January 27, 2025 • 35 mins
Watch Tom and Paul LIVE every day on YouTube: http://bit.ly/3vTiACF.
Bloomberg Surveillance hosted by Tom Keene & Paul SweeneyJanuary 27th, 2025
Featuring:
- Anurag Rana and Mandeep Singh, Bloomberg Intelligence analysts, on how DeepSeek is panicking markets
- Joe Weisenthal, Bloomberg Odd Lots co-host
- Dr Stacy Rasgon, Bernstein analyst covering US Semiconductors and Semiconductor Capital Equipment
- Lisa Mateo on newspapers
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, radio news. This is the Bloomberg
Surveillance Podcast. Catch us live weekdays at seven am Eastern
on Apple CarPlay or Android Auto with the Bloomberg Business app.
Listen on demand wherever you get your podcasts, or watch
(00:25):
us live on YouTube right now.
Speaker 2 (00:28):
In joining us.
Speaker 3 (00:29):
And you know last night when this was blowing up,
I said get me Mandeep saying.
Speaker 2 (00:32):
You'll join us in a bit here.
Speaker 3 (00:34):
But aniog Rana joins us now from Chicago with a
little different take than Mandeep. More on the cloud, more
on the systems analysis, if you will. Of these big companies,
man Anerog, thank you so much for joining.
Speaker 2 (00:46):
You don't do biholed sell.
Speaker 3 (00:48):
But would you suggest the cell side will shift opinion today.
Speaker 4 (00:54):
For the cloud providers?
Speaker 5 (00:55):
I don't think so because, as you know Paul just
mentioned about the tweet from my ACOFT CEO, he's probably
a little bit happier because for him, if the cost
curve goes down in the long run, and we don't
know if that's going to happen, you know, his adoption
rate goes up and his CAPEX goes down.
Speaker 4 (01:11):
So it's it's you know, in a roundabout way.
Speaker 5 (01:15):
Now, we don't think there is going to be an
impact on CAPEX in the near term, but in the
long run, if that just shifts, you know, it helps
their profitability. But more important to that the enterprise you know,
you could say take create on AI adoption goes up the.
Speaker 3 (01:29):
Theme this morning, Paul's been great about this is like
drunken sailors from where in Anerich you're just world class
at this. Are they drunken sailors in terms of CAPEX?
Or is there a plan? Is there a sequence? Is
there a pace to spending billions?
Speaker 5 (01:47):
So one of the things, if you would say, is
if if this AI adoption curve, which is absolutely in
infancy right now, the only thing we have seen is
in the consumer world, if enterprises start to put this
in their core app locations, then they would need a
lot more you could say, firepower to run it. Most
of the things that we have discussed is people would
(02:08):
not want to do this on their own premise. They're
bigger companies will but most of them will go to
a handful of cloud providers, and that would be Microsoft, Aws,
Google and then Oracle. So for all these companies to
increase their capacity now they are only seeing the orders
and based on that they are expanding their data center footwork.
Speaker 4 (02:28):
The question is whether do.
Speaker 5 (02:29):
They need the expensive chips to do with this or
can they run this in with cheaper chips. And that's
really I think the biggest discussion going to be over
the next twelve months.
Speaker 6 (02:38):
An raglick Tom has often want to say, my head
is spinning here. Just since Friday. I've got a big,
big tech company called Meta I call it Facebook given
the IPO. We spent a lot of time thinking about
that name in the tech the ticker FB. They up
their cap x from fifty billion to sixty five billion,
a large part on AI. Now I've got this Deep
(03:00):
Seek thinking about I can do this in.
Speaker 7 (03:02):
Much lower costs. How do we square those two things?
Speaker 5 (03:06):
Yeah, but one thing is for sure, this Deep Seek
news of the paper has been out for a while,
so it does got It did get publicized over the
last few days. But I'd be very surprised if Meta
AWS and Microsoft did not know about this. It has
been around for a while, so both of them, both
Meta and Microsoft talked about their capex last week after
knowing this news, so you know they have moved.
