WEBVTT - BI's Mandeep Singh Talks DeepSeek and China's AI Environment

0:00:02.520 --> 0:00:05.439
<v Speaker 1>Bloomberg Audio Studios, podcasts.

0:00:05.760 --> 0:00:10.479
<v Speaker 2>Radio news man Deep Singh joins us, Right now, did

0:00:10.520 --> 0:00:14.280
<v Speaker 2>you know what deep seek was before twelve or eighteen

0:00:14.400 --> 0:00:15.040
<v Speaker 2>hours ago?

0:00:15.400 --> 0:00:18.160
<v Speaker 3>I did, yes, And so what did you know about

0:00:18.200 --> 0:00:19.280
<v Speaker 3>them two days ago?

0:00:20.079 --> 0:00:23.160
<v Speaker 1>That it is a competitor in the LM space. They've

0:00:23.160 --> 0:00:26.520
<v Speaker 1>been training their own model and it's used mostly in

0:00:26.560 --> 0:00:27.800
<v Speaker 1>the East Asian region.

0:00:27.960 --> 0:00:30.560
<v Speaker 2>Okay, what I learned, thank you zero head for this

0:00:30.600 --> 0:00:35.720
<v Speaker 2>from Morgan Brown at dropbox is basically they've said, we're

0:00:35.720 --> 0:00:38.199
<v Speaker 2>not going to be as perfect as pristine as the

0:00:38.280 --> 0:00:42.040
<v Speaker 2>others out to thirty two decimal points. We're just going

0:00:42.080 --> 0:00:47.560
<v Speaker 2>to get it done out to point eight decimal points.

0:00:47.680 --> 0:00:49.239
<v Speaker 2>Is that what this is all about is they're not

0:00:49.240 --> 0:00:50.120
<v Speaker 2>going to be perfect.

0:00:50.479 --> 0:00:52.760
<v Speaker 1>I think there's more to it than just you know,

0:00:53.000 --> 0:00:57.240
<v Speaker 1>using the floating point fewer floating point operations than open

0:00:57.280 --> 0:01:01.279
<v Speaker 1>AI and Entropic and others. And in this case, they

0:01:01.800 --> 0:01:06.679
<v Speaker 1>basically focused on hardware efficiency, using the hardware in the

0:01:06.720 --> 0:01:10.399
<v Speaker 1>most efficient fashion. They had the benefit of all these

0:01:10.600 --> 0:01:14.160
<v Speaker 1>lms being out there. They use meta Llama as a

0:01:14.200 --> 0:01:17.880
<v Speaker 1>reference point. It's another open source model, and they figured

0:01:17.880 --> 0:01:21.040
<v Speaker 1>out a way to do this most efficiently than everyone else.

0:01:21.080 --> 0:01:24.360
<v Speaker 1>I mean, everyone is focused on scale right now, they

0:01:24.560 --> 0:01:26.240
<v Speaker 1>focused on hardware efficiency.

0:01:26.240 --> 0:01:27.720
<v Speaker 3>Call the Lama meta.

0:01:27.760 --> 0:01:31.800
<v Speaker 2>It's Facebook is the grandson of the original Lama on

0:01:31.840 --> 0:01:32.360
<v Speaker 2>Saturday Night.

0:01:32.680 --> 0:01:36.280
<v Speaker 4>There you go, right, very good man, Deep You've got

0:01:36.280 --> 0:01:39.440
<v Speaker 4>an analyst Bloomberg Intelligence in Hong Kong, Robert Lee. I'm

0:01:39.480 --> 0:01:42.480
<v Speaker 4>reading his research literally as we speak. And for folks

0:01:42.880 --> 0:01:45.600
<v Speaker 4>on the Bloomberg terminal, go to big and check out

0:01:45.600 --> 0:01:46.920
<v Speaker 4>if you want to need to figure out what's going

0:01:46.959 --> 0:01:48.960
<v Speaker 4>on with Deep seek and what China's doing with AI.

