January 27, 2025 • 10 mins
Bloomberg Intelligence's Mandeep Singh explains DeepSeek's focus on hardware efficiency, and competition with other LLM's. He speaks with Bloomberg's Tom Keene and Paul Sweeney.
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:02):
Bloomberg Audio Studios, podcasts.
Speaker 2 (00:05):
Radio news man Deep Singh joins us, Right now, did
you know what deep seek was before twelve or eighteen
hours ago?
Speaker 3 (00:15):
I did, yes, And so what did you know about
them two days ago?
Speaker 1 (00:20):
That it is a competitor in the LM space. They've
been training their own model and it's used mostly in
the East Asian region.
Speaker 2 (00:27):
Okay, what I learned, thank you zero head for this
from Morgan Brown at dropbox is basically they've said, we're
not going to be as perfect as pristine as the
others out to thirty two decimal points. We're just going
to get it done out to point eight decimal points.
Is that what this is all about is they're not
(00:49):
going to be perfect.
Speaker 1 (00:50):
I think there's more to it than just you know,
using the floating point fewer floating point operations than open
AI and Entropic and others. And in this case, they
basically focused on hardware efficiency, using the hardware in the
most efficient fashion. They had the benefit of all these
(01:10):
lms being out there. They use meta Llama as a
reference point. It's another open source model, and they figured
out a way to do this most efficiently than everyone else.
I mean, everyone is focused on scale right now, they
focused on hardware efficiency.
Speaker 3 (01:26):
Call the Lama meta.
Speaker 2 (01:27):
It's Facebook is the grandson of the original Lama on
Saturday Night.
Speaker 4 (01:32):
There you go, right, very good man, Deep You've got
an analyst Bloomberg Intelligence in Hong Kong, Robert Lee. I'm
reading his research literally as we speak. And for folks
on the Bloomberg terminal, go to big and check out
if you want to need to figure out what's going
on with Deep seek and what China's doing with AI.
Bloomerg Intelligence has got the data and the research there.
(01:55):
Can China support a real competitive AI environment? It seems
like they're government is more restricting some of the technology there.
Speaker 1 (02:04):
Well, so now we know the export controls probably weren't
as effective as they were supposed to be here, you know,
in terms of restricting the access to Nvidia latest chips,
and some people are saying they did use you know,
the latest h one hundred chips, albeit they were not
of the same scale as were available to Open AI
(02:25):
and the others. But look, we don't know the extent
of the hardware that was used for training. All I
can say, you know, based on this the fact that
they have a comparable model in performance to the others.
It shows that they clearly have a combination of algorithms
and compute to compete with the others.
Speaker 2 (02:45):
Most of the financials are the revenue growth models that
you and b I have. Of all these fancy AI people,
are they now at risk?
Speaker 3 (02:53):
Absolutely?
Speaker 1 (02:54):
I mean, look at open ais oh one pro model
they're charging two hundred dollars per month because it's the
best model they have to offer. The fact that you
have an open source model that is comparable to the
model that Opening Eyes chargeable.
Speaker 3 (03:09):
You're saying it's comparable.
Speaker 1 (03:11):
I mean it is, yeah, I mean look at the
benchmarks and that's why you know all these models have
common benchmarks. This is within one to two percentage points
of that benchmark.
Speaker 3 (03:21):
I know you don't do buy hole cell, but are
we going to sell here? Are we going to see
Wall Street put a cell on an video? I think
predict that.
Speaker 1 (03:28):
I mean right now, the fact that Meta raised their
Capex on Friday to me earning season is where we
will find out what all these companies end up doing.
So you know, all this has happened in the last
forty eight hours, where Meta raised their CAPEX Deep Sea
came out it suddenly everyone is going crazy.
Speaker 2 (03:46):
Paul the President lined up with the son of Japan
to do a ginormous tech deal as well.
Speaker 3 (03:52):
You okay, there, you look at you.
Speaker 4 (03:55):
I'm doing social Here are you doing social words? Doing
that out the social I'm saying, take a look at that. So, Mandy,
what do we do here? As we think about AI's
obviously from technology perspective, it has been the story for
the last two years at least. How do you think
about it now? Has anything changed in the last twenty
(04:15):
four hours for you?
