# Oracle’s Data Center Challenge as Trump, Xi Talk AI

An episode of [Bloomberg Tech](https://invisiblenetwork.ai/podcast/1611767434), 24 Sep 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/7f21ffd1-a353-4c5e-9086-b4cf011284f3/audio.mp3?utm_source=Podcast&in_playlist=9caafc41-289c-4115-995d-ae7c0030ab75

Bloomberg’s Ed Ludlow breaks down Oracle's latest move to protect itself from mounting costs on a massive New Mexico data center facing regulatory setbacks and local opposition. Plus, Meta comes out with a new palm-sized device for using its popular new AI assistant, Muse; and all eyes are on President Trump and China's Xi Jinping as they meet in the White House and discuss AI. See omnystudio.com/listener for privacy information.

## Transcript

[0:02] Speaker 1: Bloomberg Audio Studios. Podcasts.
[0:05] Speaker 2: Radio. News.
[0:13] Speaker 3: Bloomberg Tech is live from the heart of Silicon Valley
[0:17] Speaker 3: with Ed Ludlow in San Francisco.
[0:22] Speaker 1: This is Bloomberg Tech. Coming up, force majeure.
[0:24] Speaker 4: Oracle moves to protect itself from mounting costs on a
[0:27] Speaker 4: massive New Mexico data center facing regulatory setbacks and local opposition. Plus,
[0:32] Speaker 4: it looks like a Tamagotchi and fits in the palm
[0:34] Speaker 4: of your hand. We're going to discuss Meta's new device
[0:37] Speaker 4: for using its popular new AI assistant, Muse. And all
[0:40] Speaker 4: eyes on President Trump and China's Xi Jinping, who are
[0:43] Speaker 4: moments away from holding a bilateral meeting where AI... is
[0:47] Speaker 4: expected to be top of mind let's get right to
[0:50] Speaker 4: our top story and shares of oracle down more than
[0:52] Speaker 4: five percent at one point in the session down seven
[0:55] Speaker 4: percent on track for their biggest decline since mid-june this
[0:59] Speaker 4: after bloomberg reports the company sent a letter to the
[1:02] Speaker 4: developer of a massive data center being built in new
[1:05] Speaker 4: mexico citing force majeure that's all according to sources The
[1:09] Speaker 4: move could help shield the tech giant from racking up
[1:12] Speaker 4: expenses on a project that has been beset by delays.
[1:15] Speaker 4: Let's get the details with Bloomberg, Sridhar Nanarajan, who is
[1:18] Speaker 4: part of the team that broke the story. Let's get
[1:21] Speaker 4: into specifics. The parties involved, this site, and I'm sure
[1:26] Speaker 4: lots of people understand force majeure, but the basic concept here.
[1:30] Speaker 5: Again, so let's set the stage here.
[1:31] Speaker 6: We're talking about this massive data center campus, this data
[1:34] Speaker 6: center build-out that's taking place in southern New Mexico, beautiful
[1:38] Speaker 6: part of the country, borders Mexico, in a rural county.
[1:41] Speaker 5: It's a 1,400-acre campus.
[1:43] Speaker 6: But with such a big development, what you've seen is
[1:46] Speaker 6: also a lot of local opposition and regulatory headaches.
[1:50] Speaker 5: Which has.
[1:52] Speaker 6: proved to be a problem for all the parties involved. Now,
[1:55] Speaker 6: who are the parties involved? Stack Infrastructure, which is a
[1:57] Speaker 6: portfolio company of Blue Owl, which is the data center developer. Oracle,
[2:02] Speaker 6: which has agreed to be the tenants, which will be
[2:04] Speaker 6: leasing the compute that will come up at this data
[2:06] Speaker 6: center facility. And the issues that these parties face is
[2:10] Speaker 6: what happens if this project continues to be delayed because
[2:13] Speaker 6: there have been a lot of financial resources already poured
[2:16] Speaker 6: into this and a lot more to come down the pike.
[2:20] Speaker 6: So it's important to clarify for our viewers and listeners
[2:22] Speaker 6: what this is and what this isn't. This isn't Oracle
[2:25] Speaker 6: throwing up his hand and saying it is walking away
[2:28] Speaker 6: from this project. I would see this as a defensive move,
[2:32] Speaker 6: as a preemptive move to send out this notice to
[2:36] Speaker 6: Stack Infrastructure, to Blue Owl, and tell them that if
[2:39] Speaker 6: the project does get delayed, it hopes to wriggle out
[2:43] Speaker 6: of certain financial obligations and commitments. It's not clear whether
[2:47] Speaker 6: this contractual maneuver would necessarily work. and they would not
[2:51] Speaker 6: be on the hook for those payments. But that is
[2:53] Speaker 6: the effort by Oracle here because the delays are visible.
[2:56] Speaker 6: There's permitting delays.
[2:58] Speaker 5: There are other issues.
[2:59] Speaker 6: And in some ways, this project, so-called Project Jupiter in
[3:02] Speaker 6: New Mexico, is a microcosm for all the problems facing
[3:05] Speaker 6: the data center build-out across the United States, whether you're
[3:08] Speaker 6: talking about the battleground states or places like Northern Virginia.
[3:11] Speaker 5: In fact, everywhere in the country right now.
[3:15] Speaker 4: Oracle did respond to our request for comment, and then
[3:18] Speaker 4: they've posted that response publicly.
[3:19] Speaker 1: Let's go through it.
[3:20] Speaker 4: I mean, what is their line on what's happening with
[3:23] Speaker 4: Project Jupyter?
[3:24] Speaker 6: And what they have posted is consistent with our understanding
[3:27] Speaker 6: of it. In an ideal world, Oracle would want this
[3:30] Speaker 6: data center to come up as soon as possible. It
[3:33] Speaker 6: remains committed to the project. It has addressed the backlash
[3:37] Speaker 6: that it has faced in the local community and tried
[3:40] Speaker 6: to put forward its own argument for why this will
[3:43] Speaker 6: be beneficial. They have pointed out the financial incentives, they've
[3:46] Speaker 6: pointed out to the workforce coming into the area and
[3:49] Speaker 6: how in general it'll be beneficial to the broader economy.
[3:52] Speaker 6: We must also point out that Blue Owl has gone
[3:54] Speaker 6: on the record and said that this notice does not
[3:57] Speaker 6: change the multi-year financial commitments. So one has to assume
[4:01] Speaker 6: that Blue Owl is trying to convey to its investors
[4:04] Speaker 6: and markets that just because a notice has been sent
[4:07] Speaker 6: doesn't mean that the certain financial obligations do not have
[4:10] Speaker 6: to be met.
[4:12] Speaker 4: Okay, so Blue Owl owns Stack, which is the developer.
[4:15] Speaker 4: I want to get into the financing very quickly because
[4:18] Speaker 4: there's a loan here backed by lots of banks.
[4:20] Speaker 1: Very quick.
[4:21] Speaker 6: And that is why we're talking about what are the
[4:23] Speaker 6: financial anxieties here. A group of about 20 banks committed
[4:27] Speaker 6: $ 18 billion in loans for this project. Now, we do
[4:30] Speaker 6: get to see some of the private market trading. Now,
[4:33] Speaker 6: this is not liquid active trading, but there is some movement.
[4:37] Speaker 6: And these loans have already been trading below 90 cents
[4:39] Speaker 6: on the dollar. That might not be distress levels, but
[4:42] Speaker 6: that certainly implies some level of stress. If you're a
[4:45] Speaker 6: lender in a project like this, which... under ordinary circumstances
[4:50] Speaker 6: is supposed to be money good, this is not the
[4:52] Speaker 6: level where you would like to see.
[4:54] Speaker 5: The loans at.