Speaker 4 (03:29):
I mean, they're far smarter than I am when it
comes to this.
Speaker 5 (03:31):
So if they're taking up their Capex, I mean there
must be a real need for it.
Speaker 3 (03:35):
What's it mean for Coopertino there having an early morning?
They don't have lotes at Coopertino? They got something, Lisa.
What's the thing with the green powder in it?
Speaker 2 (03:45):
The machi?
Speaker 3 (03:46):
Yeah matter, they're having the machiat Coopertino some undrinkable bilge.
What are they saying at Apple today?
Speaker 5 (03:53):
They're actually smiling and saying, you guys fight it out
and tell me what's the best large language model out there.
I'll just put it in my phone and I won't
even pay for it. So they if you just look
at you know, Apple right now, that's kind of the
safety haven for people is they're not spending that capex.
If you look at that Capex line, it hasn't moved
at all in two years, and the free cash slow
just keep on coming. Their issue, honestly, is more so
(04:16):
on the you know, the growth side. They are having
challenges in China that needs to you know, smoothened out
over the next twelve months. Otherwise this is going to
be another year for them. To grow only four five percent.
That's the big issue for them. Nothing to do with AI.
Speaker 6 (04:30):
All right, Well, let's just stay on the Apple story.
It stocks down eleven percent year to date this year,
it's only a fifteen percent over the trailing twelve months,
which is radical on the performance. And for a company
like Apple, is there a bulkcase here short term?
Speaker 7 (04:45):
For Apple?
Speaker 5 (04:47):
In the short term, it's very difficult to come up
with the bulkcase because, as we talked about it, even
you know a few months ago, Apple Intelligence is not
going to be out in other parts of the world.
It's not out in China, and China is where they
are struggling honestly at this and they need to come
out and pacify investors that you know, we have a
strategy in China, we will gain market share back. Otherwise
(05:08):
this is going to be another mut the year for them.
We think the next phone, which is the one that's
coming out in September, I know that's fa I think
that's the big catalyst, but that's like, you know, that's
eight months out, nine months out.
Speaker 3 (05:19):
Right, and Agrana with us in Chicago right now. We
welcome all of you on your commute across the Nation.
For those of you in the office at home on YouTube,
thank you for being on youtubday. Just brilliant live chat
of learning a lot from people smarter than me on
the YouTube live chat this morning. Subscribe at Bloomberg Podcast.
That's the best way to get to the show out there.
(05:42):
This is a rare privilege. And for senior management listening
on radio right now. We adhere to their policy which
is Ana Agarana and man Deep Sing can never be
in the same ruct together. It's too much worse to
the to the franchise. But Paul, why don't you bring
in man Deep Sing?
Speaker 2 (05:58):
As we talked to both and Ruana and.
Speaker 6 (06:00):
Mandy man Deep Seeing covers the technology space along with
Ana rog Rana, there are two leaders globally running our
technology franchise and it is a global franchise analyst in Europe,
Asia and North America. Man Deep, But I guess what
we're hearing such Antela from Microsoft put out something that hey,
(06:20):
cost coming down is pretty natural for the technology space.
That happens, and it's a good thing because it helps adoption.
Speaker 8 (06:28):
Does that make sense to you, Well, clearly, I think
the investor expectations were different, which is why you see
this sort of stock reaction.
Speaker 7 (06:37):
So yes, we.
Speaker 8 (06:38):
Have Moore's law and we have other laws in terms
of figuring out how the cost curve may look like
over time. But clearly this is a step change in
terms of, you know, what you can do with the
existing chips that are out there, as well as what
you will need going forward. So I think this is
a big moment in terms of just that we can
(07:01):
see with these chips and the kind of scale that
is required for the next versions of these foundational models.
But there is no doubt that this makes generative AI
more accessible to software companies as well as you know,
overall users.
Speaker 6 (07:15):
Anrak, what do you think this means to you know,
I know you guys at Bloomberg Intelligence you have access
to this I d C data, which is some of
the best data on technology spending in the marketplace, and
you guys feature that in your research. Is this going
to change i DC's forecast of tech spending?