Speaker 1 (04:16):
A lot of software companies will feel, you know, they
can do their own AI now and not having to
rely on the big hyperscalers because the narrative was if
you can't spend upfront on Capex, then you don't have
a play. And it comes to the foundational model there
and so suddenly everyone feels empowered that they can do
(04:38):
their own AI with this development, I.
Speaker 4 (04:40):
See Nvidio, Broadcom, some of these big chip makers that
have had such a huge run on the IP side,
they're down ten to eleven twelve percent here today? Is
that realistic? Is that does that seem reasonable to you?
Speaker 1 (04:53):
I mean with semi companies we know the risk is
always that cyclical element. And if we are calling that
this is the the top of the cycle in terms
of semi demand and estimates aren't going up anymore, then
you know you will see this sort of stock reaction.
But I go back to my point about Meta raising
their capex projects target at it one hundred billion.
Speaker 5 (05:13):
Of that, we haven't talked to you about that fifty
billion they were Meta Cappex was twenty five billion in
twenty nineteen, then it was fifty billion for this year.
Speaker 4 (05:23):
Now they're saying to go on a sixty five I
feel like it when I look at those numbers, it
feels like drunken sale or time. What Did they have
a real strategy behind that spending or are they just
saying we need to be in this game, We need
to spend whatever we need to spend, typical Meta type
of spending.
Speaker 1 (05:40):
Yeah, and I think there is that aspect where you
want to be that front end player when it comes
to the lllms. But in the case of Meta, the challenge,
in addition to the fact that you know deep sea
model is out there, is they don't have a cloud business.
You look at Microsoft, you look at Amazon, you look
at Google, they don't they have a cloud business to
monetize those GPUs. Even if let's say deep Sea came
(06:01):
up with a better model, Microsoft can use that capacity
for inferencing on the cloud.
Speaker 3 (06:06):
They can generate cloud revenue. Mandy, what are the meetings
like in Silicon Valley? Now?
Speaker 2 (06:13):
I mean, these guys never get up before nine am,
But today they're gonna get up at six am their time.
They're gonna be rock at nine am surveillance time. What
are the meetings gonna be like at Google? What's the
meeting gonna be like for Zuckerberg?
Speaker 1 (06:27):
Oh? So I think again, Given these companies are reporting
earnings next week right now, they have to figure out
what is it that they relate to the investors in
terms of the scaling laws. Like up until now, the
biggest debate was will the scaling laws hold in twenty
twenty five? And based on this development, we don't talk
(06:47):
about scaling laws.
Speaker 3 (06:48):
Do they replicate what deep seek is doing?
Speaker 1 (06:51):
Absolutely? I think they will use some of the innovations
that deep Seek has showed in their paper, and given
they have open source it. My bed is everyone cares
about hardware efficiency, even if Microsoft is spending eighty billion
on AI Capex. They want to use that cap ex efficiency.
Speaker 2 (07:07):
The smartest thing I've heard of from the technologist Paul Sweeney,
your Ted talk on this right, your Ted talk.
Speaker 3 (07:14):
You almost wore a tie to your Ted thing. Exactly.
They're drunken sailors.
Speaker 2 (07:18):
I mean, that's the smartest thing they've heard. Do you
see capex responsibility?
Speaker 3 (07:23):
No? I think.
Speaker 1 (07:24):
I mean to my mind, all this Capex can be used,
you know, for the future, and it's quite fungible. It's
not like laying a fiber network and oh it's gonna
go waste. Compute is always used in some form. Now, yes,
they may have overpaid for the GPUs. You could argue
was it worth paying thirty thousand dollars for a GPU?
(07:46):
Probably not, but still this compute is still useful. Man.
Speaker 2 (07:48):
Deep single those folks. We're gonna go around equities, bond's
currencies come out of these again. We're off four percent.
Speaker 3 (07:53):
No, sit down, you're not done yet. You know where
you're going.
Speaker 4 (07:58):
Two segments you should.
Speaker 3 (07:59):
See as an englishakfast on Monday. It's like this huge
kind of a thing.
Speaker 2 (08:02):
The bond market yields in big time, which you'd expect
here the screening yield four point one eight percent in
nine basis points, the tenure yield in eleven basis points.
First thing, I looked at the real yield crushes in
ten basis points from a two point two zero to
two point one zero some of the angst there in
(08:23):
the market.
Speaker 3 (08:23):
And what I look at, and it's not a good
number yet.