0:01:49.520 --> 0:01:53.920
<v Speaker 4>Bloomerg Intelligence has got the data and the research there.

0:01:55.080 --> 0:01:59.520
<v Speaker 4>Can China support a real competitive AI environment? It seems

0:01:59.520 --> 0:02:04.000
<v Speaker 4>like they're government is more restricting some of the technology there.

0:02:04.360 --> 0:02:08.080
<v Speaker 1>Well, so now we know the export controls probably weren't

0:02:08.200 --> 0:02:11.480
<v Speaker 1>as effective as they were supposed to be here, you know,

0:02:11.560 --> 0:02:14.440
<v Speaker 1>in terms of restricting the access to Nvidia latest chips,

0:02:14.800 --> 0:02:17.799
<v Speaker 1>and some people are saying they did use you know,

0:02:17.960 --> 0:02:21.680
<v Speaker 1>the latest h one hundred chips, albeit they were not

0:02:21.800 --> 0:02:25.000
<v Speaker 1>of the same scale as were available to Open AI

0:02:25.160 --> 0:02:28.959
<v Speaker 1>and the others. But look, we don't know the extent

0:02:29.120 --> 0:02:32.280
<v Speaker 1>of the hardware that was used for training. All I

0:02:32.320 --> 0:02:34.680
<v Speaker 1>can say, you know, based on this the fact that

0:02:34.720 --> 0:02:37.760
<v Speaker 1>they have a comparable model in performance to the others.

0:02:38.240 --> 0:02:42.279
<v Speaker 1>It shows that they clearly have a combination of algorithms

0:02:42.280 --> 0:02:45.239
<v Speaker 1>and compute to compete with the others.

0:02:45.320 --> 0:02:48.240
<v Speaker 2>Most of the financials are the revenue growth models that

0:02:48.320 --> 0:02:52.040
<v Speaker 2>you and b I have. Of all these fancy AI people,

0:02:52.160 --> 0:02:53.440
<v Speaker 2>are they now at risk?

0:02:53.760 --> 0:02:54.359
<v Speaker 3>Absolutely?

0:02:54.360 --> 0:02:57.120
<v Speaker 1>I mean, look at open ais oh one pro model

0:02:57.160 --> 0:03:01.560
<v Speaker 1>they're charging two hundred dollars per month because it's the

0:03:01.560 --> 0:03:04.520
<v Speaker 1>best model they have to offer. The fact that you

0:03:04.600 --> 0:03:07.639
<v Speaker 1>have an open source model that is comparable to the

0:03:07.760 --> 0:03:09.600
<v Speaker 1>model that Opening Eyes chargeable.

0:03:09.840 --> 0:03:11.720
<v Speaker 3>You're saying it's comparable.

0:03:11.400 --> 0:03:13.480
<v Speaker 1>I mean it is, yeah, I mean look at the

0:03:13.520 --> 0:03:16.799
<v Speaker 1>benchmarks and that's why you know all these models have

0:03:16.919 --> 0:03:20.480
<v Speaker 1>common benchmarks. This is within one to two percentage points

0:03:20.560 --> 0:03:21.400
<v Speaker 1>of that benchmark.

0:03:21.480 --> 0:03:23.120
<v Speaker 3>I know you don't do buy hole cell, but are

0:03:23.120 --> 0:03:24.760
<v Speaker 3>we going to sell here? Are we going to see

0:03:24.800 --> 0:03:27.560
<v Speaker 3>Wall Street put a cell on an video? I think

0:03:27.639 --> 0:03:28.080
<v Speaker 3>predict that.

0:03:28.800 --> 0:03:31.040
<v Speaker 1>I mean right now, the fact that Meta raised their

0:03:31.120 --> 0:03:35.040
<v Speaker 1>Capex on Friday to me earning season is where we

0:03:35.080 --> 0:03:37.600
<v Speaker 1>will find out what all these companies end up doing.

