending from cybersecurity customers?

9:43I'm thinking about CrowdStrike and Zscaler's customers. I mean, the costs are going to have to get passed on, right? So are prices going to go up or are we going to see enterprises spending a whole lot more on security? What's our best guess there? Yeah, I spoke to customers as well, chief information security officers. And essentially what they are saying is they also expect to spend more. You know, whether that's actually using Mythos or software powered by Mythos, or in some cases, you know, companies have started to just use the models that are already available, like, you know, Anthropics, Claude, Opus models, which are also, you know, not as good as Mythos, but pretty good at finding these vulnerabilities.

10:24And they say that they're budgeting more, sometimes in the millions of dollars per year, just to run those types of models to constantly scan their software and find potential vulnerabilities. So what I'm hearing across the board is that companies see this potential wave of AI threats coming, and they're willing to spend ahead of it to make sure that they're not the victim of an embarrassing hack. Is Mythos going to be available at all to a wider slate of customers now? Initially, when it came out, they said, we're only giving it to the top banks and cybersecurity companies. Is it looking like it's going to have a wider release at all?

11:05Yeah, so Anthropic has said that in the coming weeks, they're going to start releasing models that are comparable to Mythos more broadly. But it's not clear if those models would actually be capable of, you know, doing the type of cyber offensive work that Mythos does. Because Anthropic also is going to put in guardrails that essentially will, you know, make the model say, I'm not going to hack into your system. There are ways for them to do that that would potentially, you know, theoretically say we're only going to hack into your system in a safe environment where you're doing it to your own systems.

11:37But that can be a little bit complicated. And Anthropic has been a little bit tight lipped about what exactly that rollout will look like. So it's not exactly clear how much of these capabilities are going to become broadly available anytime soon. Got it. Okay, I want to go back to Microsoft for just a second. This week is the Build Conference. You're going to be there on the ground in San Francisco. What are the big questions that you're planning to ask all the attendees and executives on site? Yeah, so Microsoft with this build conference, you know, one theme that I have been told by sources that's going to come up is Microsoft trying to essentially pitch its own homegrown AI models as appealing to developers.

12:16and specifically they're releasing their own in-house coding models as well as models for reasoning and things like, you know, transcription and image generation. And, you know, my question for developers is just, you know, what would it take for you to start using Microsoft's models more broadly? Is it about cost? Is it about performance? Is it just having, you know, different options besides the anthropic and open AI sort of duopoly on models? So I'm going to try to get a sense of how appealing that pitch that Microsoft is making is with developer customers. Well, and also I was going to say, piggybacking off our reporting, I think it was last week, I am curious to see how customers might be thinking about the chips that Microsoft could be looking to unveil.

13:02We hadn't thought that Microsoft was going to be a serious competitor in the chip category, but based on our reporting, I mean, they are going for maybe more cheaper, specialized chips that could at least make a dent in things. So it's certainly an interesting time to be there. Aaron, I want to thank you for coming on. That is Aaron Holmes, our Microsoft reporter, here at The Information. Forward deployed engineers have become all the rage as of late, and my colleagues Laura Bratton and Kevin McLaughlin went deep on how the new role is starting to shape Silicon Valley in this weekend's Sunday Insights column.

13:38I want to bring on Laura to talk about what she found. Laura, welcome back to the show. It's great to have you here. Hey, Akash. Okay, forward deployed engineers. Just remind us, is it just a fancy term for consultants? What exactly is the job description nowadays? Historically, consultants and software engineers would work together but do separate roles, and the role of a forward deployed engineer is to do both. They are both a consultant who will sit with the customer and figure out what their need for AI is, which models they should use, what kind of applications they might need, which is traditionally the role of a consultant.

14:17And then they'll actually help build that product or implement that product. Okay. So what made you want to write specifically about it this weekend? Because we've been talking about this for quite a while now. Did something change with the landscape? Yeah. I think it's really just that more companies are talking about the fact that they're hiring forward-deployed engineers in their earnings calls. Workday mentioned it in their latest earnings call, and they started hiring forward-deployed engineers this year and have about 25 on staff now. And then in bigger news, Meta and Google have said that they're going to bring on forward-deployed engineers, and those are obviously the bigger names that we cover.

14:55So I would say it's really just that the biggest of big tech is now copying what was once just Palantir's model of having four deployed engineers. And it's really a big part of their push to scale up enterprise adoption of AI. Well, one of the most interesting parts of your column, though, is, I mean, this is like the role that people want to get at these companies. I mean, it seems like it's highly competitive. It seems like people are getting paid top dollar, you know, second to maybe being a top AI researcher. I mean, the salary bands that you talked about in your story are pretty impressive in some cases.