[4:55] Speaker 4: The question from this story is, when a data center
[4:57] Speaker 4: is delayed, who picks up the bill in the interim? Bloomberg, Shredder, Natarajan,
[5:01] Speaker 4: part of the team that broke that market-moving story. Let's
[5:04] Speaker 4: turn to another side of AI. Meta has unveiled a
[5:07] Speaker 4: dedicated gadget for using its popular new AI assistant, Muse.
[5:11] Speaker 4: The palm-sized device called Muse Charm was born out of
[5:14] Speaker 4: the company's new design lab and its super.
[5:16] Speaker 1: Intelligence AI group.
[5:18] Speaker 4: Bloomberg's Riley Griffin was at the company's MetaConnect event last night,
[5:21] Speaker 4: joins us now.
[5:23] Speaker 1: I see a Tamagotchi.
[5:25] Speaker 5: I see a Tamagotchi.
[5:26] Speaker 4: You and I are probably the cutoff of the generation.
[5:28] Speaker 4: Tamagotchi released in the late 90s still exists today. What
[5:32] Speaker 4: is the idea here?
[5:33] Speaker 3: Yeah, so Mark Zuckerberg has been talking about personal superintelligence,
[5:38] Speaker 3: this idea of taking AI and bringing it into the
[5:41] Speaker 3: hands of billions. Meta, formerly Facebook, is notoriously good at
[5:47] Speaker 3: scaling things to consumers, and they're hoping to do that here.
[5:50] Speaker 3: But who knew that personal super intelligence would look like
[5:53] Speaker 3: a Tamagotchi? They're leaning into the cute factor. They're trying
[5:56] Speaker 3: to get off of phones and into other forms of hardware.
[5:59] Speaker 3: Most of the event was dedicated to talking about glasses.
[6:02] Speaker 1: There it is. But this was a big reveal. There
[6:03] Speaker 1: it is.
[6:04] Speaker 4: I mean, I listened to Mark Zuckerberg very closely. You know,
[6:06] Speaker 4: the warnings out there are, look what happened to the
[6:09] Speaker 4: Humane AI pin.
[6:10] Speaker 1: Look at the rabbit device. Massive flops.
[6:13] Speaker 4: But he's basically saying, this device that we're looking at now,
[6:16] Speaker 4: maybe explain some of its features. is easier than reaching
[6:20] Speaker 4: for your smartphone.
[6:21] Speaker 3: Yeah, he thinks that in the future, we are not
[6:24] Speaker 3: going to want to look down at our phone to
[6:26] Speaker 3: experience the world.
[6:27] Speaker 7: This.
[6:29] Speaker 3: Tamagotchi-like AI assistant has a camera front and back. It
[6:32] Speaker 3: can listen to you. There are other Muse features that
[6:37] Speaker 3: include actually FaceTiming with a Muse. Something that's dawning on
[6:41] Speaker 3: me right now as we're having this conversation is there's
[6:43] Speaker 3: a lot of fear around AI and how to use it.
[6:46] Speaker 3: This cute factor is a bit more accessible.
[6:50] Speaker 5: It might be consumer-friendly.
[6:52] Speaker 4: It's been a massive week for Muse, which is Meta's
[6:56] Speaker 4: consumer-focused AI agent, right? It's moved markets. It's topped the
[7:00] Speaker 4: charts on the downloads of the App Store. They also
[7:03] Speaker 4: took Muse into a new generation of VR glasses and
[7:06] Speaker 4: existing glasses.
[7:07] Speaker 1: What did you hear last night?
[7:09] Speaker 3: Well, really, the plan is to scale this into all
[7:12] Speaker 3: of their products, into the VR glasses, into the MetaRay bands,
[7:15] Speaker 3: into this charm device. Muse itself is the cornerstone of
[7:19] Speaker 3: their AI strategy. We saw Alexander Wang take the stage
[7:22] Speaker 3: to talk about some of those developments. And they're riding
[7:25] Speaker 3: a high right now because the app is high on
[7:29] Speaker 3: the Apple downloads. Same with Google. It's performing really well.
[7:34] Speaker 3: But one thing is they've got to develop a sense
[7:36] Speaker 3: of trust because in order to use this tool effectively,
[7:39] Speaker 3: you basically have to hand it the keys to your life.
[7:42] Speaker 4: We've got Riley Griffin, late night for you, back on
[7:44] Speaker 4: the show early morning. Thank you very much indeed. With
[7:46] Speaker 4: the popularity of the Muse Assistant and the introduction of
[7:49] Speaker 4: new devices, what does that mean for how Meta positions
[7:52] Speaker 4: itself as a company moving forward? Minda Smiley, Senior Analyst
[7:56] Speaker 4: at eMarketers with us. What is Meta now?
[8:00] Speaker 8: Yeah, I mean, it's a great question. I have to say,
[8:02] Speaker 8: you know, echoing what you and Riley said, I definitely
[8:05] Speaker 8: did not, a Tamagotchi was certainly not on my bingo
[8:08] Speaker 8: card for Meta this week, but it's not surprising. I mean,
[8:12] Speaker 8: I feel like experimentation has been the name of the
[8:14] Speaker 8: game for meta over the past several years, starting with
[8:18] Speaker 8: the metaverse, which we saw didn't exactly work out. But
[8:21] Speaker 8: now we're looking at everything from smart glasses to AI
[8:24] Speaker 8: to subscriptions to now charm. And so I think it's
[8:29] Speaker 8: clear that meta is really trying to chart its next chapter,
[8:32] Speaker 8: I think, both from a hardware perspective, but also kind
[8:36] Speaker 8: of really moving beyond what is really its bread and butter,
[8:39] Speaker 8: which is Instagram, Facebook, WhatsApp, That's still the core part
[8:42] Speaker 8: of its business, but I don't think moving forward it
[8:45] Speaker 8: wants to be so associated with social media in general.
[8:50] Speaker 4: This week the world said Muse is really good. It's
[8:53] Speaker 4: topping the charts on the Apple Store free downloads chart.
[8:57] Speaker 1: It moved markets.
[9:00] Speaker 4: I've been playing with it a lot, trying to get
[9:02] Speaker 4: some actual useful stuff. There's some limits on what you
[9:04] Speaker 4: can get into through APIs. Is Meta good at personal AI?
[9:08] Speaker 9: Yeah.
[9:10] Speaker 8: Yeah, I mean, I think coming out of the gate,
[9:13] Speaker 8: it has it has gotten a lot of great feedback
[9:15] Speaker 8: and positive feedback. I think, you know, I think Meta
[9:19] Speaker 8: over the past few years has gotten a lot of
[9:20] Speaker 8: flack for and criticism for its AI strategy. This seems
[9:23] Speaker 8: to potentially be.
[9:24] Speaker 1: A turning point.
[9:25] Speaker 8: But I think, you know, I think we all need
[9:27] Speaker 8: to pump the brakes a little bit.
[9:29] Speaker 4: I'm going to jump in here because I think we've
[9:31] Speaker 4: lost Linda's shot for a second.
[9:33] Speaker 1: I think you might be back. Oh, technology. Am I
[9:37] Speaker 1: still there? Yes, you are. I think I'm still here.
[9:39] Speaker 5: Okay.
[9:40] Speaker 8: Anyway, I was saying that pretty much I think we
[9:43] Speaker 8: need to kind of pump the brakes a little bit.
[9:45] Speaker 8: I mean, we saw it only came out a few
[9:48] Speaker 8: weeks ago, and I'm going to draw a parallel to
[9:51] Speaker 8: last year when Vibes and Sora came out, which were
[9:54] Speaker 8: the social AI video features. There was a lot of
[9:58] Speaker 8: hype around that. What does this all mean? And then now,
[10:01] Speaker 8: I mean, Sora doesn't even exist anymore. And Vibes, I mean,
[10:04] Speaker 8: I never hear anything about Vibes anymore. And so this
[10:08] Speaker 8: does seem.