Speaker 7 (07:34):
AI spending? Do you think so?
Speaker 5 (07:37):
AI spending has been going up, Paul, We've talked about
this a few times, but it's the other non AI
spending that has actually not done so well in the
last two years and over the next I would say month,
month and a half, we will get a much clearer
picture if corporations or enterprises are moving up there over
our tech budgets or not, because right now AI spending
(07:57):
is coming at the expense of areas you know, normal
software or you know, just bording hardware upgrades or consulting
that really needs to pick up because a large portion
of the tech ecosystem is some of those companies.
Speaker 7 (08:12):
Mande.
Speaker 6 (08:12):
If I think when I think of this news today again,
I'm like everybody else. Our listeners and viewers are learning
about deep Seek and what it means. I can't help
but think about TikTok at some point. Again, I feel
like by the end of business today, we're gonna have
a story coming into Washington. There's gonna be an opinion
coming from Washington as it relates to deep Seek. Does
oppose a similar risks? Do you think to TikTok from
(08:33):
a regulator's perspective, Well.
Speaker 8 (08:35):
It's still the app is still you know, early in
terms of adoption, and look, I'm sure it's got millions
of users in East Asia and you know, China and whatnot.
But I think when it comes to AI consumers, just
look at the functionality, and you know the fact that
(08:57):
they have an app wage is accessible for free as
opposed to paying two hundred dollars a month, they'll end
up using it. So it's the same argument with TikTok.
You know they will go with the functionality. They don't
care about the privacy and.
Speaker 3 (09:09):
The data man keep saying with us and ana ug Runa,
we continue anerog, let's go to earnings. Let's like remove
ourselves from deep seek. What is the character of the
cloud mystery that you see in the earning season, Aniog.
Speaker 5 (09:24):
Yeah, we are anticipating in improvement in growth rates for
both Microsoft and AWS going in. In fact, Microsoft, we
are hoping they come out and increase their guidance for
Azure numbers for the next two quarters. They have been
struggling with supply challenges where they are not able to
fulfill the demand that's coming in in their data centers,
and if they give that signal that you know, that
(09:47):
should revive or pacify some of the fears that we
are seeing right now in the market. Same thing for
a WS AWS has been talking about a recovery in
consumption for their core business looking forward to some comments
sometimes as well from them.
Speaker 4 (10:03):
So I'm hopeful that.
Speaker 5 (10:04):
Both companies will come up with good numbers over the
next you know, ten days.
Speaker 3 (10:08):
But let's say it's not like you know, Procter and Gamble.
It's not a unit and price store hit the revenue line.
But where is the fear in the earning season analog
on the income statement. Is it a margin fear, a
revenue fear, or something I don't understand.
Speaker 5 (10:24):
I think it's going to be a backlock slash revenue
fear more than the margins, because everybody understands this is
a very you know, unique business. When you have a
fixed cost business like this, you can scale it up
in the near town, but once you reach a mature face,
you know, all the revenue comes back, or you can
distribute that over a very large platform.
Speaker 4 (10:43):
That has been the case for Cloud over the last
you know, ten years.
Speaker 5 (10:46):
So in an expansion of capacity, people don't really you know,
see that as a big issue in the near town
man Deep.
Speaker 6 (10:52):
Last Friday, Meta took their capex guidance from fifty billion,
which was consensus, which was kind of guidance, to sixty
five billion. I've never seen that before. In order of magnitude,
what's going on there.
Speaker 8 (11:06):
Well, and I guess they probably knew this deep Seek advancement,
so it's not as if, you know, they were cut
off guard with the deep Seek release, so they did
it despite the you know, the whole thing around deep Seek.
So to my mind, you know, every company is doubling
down on AI right now, even though we are talking
about pricing, compression and whatnot. But this is the moment
(11:29):
where generative AI will be incorporated in all software, all applications,
and every company wants a piece of it, so they
don't mind spending more on County in.
Speaker 3 (11:38):
The room a ROD does a public one AI. Is
there any interest that Lisa Mateo is waiting and waiting
and waiting for AI.