15:34So what, this role has become like the most popular thing since, you know, wanting to be a product manager at Meta or something like that? Yeah, I think it's still early days. So it's hard to say because there were so few forward deployed engineers just a couple of years ago and the role is growing really rapidly. So LinkedIn told me that their data shows 4 ,000 forward deployed engineers have been added in the U.S. job market as role postings so far this year. That includes from companies like Cursor, OpenAI, Anthropic, Notion, SoftBank. Companies are hiring these roles, and that's much higher than about 3.5 times the amount of new job listings for forward-deployed engineers we saw last year.

16:20So I think it's definitely more popular and more companies are taking on this role. And I think that, you know, it's also higher paid than a typical software engineering role, about 10 to 15 percent higher in terms of base salary. I get the sense that maybe the definition of a forward-to-blood engineer could vary across different companies. Does that align with what you're seeing? Some consultants I spoke to said that the role is technically the same. You really just need a deep technical understanding of AI. but obviously the role is going to look different if you work at Palantir versus if you work at Salesforce or ServiceNow or one of these more traditional SaaS companies.

17:00And the argument can be made that if you work as a four-deployed engineer for a model provider like Anthropic or OpenAI, you might have a better technical understanding of their models because you really have a front row seat to see how they're developing. Right. I want to pivot to a profile that you wrote with our OpenAI and Anthropic reporter, Sri Mupiti, over the weekend. You both profiled Denise Dresser, who is the chief revenue officer at OpenAI. It was our weekend big read, and you both went deep on who it is, what background she's coming into the role with, what it is OpenAI is relying on her for.

17:38Why did you decide to focus on Denise at this moment for OpenAI's business? Well, I think we're really at a point where Anthropics AI tools have taken off with enterprise customers and enterprises, you know, like the rise of forward-to-plate engineers speaks to, enterprises are really beginning to adopt AI in a much more serious way than we saw the same time last year. And what that means is that the model providers are starting to want to capture a larger share of their software budgets as enterprises shift towards using these AI tools. And Anthropik's really been successful at doing that. About 80 % of their revenue comes from business customers.

18:20And OpenAI wants to be that too. They want to change from being a super popular consumer-focused tech company to a company for enterprise customers. So Denise was brought on to help them with that strategy. She has, you know, 14 years of experience at Salesforce. She's pretty much the perfect person for this job because you really want somebody who's been focused on selling SaaS to enterprises. She was the CEO of Slack, right? She was the CEO of Slack, yes. She spent more than a decade at Salesforce basically managing a lot of the relationships with larger profile tech companies. She was credited by a lot of folks I spoke to for, for example, helping build a relationship with AWS and managing those relationships.

19:11And then towards the last couple of years of her tenure, she transitioned after Salesforce acquired Slack to be the CEO of Slack and was really credited with helping successfully integrate Slack into Salesforce because there was sort of a culture clash among those two companies that were acquired. So tell us, what did you learn about her time at Salesforce and at Slack that you think could inform the way that she might approach this role at OpenAI? You just talked about the relationships that she had with customers, big cloud providers like AWS, et cetera. OpenAI is in this moment where, like you said, they've – it's funny because, you know, based on our reporting, actually, The enterprise business in aggregate, and we have to look at the latest numbers for this, but at one point it was still larger than Anthropics enterprise business just based on the enterprise subscriptions.

20:06But we obviously know that Codex, the coding model is where they're trying to make a lot of inroads, getting companies to buy that suite of products. So, I mean, I just wonder, you know, do we know anything about Denise's leadership style, how she approached the role, you know, how she might structure the company to compete in enterprises at OpenAI? Yeah, I mean, I can, Sri knows more about the OpenAI side of things. And my real background for the story was diving into her history at Salesforce and how it set her up for her time at OpenAI. And what I would say, two things. Her colleagues told me over and over that she is super focused on customers and very tough but fair with her colleagues.

20:52And so I think one really stark example was that, you know, it was the end of the fiscal year and she was about to close a deal. It didn't for some reason close by midnight and she was no longer eligible for commission on the deal. But she still kept her colleagues there until like three or four in the morning to close the deal. She also is the type of person to call a team into a meeting on a weekend or at 10 p.m. or text her colleagues at 5 in the morning if a customer is unhappy. So she's really, really focused on customers in a way that I think is ideal for a company like OpenAI, where historically you have more AI researchers that are just focused on building bigger and better models.