[10:08] Speaker 1: To be a little bit different.
[10:10] Speaker 8: But I think at this point, I think There are
[10:13] Speaker 8: some clear use cases that people are finding valuable, but
[10:16] Speaker 8: I think there is also a world where a lot
[10:18] Speaker 8: of this kind of goes awry or is somewhat of
[10:21] Speaker 8: a fad, to be honest.
[10:23] Speaker 4: It's interesting because actually for much of the pre-market, Meta
[10:26] Speaker 4: was down. It's now up almost 3%, whether that's directly
[10:29] Speaker 4: correlated to last night or not. Look, when Ray-Ban Metas
[10:32] Speaker 4: came out, Everyone was a bit surprised. Even in my reporting,
[10:36] Speaker 4: Meta was surprised at the success of that. Now they
[10:40] Speaker 4: have a range of devices that they believe are the
[10:43] Speaker 4: right form factors to interact with AI. At the end
[10:48] Speaker 4: of the day, have they got it right or not?
[10:51] Speaker 1: Yeah, I think it's hard to say at this point.
[10:54] Speaker 8: I'm not completely sold that smart glasses are the future.
[10:57] Speaker 8: I mean, it's interesting. I was even thinking about this
[10:59] Speaker 8: last night with the announcement of Charm. I mean, Meta
[11:02] Speaker 8: has put so much stock in smart glasses, right? And
[11:05] Speaker 8: a lot of these companies have, saying they're the future.
[11:07] Speaker 8: They're going to be the new smartphone. It's going to
[11:09] Speaker 8: be the way we interact with technology. And then to
[11:12] Speaker 8: kind of come out of left field with Charm, this
[11:15] Speaker 8: Tamagotchi-esque device, it's sort of like, wait, I thought smart
[11:18] Speaker 8: glasses were the next big thing. And I think there's
[11:20] Speaker 8: a lot of questions around like, well, wait, what do
[11:22] Speaker 8: I use my phone for versus my smart glasses versus
[11:24] Speaker 8: my charm if I want to buy that? So I
[11:27] Speaker 8: do wonder if by leaning into this experimentation with the
[11:30] Speaker 8: hardware so heavily, they're kind of sending some mixed messages
[11:33] Speaker 8: to the market in terms of, well, how do you
[11:36] Speaker 8: use these different devices and what one is best for
[11:38] Speaker 8: certain tasks versus others?
[11:40] Speaker 1: And maybe just moving a bit.
[11:41] Speaker 5: Too fast, to be honest.
[11:42] Speaker 9: Yeah.
[11:43] Speaker 4: The question for me, we'll discuss it next time, is
[11:45] Speaker 4: what are you willing to pay for, the device, or
[11:47] Speaker 4: are you willing to pay for the AI? Because it's good.
[11:49] Speaker 4: That will be the question going forward. Minda Smiley, Senior
[11:51] Speaker 4: Analyst at eMarketer, thank you very much indeed. Coming up,
[11:54] Speaker 4: AI tops the agenda as President Trump meets with China's
[11:57] Speaker 4: Xi Jinping. Why Trump says new guardrails are off the table.
[12:02] Speaker 1: This is Bloomberg Tech.
[12:16] Speaker 10: I see two main categories of risk.
[12:19] Speaker 1: The first is misuse of AI models.
[12:21] Speaker 11: For example, misuse by bioterrorists to create biological weapons.
[12:25] Speaker 1: The second risk is loss of control.
[12:27] Speaker 10: We need to understand what these systems are doing and
[12:30] Speaker 10: have strong evidence that they will do what people intend,
[12:32] Speaker 10: even as they get very, very smart. If managed poorly,
[12:36] Speaker 10: I even believe that AI could be a risk to
[12:38] Speaker 10: humanity as a whole. It doesn't matter whether people put
[12:41] Speaker 10: their risk of catastrophe at 10% or 1% or 12%
[12:44] Speaker 10: or 0.1%. None of these levels are remotely acceptable.
[12:47] Speaker 12: We called for global cooperation across the world between governments
[12:52] Speaker 12: to set international standards.
[12:54] Speaker 10: As AI becomes more capable, people must remain at the
[12:57] Speaker 10: center of AI decision-making.
[13:01] Speaker 4: That was Anthropic CEO Dario Amodei and OpenAI CEO Sam
[13:04] Speaker 4: Altman addressing the world leaders at the UN stage. But
[13:07] Speaker 4: President Trump is pushing back on those calls for global cooperation.
[13:12] Speaker 4: Ahead of his meeting with China's Xi Jinping, Trump says
[13:14] Speaker 4: superintelligence will be a big topic of discussion, but that
[13:18] Speaker 4: he wants to, quote, leave it exactly where it is, adding,
[13:22] Speaker 4: our guardrail is the DOJ. Bloomberg's Tyler Kendall with us
[13:26] Speaker 4: for more. Okay, that's the scene setter. The president quite
[13:30] Speaker 4: clearly has his position on how AI should be overseen.
[13:36] Speaker 5: Right.
[13:36] Speaker 7: And at this point, it's very clear that the White
[13:38] Speaker 7: House is prioritizing competition, especially when it comes to China,
[13:41] Speaker 7: over any sort of proposed guardrails, saying in this post
[13:44] Speaker 7: on Truth Social that we shouldn't expect any announcement related
[13:48] Speaker 7: to safeguards during this summit. Of course, that does echo
[13:51] Speaker 7: what we heard from the president in his speech to
[13:52] Speaker 7: the U.N. General Assembly earlier this week, seeming to reject
[13:55] Speaker 7: the idea. that we could see some sort of international
[13:58] Speaker 7: pact about how to regulate the technology going forward. Both
[14:01] Speaker 7: sides here today acknowledging the fierce competition when it does
[14:05] Speaker 7: come to AI development. We heard from Chinese President Xi
[14:08] Speaker 7: Jinping himself saying that China and the U.S. are capable
[14:11] Speaker 7: to manage AI on their own, going on to add
[14:14] Speaker 7: that the countries have the ability to, quote, ensure the
[14:17] Speaker 7: development of AI is always under human control. He said
[14:20] Speaker 7: competition should be positive, not winning or losing. We'll have
[14:24] Speaker 7: to see where the dialogue goes from here. We know
[14:26] Speaker 7: that this White House has already floated the idea that
[14:28] Speaker 7: we could see some sort of hotline mechanism established that
[14:32] Speaker 7: would essentially keep the line of communication open between the U.S.
[14:34] Speaker 7: Treasury Secretary Scott Besant and his Chinese counterpart if issues
[14:38] Speaker 7: were to arrive. I don't have to tell you, though,
[14:41] Speaker 7: that there are so many issues underpinning this conversation when
[14:44] Speaker 7: it comes to the two economies, including national security agencies
[14:48] Speaker 7: finding here in the U.S. that Chinese firms are conducting
[14:51] Speaker 7: these distillation attacks. We do know export curbs on critical
[14:55] Speaker 7: minerals remain a top issue for this meeting, as well
[14:57] Speaker 7: as the U.S.
[14:58] Speaker 6: 's own curbs when it.
[14:59] Speaker 7: Comes to high-tech exports. Ed, at this point, we are
[15:03] Speaker 7: going to see it front and center for the state
[15:05] Speaker 7: dinner tonight. As you well know, there is a long
[15:07] Speaker 7: list of CEOs that will be in attendance, including NVIDIA's
[15:11] Speaker 7: Jensen Wong.
[15:12] Speaker 1: The most Tyler Kendall back in Washington, D.C. from the U.N.
[15:15] Speaker 4: General Assembly yesterday in New York City. Thank you very much.