Speaker 5 (11:48):
See, the thing is that the consumers have already embraced
it through chat, GPT, through Gemini, through Perplexity. But when
you look at the enterprise use case, we are still
in the building phase. And that's probably this year is
going to be like next year, we should see a
much fostered adoption by companies in terms of and that
you said will show up in cloud numbers right now.
(12:09):
Even when you look for somebody like a Microsoft, most
of their AI revenue is from consumer app, which is
basically open AI is shock GPT.
Speaker 7 (12:18):
So all right, Tom, we like to talk to anog.
Speaker 2 (12:20):
Do you like Ludlow this morning?
Speaker 8 (12:22):
No?
Speaker 6 (12:22):
You just I mean we're likes will be Bloomberg Tech exactly.
Speaker 2 (12:26):
I mean, can you get the shoes with the White Souls?
Speaker 7 (12:29):
Yes, that's what David shoes.
Speaker 2 (12:33):
No, you weren't shoes of the White Souls.
Speaker 5 (12:36):
No, No, old school.
Speaker 6 (12:39):
And Tom, you know Man Deep and Hon they're pretty good.
But the research, the real investment research on this whole
Deep seek Chinese AI.
Speaker 7 (12:48):
Robert Lee. These are Bloomberg intelligence analysts in Hong Kong.
Speaker 6 (12:51):
So folks who have access to the Bloomberg terminal go
there because Man Deep and they're just he was out
with research a couple of months ago on this and
he says, hey, this stuff is there.
Speaker 2 (13:01):
Okay.
Speaker 3 (13:02):
You know Lea in Uh in Hong Kong was way
out front of this. Why are we stunned this morning?
Speaker 8 (13:09):
Well because he didn't found the table in terms of saying,
my god, the table we're negative?
Speaker 3 (13:15):
Where are we negative? Three point five percent? In YAC
right now? What I mean, let's bring it back to
what happened. Why were we surprised.
Speaker 8 (13:27):
By this, No, I don't think you are surprised because
open source has been talked about a lot. We have
been talking about how companies are looking to you know,
drive scaling uh at inference time, and how pricing was
a constraint, and so look, this is the way the
industry was headed. It's just that Deep Sea came out
with a you know, a long paper that describes what
(13:49):
they did in detail, and everyone else is kind of
closed source. So being open source has its advantages, and
that's what we're seeing with deep Seat here.
Speaker 6 (13:58):
So the baniers to entry here on ROD just in general,
on AI are the banished entry low for AI.
Speaker 5 (14:08):
I think that's exactly what this paper that Mandep is
alluding to is talking about. That you and I can
go out and build some of this thing without much capital.
Now the question is how are you going to distribute it?
And which is why I go back to the cloud
providers that those are the guys who are going to
work with enterprises to embed some of this technology. If
you look at somebody like an AWS, they work with
all these providers, they work with Meta Islama model or
(14:31):
Anthropics cloud models. So it's really up to the customer
which model they choose or sometimes they would write the
reroute the inquiry based on what's better for that particular question.
Speaker 4 (14:43):
So it's really the distribution.
Speaker 5 (14:45):
What I'm saying is is the value here which takes
us back to the cloud providers and also do Apple.
Speaker 3 (14:50):
This has been fun erragran and your trooper to get
up this early in Chicago aner Agroana with Bloomberg Intelligence
at Man Deep saying, can we say on behalf of
all of us?
Speaker 2 (14:59):
Mandep, thank you much.
Speaker 3 (15:01):
Okay, we'll try not to do it three times tomorrow,
maybe once, or maybe we'll see.
Speaker 2 (15:05):
A Wednesday or whatever.
Speaker 3 (15:06):
Man Deep sing and enter Agrana. Is what the show
is about, just world class on tech analogy.
Speaker 1 (15:18):
You're listening to the Bloomberg Surveillance Podcast. Catch us live
weekday afternoons from seven to ten am Eastern Listen on
Applecarplay and Android Otto with the Bloomberg Business app, or
watch us live on YouTube.
Speaker 2 (15:31):
We Get Lucky.
