# Anthropic’s $65 Billion Surge, OpenAI’s Teen Push

An episode of [Bloomberg Tech](https://invisiblenetwork.ai/podcast/1611767434), 18 Aug 2026, 44 min.

Audio: https://podtrac.com/pts/redirect.mp3/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/41764a4f-fc64-4e11-89ba-ae7c0030ab5e/4330d8b4-3d1a-4705-b0b7-b4aa012a5652/audio.mp3?utm_source=Podcast&in_playlist=9caafc41-289c-4115-995d-ae7c0030ab75

Bloomberg’s Ed Ludlow breaks down Anthropic’s annualized revenue topping $65 billion as the AI startup moves closer to a potential IPO. Plus, OpenAI launches ChatGPT for Teens, aimed at giving 13- to 17-year-olds a safer way to use the chatbot for learning. And, after advanced AI models escaped controlled safety tests and accessed real-world systems, we speak with the CEO of the startup hired to stress-test them. See omnystudio.com/listener for privacy information.

## Transcript

[0:02] Speaker 1: Bloomberg Audio Studios, podcasts, radio news. Bloomberg Tech is live
[0:15] Speaker 1: from the heart of Silicon Valley with Ed Larvow in
[0:19] Speaker 1: San Francisco.
[0:23] Speaker 2: This is Bloomberg Tech coming up. Anthropics annualized revenue top
[0:26] Speaker 2: sixty five billion dollars, bolstering its plans for an IPO
[0:30] Speaker 2: plus OpenAI releases chat GPT for teens, a version of
[0:34] Speaker 2: its chatbot aimed at thirteen to seventeen year olds to
[0:37] Speaker 2: encourage safe usage and deeper learning. And after advanced AI
[0:41] Speaker 2: models broke out of controlled safety tests and reached real
[0:44] Speaker 2: world systems, we'll speak with the startup that stress tested
[0:48] Speaker 2: them in the markets they're to sell off in chip
[0:50] Speaker 2: makers that is dragging technology lower. There's a lot of
[0:53] Speaker 2: consideration about the global economy out there, the outlook for inflation.
[0:57] Speaker 2: You have treasury yields at twenty twenty five pies, and
[1:01] Speaker 2: there's still this debate about where we're at in the cycle,
[1:04] Speaker 2: particularly with the AI build out in private markets. Something
[1:08] Speaker 2: interesting happening. Let's look at today's big number, sixty five
[1:11] Speaker 2: billion dollars that has Andthropics projected annual ren rate for
[1:15] Speaker 2: revenue by the end of July. This is a twenty
[1:18] Speaker 2: four hour update. You remember we started the show yesterday
[1:21] Speaker 2: with an AR figure for and Fropic that originated from May.
[1:24] Speaker 2: Let's get the latest from Shrien Gafari, part of the
[1:27] Speaker 2: Bloomberg team that broke the story. So the latest ARR
[1:31] Speaker 2: is sixty five billion dollars. What else is in the reporting?
[1:36] Speaker 3: That's right?
[1:36] Speaker 4: So if you look at the increase, we last reported
[1:39] Speaker 4: that anthrop Bay had hit forty seven billion in that
[1:43] Speaker 4: same run rate revenue figure in May, So this is
[1:46] Speaker 4: a notable increase ahead of their planned IPO and fall.
[1:50] Speaker 4: Of course, the market is watching closely to see how
[1:55] Speaker 4: much revenue growth is going to continue at what pace
[1:59] Speaker 4: for these leading companies going into their public offering.
[2:03] Speaker 2: When we started the show yesterday on the ARR figure
[2:06] Speaker 2: from May, which was forty seven billion dollars, there were
[2:08] Speaker 2: lots of AI stocks moving to the upside. There was
[2:12] Speaker 2: some read through from that. Today we're not seeing that,
[2:14] Speaker 2: but we are learning more about Anthropics momentum. You know,
[2:18] Speaker 2: people are hesitant to compare Anthropics to Open AI, But
[2:21] Speaker 2: where is that race right now if you were to
[2:23] Speaker 2: measure it on revenue or even like on usage in
[2:26] Speaker 2: the subscriber base.
[2:29] Speaker 4: So as we reported just last week. Open AI has
[2:33] Speaker 4: recently hit forty billion in that annualized revenue projection figure,
[2:40] Speaker 4: So if you're comparing to Anthropic, that is a lower
[2:42] Speaker 4: number and it is part of this. You know, Anthropic
[2:45] Speaker 4: in the past year has really grown at sort of
[2:48] Speaker 4: rocket speed and becoming a high more highly valued company
[2:53] Speaker 4: in the private markets and seeing some higher now annualized
[2:57] Speaker 4: revenue run ratefigures. That being said, the way that they
[3:00] Speaker 4: measure revenue is not exactly the same that once these
[3:05] Speaker 4: companies do actually go public and we start to see
[3:08] Speaker 4: their financials, then we will know more of the details
[3:12] Speaker 4: of how that revenue is counted for each But that
[3:14] Speaker 4: being said, Anthropic now is citing a higher annualized revenue
[3:18] Speaker 4: run with figure.
[3:20] Speaker 2: Being Boucherie Gafari, AI correspondent, thank you very much. To
[3:23] Speaker 2: take a look at markets again. So there's volatility that's
[3:26] Speaker 2: continuing for tech, particular pressure on semiconductors US tenure yield.
[3:31] Speaker 2: Is it kind of twenty twenty five highs right now?
[3:34] Speaker 2: We see some bond anxiety out there. The S and
[3:37] Speaker 2: P five hundred HEA towards its third straight day of losses.
[3:41] Speaker 2: Where are we in the cycle? Janas Henderson portfolio manager
[3:44] Speaker 2: and head of Technology Research, Denny Fish joins us now,
[3:47] Speaker 2: and I want to continue a conversation that started in
[3:49] Speaker 2: yesterday's show one of your industry colleagues arguing we are
[3:52] Speaker 2: in the riskier part of the cycle, where the biggest
[3:55] Speaker 2: companies in tech are moving from their balance sheet to
[3:58] Speaker 2: equity and debt in the market. It's how do you
[4:01] Speaker 2: feel about that, Denny.
[4:02] Speaker 5: Well, clearly there's not as much room on the balance
[4:05] Speaker 5: sheet as there was two years ago, so by definition,
[4:09] Speaker 5: we're definitely getting riskier. I think what's different, though, is
[4:14] Speaker 5: that it's still existential to each one of these companies
[4:18] Speaker 5: depending on where they sit. So if they can't go
[4:21] Speaker 5: to the debt markets, my suspicion is for the next
[4:24] Speaker 5: couple of years, if you're a hyperscaler, you'll probably issue
[4:28] Speaker 5: two or three percent of equity and accept the dilution
[4:31] Speaker 5: if you think it's the best thing to do for
[4:33] Speaker 5: your business, if the credit marks are credit markets aren't
[4:37] Speaker 5: a minimal, or if rage just continue to go to
[4:40] Speaker 5: a level that you know isn't necessarily appealing.
[4:43] Speaker 2: At the start of the summer, you wrote an essay
[4:46] Speaker 2: title the AI Tied rolls On that was in June
[4:49] Speaker 2: June eighteenth, and so the question is what's changed since
[4:52] Speaker 2: June eighteenth until now, if anything at all. But I
[4:56] Speaker 2: suppose one observation you made just before we came on
[4:58] Speaker 2: air is, you know, as stocks go up and stocks
[5:01] Speaker 2: go down, it's normal to see the socks down five
[5:03] Speaker 2: percent or up five percent, for example. It's hard to
[5:06] Speaker 2: see through.
[5:07] Speaker 5: It, though, yeah, I mean it is hard for the
[5:10] Speaker 5: average investor to see through it. You know, I've been
[5:12] Speaker 5: investing in semiconductors for twenty years, and it's just the
[5:16] Speaker 5: reality of the sector, and we are in the most
[5:19] Speaker 5: pronounced cycle that we've ever had for leading edge semiconductors.