[15:18] Speaker 4: Let's stay with those U.S.-China tensions. Reva Goujon, director at
[15:22] Speaker 4: Rhodium Group, joins us. And look, in the AI context,
[15:26] Speaker 4: China is still an economic adversary. How often do we
[15:29] Speaker 4: compare the performance of a Chinese model to a U.S.
[15:33] Speaker 4: model benchmark against each other? How would that get reconciled
[15:36] Speaker 4: in this meeting?
[15:39] Speaker 13: Well, it's not going to be reconciled as such, right?
[15:41] Speaker 13: We already have a very brief AI dialogue, if you
[15:44] Speaker 13: can call it a dialogue. It's really a commitment to
[15:47] Speaker 13: set up a regular mechanism for talks. But there is
[15:51] Speaker 13: a lot of tension still feeding into the summit and
[15:54] Speaker 13: now moving into this sort of mini extension of the
[15:58] Speaker 13: truce toward January 10th. And so as Chinese open weight
[16:03] Speaker 13: models are just on the heels of U.S. frontier labs,
[16:08] Speaker 13: there's still a lot of anxiety over just how close
[16:11] Speaker 13: this competition is. And as this winds into the pacing
[16:16] Speaker 13: frontier debate as well, that inevitably swings back to the
[16:20] Speaker 13: export control debate.
[16:22] Speaker 4: Okay, so what's happened in the last 24 hours is
[16:24] Speaker 4: the frontier labs in calling for pacing at the frontier
[16:27] Speaker 4: have also said there needs to be global cooperation with
[16:30] Speaker 4: government involvement and independent evaluation. The President of the United
[16:34] Speaker 4: States has a very different approach.
[16:36] Speaker 1: Let's listen.
[16:37] Speaker 9: I'm not going to stifle growth of something that will
[16:40] Speaker 9: be bigger than the Industrial Revolution, but we will only
[16:43] Speaker 9: encourage superintelligence.
[16:46] Speaker 2: We're going to encourage it, not... We're going to watch
[16:49] Speaker 2: it closely through the Department of Justice.
[16:53] Speaker 4: The president also said that they were against what he
[16:56] Speaker 4: called a global scheme to control AI. What is the
[17:00] Speaker 4: net result when you have an administration with one point
[17:03] Speaker 4: of view and private industry pursuing a different avenue?
[17:07] Speaker 13: Yeah, here's the dynamic. It's basically a White House that
[17:10] Speaker 13: has pursued a very laissez-faire approach toward AI regulation. And
[17:16] Speaker 13: the advent of highly powerful models is forcing the hand
[17:20] Speaker 13: to say, well, there have to be some kind of
[17:23] Speaker 13: guardrails when we have these rogue AI incidents proliferating by
[17:27] Speaker 13: the day. And so there's no singular point of view
[17:31] Speaker 13: in the White House on this. You can see The
[17:34] Speaker 13: president is animated by the data center moratorium, the idea
[17:40] Speaker 13: of a hoax, the idea that there are folks out
[17:43] Speaker 13: there trying to halt U.S.
[17:45] Speaker 5: Progress.
[17:46] Speaker 13: And so clearly that's where he's leaning into this and saying, look,
[17:50] Speaker 13: we don't need to lean in with more restrictions. We
[17:52] Speaker 13: need to let our companies run. On the other hand,
[17:55] Speaker 13: you have Scott Besant, who is holding the reins of
[17:59] Speaker 13: AI policy quite tightly. He led this dialogue with the Chinese,
[18:03] Speaker 13: and he does have some serious conviction about where AI safety, guardrails, coordination,
[18:09] Speaker 13: kind of rational third-party testing, things like this do need
[18:13] Speaker 13: to be seriously considered. So we'll need to see where
[18:17] Speaker 13: the influencers fall on this, but it'll be very start-stop.
[18:22] Speaker 1: Reva, let's talk more about that.
[18:23] Speaker 4: The reporting is that the president will appoint a new
[18:26] Speaker 4: AI czar, a new AI task force. Kind of what
[18:31] Speaker 4: I hear is that they will look to someone who
[18:33] Speaker 4: has no company affiliation. You just said that Treasury Secretary
[18:37] Speaker 4: Besant has the reins. What do you mean by that
[18:40] Speaker 4: more explicitly, please?
[18:42] Speaker 13: Well, it's interesting. Dual hatting, triple hatting, quadruple hatting has
[18:47] Speaker 13: been very common in this administration. Besant certainly has his
[18:50] Speaker 13: hands full keeping bond markets calm while driving AI policy.
[18:55] Speaker 13: But since the exit of AI czar David Sachs late
[18:59] Speaker 13: last year, you've seen Besant lean in more on AI policy, again,
[19:04] Speaker 13: on this US-China dialogue as well. When it comes to
[19:08] Speaker 13: vetting models and having these discussions with Frontier Labs, he's
[19:11] Speaker 13: heavily involved. So you do see him kind of holding
[19:15] Speaker 13: this portfolio. Whether that makes him AI czar, we'll have
[19:18] Speaker 13: to see. But it's that camp that's going to be
[19:22] Speaker 13: important to test with Besant, with Kratios at OSTP, with
[19:28] Speaker 13: Jacob Helberg at State. And this is where Some pragmatic
[19:33] Speaker 13: guardrails could could come into play where this gets thorny again, though,
[19:38] Speaker 13: is when it gets into to export controls, because that's
[19:41] Speaker 13: what inflames the U.S. China track. But also where you
[19:44] Speaker 13: see some companies that are making the argument to shore
[19:47] Speaker 13: up US defenses, don't do anything to further enable Chinese
[19:51] Speaker 13: capabilities when we're seeing such powerful model capabilities deployed.
[19:56] Speaker 4: Reva, AI is not the only thing on the table. Minerals,
[20:01] Speaker 4: global security, the issue of Taiwan. Where can the US
[20:04] Speaker 4: make some progress?
[20:07] Speaker 13: So maybe the least interesting part of this summit is
[20:10] Speaker 13: where they fall on tariffs because there's a consensus on that.
[20:13] Speaker 13: The effective U.S. tariff rate on most Chinese goods is
[20:16] Speaker 13: around 20%. That seems to be pretty much locked in.
[20:21] Speaker 13: Then on Taiwan, this is really where we're looking at
[20:25] Speaker 13: Chinese quiet cooperation over Iran to try to get to
[20:28] Speaker 13: some sort of exit strategy, but it doesn't come for free.
[20:31] Speaker 13: So watch for Chinese demands over U.S. restraint on Taiwan
[20:36] Speaker 13: as it tries to neutralize a pending arms package. I
[20:38] Speaker 13: think the U.S. can get around that because the arms
[20:41] Speaker 13: deliveries are delayed anyway, given just depleted munitions and backlogs.
[20:46] Speaker 13: The really thorny part is still on technology measures, where
[20:50] Speaker 13: there are a number of pending measures coming out of
[20:52] Speaker 13: the FCC, blunt import restrictions, etc. We're seeing pending moves
[20:58] Speaker 13: out of commerce with a semiconductor 232 still kicking around.
[21:01] Speaker 13: That would apply a punitive tariff on China-sourced chips, as
[21:05] Speaker 13: well as other boundaries that are not invested in front-end
[21:09] Speaker 13: manufacturing in the U.S. And then you have a whole
[21:12] Speaker 13: package of export controls that are gearing up for the
[21:16] Speaker 13: National Defense Authorization Act end year 2021. And that is
[21:19] Speaker 13: exactly what Xi Jinping wants to neutralize during the summit.
[21:24] Speaker 13: And now that's going to be tested in this interim
[21:26] Speaker 13: with the kind of mini truce extension that we have.
[21:30] Speaker 1: Riva Goujon of the Rhodium Group, thank you very much
[21:32] Speaker 1: for your time.
[21:33] Speaker 4: Coming up, Qualcomm renews its global licensing pact with Apple.