0:03:37.680 --> 0:03:40.560
<v Speaker 1>So you know, all this has happened in the last

0:03:40.920 --> 0:03:44.080
<v Speaker 1>forty eight hours, where Meta raised their CAPEX Deep Sea

0:03:44.200 --> 0:03:46.880
<v Speaker 1>came out it suddenly everyone is going crazy.

0:03:46.960 --> 0:03:50.200
<v Speaker 2>Paul the President lined up with the son of Japan

0:03:50.720 --> 0:03:52.760
<v Speaker 2>to do a ginormous tech deal as well.

0:03:52.920 --> 0:03:54.440
<v Speaker 3>You okay, there, you look at you.

0:03:55.000 --> 0:03:59.080
<v Speaker 4>I'm doing social Here are you doing social words? Doing

0:03:59.120 --> 0:04:02.120
<v Speaker 4>that out the social I'm saying, take a look at that. So, Mandy,

0:04:02.200 --> 0:04:05.680
<v Speaker 4>what do we do here? As we think about AI's

0:04:05.720 --> 0:04:09.600
<v Speaker 4>obviously from technology perspective, it has been the story for

0:04:09.640 --> 0:04:12.680
<v Speaker 4>the last two years at least. How do you think

0:04:12.680 --> 0:04:15.600
<v Speaker 4>about it now? Has anything changed in the last twenty

0:04:15.640 --> 0:04:16.359
<v Speaker 4>four hours for you?

0:04:16.960 --> 0:04:20.679
<v Speaker 1>A lot of software companies will feel, you know, they

0:04:20.800 --> 0:04:23.479
<v Speaker 1>can do their own AI now and not having to

0:04:23.560 --> 0:04:28.120
<v Speaker 1>rely on the big hyperscalers because the narrative was if

0:04:28.160 --> 0:04:32.120
<v Speaker 1>you can't spend upfront on Capex, then you don't have

0:04:32.200 --> 0:04:34.640
<v Speaker 1>a play. And it comes to the foundational model there

0:04:35.000 --> 0:04:38.240
<v Speaker 1>and so suddenly everyone feels empowered that they can do

0:04:38.320 --> 0:04:40.960
<v Speaker 1>their own AI with this development, I.

0:04:40.920 --> 0:04:44.039
<v Speaker 4>See Nvidio, Broadcom, some of these big chip makers that

0:04:44.080 --> 0:04:46.000
<v Speaker 4>have had such a huge run on the IP side,

0:04:46.120 --> 0:04:49.760
<v Speaker 4>they're down ten to eleven twelve percent here today? Is

0:04:49.839 --> 0:04:52.680
<v Speaker 4>that realistic? Is that does that seem reasonable to you?

0:04:53.080 --> 0:04:55.640
<v Speaker 1>I mean with semi companies we know the risk is

0:04:55.680 --> 0:04:59.159
<v Speaker 1>always that cyclical element. And if we are calling that

0:04:59.279 --> 0:05:01.640
<v Speaker 1>this is the the top of the cycle in terms

0:05:01.680 --> 0:05:05.479
<v Speaker 1>of semi demand and estimates aren't going up anymore, then

0:05:05.600 --> 0:05:07.800
<v Speaker 1>you know you will see this sort of stock reaction.

0:05:07.880 --> 0:05:09.920
<v Speaker 1>But I go back to my point about Meta raising

0:05:09.960 --> 0:05:13.279
<v Speaker 1>their capex projects target at it one hundred billion.

0:05:13.279 --> 0:05:15.960
<v Speaker 5>Of that, we haven't talked to you about that fifty

0:05:16.000 --> 0:05:20.080
<v Speaker 5>billion they were Meta Cappex was twenty five billion in

0:05:20.160 --> 0:05:23.480
<v Speaker 5>twenty nineteen, then it was fifty billion for this year.