Speaker 2 (08:26):
With the Vicks twenty one point five four, not thirty,
but from fifteen to twenty one.
Speaker 4 (08:31):
That gets your attention, poll Man, Deep is a concern here.
The deep Seak product could be as good an AI
solution as what some of the Western companies are providing,
but at a lower cost. Is that the bottom line here?
Speaker 1 (08:44):
Yeah? I mean look at the app stores that deep
Seak is the top appro right now here in the US,
and everyone is using it multitude chat.
Speaker 3 (08:54):
No, I don't mean ititer up, but this is important.
Speaker 2 (08:56):
Lesa Matao's not using it, Tom Keene is not using it.
An adult like you or Joe Wisenthal who's knee deep
into this stuff.
Speaker 3 (09:05):
Good morning, Joe, Joey. Joe's not up yet, Okay, Joe
Wisenthal is going to bring up deep seek and bring up.
Speaker 2 (09:13):
Chet GPT of one of the different managed levels. Can
you tell the difference?
Speaker 1 (09:18):
Look, I mean there is an element of personalization that
these apps pro wide when it comes to asking the
type of questions you're asking. But if you have a
generic query, deep seek is going to give you an
answer that's comparable to Chat GPT.
Speaker 4 (09:32):
I'm looking Tom, tell me you're on deep Seek Now.
I google the top app number one, deep Seek, number
two Chat GPT, and number three Paramount plus for land
Man and for Yellowstone. That's what I'm talking about.
Speaker 2 (09:49):
That we had a land Man weekd at home. I said,
Sweeney's liked Deep into.
Speaker 4 (09:53):
The number seven.
Speaker 3 (09:54):
Fox Sports get one more in here, one more in here,
because man, Deep's coming back at eight.
Speaker 4 (09:59):
What do you need? What do you think you're gonna
hear from some of these technology companies that next week
as they talk about China, as they talk about this,
what is a deep Seek thing?
Speaker 3 (10:09):
Yeah?
Speaker 1 (10:10):
I mean the whole aspect around CAPEX and scaling laws
and what to do with the latest GPUs. Are you
spending more on training or inferencing? Everything is on the
table in terms of hardware efficiency.
Speaker 2 (10:24):
I have no idea what inferences is a clinic for
the man deep sing of Bloomberg intelligence
Bloomberg Audio Studios, podcasts.
Speaker 2 (00:05):
Radio news man Deep Singh joins us, Right now, did
you know what deep seek was before twelve or eighteen
hours ago?
Speaker 3 (00:15):
I did, yes, And so what did you know about
them two days ago?
Speaker 1 (00:20):
That it is a competitor in the LM space. They've
been training their own model and it's used mostly in
the East Asian region.
Speaker 2 (00:27):
Okay, what I learned, thank you zero head for this
from Morgan Brown at dropbox is basically they've said, we're
not going to be as perfect as pristine as the
others out to thirty two decimal points. We're just going
to get it done out to point eight decimal points.
Is that what this is all about is they're not
(00:49):
going to be perfect.
Speaker 1 (00:50):
I think there's more to it than just you know,
using the floating point fewer floating point operations than open
AI and Entropic and others. And in this case, they
basically focused on hardware efficiency, using the hardware in the
most efficient fashion. They had the benefit of all these
(01:10):
lms being out there. They use meta Llama as a
reference point. It's another open source model, and they figured
out a way to do this most efficiently than everyone else.
I mean, everyone is focused on scale right now, they
focused on hardware efficiency.
Speaker 3 (01:26):
Call the Lama meta.
Speaker 2 (01:27):
It's Facebook is the grandson of the original Lama on
Saturday Night.
Speaker 4 (01:32):
There you go, right, very good man, Deep You've got
an analyst Bloomberg Intelligence in Hong Kong, Robert Lee. I'm
reading his research literally as we speak. And for folks
on the Bloomberg terminal, go to big and check out
if you want to need to figure out what's going
on with Deep seek and what China's doing with AI.
Bloomerg Intelligence has got the data and the research there.
(01:55):
Can China support a real competitive AI environment? It seems
like they're government is more restricting some of the technology there.