[21:37] Speaker 1: A bit of a surprise story this morning. We've got
[21:38] Speaker 1: the deets. This is Bloomberg Tech. Okay, time for Talking Tech.
[21:47] Speaker 4: And first up, Thailand approved its first national chip, an
[21:51] Speaker 4: advanced electronic strategy. The country will target about $ 80 billion
[21:55] Speaker 4: in new investment by 2050, according to a board of
[21:58] Speaker 4: investment statement. And the plan aims to generate $ 150 billion
[22:02] Speaker 4: in annual industry revenue. Plus, AI video startup Higgs Field
[22:06] Speaker 4: is on track to generate annualized revenue of more than
[22:09] Speaker 4: $ 1 billion based on current performance. The rapid growth is
[22:12] Speaker 4: driven by strong demand from direct-to-consumer businesses, which are increasingly
[22:17] Speaker 4: adopting AI-generated videos. And Qualcomm has renewed its global licensing
[22:22] Speaker 4: pact with Apple, effective in April 2027, ensuring the iPhone
[22:26] Speaker 4: maker will remain a key customer in the near term.
[22:29] Speaker 4: As Qualcomm's largest customer data compiled by Bloomberg shows, Apple
[22:33] Speaker 4: accounts for more than a fifth of its revenue. Of course,
[22:35] Speaker 4: Apple doing its own modems in some cases and in
[22:38] Speaker 4: others still relying on Qualcomm for modem. Coming up, AI
[22:42] Speaker 4: agents gaining more autonomy inside the enterprise. But do we
[22:45] Speaker 4: have any idea what they're up to? This is Bloomberg Tech.
[22:56] Speaker 4: Welcome back to Bloomberg Tech. Oracle issuing a force majeure
[22:59] Speaker 4: notice to a developer on its New Mexico site is
[23:02] Speaker 4: a big story from Bloomberg. The stock's down 5% off
[23:05] Speaker 4: its session low, and there's a lot of sell-side reaction now.
[23:08] Speaker 4: So William Blair has a note saying this is logical
[23:11] Speaker 4: to issue force majeure after known delays on that New
[23:14] Speaker 4: Mexico site, 2.45 gigawatts. KeyBank has a note out explaining, like, look...
[23:20] Speaker 4: In this context, Oracle's the core tenant and they don't
[23:23] Speaker 4: want to have to start making payments as the core
[23:25] Speaker 4: tenant on that site if it's not up and operational.
[23:28] Speaker 4: Both of them reiterate the statement from Oracle that they're
[23:31] Speaker 4: committed to the site and actually it's going to plan.
[23:34] Speaker 4: Recap later in the hour on that. Meta's a big story.
[23:38] Speaker 4: The stock is now up significantly. And again, this is
[23:41] Speaker 4: all about the devices that were launched last night, MetaConnect.
[23:45] Speaker 4: We discussed it earlier in the program.
[23:47] Speaker 1: On the whole.
[23:48] Speaker 4: SellSide saying this is a positive conduit for the consumer
[23:53] Speaker 4: to use Meta's work in AI agents, which is through Muse.
[23:58] Speaker 4: AI fears are fueling a huge rally in cybersecurity stocks
[24:01] Speaker 4: as well. A basket of cybersecurity stocks tracked by Goldman
[24:04] Speaker 4: Sachs has more than doubled since its April low with CrowdStrike,
[24:08] Speaker 4: Palo Alto, Fortinet, all up more than 130%. But those
[24:12] Speaker 4: gains are also pushing valuations to extremes. Bloomberg Equities reporter
[24:16] Speaker 4: Ryan Veselica says, is tracking the trade.
[24:19] Speaker 1: What's going on in cyber?
[24:21] Speaker 14: Hey, well, it's been a big year for cybersecurity stocks,
[24:24] Speaker 14: especially relative to the rest of software. So while the
[24:27] Speaker 14: whole group came under a lot of pressure earlier this
[24:30] Speaker 14: year on concerns about AI disruption, there's been this growing
[24:33] Speaker 14: appreciation that the power of AI represents a potentially huge
[24:38] Speaker 14: opportunity for cybersecurity. We've seen all kinds of AI-assisted hacks
[24:43] Speaker 14: and data breaches and so forth, which really just drives
[24:46] Speaker 14: the idea that we're going to need a lot more
[24:48] Speaker 14: security services going forward. That's really powered the group higher
[24:52] Speaker 14: this year. The question is, have they just gotten ahead
[24:55] Speaker 14: of their skis? Are they too expensive now? Where does
[24:58] Speaker 14: the trade go from here? That's a valuation story. Take
[25:02] Speaker 14: me back to the first half of this year and
[25:04] Speaker 14: the SaaSpocalypse. Were the cybersecurity names caught up in that,
[25:08] Speaker 14: in the idea that they would be displaced by AI?
[25:12] Speaker 14: there was a lot of indiscriminate selling. Now, this group
[25:15] Speaker 14: was able to rebound faster than a lot of the
[25:17] Speaker 14: rest of the space. We've only started to see other
[25:19] Speaker 14: parts of the software ecosystem rebounding over the past couple
[25:23] Speaker 14: of months, especially helped by the most recent earnings season.
[25:25] Speaker 14: Did seem like people never quite soured on cybersecurity in
[25:28] Speaker 14: the same way, but things were so widespread, the selling,
[25:32] Speaker 14: that they did really get caught up in that.
[25:35] Speaker 1: Bluebears Ryan for Celica. Thank you very much.
[25:37] Speaker 4: An open AI model hacked an Australian government website earlier
[25:40] Speaker 4: this year, gaining unauthorized access to health care statistics files.
[25:45] Speaker 4: Officials say there's no evidence personal information was compromised, but
[25:48] Speaker 4: open AI took three months to notify the government. Australia's
[25:52] Speaker 4: deputy prime minister called the breach utterly unacceptable. AI is
[25:57] Speaker 4: rapidly shifting from a tool employees use to an autonomous
[26:00] Speaker 4: actor inside the enterprise, creating an entirely new security challenge.
[26:06] Speaker 4: Darktrace says companies often know where AI has been deployed,
[26:10] Speaker 4: but not necessarily what those agents are doing in real time.
[26:13] Speaker 4: Darktrace CEO Ed Jennings joins us as the company launches
[26:16] Speaker 4: new tools to track everything from shadow AI to agent
[26:20] Speaker 4: identities and behavior.
[26:21] Speaker 1: And this is fascinating.
[26:23] Speaker 4: A lot of people would say this is in the
[26:24] Speaker 4: bucket of observability, but you're saying we sometimes don't know
[26:28] Speaker 4: what it is an agent's up to.
[26:31] Speaker 2: Well, Ed, thank you for having me. We do know,
[26:33] Speaker 2: we can see.
[26:34] Speaker 15: You know, Darktrace for the last 10 years has been
[26:37] Speaker 15: both detecting, both observing, and actually remediating using AI itself,
[26:43] Speaker 15: different actions.
[26:44] Speaker 2: I like your comparison.
[26:45] Speaker 15: We've oftentimes said that the agents are this new insider
[26:49] Speaker 15: threat because it's more of watching what they're doing.
[26:51] Speaker 2: When do they behave unnaturally?
[26:53] Speaker 15: When do they start doing things that are beyond what
[26:55] Speaker 15: they should be doing? And that's something that we've been
[26:57] Speaker 15: observing with our own AI capability.
[26:59] Speaker 2: For quite a while.
[27:01] Speaker 4: Ed, for a very broad audience, this week it became
[27:03] Speaker 4: very real. If the consumer sees the benefit of using
[27:07] Speaker 4: Muse or Grokbot because they can actually do something useful,
[27:11] Speaker 4: how much do people now need to start thinking about, well,
[27:13] Speaker 4: what if that agent behaves in a way it was
[27:15] Speaker 4: not intended to?