0:05:23.920 --> 0:05:26.680
<v Speaker 4>Now they're saying to go on a sixty five I

0:05:26.760 --> 0:05:28.600
<v Speaker 4>feel like it when I look at those numbers, it

0:05:28.600 --> 0:05:32.240
<v Speaker 4>feels like drunken sale or time. What Did they have

0:05:32.320 --> 0:05:34.760
<v Speaker 4>a real strategy behind that spending or are they just

0:05:34.760 --> 0:05:36.960
<v Speaker 4>saying we need to be in this game, We need

0:05:37.000 --> 0:05:39.520
<v Speaker 4>to spend whatever we need to spend, typical Meta type

0:05:39.520 --> 0:05:40.000
<v Speaker 4>of spending.

0:05:40.200 --> 0:05:42.960
<v Speaker 1>Yeah, and I think there is that aspect where you

0:05:43.120 --> 0:05:45.559
<v Speaker 1>want to be that front end player when it comes

0:05:45.560 --> 0:05:48.440
<v Speaker 1>to the lllms. But in the case of Meta, the challenge,

0:05:48.880 --> 0:05:50.880
<v Speaker 1>in addition to the fact that you know deep sea

0:05:51.000 --> 0:05:53.440
<v Speaker 1>model is out there, is they don't have a cloud business.

0:05:53.520 --> 0:05:55.800
<v Speaker 1>You look at Microsoft, you look at Amazon, you look

0:05:55.839 --> 0:05:58.279
<v Speaker 1>at Google, they don't they have a cloud business to

0:05:58.400 --> 0:06:01.680
<v Speaker 1>monetize those GPUs. Even if let's say deep Sea came

0:06:01.760 --> 0:06:05.080
<v Speaker 1>up with a better model, Microsoft can use that capacity

0:06:05.080 --> 0:06:06.440
<v Speaker 1>for inferencing on the cloud.

0:06:06.600 --> 0:06:10.000
<v Speaker 3>They can generate cloud revenue. Mandy, what are the meetings

0:06:10.200 --> 0:06:12.880
<v Speaker 3>like in Silicon Valley? Now?

0:06:13.279 --> 0:06:15.600
<v Speaker 2>I mean, these guys never get up before nine am,

0:06:15.839 --> 0:06:18.160
<v Speaker 2>But today they're gonna get up at six am their time.

0:06:18.400 --> 0:06:21.760
<v Speaker 2>They're gonna be rock at nine am surveillance time. What

0:06:21.800 --> 0:06:24.599
<v Speaker 2>are the meetings gonna be like at Google? What's the

0:06:24.640 --> 0:06:26.440
<v Speaker 2>meeting gonna be like for Zuckerberg?

0:06:27.160 --> 0:06:30.279
<v Speaker 1>Oh? So I think again, Given these companies are reporting

0:06:30.320 --> 0:06:33.400
<v Speaker 1>earnings next week right now, they have to figure out

0:06:33.839 --> 0:06:35.960
<v Speaker 1>what is it that they relate to the investors in

0:06:36.040 --> 0:06:39.520
<v Speaker 1>terms of the scaling laws. Like up until now, the

0:06:39.560 --> 0:06:43.159
<v Speaker 1>biggest debate was will the scaling laws hold in twenty

0:06:43.200 --> 0:06:47.280
<v Speaker 1>twenty five? And based on this development, we don't talk

0:06:47.320 --> 0:06:48.200
<v Speaker 1>about scaling laws.

0:06:48.279 --> 0:06:51.159
<v Speaker 3>Do they replicate what deep seek is doing?

0:06:51.839 --> 0:06:55.080
<v Speaker 1>Absolutely? I think they will use some of the innovations

0:06:55.120 --> 0:06:57.400
<v Speaker 1>that deep Seek has showed in their paper, and given

0:06:57.440 --> 0:07:00.720
<v Speaker 1>they have open source it. My bed is everyone cares

0:07:00.720 --> 0:07:04.520
<v Speaker 1>about hardware efficiency, even if Microsoft is spending eighty billion

0:07:04.560 --> 0:07:07.640
<v Speaker 1>on AI Capex. They want to use that cap ex efficiency.