Speaker 3 (15:31):
Joe Wisenthal has done so much for Bloomberg, I should say,
for business journalism, and of course it's work here at
Bloomberg with Tracy Alloway. But one of the great things
is he did Paul, he you know, I'm like a
I E I E I O. Joe's actually doing it,
he said, no, but but nobody. He looked at the
summer tour of Light Sweet Crude and said, I can't
(15:54):
do the Light Sweet Crude analysis without AI. You three
days ago you nailed deep Seek. You're using it at home. Yeah,
what's the experience of deep Sea?
Speaker 9 (16:07):
So I'll just see the main There are two main
takeaways for me. One is I have never experienced any
technology in the history of any tech, computers, apps, whatever,
that have zero switching.
Speaker 7 (16:19):
Costs the way AI chatbots have.
Speaker 9 (16:21):
I used Chad GBT for a long time. Claude came out,
the one from Anthropic, I was like, oh, this works Gemini.
I used Gemini now that for some reason, that one
hasn't sucking as much then I. But then deep Sea
came out and I said, or I saw the chap
out last week online and I was like, I'm gonna
try this out. It took me thirty seconds to register,
and I was like, this is just as good as
(16:42):
what I'm getting from Chad. You know, I ask at
random things during the day and.
Speaker 2 (16:47):
It's just zerop radio. It's all, you know.
Speaker 9 (16:52):
I asked a random historical questions or linguistic I'm interested
in Lenco, like.
Speaker 2 (16:55):
Why did Texas lose to Ohio? No, I can't. That's
too painful. I don't ask that.
Speaker 9 (17:00):
But then the other thing is I made a twenty
twenty five one of my New Year's resolutions. Yep, I've
never I said I'm going to try and code something
with AI one of my goals. Yeah, And so I
started like building a little software. But I just asking
to create this code, and uh, I was able to
plug my app that I'm building on my home. It's
a little linguistics app into the Deepseek API, and it's
(17:24):
like a fraction of the cost of what I use
for the open AI API. So, like I just played
around with it. I was like, Wow, here's this thing
that seems just as good as changchipyt. Sorry, here's this thing.
Speaker 2 (17:35):
That seems just good changing formosa.
Speaker 6 (17:39):
So yes, I'm not surprised that, Like, all right, the
dumb question there, Yeah, are we not using Google anymore
because of this?
Speaker 10 (17:45):
No?
Speaker 9 (17:45):
I still well, look it is you know, like I
still do. I still use Google all the time. I
still use Wikipedia all the time. They're just different.
Speaker 7 (17:53):
But what's your initial go to, oh, I need to
know about this.
Speaker 9 (17:56):
I think it really depends on the context. I say,
I really like using AI for things like understanding a
term within a specific context or something like that or something.
But you know, like you don't know if it's true,
So you know, I always do a backup. I see
something and then I say, oh, look at Google is
like is this actually true?
Speaker 3 (18:16):
Morgan Brown, Dropbox Imperial w He's out with the smartest
note I've seen where he's talking about decimal points. I
don't want to go into that. You and Tracy at
audience talk a lot more to techy people about this.
What's their message on the future of American AI versus
(18:36):
these shock threats.
Speaker 9 (18:39):
I mean, you know what, here's what my my take,
which is that they're already I mean, we saw the
Trump announcement last week with Stargate right half a trillion dollar.
Speaker 3 (18:49):
I thought that was I thought that was the movie
with what's happening now?
Speaker 9 (18:54):
I think because of the arrival of this very compelling
Chinese model, is it just cements the notion that there
is a global arms race allah the bomb And I
don't know if achieving AGI or whatever people talk about
is a breakthrough on par or is strategic as a
nuclear bomb. But now everyone is just going to be
we have to beat them, we have to beat them,
and there's gonna be so much money flowing to tech
(19:16):
companies now in this idea that there is a global race, Joe.
Speaker 6 (19:19):
I feel like by the end of business today, the
top star on the Bloomberg terournal will be US government
looking at deep Seek as it's looking at TikTok TikTok.