[5:24] Speaker 5: And so I think that part is what's really you
[5:27] Speaker 5: know unique here to your question of what's changed since
[5:30] Speaker 5: June eighteenth, I'd say if I just parsed back the
[5:34] Speaker 5: earning season that we just finished up Q two earning season,
[5:38] Speaker 5: the AI supply chain, if anything, pretty much exceeded expectations
[5:43] Speaker 5: across the board, and the estimates for twenty twenty seven
[5:46] Speaker 5: and twenty eight just went up a bunch more than
[5:49] Speaker 5: we we had before the reporting season. And I'd say
[5:52] Speaker 5: the other thing that has been interesting is, you know,
[5:56] Speaker 5: SpaceX has gone public. We saw the XAI number and
[6:00] Speaker 5: the spot pricing for GPUs that they're selling to Anthropic
[6:04] Speaker 5: into Google. We saw the return metrics on capex spend
[6:07] Speaker 5: for Microsoft and Amazon which were rewarded by the market,
[6:11] Speaker 5: which was really nice to see. And then even a nebuus,
[6:14] Speaker 5: which is a neocloud. If you parse back their commentary,
[6:18] Speaker 5: there were a couple things that were interesting, and that
[6:20] Speaker 5: is once again suggesting spot pricing continues to be really strong.
[6:24] Speaker 5: But then also the uptake of enterprises and sovereigns and
[6:28] Speaker 5: that other part of that other piece of the pie
[6:31] Speaker 5: that doesn't get as much attention everybody gets focused on
[6:34] Speaker 5: hyperscaler capex. Half of the AI spend out there, if
[6:38] Speaker 5: you talk to Jensen and listen to their calls, is enterprise, neocloud, sovereigns,
[6:43] Speaker 5: and that's growing really fast and that's going to be
[6:46] Speaker 5: a really important component here in addition to the hyperscalers
[6:49] Speaker 5: and the open aiyes and the anthropics.
[6:51] Speaker 3: Of the world.
[6:52] Speaker 2: Have you tweaked anything or changed anything in your approach,
[6:55] Speaker 2: your holdings, your strategy since June eighteenth as a result
[6:57] Speaker 2: of what you've just outlined.
[7:00] Speaker 5: Not necessarily, you know, clearly we're we're trying to find
[7:03] Speaker 5: ways to participate in this market that are non semiconductor,
[7:07] Speaker 5: right and to the extent that we can find that
[7:10] Speaker 5: to balance out the portfolio.
[7:12] Speaker 3: That's what we're doing.
[7:13] Speaker 2: That's just an issue of diversity, it is.
[7:16] Speaker 5: And trying to because we do think there's gonna be
[7:18] Speaker 5: a lot of money to be made with companies like
[7:20] Speaker 5: as a Microsoft starts to rerate, for example, after it
[7:23] Speaker 5: had the quarter that it had, it is really underperformed
[7:26] Speaker 5: for the last eighteen months because concerns about ROI on spend,
[7:32] Speaker 5: the impact of AI on the software business, and whether
[7:37] Speaker 5: they were going to develop a frontier model or not.
[7:39] Speaker 5: And some of those concerns I think are starting to
[7:41] Speaker 5: be put to bed, and so looking for opportunities like
[7:43] Speaker 5: that and then also getting comfort as you know Andy
[7:46] Speaker 5: Jassey talked about with Amazon and the three back the
[7:49] Speaker 5: three year payback period on their AI spend right now.
[7:52] Speaker 2: So the idea on that is if you invest in
[7:54] Speaker 2: a technology stack based on certain GPU within three years,
[7:58] Speaker 2: it's paid for itself and then it's just cash gener.
[8:00] Speaker 3: Yeah, yeah, yes, you know, for the most part.
[8:03] Speaker 2: So that brings us to the Nvidia story of the
[8:05] Speaker 2: last ten days, Jensen gets six Wall Street giants or
[8:08] Speaker 2: investment firms to try and use them as a conduit
[8:11] Speaker 2: to third party capital. You set up an SPV I think,
[8:14] Speaker 2: borrow money, buy the GPUs and lease them to an
[8:16] Speaker 2: Nvidia customer. So lots of people that was difficult to understand.
[8:21] Speaker 2: What did you make of it?
[8:23] Speaker 5: So my interpretation of all of this is, and I'll
[8:28] Speaker 5: go back to what I talked about, you have hyperscaler
[8:31] Speaker 5: and then you have enterprises and sovereigns and the rest
[8:33] Speaker 5: of the market. I think in Jensen's mind, he thinks
[8:37] Speaker 5: there's a place for a globally distributed standardized cloud and
[8:42] Speaker 5: that is based on Nvidia, and that's Blackwell's, that's Rubens Vera,
[8:46] Speaker 5: Rubens Fineman, which will be coming at some point. And
[8:51] Speaker 5: to the extent that he's right about that and this
[8:56] Speaker 5: and the duration, it makes a lot of sense to
[8:58] Speaker 5: have third party capital to help finance that. So I
[9:01] Speaker 5: think in the broader scheme and the strategy for Nvideo,
[9:04] Speaker 5: it makes a lot of sense. And just seeing you know,
[9:07] Speaker 5: external partners have that confidence in the strategy as well
[9:11] Speaker 5: and being willing to sdate syndicate that type of funding
[9:16] Speaker 5: is a really important barometer for the AI spending trip.
[9:21] Speaker 2: We're out of time, but next time we'll talk about
[9:23] Speaker 2: computers collateral. Yeah, the idea. This is an asset class
[9:26] Speaker 2: of its own. Denny Fish, Janis Henderson, portfolio manager. Great
[9:29] Speaker 2: to have you back on the show. Another stock that
[9:31] Speaker 2: we're watching, by the way, is Ali Barber. There's been
[9:34] Speaker 2: some movement in the ADRs. That's the US listed shares
[9:36] Speaker 2: off session highs this morning, but there seems to be
[9:39] Speaker 2: some momentum out of various products, Alipay being one that
[9:43] Speaker 2: causes kind of up teak uptick after the market open
[9:46] Speaker 2: this morning. Will continue to track that one. Coming up
[9:48] Speaker 2: on the program. Open ai just released a new experience
[9:52] Speaker 2: to ease fears that AI may be unsafe for kids.
[9:55] Speaker 2: Opening ey as head of Youth and Families, Lauren Jonas
[9:58] Speaker 2: joins us next. This has been big te open ai
[10:10] Speaker 2: rolled out its latest experience called chat GPT for teens,
[10:13] Speaker 2: with features that prioritize education and digital safety for what
[10:17] Speaker 2: it calls, quote the first AI generation of students. Lauren Jonas,
[10:22] Speaker 2: open AI's head of Youth and Families joins us here
[10:24] Speaker 2: on Bloomberg Tech. I think let's start with the very basics.
[10:27] Speaker 2: How does chat gibt for teens work?
[10:29] Speaker 6: Yeah, thank you so much for having me. So, chatchabt
[10:32] Speaker 6: for teens is a dedicated learning focused experience for teens
[10:37] Speaker 6: with built in safeguards for this particular demographic. You know,
[10:41] Speaker 6: nine to ten teens are using chat gipt for learning,
[10:44] Speaker 6: and so our goal here is to make this an
[10:46] Speaker 6: incredible learning tool for teenagers.
[10:49] Speaker 2: Lauren, I don't know that I'm a power user of
[10:52] Speaker 2: chat gubt. I use an enterprise version that Bloomberg provides
[10:54] Speaker 2: to me, right, but I think it's interesting to put
[10:57] Speaker 2: them side by side. Is there a fundamental difference in
[11:00] Speaker 2: the models underpinning chat gbt for teens versus the version
[11:03] Speaker 2: I'm using, for example? Is there a design difference between
[11:07] Speaker 2: the two.