0:07:07.720 --> 0:07:11.240
<v Speaker 2>The smartest thing I've heard of from the technologist Paul Sweeney,

0:07:11.400 --> 0:07:13.880
<v Speaker 2>your Ted talk on this right, your Ted talk.

0:07:14.120 --> 0:07:16.680
<v Speaker 3>You almost wore a tie to your Ted thing. Exactly.

0:07:17.080 --> 0:07:18.520
<v Speaker 3>They're drunken sailors.

0:07:18.680 --> 0:07:21.080
<v Speaker 2>I mean, that's the smartest thing they've heard. Do you

0:07:21.120 --> 0:07:22.880
<v Speaker 2>see capex responsibility?

0:07:23.360 --> 0:07:24.240
<v Speaker 3>No? I think.

0:07:24.920 --> 0:07:29.160
<v Speaker 1>I mean to my mind, all this Capex can be used,

0:07:29.240 --> 0:07:32.000
<v Speaker 1>you know, for the future, and it's quite fungible. It's

0:07:32.040 --> 0:07:34.880
<v Speaker 1>not like laying a fiber network and oh it's gonna

0:07:34.920 --> 0:07:39.880
<v Speaker 1>go waste. Compute is always used in some form. Now, yes,

0:07:39.960 --> 0:07:43.040
<v Speaker 1>they may have overpaid for the GPUs. You could argue

0:07:43.560 --> 0:07:45.920
<v Speaker 1>was it worth paying thirty thousand dollars for a GPU?

0:07:46.080 --> 0:07:48.840
<v Speaker 1>Probably not, but still this compute is still useful. Man.

0:07:48.880 --> 0:07:51.120
<v Speaker 2>Deep single those folks. We're gonna go around equities, bond's

0:07:51.160 --> 0:07:53.400
<v Speaker 2>currencies come out of these again. We're off four percent.

0:07:53.840 --> 0:07:56.960
<v Speaker 3>No, sit down, you're not done yet. You know where

0:07:57.040 --> 0:07:57.520
<v Speaker 3>you're going.

0:07:58.240 --> 0:07:59.320
<v Speaker 4>Two segments you should.

0:07:59.080 --> 0:08:01.720
<v Speaker 3>See as an englishakfast on Monday. It's like this huge

0:08:01.760 --> 0:08:02.800
<v Speaker 3>kind of a thing.

0:08:02.960 --> 0:08:06.640
<v Speaker 2>The bond market yields in big time, which you'd expect

0:08:06.720 --> 0:08:10.120
<v Speaker 2>here the screening yield four point one eight percent in

0:08:10.320 --> 0:08:14.520
<v Speaker 2>nine basis points, the tenure yield in eleven basis points.

0:08:14.560 --> 0:08:17.360
<v Speaker 2>First thing, I looked at the real yield crushes in

0:08:17.520 --> 0:08:20.400
<v Speaker 2>ten basis points from a two point two zero to

0:08:20.400 --> 0:08:23.000
<v Speaker 2>two point one zero some of the angst there in

0:08:23.040 --> 0:08:23.680
<v Speaker 2>the market.

0:08:23.880 --> 0:08:25.560
<v Speaker 3>And what I look at, and it's not a good

0:08:25.600 --> 0:08:26.240
<v Speaker 3>number yet.

0:08:26.120 --> 0:08:29.000
<v Speaker 2>With the Vicks twenty one point five four, not thirty,

0:08:29.080 --> 0:08:31.040
<v Speaker 2>but from fifteen to twenty one.

0:08:31.040 --> 0:08:33.719
<v Speaker 4>That gets your attention, poll Man, Deep is a concern here.

0:08:34.120 --> 0:08:37.760
<v Speaker 4>The deep Seak product could be as good an AI

0:08:37.880 --> 0:08:40.920
<v Speaker 4>solution as what some of the Western companies are providing,

0:08:41.280 --> 0:08:43.880
<v Speaker 4>but at a lower cost. Is that the bottom line here?