Speaker 7 (19:29):
Well, you know what, so this is interesting.
Speaker 9 (19:30):
The other thing, and thanks for reminding me of this,
which is that deep Seek is a piece of open,
open source software. It is it's unbannable. Now, there is
an app that theoretically could be banned from the app store,
but I think it'll be high. But it's a piece
of open source software.
Speaker 3 (19:44):
I have a friend.
Speaker 9 (19:45):
You know, you can run it on your own computer
if you have the And so this is fundamentally different
from TikTok in the sense that anyone can run this
piece of code, just download it, and so it's fundamentally unbannable.
Speaker 3 (19:56):
You've been on fire, Adam Tooms this weekend. Very nice,
come your euro dollar three part series. You have Howard
Marx on Good morning, mister Mark, thank you so much
for your support over the years. What's your message from
Howard Marx about this?
Speaker 9 (20:11):
Odd He doesn't think we're in an AYE bubble yet,
he said, you know, he like doesn't zuberans he doesn't see,
which I thought was really interesting because to me, it
feels like exuberance. But the idea of me questioning Howard
marks on questions like what is exuberance really feels like
someone who has experienced multiple cycles over the years, right,
I don't think I would go with my take versus take.
Speaker 2 (20:31):
Okay, did you go to NAM were you out?
Speaker 4 (20:32):
And well?
Speaker 3 (20:35):
Well, I mean the endorsements like sweet crudez. Should I
go with the Gratch, pink Rander Penguin?
Speaker 2 (20:42):
I mean, that's what's got the fishman pick you got it?
You got it? I mean it's five forty nine and Sweetwater.
I'll tell missus that's like you got to it's free
and it's pink.
Speaker 7 (20:54):
Is that done with the Jeff?
Speaker 4 (20:56):
Oh?
Speaker 3 (20:56):
Yeah, they did, Joe bid But I'm sorry, Wretches on fire,
Retches on fire, that's the real break. Fender bought Gretch
and you know what's light sweet crew?
Speaker 2 (21:07):
Doing you playing.
Speaker 9 (21:08):
We have a show this Wednesday the eleventh Street Bar
in Manhattan. It's free, eight o'clock East Village. Everyone should
come out. We're gonna play some new songs.
Speaker 2 (21:18):
Are you really for?
Speaker 3 (21:19):
And I like, you know, I'm in bed at a thirty?
Does it really start at eight? Or are you doing
the rock star?
Speaker 9 (21:25):
Do we have an opening act at eight? We're going
to be on just after nine? But you know after
come out one night. It's one night, yeah.
Speaker 7 (21:32):
But the next morning have one.
Speaker 9 (21:35):
You have one off day the next day.
Speaker 2 (21:37):
Okay. So you got Howard Marks coming up online out today.
Check it out?
Speaker 3 (21:41):
Okay, Joe Wisenthal, thank you so much and really folks
followed Joe the Stalwart and uh on Twitter, and also
followed Joe on LinkedIn where he's really going step by
step through his experience with AI, which I think is
really cool.
Speaker 1 (22:00):
This is the Bloomberg Surveillance Podcast. Listen live each weekday
starting at seven am Eastern on Apple Corplay and Android
Auto with the Bloomberg Business app. You can also listen
live on Amazon Alexa from our flagship New York station,
Just Say Alexa play Bloomberg eleven thirty.
Speaker 3 (22:18):
Stacey Raskin is double chemical engineering definitive out of UCLA
and mit he wandered back to the Avenue of the
Stars where he is a star for Bernstein. What you
need to know is in the old days, you had
a black covered book from Sanford Bernstein and you knew
good morning Brad Hints that you were dealing with the
(22:40):
best Stacy honored that you could join us this morning.
Will you change by hold sell on the American tech
companies because of this Deep Seak announcement, I mean.
Speaker 2 (22:52):
Lok, so we're not change.
Speaker 10 (22:54):
We haven't changed anything. We get killed ourselves over the
weekend to go through some of this stuff and put
some views in places like I had three kind of
broad takeaways please from yeah, from from from the work.