[27:16] Speaker 15: I think they very much do. We work with enterprises.
[27:21] Speaker 15: many of which are in financial services. And nobody is,
[27:24] Speaker 15: in all the discussion of what's slowing down on sort
[27:26] Speaker 15: of the advancements and innovation and models themselves, nobody is
[27:30] Speaker 15: slowing down on the use of these technologies within their companies.
[27:33] Speaker 15: We see even more and more push for automation, for simplification,
[27:37] Speaker 15: for using these tools to to do new ways of
[27:39] Speaker 15: doing business. And so, but our CISO, our Chief Information
[27:42] Speaker 15: Security Officer customers are behind the curve.
[27:45] Speaker 2: They're trying to catch up.
[27:46] Speaker 15: It feels a little bit like the cloud rush of
[27:48] Speaker 15: a few years back when cloud innovation got ahead of
[27:51] Speaker 15: security and security had to catch up to it.
[27:53] Speaker 2: And it feels like we're in one of those moments
[27:54] Speaker 2: right now.
[27:56] Speaker 1: Did you ever watch the show Silicon Valley?
[27:58] Speaker 2: Yeah, well, of course I watched the show Silicon Valley.
[28:00] Speaker 4: Yeah, so for lots of people, this is like son
[28:02] Speaker 4: of Anton. You have Son of Anton, you instruct it
[28:05] Speaker 4: to do something, and.
[28:07] Speaker 1: The reward is not correctly specified.
[28:09] Speaker 4: So you come back and it's deleted your entire code
[28:11] Speaker 4: base and ordered 4,000 pounds of hamburger meat. You know,
[28:15] Speaker 4: is that the reality of what we're talking about here?
[28:18] Speaker 2: Yeah, very much.
[28:18] Speaker 15: I mean, because of our perspective, we just are today
[28:21] Speaker 15: announcing the AI labs, our signal labs. And it's not
[28:27] Speaker 15: the traditional threat intelligence type lab.
[28:28] Speaker 2: It's specifically built.
[28:30] Speaker 15: in a secure environment, sort of under the protection of
[28:32] Speaker 15: our own capabilities, to look at models and look at agents.
[28:36] Speaker 2: And what do they do when they're pushed? What do
[28:39] Speaker 2: you do when you ask.
[28:40] Speaker 15: Them to do things that are not necessarily within the,
[28:42] Speaker 15: how do you force misalignment? And we found some very
[28:45] Speaker 15: interesting things when we ran in this environment. And maybe
[28:48] Speaker 15: not 4,000 pounds of hamburger meat, but definitely agents acting
[28:53] Speaker 15: way above and beyond what they were intended to do.
[28:56] Speaker 4: Your job is to convince customers to spend, to make
[29:00] Speaker 4: sure that they have the technology in place to protect,
[29:04] Speaker 4: that agents do what they were designed to do. How
[29:08] Speaker 4: receptive is your customer base to this idea? How are
[29:10] Speaker 4: they weighing the economic benefit of managing a fleet of
[29:14] Speaker 4: agents versus the need to spend to protect against the risks?
[29:18] Speaker 15: Yeah, having been in cyber for a long time, I mean,
[29:20] Speaker 15: this is the most sort of demand signals I've ever
[29:22] Speaker 15: remember seeing. And we talk about that amongst our team.
[29:25] Speaker 15: People are open, they're curious, they're trying to understand what
[29:28] Speaker 15: they can do. They know that the traditional approaches of
[29:30] Speaker 15: signatures and looking for known malware, it does not work
[29:34] Speaker 15: when agents don't conform to that. And they're dynamically finding
[29:37] Speaker 15: new attacks and new new paths and new sometimes unintended
[29:40] Speaker 15: they're not always malicious sometimes they're just reckless in their
[29:43] Speaker 15: behavior and so the demand market i think why earlier
[29:46] Speaker 15: you were talking about the rise of cyber stocks and
[29:48] Speaker 15: so much investment going into to different private companies in
[29:51] Speaker 15: the space because it is a tide rising the boats moment.
[29:54] Speaker 4: I've posed this question to the sell side and to
[29:56] Speaker 4: the Frontier Labs, and now I'm posing it to you,
[29:58] Speaker 4: which is what is the greater risk? The model behaves
[30:02] Speaker 4: in a way that it wasn't intended to, misalignment, or
[30:05] Speaker 4: that it falls into the hands of bad actors, which
[30:07] Speaker 4: is traditionally what cybersecurity defends against.
[30:10] Speaker 2: To be honest, I think we have to think about both.
[30:13] Speaker 15: We talk about building safe models and we talk about
[30:16] Speaker 15: operating safe models. We're mostly helping our company's operate safe models.
[30:20] Speaker 15: But there is a long legacy in cyber of governments
[30:23] Speaker 15: working together, something like Five Eyes or the UK and
[30:25] Speaker 15: the US. There's many examples, even competitors in financial services
[30:29] Speaker 15: building associations to share threat intelligence, to share indications of compromise.
[30:33] Speaker 15: And there's a lot of examples of government working with
[30:35] Speaker 15: private industry to make.
[30:36] Speaker 2: Sure we're in front of these things.
[30:37] Speaker 15: I think as we more mature in this category of
[30:40] Speaker 15: cyber for AI, you're gonna see those kind of alignments
[30:43] Speaker 15: I think are important.
[30:43] Speaker 2: I think it's gonna keep us all of us safer.
[30:47] Speaker 1: Very quick.
[30:48] Speaker 4: Your consideration of existential threat to human life, where does
[30:51] Speaker 4: Ed Jennings sit?
[30:54] Speaker 15: I think that some of what we're seeing now is immaturity.
[30:56] Speaker 15: I think some of the way these tests were done,
[30:58] Speaker 15: some of these elements of what we've seen in breaking
[31:01] Speaker 15: sort of sandboxes could have been somewhat avoided. I do
[31:04] Speaker 15: think these models need to mature. And they need to
[31:07] Speaker 15: think that security is part of their design construct. I
[31:09] Speaker 15: think some of the things we're finding in our own
[31:11] Speaker 15: labs would be fairly easily avoidable controls that could be
[31:15] Speaker 15: put in place.
[31:15] Speaker 2: And I think let's start with those.
[31:18] Speaker 15: Let's start with a behavioral approach to security where you're
[31:20] Speaker 15: really watching in real time and monitoring these. And then
[31:22] Speaker 15: I think we can take advantage of all that AI
[31:24] Speaker 15: has to bring.
[31:26] Speaker 4: Darktrace CEO, Ed Jennings, thank you very much. Coming up,
[31:29] Speaker 4: China's deep-seek detailed and innovative method for training AI agents.
[31:33] Speaker 4: We'll discuss the details and the potential impact next. This
[31:37] Speaker 4: is Bloomberg Tech. Microsoft says AI safety can't be left
[31:48] Speaker 4: to just one or two companies. President Brad Smith is
[31:51] Speaker 4: backing independent evaluators and says safeguards need to be built
[31:55] Speaker 4: across the entire AI ecosystem. I spoke with Brad Smith
[31:58] Speaker 4: yesterday about who should be responsible as AI gets more powerful.
[32:04] Speaker 11: Well, if we think about AI safety, and it is
[32:06] Speaker 11: just a critical topic, I think the first thing we
[32:08] Speaker 11: should recognize is we're not going to advance safety if
[32:12] Speaker 11: we rely on only one or two companies or one
[32:14] Speaker 11: slice of the AI sector. The model companies are obviously critical.