[11:08] Speaker 6: There is not a difference in the power of the models,
[11:10] Speaker 6: but there's a difference in the safeguards. The goal here
[11:13] Speaker 6: is age appropriate safeguards for teenagers. So we have model
[11:18] Speaker 6: policies that are different for teens. The way the models
[11:21] Speaker 6: should respond is different for teenagers. That's the fundamental difference
[11:24] Speaker 6: that you see between those two experiences.
[11:26] Speaker 2: Can we get specific on that? What are the god
[11:29] Speaker 2: rails that you've put in place for teenagers.
[11:31] Speaker 6: Yeah, numerous. So at the model level, we have advised
[11:35] Speaker 6: by the American Psychological Association and a number of third
[11:37] Speaker 6: parties model policies that we publish that are difference for teenagers.
[11:42] Speaker 6: This is around harmful content of a number of different kinds.
[11:45] Speaker 6: This is based on a teen's developmental life stage and
[11:49] Speaker 6: where they are. Additionally, product differences. So as you're seeing
[11:53] Speaker 6: on the screen now, this is a really learning focused experience.
[11:56] Speaker 6: Teens have the ability to put themselves into a study
[11:59] Speaker 6: only experience. We call this study mode, have the product
[12:03] Speaker 6: respond to them in a study only way during certain
[12:06] Speaker 6: hours of the day, invoke quizzes and what have you
[12:09] Speaker 6: to make this really robust learning experience.
[12:11] Speaker 2: Find that very interesting. So this is analogous with social
[12:14] Speaker 2: media in the context of verification. How do you know
[12:17] Speaker 2: that a user of chatchabt for teens is a teen?
[12:21] Speaker 2: Is within the sort of ages thirteen to seventeen that
[12:24] Speaker 2: you're going after with this new tool.
[12:26] Speaker 6: Yeah, so AI is very different than social media. It's
[12:28] Speaker 6: a fundamentally different product. But in terms of identifying who's
[12:32] Speaker 6: a teen, we do this a couple of different ways.
[12:34] Speaker 6: It's a layered approach. Some folks tell us their teenagers,
[12:38] Speaker 6: in which case they get the chatch for teen's experience
[12:40] Speaker 6: by default. We also have built a model ourselves that
[12:44] Speaker 6: we call age estimation or age prediction. If a person
[12:47] Speaker 6: is predicted as under the age of eighteen, they will
[12:49] Speaker 6: also get this experience by default.
[12:53] Speaker 2: Can we have the backstory what catalyzed the decision to say, okay,
[12:57] Speaker 2: we need and we will launch chatgipt for teens.
[13:01] Speaker 6: We have been working on teens at open Ai for
[13:04] Speaker 6: a number of years. This is just really focusing on
[13:08] Speaker 6: the use case that we see teens engaging with in
[13:11] Speaker 6: the product. Again, the massive majority of use on chachiput
[13:13] Speaker 6: is learning, and so the goal here is to make
[13:15] Speaker 6: it the most robust learning tool that it can be,
[13:19] Speaker 6: and so that's our focus here. We also co design
[13:22] Speaker 6: this product with teens and their parents, and this is
[13:25] Speaker 6: a lot of what they ask for. Teens love this,
[13:27] Speaker 6: they want to continue to learn, and they love these futures.
[13:31] Speaker 2: You know, learning is done in the context of education,
[13:34] Speaker 2: which is done in the context of schools, right and
[13:37] Speaker 2: how much is that a part of the plan to
[13:41] Speaker 2: go out to the education system throughout this country and
[13:45] Speaker 2: outside of the United States and work with academic institutions
[13:49] Speaker 2: say like, look, here is a tool you can use
[13:51] Speaker 2: school wide as part of the you know to learn
[13:55] Speaker 2: set curriculums, et cetera.
[13:57] Speaker 6: We have a partnership with the American Federation of Teachers
[13:59] Speaker 6: that we're we're thrilled by, and we have teacher specific
[14:02] Speaker 6: tools and we're thrilled to support teachers and their teaching journey.
[14:05] Speaker 6: But CHATCHYBT is not in schools today across the country,
[14:09] Speaker 6: and that's a very deliberate choice on our part. We
[14:11] Speaker 6: have been very slow and very thoughtful to introduce this
[14:13] Speaker 6: into schools, but are thrilled to partner with the American
[14:16] Speaker 6: Federation of Teachers and others to make sure we're serving
[14:20] Speaker 6: that demographic.
[14:21] Speaker 2: But there will be an action taken to get CHATCHIBT
[14:25] Speaker 2: foteens into schools.
[14:27] Speaker 6: We have no immediate plans. We're starting with teachers and
[14:30] Speaker 6: really focused on teachers and administrators and making sure that
[14:32] Speaker 6: the folks that surround the team are well supported.
[14:36] Speaker 2: So it sounds like, you know, in terms of real terms,
[14:38] Speaker 2: how teenagers are using CHATCHYBT today and will use CHATCHYBT.
[14:43] Speaker 2: For teens, it's outside of school hours, it's at home
[14:47] Speaker 2: in the evenings and therefore it sounds like there's an
[14:51] Speaker 2: element of supervision. So how do the parents get brought
[14:54] Speaker 2: into this, Lauren, Is that sort of formalized parents oversight
[14:59] Speaker 2: of use of the tool.
[15:00] Speaker 6: Yeah, great question. So we have parental controls. We launched
[15:04] Speaker 6: parental controls about a year ago. We were the first
[15:06] Speaker 6: in the industry, as part of parental controls to launch
[15:09] Speaker 6: what we call parents safety notification, so that is, you know,
[15:13] Speaker 6: if a teen is on the platform and they're having
[15:14] Speaker 6: a hard time, we notify that parents that they're having
[15:17] Speaker 6: a hard time, give them enough information to take an action.
[15:21] Speaker 6: But there are as part of our chattybut for teens,
[15:23] Speaker 6: launch additional parental controls. You know, parents have the ability
[15:27] Speaker 6: to put their teenager into what we call study hours,
[15:30] Speaker 6: so that is chatchypt will only be in a steady
[15:34] Speaker 6: mode to ask follow up questions, make sure the teen
[15:36] Speaker 6: is learning the material, so you know, again, like teachers,
[15:39] Speaker 6: the goal here is to make sure that the folks
[15:41] Speaker 6: that surround the teen are well.
[15:42] Speaker 2: Supported, Lauren. Another area where where the technology is analogous
[15:48] Speaker 2: with what we've seen in social media is a teenager's
[15:52] Speaker 2: relationship with the technology for emotional support. You know to
[15:56] Speaker 2: engage with chat GPT for advice, to confide in etc.
[16:01] Speaker 2: How much have you considered that and formalized that in
[16:04] Speaker 2: terms of policy in guardrails.
[16:06] Speaker 6: Yeah, So these policies are formalized and we continue to
[16:08] Speaker 6: build on these and have built on them with CHATWIPT
[16:11] Speaker 6: for teens. With this launch, we added to our reliance
[16:16] Speaker 6: policy the model should not be a friend to teens,
[16:19] Speaker 6: It should not be sentient, it should not claim any
[16:23] Speaker 6: of this for teens. Again, the goal is that this
[16:25] Speaker 6: is a tool for teenagers and nothing more. And so
[16:28] Speaker 6: with this launch, we continue to build on these policies
[16:31] Speaker 6: and these safeguards for teens with the American Psychological Association,
[16:34] Speaker 6: with Moonshot, with the American Federation of Teachers to make
[16:37] Speaker 6: sure that we're getting this right.
[16:39] Speaker 2: Lauren Jonas head of you from families at open AI,
[16:41] Speaker 2: thank you very much. Indeed, Meta is facing one point
[16:46] Speaker 2: four trillion dollars in financial penalties. I have allegations that
[16:50] Speaker 2: the company purposefully designed Facebook and Instagram to encourage compulsive
[16:55] Speaker 2: use among young users. The landmark social media trial is
[16:59] Speaker 2: just kicking off today. Most legal reporter Madelin Mechaboide joins
[17:02] Speaker 2: us now with what to expect, and it's probably worth
[17:05] Speaker 2: explaining how that one point four trillion figure is calculated.