Speaker 1 (02:04):
Well, so now we know the export controls probably weren't
as effective as they were supposed to be here, you know,
in terms of restricting the access to Nvidia latest chips,
and some people are saying they did use you know,
the latest h one hundred chips, albeit they were not
of the same scale as were available to Open AI
(02:25):
and the others. But look, we don't know the extent
of the hardware that was used for training. All I
can say, you know, based on this the fact that
they have a comparable model in performance to the others.
It shows that they clearly have a combination of algorithms
and compute to compete with the others.
Speaker 2 (02:45):
Most of the financials are the revenue growth models that
you and b I have. Of all these fancy AI people,
are they now at risk?
Speaker 3 (02:53):
Absolutely?
Speaker 1 (02:54):
I mean, look at open ais oh one pro model
they're charging two hundred dollars per month because it's the
best model they have to offer. The fact that you
have an open source model that is comparable to the
model that Opening Eyes chargeable.
Speaker 3 (03:09):
You're saying it's comparable.
Speaker 1 (03:11):
I mean it is, yeah, I mean look at the
benchmarks and that's why you know all these models have
common benchmarks. This is within one to two percentage points
of that benchmark.
Speaker 3 (03:21):
I know you don't do buy hole cell, but are
we going to sell here? Are we going to see
Wall Street put a cell on an video? I think
predict that.
Speaker 1 (03:28):
I mean right now, the fact that Meta raised their
Capex on Friday to me earning season is where we
will find out what all these companies end up doing.
So you know, all this has happened in the last
forty eight hours, where Meta raised their CAPEX Deep Sea
came out it suddenly everyone is going crazy.
Speaker 2 (03:46):
Paul the President lined up with the son of Japan
to do a ginormous tech deal as well.
Speaker 3 (03:52):
You okay, there, you look at you.
Speaker 4 (03:55):
I'm doing social Here are you doing social words? Doing
that out the social I'm saying, take a look at that. So, Mandy,
what do we do here? As we think about AI's
obviously from technology perspective, it has been the story for
the last two years at least. How do you think
about it now? Has anything changed in the last twenty
(04:15):
four hours for you?
Speaker 1 (04:16):
A lot of software companies will feel, you know, they
can do their own AI now and not having to
rely on the big hyperscalers because the narrative was if
you can't spend upfront on Capex, then you don't have
a play. And it comes to the foundational model there
and so suddenly everyone feels empowered that they can do
(04:38):
their own AI with this development, I.
Speaker 4 (04:40):
See Nvidio, Broadcom, some of these big chip makers that
have had such a huge run on the IP side,
they're down ten to eleven twelve percent here today? Is
that realistic? Is that does that seem reasonable to you?
Speaker 1 (04:53):
I mean with semi companies we know the risk is
always that cyclical element. And if we are calling that
this is the the top of the cycle in terms
of semi demand and estimates aren't going up anymore, then
you know you will see this sort of stock reaction.
But I go back to my point about Meta raising
their capex projects target at it one hundred billion.
Speaker 5 (05:13):
Of that, we haven't talked to you about that fifty
billion they were Meta Cappex was twenty five billion in
twenty nineteen, then it was fifty billion for this year.
Speaker 4 (05:23):
Now they're saying to go on a sixty five I
feel like it when I look at those numbers, it
feels like drunken sale or time. What Did they have
a real strategy behind that spending or are they just
saying we need to be in this game, We need
to spend whatever we need to spend, typical Meta type
of spending.
Speaker 1 (05:40):
Yeah, and I think there is that aspect where you
want to be that front end player when it comes
to the lllms. But in the case of Meta, the challenge,
in addition to the fact that you know deep sea
model is out there, is they don't have a cloud business.
You look at Microsoft, you look at Amazon, you look
at Google, they don't they have a cloud business to
monetize those GPUs. Even if let's say deep Sea came
(06:01):
up with a better model, Microsoft can use that capacity
for inferencing on the cloud.
Speaker 3 (06:06):
They can generate cloud revenue. Mandy, what are the meetings
like in Silicon Valley? Now?
Speaker 2 (06:13):
I mean, these guys never get up before nine am,
But today they're gonna get up at six am their time.
They're gonna be rock at nine am surveillance time. What
are the meetings gonna be like at Google? What's the
meeting gonna be like for Zuckerberg?
Speaker 1 (06:27):
Oh? So I think again, Given these companies are reporting
earnings next week right now, they have to figure out
what is it that they relate to the investors in
terms of the scaling laws. Like up until now, the
biggest debate was will the scaling laws hold in twenty
twenty five? And based on this development, we don't talk
(06:47):
about scaling laws.