[32:19] Speaker 11: But so are companies like ours that produce the layer
[32:23] Speaker 11: of software that controls AI agents, that monitors what they do,
[32:28] Speaker 11: that lets companies manage not just what they're spending, but
[32:32] Speaker 11: where they're going and where they're not. And there's also,
[32:35] Speaker 11: I think, an important role for hyperscalers. Again, a company
[32:39] Speaker 11: like Microsoft plays that role. And you really want safety
[32:43] Speaker 11: to be built in in multiple layers. At times, you
[32:46] Speaker 11: may want a government to be able to have the
[32:48] Speaker 11: on-off switch. But why would you put that switch in
[32:53] Speaker 11: only one place when you can put it in three
[32:55] Speaker 11: or four? The analogy I would offer is think about
[33:00] Speaker 11: what happened when cars reached the point where they were
[33:03] Speaker 11: going fast enough that there were starting to be collisions
[33:06] Speaker 11: in an intersection. So what did people do?
[33:09] Speaker 2: They said, well, wait a second.
[33:11] Speaker 11: We need stoplights, stop signs, better pavement on the roads, guardrails,
[33:16] Speaker 11: speed limits, a police force, and a speedometer in the car.
[33:20] Speaker 11: But of that list I just had, the only thing
[33:23] Speaker 11: that changed the car was the speedometer. Whenever things are
[33:28] Speaker 11: going too fast, you almost always need to look at
[33:31] Speaker 11: the ecosystem as a whole, ensure it's robust and healthy,
[33:35] Speaker 11: and bring the ecosystem together to put in place the
[33:38] Speaker 11: safety framework that's needed.
[33:41] Speaker 4: The ecosystem idea is interesting. You know, what's happening right
[33:45] Speaker 4: now as you and I speak is Sam Altman is
[33:47] Speaker 4: briefing the United Nations Security Council, right? And that has
[33:50] Speaker 4: dominated the news cycle, the point of view of the
[33:53] Speaker 4: frontier labs of not just pacing the frontier, but what
[33:56] Speaker 4: they believe should happen in terms of independent evaluation, et cetera.
[34:00] Speaker 4: You know, Microsoft is OpenAI's leader. most important strategic partner.
[34:05] Speaker 4: You have massive economic exposure to OpenAI as well. Does
[34:10] Speaker 4: Microsoft get to be a part of those conversations in
[34:13] Speaker 4: terms of what the frontier labs will do, what they
[34:15] Speaker 4: advocate for in pacing the frontier?
[34:18] Speaker 2: Well, I would say a few things.
[34:19] Speaker 11: First, yes, Microsoft is a part of that conversation by
[34:22] Speaker 11: definition because there's a joint deployment safety board that includes
[34:28] Speaker 11: OpenAI and Microsoft people that have to weigh in together
[34:33] Speaker 11: before the next major model, for example, is released. So indisputably,
[34:39] Speaker 11: we're part of that and part of that decision-making. And
[34:43] Speaker 11: I give a lot of credit both to Sam Altman,
[34:46] Speaker 11: to Dario. They're providing information today, and I think they're
[34:49] Speaker 11: asking many of the right questions. We should have independent evaluators.
[34:54] Speaker 11: They need to be capable evaluators. They need to have
[34:57] Speaker 11: access to the right data. Ultimately, the world has to
[35:00] Speaker 11: have confidence in these evaluators. They better be chosen properly.
[35:04] Speaker 11: They need to have genuine independence.
[35:09] Speaker 4: That was Microsoft Vice Chair and President Brad Smith. As
[35:12] Speaker 4: AI companies increasingly turn to outside groups to test their models,
[35:16] Speaker 4: some in Washington are raising questions about just how independent
[35:20] Speaker 4: those evaluators really are. Bloomberg's Maggie Eastland has the reporting,
[35:23] Speaker 4: and the concern is on ideological bias.
[35:26] Speaker 1: Explain it.
[35:29] Speaker 16: Yeah, so what we found in our reporting is when
[35:31] Speaker 16: you actually take a look at the ecosystem of groups,
[35:35] Speaker 16: mainly nonprofits, that are capable of evaluating and assessing AI models,
[35:40] Speaker 16: a lot of them have at one time or another
[35:42] Speaker 16: received money from groups that are tied to this effective
[35:45] Speaker 16: altruist movement. Now that's something that officials here in Washington
[35:49] Speaker 16: have been extremely skeptical of and averse to, particularly because
[35:54] Speaker 16: this movement for many years has championed AI safety and
[35:57] Speaker 16: in particular talked about potentially the need to slow down
[36:00] Speaker 16: or even stop AI if it gets out of human control.
[36:04] Speaker 4: You just heard from Microsoft President Brad Smith, right? The
[36:06] Speaker 4: world needs to have confidence in those independent evaluators, is
[36:09] Speaker 4: what he said. But take us to the other side
[36:11] Speaker 4: of the table, which is the administration, existing regulatory bodies.
[36:15] Speaker 4: They are concerned about that ideological bias.
[36:19] Speaker 2: Absolutely, yes.
[36:20] Speaker 16: They do have concerns, but at the same time, there's
[36:23] Speaker 16: so much interest and pressure and just an attitude that
[36:28] Speaker 16: there needs to be more security around AI. And even
[36:33] Speaker 16: for companies that are deploying AI, like Target and Mount
[36:37] Speaker 16: Sinai Health System we reference in our reporting, are interested
[36:40] Speaker 16: in AI standards because no one who's deploying AI wants
[36:43] Speaker 16: to have an agent go and do.
[36:44] Speaker 5: Something that was unauthorized.
[36:45] Speaker 16: So there's so much attention around standards that The Center
[36:49] Speaker 16: for AI Innovation in the Commerce Department, they do some
[36:54] Speaker 16: model evaluations themselves, but they've also been meeting with some
[36:56] Speaker 16: of these outside groups, including METER, one that we reference
[36:59] Speaker 16: in our reporting, to talk about AI standards. Again, though,
[37:03] Speaker 16: the problem is going to be how well officials certify
[37:06] Speaker 16: that these evaluators are actually independent.
[37:10] Speaker 4: Bloomberg's Maggie Eastlander, along with Bloomberg's Rachel Metz. That report
[37:14] Speaker 4: you should read on the Bloomberg terminal or Bloomberg.com. Thank
[37:16] Speaker 4: you very much. China's deep seek has laid out a
[37:18] Speaker 4: new method for training AI agents that could allow them
[37:22] Speaker 4: to learn more efficiently while minimizing the kind of misbehavior
[37:26] Speaker 4: that has fueled global concerns. Bloomberg's executive editor, Peter Elstrom,
[37:30] Speaker 4: has the details.
[37:31] Speaker 1: Fascinating. Explain this to us. Hi, Ed.
[37:35] Speaker 12: Yeah, DeepSeek, of course, made its name at the beginning
[37:38] Speaker 12: of last year when it came out with this reasoning model, R1,
[37:43] Speaker 12: that performed as well as some of the Silicon Valley models,
[37:46] Speaker 12: but it was done with much, much fewer resources. So
[37:49] Speaker 12: the research that comes out at DeepSeek is very closely tracked.
[37:52] Speaker 5: In this case, they dropped.
[37:54] Speaker 12: A 10,000-word research paper with something like 130 co-authors, including
[37:59] Speaker 12: the CEO, the founder of the company, And what the
[38:04] Speaker 12: paper says is, as you alluded to, a couple of
[38:07] Speaker 12: different things that are important. One, it's a platform for
[38:11] Speaker 12: training AI agents in a way that they believe is
[38:14] Speaker 12: more efficient than what's out there right now. It's called
[38:17] Speaker 12: DeepSeek Elastic Compute. And the idea is that they can
[38:21] Speaker 12: set up these sandboxes to train agents, give them some
[38:24] Speaker 12: tasks to practice, and multiply those sandboxes quite a bit
[38:27] Speaker 12: and use their algorithm to be able to allocate computing
[38:31] Speaker 12: resources so that you can do as many as 3
[38:33] Speaker 12: million sandboxes during a single day. So these agents can
[38:37] Speaker 12: train and move on to new sandboxes. At the same time,
[38:40] Speaker 12: they want to impose some kind of safety restrictions. They
[38:43] Speaker 12: detail quite a bit of misbehavior that the agents have
[38:46] Speaker 12: been involved in, including in their training practices, like corrupting
[38:50] Speaker 12: files or jumping outside of the sandboxes in certain ways.