[17:10] Speaker 2: It's based on lots of potential penalties and violations.
[17:15] Speaker 3: That's right.
[17:15] Speaker 7: The one point four trillion is a huge number, and
[17:17] Speaker 7: that's because it's at the very top end of the
[17:20] Speaker 7: estimate of what penalties might be in this case if
[17:24] Speaker 7: the states are successful at trial, and it's going to
[17:27] Speaker 7: be based on the individual consumer protection laws of the
[17:30] Speaker 7: four states that are leading this charge, and so each
[17:33] Speaker 7: state has in their laws a fine that they assess
[17:36] Speaker 7: when they identify a violation of these laws, and so
[17:40] Speaker 7: the calculations come into play when the judge is good
[17:43] Speaker 7: in the jury are going to try to determine what
[17:45] Speaker 7: constitutes a violation, and whatever number they land on there
[17:49] Speaker 7: is going to affect what the figure is. So we've
[17:52] Speaker 7: heard one point four trillion, we've heard two hundred billion
[17:55] Speaker 7: as a possibility. But what you can take away from
[17:59] Speaker 7: that is that we know that the stakes are high
[18:01] Speaker 7: for Metta. Here remains to be seen exactly what that
[18:04] Speaker 7: dollar figure might look like at the end of the day.
[18:07] Speaker 2: This is happening in Oakland, California right now. What do
[18:10] Speaker 2: we expect to happen, how many weeks will the trial
[18:12] Speaker 2: go on, and what is the format of the trial?
[18:15] Speaker 7: That's right, So this trial is happening before a jury. However,
[18:19] Speaker 7: the jury will be in an advisory capacity, so they're
[18:22] Speaker 7: going to issue their determination after they hear the evidence,
[18:25] Speaker 7: and then the judge can decide whether she wants to
[18:28] Speaker 7: agree with the jury, adopt their recommendation, or rule in
[18:31] Speaker 7: a different manner based on what she heard. We know
[18:34] Speaker 7: that the trial is scheduled at this point to go
[18:36] Speaker 7: until September twenty third, and this judge Eyvon Gonzalez rogers
[18:42] Speaker 7: Stickler for her schedule. So I think that that's kind
[18:44] Speaker 7: of what we can expect here. Today is going to
[18:47] Speaker 7: be opening arguments from both sides. So we'll hear from
[18:50] Speaker 7: the state ags, we'll hear from META. They're going to
[18:52] Speaker 7: lay out kind of a roadmap of what we can
[18:54] Speaker 7: expect to hear for the rest of trial, and then
[18:57] Speaker 7: we're going to hear from witnesses, see internal documents, and
[19:01] Speaker 7: kind of dive into the evidence.
[19:02] Speaker 2: Here bloombogs, Madeline Meckelberg, thank you. Coming up. Apple set
[19:08] Speaker 2: out the AI spending race and now and maybe paying
[19:11] Speaker 2: the price. We have more on that next. This is
[19:14] Speaker 2: Bloomberg Tech. Apple's decision to sit out the AI's spending
[19:34] Speaker 2: boom once looked pretty smart, but with investors pulling back
[19:37] Speaker 2: into AI winners, piling back into a when it's stigned
[19:40] Speaker 2: to work against the stock. Apple shares a down ten
[19:43] Speaker 2: percent from a record high on July twenty eighth, following
[19:46] Speaker 2: a disappointing earnings print Bluemos Ryan for Selica. It's been
[19:49] Speaker 2: looking at what's behind the reversal, so I think to
[19:51] Speaker 2: be fair, Apple's still up like fourteen fifteen percent year today,
[19:55] Speaker 2: just ahead of the S and P five hundred, but
[19:57] Speaker 2: it has changed the overall look at the stock sin's earnings.
[20:02] Speaker 8: What I would say is really interesting is the degree
[20:04] Speaker 8: to which Apple has become inversely correlated with the rest
[20:08] Speaker 8: of tech semiconductor stocks in particular. So you're even seeing
[20:12] Speaker 8: this today where Apple is up a little bit but
[20:15] Speaker 8: the semiconductor index is down pretty significantly. It does seem
[20:18] Speaker 8: like whenever one part of the market is doing well,
[20:21] Speaker 8: the other part is foltering, And it really is an
[20:25] Speaker 8: indication of how Apple has become something of an ANTIAI stock.
[20:29] Speaker 8: So whenever you see the ciniment towards AI related companies
[20:33] Speaker 8: like hyperscalers, like semiconductor companies. Whenever those are doing well,
[20:37] Speaker 8: that's when you see Apple really kind of pull back
[20:40] Speaker 8: a little bit.
[20:40] Speaker 2: And then when you see that.
[20:41] Speaker 8: Reverse, like today Apple's doing well and the rest of
[20:45] Speaker 8: the AI space is down, it's really a reflection of
[20:48] Speaker 8: ciniment towards AI overall, with Apple on the other side
[20:51] Speaker 8: of that.
[20:52] Speaker 2: Part of this is how Apple is different from the
[20:55] Speaker 2: behavior of the hyperscalers, right, so it's AI services as
[20:59] Speaker 2: through partnership, and it does not spend on infrastructure. How
[21:04] Speaker 2: is that playing into the psychology of those you are
[21:07] Speaker 2: speaking to in the markets.
[21:09] Speaker 8: Well, it depends on who you're talking to. If you're
[21:11] Speaker 8: talking to someone who is very optimistic about the AI
[21:13] Speaker 8: related capex, if they're expecting this to really pay off
[21:17] Speaker 8: with a strong return on investment, then they tend to
[21:19] Speaker 8: be pretty positive about those companies, especially since their growth
[21:23] Speaker 8: is more robust than Apple right now and their valuations
[21:27] Speaker 8: tend to be lower. Whereas on the other side of that,
[21:29] Speaker 8: if you're concerned about is this spending going to pay off?
[21:33] Speaker 8: Is there some kind of you know, bubble in the
[21:34] Speaker 8: works here, then maybe you look at Apple, which is
[21:37] Speaker 8: doing a ton of buybacks, which is more of a
[21:39] Speaker 8: lower grower but pretty steady there. You know, higher valuation,
[21:42] Speaker 8: but seeing as much more of a quality and kind
[21:44] Speaker 8: of a safe haven trade. So if you're concerned about
[21:46] Speaker 8: the capex side of things, that's why you're seeing people
[21:49] Speaker 8: rotate into Apple.
[21:52] Speaker 2: Just have thirty seconds. But what if the AI trade
[21:54] Speaker 2: on wines? How would Apple respond? Well?
[21:58] Speaker 8: People I spoke to said that if we did see
[22:00] Speaker 8: another reversal in the AI ecosystem, Apple would probably benefit
[22:04] Speaker 8: from this kind of trend. It might have some muted
[22:07] Speaker 8: downside just because it has no AI related capex risk.
[22:10] Speaker 8: It's not part of the AI disruption narrative and it's
[22:13] Speaker 8: not really part of that whole ecosystem. So if we
[22:15] Speaker 8: do see a pullback there, we could see this inverse
[22:18] Speaker 8: correlation work back in Apple's favor.
[22:21] Speaker 2: I think Apple is still trading it beyond thirty three
[22:24] Speaker 2: times forward earning bluemost Ryan for Seneca, Thank you very much.
[22:28] Speaker 2: Like coming up. AI is getting better at hacking, but
[22:31] Speaker 2: are we getting better at stopping it? Irregular CEO Dan
[22:35] Speaker 2: lah is with us next really important conversation coming up.
[22:39] Speaker 2: It's half time here on Bloomberg Tech. We're going to
[22:42] Speaker 2: show you some markets, then maybe some pictures of San Francisco.