Speaker 3 (06:48):
Do they replicate what deep seek is doing?
Speaker 1 (06:51):
Absolutely? I think they will use some of the innovations
that deep Seek has showed in their paper, and given
they have open source it. My bed is everyone cares
about hardware efficiency, even if Microsoft is spending eighty billion
on AI Capex. They want to use that cap ex efficiency.
Speaker 2 (07:07):
The smartest thing I've heard of from the technologist Paul Sweeney,
your Ted talk on this right, your Ted talk.
Speaker 3 (07:14):
You almost wore a tie to your Ted thing. Exactly.
They're drunken sailors.
Speaker 2 (07:18):
I mean, that's the smartest thing they've heard. Do you
see capex responsibility?
Speaker 3 (07:23):
No? I think.
Speaker 1 (07:24):
I mean to my mind, all this Capex can be used,
you know, for the future, and it's quite fungible. It's
not like laying a fiber network and oh it's gonna
go waste. Compute is always used in some form. Now, yes,
they may have overpaid for the GPUs. You could argue
was it worth paying thirty thousand dollars for a GPU?
(07:46):
Probably not, but still this compute is still useful. Man.
Speaker 2 (07:48):
Deep single those folks. We're gonna go around equities, bond's
currencies come out of these again. We're off four percent.
Speaker 3 (07:53):
No, sit down, you're not done yet. You know where
you're going.
Speaker 4 (07:58):
Two segments you should.
Speaker 3 (07:59):
See as an englishakfast on Monday. It's like this huge
kind of a thing.
Speaker 2 (08:02):
The bond market yields in big time, which you'd expect
here the screening yield four point one eight percent in
nine basis points, the tenure yield in eleven basis points.
First thing, I looked at the real yield crushes in
ten basis points from a two point two zero to
two point one zero some of the angst there in
(08:23):
the market.
Speaker 3 (08:23):
And what I look at, and it's not a good
number yet.
Speaker 2 (08:26):
With the Vicks twenty one point five four, not thirty,
but from fifteen to twenty one.
Speaker 4 (08:31):
That gets your attention, poll Man, Deep is a concern here.
The deep Seak product could be as good an AI
solution as what some of the Western companies are providing,
but at a lower cost. Is that the bottom line here?
Speaker 1 (08:44):
Yeah? I mean look at the app stores that deep
Seak is the top appro right now here in the US,
and everyone is using it multitude chat.
Speaker 3 (08:54):
No, I don't mean ititer up, but this is important.
Speaker 2 (08:56):
Lesa Matao's not using it, Tom Keene is not using it.
An adult like you or Joe Wisenthal who's knee deep
into this stuff.
Speaker 3 (09:05):
Good morning, Joe, Joey. Joe's not up yet, Okay, Joe
Wisenthal is going to bring up deep seek and bring up.
Speaker 2 (09:13):
Chet GPT of one of the different managed levels. Can
you tell the difference?
Speaker 1 (09:18):
Look, I mean there is an element of personalization that
these apps pro wide when it comes to asking the
type of questions you're asking. But if you have a
generic query, deep seek is going to give you an
answer that's comparable to Chat GPT.
Speaker 4 (09:32):
I'm looking Tom, tell me you're on deep Seek Now.
I google the top app number one, deep Seek, number
two Chat GPT, and number three Paramount plus for land
Man and for Yellowstone. That's what I'm talking about.
Speaker 2 (09:49):
That we had a land Man weekd at home. I said,
Sweeney's liked Deep into.
Speaker 4 (09:53):
The number seven.
Speaker 3 (09:54):
Fox Sports get one more in here, one more in here,
because man, Deep's coming back at eight.
Speaker 4 (09:59):
What do you need? What do you think you're gonna
hear from some of these technology companies that next week
as they talk about China, as they talk about this,
what is a deep Seek thing?
Speaker 3 (10:09):
Yeah?
Speaker 1 (10:10):
I mean the whole aspect around CAPEX and scaling laws
and what to do with the latest GPUs. Are you
spending more on training or inferencing? Everything is on the
table in terms of hardware efficiency.
Speaker 2 (10:24):
I have no idea what inferences is a clinic for
the man deep sing of Bloomberg intelligence
Popular Podcasts
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com