[38:53] Speaker 12: So as they are trained in these sandboxes, they get
[38:56] Speaker 12: tighter and tighter with their restrictions, and the monitoring gets
[39:00] Speaker 12: more severe, too. So this is a paper that they've
[39:02] Speaker 12: dropped at this point. They explain it in a lot
[39:04] Speaker 12: of detail. These are kind of the highlights.
[39:06] Speaker 2: At this point.
[39:07] Speaker 4: What does the market for real AI agents look like
[39:10] Speaker 4: in China and in Asia more broadly?
[39:13] Speaker 12: Well, they adopted agents quite early. They were big adopters
[39:18] Speaker 12: of OpenClaw, for example. And you've seen some of the
[39:22] Speaker 12: big tech giants really adopt these kinds of practices and
[39:25] Speaker 12: see whether they can implement them. They have a number
[39:28] Speaker 12: of different technology models.
[39:29] Speaker 1: That we don't have.
[39:30] Speaker 12: For example, Tencent has infused its WeChat messaging service with agents.
[39:37] Speaker 12: That's been quite popular at this point, and they're adopting
[39:39] Speaker 12: them in other areas. Certainly, DeepSeq wants to be a
[39:42] Speaker 12: player there. And you see companies like ByteDance and Manus,
[39:46] Speaker 12: of course, came out of China. That was the company
[39:48] Speaker 12: that Meta owned briefly before they had to return it
[39:52] Speaker 12: and make it independent. So Manus is another big player
[39:55] Speaker 12: in agentic AI in China. It's been quite successful.
[39:58] Speaker 4: Bloomberg's Peter Elstrom, who's in London, but leads our coverage
[40:01] Speaker 4: of technology companies and the field in Asia.
[40:04] Speaker 1: Great to have you back on the show. Thank you
[40:05] Speaker 1: very much.
[40:06] Speaker 4: Now, coming up, we're going to go back to Washington
[40:08] Speaker 4: and get the latest on the meeting between President Trump
[40:11] Speaker 4: and China's Xi Jinping. All the headlines you need. This
[40:14] Speaker 4: is Bloomberg Tech. Chinese President Xi Jinping is at the
[40:28] Speaker 4: White House holding a bilateral meeting with President Trump. Bloomberg's
[40:32] Speaker 4: Mike Shepard, our tech editor in Washington, D.C., is with us.
[40:36] Speaker 4: There have been so many headlines throughout the morning. Have
[40:40] Speaker 4: we heard anything on AI in particular?
[40:43] Speaker 5: Yes, we have.
[40:44] Speaker 9: You know, the president welcomed Xi Jinping about 10 a.m.
[40:48] Speaker 9: And they went through some pomp and circumstance, including a
[40:51] Speaker 9: military review and an exchange of gifts. And they also
[40:54] Speaker 9: deliver brief remarks at where President Donald Trump indicated that
[40:58] Speaker 9: they would touch on a range of sensitive issues, including
[41:01] Speaker 9: artificial intelligence, which, of course, he has in recent days
[41:05] Speaker 9: taken to calling superintelligence.
[41:07] Speaker 5: He didn't promise any significant breakthroughs.
[41:10] Speaker 9: Xi Jinping, during his statement, said that they would discuss
[41:13] Speaker 9: AI as well. He called for using AI for the
[41:17] Speaker 9: good and said that he thinks that artificial intelligence should
[41:21] Speaker 9: remain under human control. Now, early this morning, before Xi's
[41:25] Speaker 9: arrival at the White House, we did see that post
[41:28] Speaker 9: on Truth Social from President Trump indicating that the meeting
[41:31] Speaker 9: would not produce any new guardrails. This despite calls from
[41:35] Speaker 9: many in Silicon Valley, and a number of world leaders
[41:37] Speaker 9: who expressed the desire for wider safety standards on AI
[41:41] Speaker 9: during appearances at the United Nations General Assembly and at
[41:46] Speaker 9: the Security Council yesterday. So we'll have to watch, Ed,
[41:50] Speaker 9: what emerges from this meeting when it comes to AI,
[41:53] Speaker 9: if it really is just the dialogue between the two sides.
[41:56] Speaker 9: Even that would be something of a breakthrough for these rivals.
[41:59] Speaker 4: It's not the only thing on the agenda, right? And
[42:02] Speaker 4: if you look at the points of discussion we expect
[42:04] Speaker 4: between President Trump and President.
[42:07] Speaker 1: Xi, the U.S.
[42:07] Speaker 4: And China, loads of them are still relevant to the
[42:10] Speaker 4: technology industry. Trade in tariffs is something that has come
[42:13] Speaker 4: up in the program already. Give us the different buckets
[42:17] Speaker 4: that we hope to hear about.
[42:19] Speaker 9: Well, when we talk about trade, we have to think
[42:21] Speaker 9: first about the trade truce, which had been in effect
[42:25] Speaker 9: and was to expire in November. They've extended it further
[42:30] Speaker 9: into January, buying a couple of months. But then negotiators
[42:33] Speaker 9: from both sides will have to sit back down at
[42:36] Speaker 9: the drawing board yet again to make sure that each
[42:40] Speaker 9: side is being satisfied with the terms and whether they
[42:43] Speaker 9: are being carried out.
[42:44] Speaker 13: The U.S.
[42:44] Speaker 9: Is concerned in particular that China may not hold up
[42:48] Speaker 9: its end of the bargain again. on continuing to allow
[42:50] Speaker 9: the flow of rare earths to U.S. buyers and then
[42:54] Speaker 9: also abide by its promise to buy more U.S. soybeans
[42:57] Speaker 9: and other goods. So that is one key area, of course.
[43:01] Speaker 9: And then there is the matter of Taiwan. That is
[43:03] Speaker 9: something that China is sure to bring up. And it's
[43:05] Speaker 9: unclear exactly how the U.S. will position itself. But Taiwan
[43:09] Speaker 9: is this locus of chip manufacturing and a tech powerhouse.
[43:14] Speaker 9: And really any disturbance in the force surrounding Taiwan is
[43:17] Speaker 9: one that would be of great concern to markets and
[43:19] Speaker 9: policymakers here.
[43:20] Speaker 4: 15 seconds, Mike, but then there's the state dinner and
[43:24] Speaker 4: tech is represented.
[43:25] Speaker 5: Oh, absolutely.
[43:26] Speaker 9: There will be a star-studded cast of Silicon Valley leaders
[43:29] Speaker 9: ranging from Jensen Wang to Mark Zuckerberg. And of course,
[43:33] Speaker 9: they're going to be bringing their interests to the table.
[43:36] Speaker 9: But more importantly, it is just a moment for them
[43:39] Speaker 9: to be seen with the president during this moment of pageantry.
[43:44] Speaker 4: Bloomberg's Mike Sheppard. Thank you very much indeed. That does
[43:47] Speaker 4: it for this edition of Bloomberg Tech. Some really big
[43:49] Speaker 4: technology stories, some of them with a markets element to recap.
[43:53] Speaker 1: And you can do that on the podcast. Loads of you.
[43:56] Speaker 4: The feedback is you like the podcast. You listen to
[43:58] Speaker 4: it on your way to work, whatever it is. So
[43:59] Speaker 4: thank you. That's where you can find it four days
[44:02] Speaker 4: through the week. This is Bloomberg Tech.

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.