21:34Now you have a leader who's focused on customers and how successfully they're able to implement OpenAI's products. One last thing that I want to mention is she also, I thought this was a really great example as well. She pitched Slackbot, which has now become central to Salesforce's strategy, you know, at least a year before it launched. And other Salesforce executives, I'm told, were prioritizing AgentForce, which kind of struggled when it was initially launched. And I think that that was a really big tell of Denise's forward-looking abilities and how she not only can manage relationships with customers, but also really see into the future about what products might be successful for customers.

22:18Great. Well, Laura, I want to thank you for coming on, and I encourage everyone to check out that profile. It is on our website. It is our weekend big read. That is Laura Bratton, our AI reporter and author of our Applied AI newsletter here on TITV. Meta's former chief technology officer, Mike Shrapfer, is now the founder of Gigascale Capital, a new fund focused on the intersection of energy, materials and sustainability. The fund raised$250 million, and I want to bring on Mike to talk more about his playbook for his fund. Mike, welcome to the show. It's great to have you here. Thank you. Glad to be here.

22:55So you founded Gigascale Capital, I think it was a little over two years ago now. You've got$250 million in new funding. Tell us about what your vision for the fund originally was and how this capital is going to help you get there. Yeah, when we started three years ago, our thesis was quite out of the norm, which was that the physical world, how we build our economy, was the most interesting place to build venture-scale companies. And that was not the common wisdom. I would say the fun thing about where we are now is it's working. I don't have to convince anyone. All of our companies are in a place where the question isn't, do people want what they're building?

23:31It's how fast can they get it there? We're talking clean power for data centers. We're talking clean materials for anything you want to build. We're talking, you know, compute in faster, better, cheaper for inference and for training. This is just like a tremendous amount of demand has been pulled forward to build all of these things that are kind of better, faster, cheaper than the way we've been doing it for the last hundred years. So of all these segments and of all these really specific technology, because energy is so big. I mean, sustainability is huge and there are problems all across the category that need to be solved.

24:10I mean, tell us, what is like the one innovation that, you know, just last seven days alone that you have just been so obsessed with that, you know, you could see yourself really making a ton of investments in this one specific category? Well, I'll just, I'll give you a specific company and a specific product, because I think that's even better. There's a company called Heron Power, and they've taken this, you know, literally a hundred year old technology, the transformer, and I'm talking about a power electric transformer, not the transformer in an AI model. We have to be specific. But it is literally the piece of equipment that is bottlenecking data center AI build out right now.

24:45It's, you know, four to five year back order. It relies on this very special kind of bespoke steel you've got to make, all the rest of it. And Heron Power, Drew Baglino said, hey, we've been for the last 10 years developing a new technology for this, power electronics. It's in every electric vehicle that's shipped. So we've done this at scale. Let's like take that technology in the EV and put it on the grid. I can build a box that's a third the size. It's software programmable. So instead of deciding when I build my piece of hardware exactly what the specs have to be, I drop it on a pad and I can use it for a data center.

25:16I can use it for a solar farm. I can use it for an EV car charger. It works in any of these conditions. So it's sort of a single Lego block that works for everything. And so this is a company that will be shipping commercial products next year. They were founded last year. And another one of these, they have more demand than they know how to deal with. And so it's an example of new technology solving an old problem. What would you say is a contrarian belief of yours that is informing the way you plan to invest at GigaScale Capital? And I ask this because, again, energy, it is an obvious problem.

25:53Everyone knows it needs to be solved. Any money going into space is good. But what is the contrarian view that you hold? Well, I think one is I bet you we'll have more compute in the ocean than we do in space by 2030. say more why oh we have this massive 10 terawatt untapped resource of wave power it is cheaper to build a ship than it is to launch a satellite uh so on first principles we should be doing a lot more compute you know in the open ocean than we should be in space and you have investments in the ocean yeah there's a company called pantalossa again early bet we made it many years ago uh contrarian bet but now peter teal has just come in and and gave them a bunch of money so they now have the sort of resources to go ahead and build this.

26:34They've had units, you know, in the ocean off the coast of Oregon working, and now they're going towards full-scale deployment and development. So they're going to be running a few payloads, hopefully next year or year after. So it sounds like you're not that bullish then on data centers in space. I think Elon Musk is going to do a great job of it. I think it's a trap for everyone else. I mean, he's got wholesale costs for satellite costs and for launch costs. It's going to be really hard to beat that. So I'm sure, I mean, and Starlink, you know, V3 is probably 20 kilowatts. So he's sort of already there.