0:08:44.000 --> 0:08:46.400
<v Speaker 1>Yeah? I mean look at the app stores that deep

0:08:46.400 --> 0:08:49.679
<v Speaker 1>Seak is the top appro right now here in the US,

0:08:49.760 --> 0:08:53.800
<v Speaker 1>and everyone is using it multitude chat.

0:08:54.360 --> 0:08:56.160
<v Speaker 3>No, I don't mean ititer up, but this is important.

0:08:56.559 --> 0:09:00.959
<v Speaker 2>Lesa Matao's not using it, Tom Keene is not using it.

0:09:01.480 --> 0:09:04.720
<v Speaker 2>An adult like you or Joe Wisenthal who's knee deep

0:09:04.720 --> 0:09:05.360
<v Speaker 2>into this stuff.

0:09:05.400 --> 0:09:09.280
<v Speaker 3>Good morning, Joe, Joey. Joe's not up yet, Okay, Joe

0:09:09.280 --> 0:09:12.960
<v Speaker 3>Wisenthal is going to bring up deep seek and bring up.

0:09:13.040 --> 0:09:16.400
<v Speaker 2>Chet GPT of one of the different managed levels. Can

0:09:16.440 --> 0:09:17.400
<v Speaker 2>you tell the difference?

0:09:18.559 --> 0:09:21.800
<v Speaker 1>Look, I mean there is an element of personalization that

0:09:21.920 --> 0:09:24.880
<v Speaker 1>these apps pro wide when it comes to asking the

0:09:24.920 --> 0:09:27.560
<v Speaker 1>type of questions you're asking. But if you have a

0:09:27.600 --> 0:09:30.400
<v Speaker 1>generic query, deep seek is going to give you an

0:09:30.440 --> 0:09:32.240
<v Speaker 1>answer that's comparable to Chat GPT.

0:09:32.679 --> 0:09:36.839
<v Speaker 4>I'm looking Tom, tell me you're on deep Seek Now.

0:09:36.880 --> 0:09:41.640
<v Speaker 4>I google the top app number one, deep Seek, number

0:09:41.640 --> 0:09:47.400
<v Speaker 4>two Chat GPT, and number three Paramount plus for land

0:09:47.480 --> 0:09:50.160
<v Speaker 4>Man and for Yellowstone. That's what I'm talking about.

0:09:49.880 --> 0:09:51.960
<v Speaker 2>That we had a land Man weekd at home. I said,

0:09:51.960 --> 0:09:53.559
<v Speaker 2>Sweeney's liked Deep into.

0:09:53.400 --> 0:09:54.120
<v Speaker 4>The number seven.

0:09:54.200 --> 0:09:57.600
<v Speaker 3>Fox Sports get one more in here, one more in here,

0:09:57.640 --> 0:09:59.080
<v Speaker 3>because man, Deep's coming back at eight.

0:09:59.280 --> 0:10:02.520
<v Speaker 4>What do you need? What do you think you're gonna

0:10:02.520 --> 0:10:04.720
<v Speaker 4>hear from some of these technology companies that next week

0:10:04.760 --> 0:10:07.680
<v Speaker 4>as they talk about China, as they talk about this,

0:10:08.360 --> 0:10:09.480
<v Speaker 4>what is a deep Seek thing?

0:10:09.840 --> 0:10:10.040
<v Speaker 3>Yeah?

0:10:10.080 --> 0:10:13.920
<v Speaker 1>I mean the whole aspect around CAPEX and scaling laws

0:10:13.960 --> 0:10:17.560
<v Speaker 1>and what to do with the latest GPUs. Are you

0:10:17.800 --> 0:10:22.200
<v Speaker 1>spending more on training or inferencing? Everything is on the

0:10:22.240 --> 0:10:24.479
<v Speaker 1>table in terms of hardware efficiency.

0:10:24.679 --> 0:10:27.520
<v Speaker 2>I have no idea what inferences is a clinic for

0:10:27.600 --> 0:10:30.240
<v Speaker 2>the man deep sing of Bloomberg intelligence