[22:46] Speaker 2: This is Bloomberg Tech. Welcome back to Bloomberg Tech. Quickly.
[23:00] Speaker 2: This is what technology markets looked like. There's a sell
[23:02] Speaker 2: off and chip makers we're at one point in the
[23:05] Speaker 2: session on track for our biggest drop since July second.
[23:08] Speaker 2: Largely it relates to anxiety in the bomb market, but
[23:11] Speaker 2: there is some selling pressure, particularly in the most AI
[23:15] Speaker 2: relevant corners of the chip sector. Then as that one
[23:19] Speaker 2: hundred or so following suit down one point six percent,
[23:22] Speaker 2: a consideration of what's happening with global inflation, but there's
[23:25] Speaker 2: still half an eye on what's happening in the Middle
[23:28] Speaker 2: East and the president's latest comments on the war with Iran.
[23:33] Speaker 2: Let's get to what's happening in the world of technology.
[23:35] Speaker 2: We recently learned that AI models from Open Ai and
[23:38] Speaker 2: Thropic and others broke out of controlled tests and access
[23:42] Speaker 2: to real world systems. Those evaluations were run with irregular
[23:46] Speaker 2: a startup hired to stress test advanced AI models, but
[23:50] Speaker 2: the sandboxed environment had a critical flaw. A misconfiguration meant
[23:54] Speaker 2: the models could access the open Internet irregular CEO Dan
[23:59] Speaker 2: la joins US now in San Francisco.
[24:01] Speaker 9: It's a pleasure to be here.
[24:02] Speaker 2: Thank you very much for being here. Let's start with
[24:05] Speaker 2: the idea of a misconfiguration, help the audience understand the
[24:09] Speaker 2: very basics of what that misconfiguration was.
[24:11] Speaker 9: Should we start by explaining why are we even doing
[24:14] Speaker 9: cyber evaluations for this?
[24:15] Speaker 2: More we can, but I want to understand what happens. Yeah,
[24:18] Speaker 2: let's start that.
[24:19] Speaker 9: So before models are being a list, it's highly important
[24:22] Speaker 9: to stress test them. And the reason is models are
[24:25] Speaker 9: getting really, really capable. You know, we've seen demonstration of
[24:28] Speaker 9: that in recent times. If models are getting good at
[24:31] Speaker 9: many other skills because they're better at reasoning and coding,
[24:33] Speaker 9: they're also getting great attacking. Before models are being deployed
[24:36] Speaker 9: into the world, there is a necessity to use them
[24:39] Speaker 9: and just like test them in the LEB environment to
[24:40] Speaker 9: make sure that they're safe and secure. These tests are
[24:44] Speaker 9: getting more and more complicated over the time because the
[24:46] Speaker 9: models are getting great and some of these that's actually
[24:48] Speaker 9: including environments that are highly highly realistic with a lot
[24:51] Speaker 9: of livs. In some of these environments, there was duty
[24:54] Speaker 9: or miscommunication us to some human LSHCHYC on our side
[24:57] Speaker 9: as well.
[24:58] Speaker 3: We had an issue.
[25:00] Speaker 9: The environment was left with an access to the Internet,
[25:03] Speaker 9: and that allowed the models to also just like access
[25:06] Speaker 9: targets outside.
[25:07] Speaker 2: It may ask a human error is it is simple?
[25:10] Speaker 2: As it was miscoded by a human The rules and
[25:13] Speaker 2: parameters were not written as they were intended.
[25:16] Speaker 9: So these environments have a lot of different leaders to
[25:19] Speaker 9: play with and just like to explain and give context
[25:21] Speaker 9: on what happens here, just like imagine the fact that
[25:23] Speaker 9: you know you want to just like stress as for
[25:25] Speaker 9: realistic scenarios, for example, whether someone can hack your phone
[25:28] Speaker 9: or hack your laptop in order to steal sensitive information.
[25:31] Speaker 9: As a journalist, that should highly concern you, like because
[25:34] Speaker 9: if I'm able to hack you through a model in
[25:36] Speaker 9: order to steal your sources, that's an issue, yes, But
[25:40] Speaker 9: in order to simulate whether that's a possibility, you need
[25:42] Speaker 9: to simulate quite a lot, right, You need to simulate
[25:44] Speaker 9: the entality of the Bloomberg network. You need to understand
[25:46] Speaker 9: that even generate fake topic in order to just get
[25:48] Speaker 9: as close as.
[25:49] Speaker 3: Possible to the world.
[25:50] Speaker 9: Because these environments got more and more and more complicated.
[25:53] Speaker 9: In some of them, there are a lot of controls.
[25:55] Speaker 9: One of these controls is access to the Internet, and
[25:57] Speaker 9: why do you even have access to the Internet as
[25:59] Speaker 9: one of the controls If you think about it a
[26:01] Speaker 9: lean world, the tack hele that maybe after you will
[26:03] Speaker 9: definitely not say, oh, you know what, the models were
[26:05] Speaker 9: tested without an Internet collection, so let's right, let's not
[26:08] Speaker 9: you know, just like have it as well. Obviously, if
[26:10] Speaker 9: I'm going to go after I'm going to go after
[26:11] Speaker 9: you with everything at my disposal.
[26:13] Speaker 2: And that's going to include Internet access. Well. In the
[26:15] Speaker 2: open Ai case, there were two models, one released and
[26:18] Speaker 2: one unreleased, but more performance model, and the instruction from
[26:21] Speaker 2: open ai was ace this test. We want to evaluate
[26:24] Speaker 2: your cyber capabilities, and the models concluded that in order
[26:28] Speaker 2: to ace the test, they could get information from a
[26:31] Speaker 2: third party platform via the Internet.
[26:34] Speaker 9: So just to clarify, yes, that case did not involve
[26:37] Speaker 9: us in any sense. Okay, that was a separate case.
[26:40] Speaker 9: And I think the reason that you're seeing multiple cases
[26:43] Speaker 9: that happened recently for us there were issues with the
[26:46] Speaker 9: UK government as well, the wations of open air.
[26:48] Speaker 2: And hugging face.
[26:49] Speaker 9: Some of the anthropic disclosure to disclosures setup is mostly
[26:53] Speaker 9: because the models have passed a threshold of competency that
[26:56] Speaker 9: allows them to actually do these elements and you know,
[26:59] Speaker 9: just like have really world effect. That shows why it's
[27:02] Speaker 9: actually so crucial to also do pre deployment testing because
[27:05] Speaker 9: if you're not going to discover that in a pre
[27:07] Speaker 9: deployment test, you're going to discover that when it happens
[27:09] Speaker 9: in the world, and that's a much greater issue. That's
[27:12] Speaker 9: also why you know, looking forward, we both included and
[27:15] Speaker 9: the Commuter to an effort of just like reallyfined best
[27:18] Speaker 9: practices for the industry because we have to find a
[27:20] Speaker 9: solution on how to do this testing in a way
[27:22] Speaker 9: that includes that but also what it's kind of like
[27:25] Speaker 9: a view on what's happening in a security more broadly
[27:28] Speaker 9: right now in an essay.
[27:29] Speaker 2: But that is so interesting, it's several points. Many people
[27:32] Speaker 2: in your industry said, this is a watershed moment for
[27:35] Speaker 2: AI in the context of how it plays in cybersecurity. Yes,
[27:38] Speaker 2: but many also said, it's not a scandal. You want
[27:41] Speaker 2: to see the frontier labs testing and evaluating the capabilities
[27:45] Speaker 2: of the models. What have you changed since you know,
[27:49] Speaker 2: how do you get that balance better.
[27:52] Speaker 3: You have the testimals before.
[27:54] Speaker 9: Because if you don't do that, the identity is way worse.
[27:57] Speaker 9: You're going to just discover these issues in the world
[27:58] Speaker 9: without any controlled volment. So the reason that we're finding
[28:01] Speaker 9: so many of them is because in the controlled environment,
[28:02] Speaker 9: even if mistakes happen, you have the ability to assess
[28:05] Speaker 9: a lot of information because you have traces and you're
[28:07] Speaker 9: able to get better. The alternative would be discovering something
[28:10] Speaker 9: that happened after the deployment, which is reaction much worse exactly,
[28:13] Speaker 9: and that's a huge pa.
[28:14] Speaker 2: It may I ask then, if you must test, then
[28:17] Speaker 2: what have you done to ensure that testing is safer?
[28:20] Speaker 2: What have you change?
[28:21] Speaker 3: So a lot.
[28:22] Speaker 9: So I think there's multiple elements here. Just like in
[28:24] Speaker 9: the one, we've expanded a lot of the monitoring and
[28:27] Speaker 9: manual efforts that are now being put in two, we've
[28:30] Speaker 9: essentially upgraded a lot of the monitoring systems that we're do.
[28:33] Speaker 9: I think a point that is really just like important
[28:35] Speaker 9: to make here is classic cyber monitoring tools only goes
[28:39] Speaker 9: so far in an AI context. And the reason is
[28:42] Speaker 9: a lot of the analysis that you need to do
[28:44] Speaker 9: is about underlyzing specific tansclipts of just like modelers having
[28:47] Speaker 9: conversations with themselves. It's just like very high philosity, So
[28:49] Speaker 9: there's a lot of upgrades, so just like how it
[28:51] Speaker 9: happens there. So like a third element is also the
[28:53] Speaker 9: process of how you're doing some of the testing. So
[28:56] Speaker 9: there are many many elements that we are now working
[28:58] Speaker 9: on and we are going to just like a new
[29:00] Speaker 9: white paper off best practices on how that's going to happen,
[29:03] Speaker 9: but a lot of us already implemented at this point.
[29:06] Speaker 2: May I ask, are there other incidents that you're aware
[29:09] Speaker 2: of that have happened in this interim period that are
[29:12] Speaker 2: not yet public. I appreciate there's a reality to what
[29:15] Speaker 2: you can and cannot say, but it's a question of
[29:17] Speaker 2: how pervasive this issue is.
[29:20] Speaker 9: So we've reported to a customer based just like everything
[29:23] Speaker 9: that we know, but obviously not going to get into
[29:26] Speaker 9: confidential information or sure interaction.
[29:27] Speaker 2: Between there are incidents that you've identical. Sorry, So so that's
[29:30] Speaker 2: not what I'm saying. I'm just saying that we've reported
[29:32] Speaker 2: to our customers everything that it's like that's on outside
[29:35] Speaker 2: and you know it just like obviously not going to
[29:36] Speaker 2: get into just like details of you know what are reporting.
[29:38] Speaker 2: But I really do think that we're kind of fact
[29:40] Speaker 2: as you've mentioned are in a watershed moment and just
[29:43] Speaker 2: like that, we need to think about what happens next.
[29:45] Speaker 2: And in a way, I want to just like make
[29:47] Speaker 2: two comments that I think are highly importantly. Like one,
[29:51] Speaker 2: you know, just like the labs.
[29:52] Speaker 9: Deeply care about security and I think you're seeing that
[29:54] Speaker 9: even by the fact that are going through you know,
[29:57] Speaker 9: the tuble of just like doing all of this testing
[29:58] Speaker 9: pre deployment because the unders and how big is this
[30:00] Speaker 9: moment And every time that we came to one of
[30:02] Speaker 9: the labs of saying, you know, we should change the
[30:04] Speaker 9: protocol hell.
[30:05] Speaker 2: SETA, they were extremely receptive to it.
[30:07] Speaker 9: And the second is that we actually are also in a
[30:09] Speaker 9: moment in time where you know, looking over the long term,
[30:12] Speaker 9: there is a huge opportunity in what AI can do
[30:15] Speaker 9: for the defender site. Right. I think we can even
[30:18] Speaker 9: be optimistic about you know, what ultimately happens the elements here,
[30:21] Speaker 9: which is, you know, we can live in a world
[30:23] Speaker 9: in a few years where AI is going to view
[30:25] Speaker 9: every line of code for security elements. We can live
[30:28] Speaker 9: in a world where there's going to be formal verification
[30:30] Speaker 9: scale your ability to actually make sure mathematically that you
[30:33] Speaker 9: know that things are getting better but one of the issues,
[30:36] Speaker 9: and that's also something that we published in the in
[30:37] Speaker 9: the essay, is something that we call the n state fallacy,
[30:40] Speaker 9: which is, you know, the long term values are occasionally
[30:44] Speaker 9: conflated with short term publems that we still need to
[30:46] Speaker 9: figure out how to.
[30:47] Speaker 2: So this is where I'd like to end our conversation,
[30:49] Speaker 2: the end state fallacy. Where is AI security headed published
[30:53] Speaker 2: in the last twenty four hours. Yes, the conclusion you
[30:55] Speaker 2: draw is that frontier models will roughly double their performance
[30:59] Speaker 2: every six months at least for the next two years,
[31:02] Speaker 2: and their effective autonomous work horizon goes with it. The
[31:08] Speaker 2: question the audience quite rightly has is does that mean
[31:11] Speaker 2: that you can or cannot contain and prevent the types
[31:14] Speaker 2: of incidence that we've reported on in the last three weeks.
[31:17] Speaker 9: So I'm very optimistic, but the world the reason that
[31:19] Speaker 9: you're quality and state fallacy is because we believe that
[31:22] Speaker 9: there is a difference on kind of like the time
[31:24] Speaker 9: horizon of just like how you look at question. I
[31:26] Speaker 9: just like to be clear, if I'm going to set
[31:28] Speaker 9: myself an objective something like getting better shape, the end
[31:32] Speaker 9: state is incredibly better than optimistic. But if you look
[31:34] Speaker 9: at me, you know, the path to getting in shape
[31:36] Speaker 9: is going to be painful, and I think that occasionally
[31:39] Speaker 9: people are conflating that the properties off the end state
[31:42] Speaker 9: that can be very valuable with the difficulty of getting them.
[31:45] Speaker 9: So over the long time horizon, we can have ais
[31:47] Speaker 9: also verifying, you know, just like sandboxers verifying that there
[31:51] Speaker 9: are no issues. A lot of AIS are going to
[31:53] Speaker 9: be useful for extended monitoring that we think should be
[31:55] Speaker 9: implemented not just in testing environments but also in the world,
[31:58] Speaker 9: but the chin well in the short term, they're going
[32:01] Speaker 9: to potentially be issues that we need to figure out
[32:03] Speaker 9: as a community how to solve. And we also just
[32:06] Speaker 9: like complete strategies on what to do.
[32:07] Speaker 2: Though, So I want to go back to what this
[32:09] Speaker 2: experience has been like for you and for the company.
[32:12] Speaker 2: You had two roles. You did the testing, provided testing
[32:15] Speaker 2: and also part of the providing and operation of the
[32:18] Speaker 2: sandbox environment. Have you seen massive demand for your services?
[32:22] Speaker 2: Have customers said we need to rethink the relationship? What
[32:26] Speaker 2: has happened?
[32:26] Speaker 9: So, you know, just like I think the most credible
[32:28] Speaker 9: thing is just like to say, just like us, what
[32:30] Speaker 9: our customers said, which if you look all them and
[32:32] Speaker 9: ended the statement by you know, it's like thinking, you know,
[32:35] Speaker 9: there's like and recognizing the complexity of what we do
[32:38] Speaker 9: and saying that they're looking forward to just like doubling
[32:40] Speaker 9: down on the relationship in the future. And the reason
[32:43] Speaker 9: is it's a dual specialty that's incredibly how to do.
[32:45] Speaker 9: In order to do that well and in order to
[32:47] Speaker 9: make sure that we work together to ensure that ultimately
[32:49] Speaker 9: you and I don't mean assarily us ad, I mean yours,
[32:51] Speaker 9: real society are going to be in a secure world.
[32:54] Speaker 9: We need people that have a really dual explotise, both
[32:56] Speaker 9: a very deep cyber background and a very deep AI background,
[32:59] Speaker 9: such a we can test environments and as you can imagine,
[33:02] Speaker 9: it's a fairly little thing, so as an outcome, it's
[33:05] Speaker 9: a learning moment for everyone in the industry. And by
[33:08] Speaker 9: the way, we are not a sandbox company, just to
[33:10] Speaker 9: be clear, that's actually not the role of what do
[33:11] Speaker 9: we specifically do. The open air hugging face is not
[33:13] Speaker 9: something that is on our side, just like but it
[33:16] Speaker 9: is still you know, kind of like a gun moment
[33:18] Speaker 9: of things changing in the industry and together there's a
[33:22] Speaker 9: role of just like taking the people that have a
[33:23] Speaker 9: deep knowledge of what's going on and being able to
[33:26] Speaker 9: work together to resolve and understand how to make sure
[33:29] Speaker 9: that things are going to be in a much better
[33:31] Speaker 9: optimistic position.
[33:32] Speaker 2: Than to have a regular CEO and co founder with
[33:35] Speaker 2: us here on Bloomberg Tech in San Francisco. Thank you
[33:37] Speaker 2: very much.
[33:38] Speaker 9: It was a pleasure being.
[33:38] Speaker 2: Add coming out Dave mccamillo, Managing partner at GB Joints
[33:43] Speaker 2: to talk about why the next big AI investing opportunities
[33:46] Speaker 2: to be in the application layer, also coming off some
[33:49] Speaker 2: recent exit success. This is Bloomberg Tech. The speed of
[33:58] Speaker 2: AI development means investors trying to stay ahead of the
[34:00] Speaker 2: curve also need to move faster than ever. For venture
[34:04] Speaker 2: firm GV, that means focusing now on the application there
[34:07] Speaker 2: having already made strategic bets on AI infrastructure chip startups.
[34:10] Speaker 2: David Nichiello, managing partner at GV, joins us now and
[34:13] Speaker 2: gv's sole limited partner, as you may know, is Alphabet,
[34:17] Speaker 2: but it invests independently without regard for Alphabet's business strategy.
[34:23] Speaker 2: It was written, so I read it. Welcome to the program. Thanks.
[34:27] Speaker 2: It's really interesting to have you here today because quite
[34:30] Speaker 2: recently we covered one of your exits with the acquirer
[34:33] Speaker 2: that was Qualcom a Moduli. I just wanted to sort
[34:36] Speaker 2: of somebody who was high up the cap table on
[34:39] Speaker 2: that company, your experience of that process.
[34:42] Speaker 10: I mean, we're fortunate to be the first check into
[34:44] Speaker 10: the company, and you know we led that first round
[34:45] Speaker 10: in a bet involved with Chris and Tim the entire
[34:48] Speaker 10: part of the journey. You know, the company Modular is
[34:52] Speaker 10: the software layer that's right top of technology, sits on
[34:55] Speaker 10: top of AI chips to make sure that any model
[34:58] Speaker 10: can run on any chip.
[34:59] Speaker 2: So I'm kind of riffing on the intro, yeah, ry
[35:01] Speaker 2: that there is this appreciation more than ever of that layer.
[35:05] Speaker 10: And Nvidia has Kuda and Chris's idea was what if
[35:08] Speaker 10: there was an alternative to Kudah that was more open
[35:11] Speaker 10: and could work for any chip and any model. And
[35:13] Speaker 10: Qualcom was an early customer. Christiano is quite visionary. And
[35:17] Speaker 10: I still remember where I was sitting when Chris Latner
[35:20] Speaker 10: called me and said, Hey, you know what, Christiano is
[35:21] Speaker 10: going to fly up from Irvine and.
[35:24] Speaker 3: He wants to have dinner with me. What do you
[35:25] Speaker 3: think he wants to talk about?
[35:26] Speaker 10: You know, they've been customers for a while and they're
[35:28] Speaker 10: excited about what they're seeing, And of course that dinner
[35:31] Speaker 10: ended up being an overture to Chris Lattner to come in.
[35:34] Speaker 3: And be a part of Qualcom.
[35:36] Speaker 10: Usually founders don't love the idea of Emina particularly did.
[35:40] Speaker 2: You like the idea? I mean, would you can be honest,
[35:43] Speaker 2: But your advice.
[35:44] Speaker 10: Was I always start with a skepticism of M and A.
[35:47] Speaker 10: The thing that I think about when I think about
[35:49] Speaker 10: you know, what is best for a company is what's
[35:51] Speaker 10: best for the founder. And here Chris and Tim saw
[35:54] Speaker 10: a huge vision of disrupting Kuda, and they wanted the
[35:57] Speaker 10: fastest path to do that. Chris Giano painted future we're
[36:01] Speaker 10: together with Qualcom's capital and focus and attention and portfolio
[36:05] Speaker 10: of chips, where they could do that together over the
[36:07] Speaker 10: next decade.
[36:08] Speaker 3: That was what Chris was enamored by.
[36:10] Speaker 2: If we zoomed out to get perspective, the environment fit
[36:13] Speaker 2: for your industry is private market. Venture money is very
[36:16] Speaker 2: happy to support companies private for longer, etc. For sure,
[36:24] Speaker 2: there is some life in the IPO market. Where does
[36:28] Speaker 2: the sort of acquisition by a bigger technology be sit
[36:31] Speaker 2: in the field right now? Yeah?
[36:33] Speaker 10: So for us for GV, the thing that we care
[36:35] Speaker 10: most about is investing in founders as early as possible, right,
[36:38] Speaker 10: so we try to invest in the very first round,
[36:40] Speaker 10: like we did it modular, like we've done at Flapping
[36:42] Speaker 10: Airplanes or so Manova Systems where we led.
[36:44] Speaker 3: The Series A.
[36:45] Speaker 10: And our objective is to help those companies grow and
[36:47] Speaker 10: realize the technology that they're building, the vision of that technology,
[36:51] Speaker 10: and get it out into the market. And sometimes that
[36:53] Speaker 10: happens through an IPO, which is a massive amount of
[36:56] Speaker 10: funding and just another milestone on that startups journey to
[37:00] Speaker 10: M and A sometimes, like in the case of Slack,
[37:02] Speaker 10: it was IPO and then M and A right, and like,
[37:04] Speaker 10: eventually that technology needs to exist in the world and
[37:07] Speaker 10: be funded by public markets, and it's either funded through
[37:11] Speaker 10: a public company or through an IPO.
[37:13] Speaker 2: It's a very interesting list. So Samonov is an example
[37:15] Speaker 2: of a company that hasn't done either yet. But where's
[37:20] Speaker 2: the activity happening? You know, many of your industry peers
[37:23] Speaker 2: and colleagues talk up San Francisco in the Bay Area
[37:27] Speaker 2: and Silicon Valley as being the main stay of intellectual
[37:30] Speaker 2: and financial capital.
[37:32] Speaker 10: So GV has thirteen billion dollars under management and offices
[37:35] Speaker 10: in London, New York, Cambridge, mass and the Bay Area.
[37:38] Speaker 10: We're trying to convince most of our London, New York
[37:40] Speaker 10: and Cambridge investors.
[37:42] Speaker 3: To spend more time in San Francisco.
[37:43] Speaker 2: Wow. We think that all of the you mean, the
[37:45] Speaker 2: investing partners of the firm that are based outside of
[37:48] Speaker 2: the Bay comes spend.
[37:49] Speaker 3: We're convincing that some more spend more time.
[37:51] Speaker 10: So a partner like Crystal Wang in New York, she
[37:53] Speaker 10: spends every other week in San Francisco because of the
[37:56] Speaker 10: sheer volume of the activity that's happening here. She's still
[37:59] Speaker 10: invested in your best companies. We still think the best
[38:02] Speaker 10: companies are built everywhere in the world, but right now
[38:05] Speaker 10: the clear concentration and the activity is happening in the valley.
[38:09] Speaker 2: The last thing I want to ask you about is proxy.
[38:12] Speaker 2: So this is an interesting idea. In the software phase,
[38:15] Speaker 2: it was a bit easier because you had publicly traded proxies.
[38:18] Speaker 2: You could sort of determine valuation and tam and where
[38:21] Speaker 2: they would sit in the field. In other areas, it's
[38:24] Speaker 2: more difficult if all these companies are staying private long
[38:27] Speaker 2: or they're being acquired, like many of your portfolio companies, have.
[38:31] Speaker 10: I mean, I think valuations are increasing exponentially, and they're
[38:35] Speaker 10: doing that because of the.
[38:36] Speaker 2: TAM tambonentially, but justifiably so I think.
[38:39] Speaker 10: In many cases, justifiably. I mean, when you have exits
[38:41] Speaker 10: that look like they're in the hundreds of billions or
[38:44] Speaker 10: trillions of dollars, like the set of IPOs that are
[38:46] Speaker 10: coming soon, then you have a lot of venture frenzy
[38:49] Speaker 10: into early stage companies. I think at GV we like
[38:52] Speaker 10: to pair deep subject matter experts with people who have conviction.
[38:56] Speaker 10: There are some firms that just look at conviction, and
[38:58] Speaker 10: so when you're just following conviction or you're just following heat,
[39:01] Speaker 10: you're going to invest in things that everybody else thinks
[39:04] Speaker 10: are interesting.
[39:04] Speaker 3: We tend not to do that. We tend to be
[39:06] Speaker 3: quite contrarian.
[39:06] Speaker 10: I think you'll see that we were investing in infrastructure
[39:09] Speaker 10: before the chat GPT moment. We were investing in in
[39:12] Speaker 10: AI applications when everyone else sort of pivoted into infrastructure.
[39:16] Speaker 3: Right now, we're really excited about science.
[39:17] Speaker 10: I think that AI for science is going to be
[39:20] Speaker 10: a fascinating place to be in the next ten years.
[39:23] Speaker 10: And this is using the best models with the best
[39:26] Speaker 10: science in the world to find breakthroughs in things like
[39:28] Speaker 10: material science, physics, math. We're seeing papers being written at
[39:32] Speaker 10: an unprecedented rate that are.
[39:34] Speaker 3: Showing human breakthroughs.
[39:35] Speaker 10: And we think that rather than the chetchipt moment, the
[39:39] Speaker 10: chat moment being the thing that connects with main Street,
[39:41] Speaker 10: it might be those scientific discoveries that connect with main
[39:44] Speaker 10: Street that increase the human condition.
[39:46] Speaker 2: That's probably the next watch then for the show in
[39:49] Speaker 2: Bloomberg Tech. Dave Minichiello, managing partner at GV, thank you
[39:52] Speaker 2: very much for being here with us. Okay, coming up,
[39:54] Speaker 2: Sources tell Bloomberg's the Justice Department is investigating Andreson Horowitz
[40:00] Speaker 2: sort of partners serving on the boards of competing startups.
[40:03] Speaker 2: We have the details of that report. Next, this is
[40:06] Speaker 2: Bloomberg Tech Benja Capital firm Andreessen Horowitz is the focus
[40:19] Speaker 2: of an antitrust pro by the Justice Department. According to sources,
[40:24] Speaker 2: the DOJ is investigating whether a sixteen partners are improperly
[40:27] Speaker 2: serving on the boards of competing AI companies Data Bix
[40:31] Speaker 2: and five Tran. Let's get the details of Bloomberg's antrust
[40:34] Speaker 2: reported Josh Cisco, who broke the story. There are specific
[40:38] Speaker 2: individuals at Andrews and Howitz also that are a part
[40:41] Speaker 2: of this story. Let's just start with the basics of
[40:43] Speaker 2: what we're reporting and what we need to know.
[40:46] Speaker 11: So there's a law, an old law dating back more
[40:51] Speaker 11: than a century that says you can't that investors can't
[40:54] Speaker 11: serve on the boards of competing companies and for injuries
[40:57] Speaker 11: in Horowitz that would be co founder Ben Horowitz, who
[41:02] Speaker 11: serves on the board of Data Bricks, and then a
[41:04] Speaker 11: gentleman named Martine Casato who serves on the board of
[41:08] Speaker 11: five tran They both are at a high level data
[41:12] Speaker 11: analytics companies, and so the Justice Department wants to make
[41:17] Speaker 11: sure that there is no sort of competitive issue with that.
[41:24] Speaker 2: We say in our report, this is a probe. What
[41:27] Speaker 2: is the process? What is it that the DOJ is
[41:30] Speaker 2: doing in terms of actions taken or not?
[41:33] Speaker 11: They ask the company's questions, They issue sort of formal
[41:39] Speaker 11: investigative demands civil subpoenas to sort of the companies themselves,
[41:45] Speaker 11: any competing companies that might have information.
[41:47] Speaker 3: Or help them bet are able to understand the market.
[41:50] Speaker 11: And then after doing an investigation, then they decide whether
[41:55] Speaker 11: to whether and how to settle it or to file
[41:58] Speaker 11: a lawsuit and the solution here the government in the
[42:03] Speaker 11: Biden administration under Jonathan Cantor, who led anti trust enforcement
[42:07] Speaker 11: then at the DOJ, they brought a lot of these
[42:10] Speaker 11: investigations and they typically were resolved by the director of
[42:15] Speaker 11: one of the companies stepping down from one of the boards.
[42:19] Speaker 11: In all of those situations was a little different because
[42:22] Speaker 11: it was the same person on both boards. Here you
[42:25] Speaker 11: have two different people serving on each board, and so
[42:30] Speaker 11: that could create a bit more tricky dynamics in terms
[42:34] Speaker 11: of like how you figure out who steps down from
[42:37] Speaker 11: which board.
[42:39] Speaker 2: It spokes people for date the vision noted spokespeople for
[42:43] Speaker 2: data breaks in the Justice Department declined to comment on
[42:46] Speaker 2: our reporting, andres and Horowitz and five Trend did not
[42:49] Speaker 2: respond to our request for comment. I guess difficult. You know,
[42:52] Speaker 2: this is an investigation as detail, and the reporting went
[42:54] Speaker 2: on for a year and is now being disclosed to
[42:57] Speaker 2: our reporting. But what happens next.
[43:00] Speaker 11: Well, the government will have to decide what its next
[43:03] Speaker 11: steps are, whether it can work out a settlement with
[43:06] Speaker 11: the company. It could also just drop the investigation altogether
[43:09] Speaker 11: if it doesn't feel like its case is strong enough
[43:11] Speaker 11: or you know, the most unlikely outcome is it could
[43:14] Speaker 11: sue to force one of the directors to step down.
[43:19] Speaker 2: Or to again fund any other solution that they that
[43:22] Speaker 2: they want. And I'll just repeat that in our reporting
[43:26] Speaker 2: all of the parties involved have either declined to comment
[43:29] Speaker 2: or haven't responded to comment, but if they do, we'll
[43:31] Speaker 2: bring the update on what it's a really big story.
[43:33] Speaker 2: But in bos justisco thank you very much. So that
[43:36] Speaker 2: does it for this addition of Bloomberg Tech. One final
[43:40] Speaker 2: check on market. It's again it's semiconductors that are under
[43:42] Speaker 2: pressure and that's taking the rest of the technology sector
[43:45] Speaker 2: with it. There's some anxiety in the bomb market. You
[43:48] Speaker 2: see the US tenure yield at twenty twenty five highs
[43:51] Speaker 2: a part of the pressure that is translating into the
[43:55] Speaker 2: tech equity space. Recap the show on the podcast you
[43:57] Speaker 2: know where to find it. This is Bloomberg Teck

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.