But number one is I think there's been a headline.
But I think this is why the circus. There was
a headline that kind of came out all is that
(23:16):
was kind of like, hey, deep Sea duplicated open AI
for five million dollars, and then went so I do
not think they duplicated open AI for five million dollars.
Number two, Look, I think the models that they built
are fantastic. They're really really good. But I don't think
they're miracles. We can talk about why and and number
three finally, like, I think the panic and there's clearly
(23:37):
a full blown on panic going on over the weekend
and this morning, I think it's overblown. I think, you know,
I do not think that deep seek is the doomsday
for for AI infrastructure.
Speaker 2 (23:48):
For those of you are for those of you are
on radio.
Speaker 3 (23:51):
On YouTube, you can see a space shuttle behind a
gentleman from m I T I.
Speaker 10 (23:56):
I didn't even know that I lived.
Speaker 3 (23:58):
That Stacey, and we were certain it worked until the
heat panels didn't work in January of nineteen eighty six.
Do we have a certitude that the stuff is going
to work for Nvidia or Microsoft? Are the other fourteen
names you follow?
Speaker 10 (24:15):
I cover ten names Microsoft and some of the others
are covered by my college cover semicaracters. When you talk
about work, like, I don't understand what you mean. But
the question here is these models look really efficient, and
the resources that were required to train them, you know,
look lower than what we've seen from some other models
in the past. And then even you look at where
(24:35):
they're actually pricing the stuff, and they're pricing and they're
pricing it at levels that are quite a bit lower
where say open ay makes their models available on if
you want to go use them. And so the question is,
given all of that that you need less, you need
fewer GPUs, that's that's the big one, and fewer infrastructure
buildouts and everything else. And so I mean, just just
to talk about it's like, what have they done. They've
(24:58):
got two flavors of models. They've got something called V three,
which is sort of like a chat model. It uses
something called a mixture of experts that makes it very efficient.
And they've got a reasoning model called one where they
used reinforcement learning and some other techniques from the V
three base model to build rivals to say to open
a eyes one reasoning model. If you look at the
(25:23):
why they're so efficient in terms of the training that
they did on and the compute resources they used to
train the V three models, it's very efficient. They compared
it to say like LAMA four or five B, and
it took you know, roughly ten percent of the computer
resource to train it then metas a LAMA four or
five B did. However, the structure of the model, would
that make sense? It uses something called a mixture of experts.
(25:46):
That's the whole point. Other mixture of experts, models that
are out there, you know, take anywhere from you know,
a six to one seventh as mount compute resources.
Speaker 2 (25:55):
This is a good one.
Speaker 10 (25:55):
It was like one tent or even even better than that.
But I don't think it's a miracle. And you got
to remember, like remember I'm a semiconductor analyst. Like for
fifty years, costs and semiconductors got cut in half, like
for every every cheers for fifty years. That was not
a bad thing for semiconductor man, that was a good thing.
We all want adoption. To get adoption, we need lower costs.
(26:19):
This is something called Jevans paradox, right, like increasing efficiency
drives actually more demand, more net demand, not less demand.
I'm a firm believer in that. I have a strong
suspicion that if we can free up more compute resources,
it will get it.
Speaker 2 (26:32):
Will be used.
Speaker 10 (26:33):
I don't think we're anywhere close to the end of
compute needs for artificial intelligence. We're gonna need this just
given I mean.
Speaker 2 (26:39):
The costs have been going up. Right next year, we
need this.
Speaker 3 (26:43):
But anytime anybody coast William Stanley, Jefvins on the show,
they get to come back.
Speaker 7 (26:47):
They get to continue.
Speaker 6 (26:48):
Stacey, what do you expect to hear from the semiconductor
companies that you cover in terms of their response to
this news, whether it's and video or others.
Speaker 10 (26:56):
Yeah, well, I mean, so we were in the middle
of earning season right now, so we'll see as they
come out, right, but we forget the semiconductor comings from it.
We've already seen, like like some of the other like responses,
and you have to remember everything that that that we've
seen deep seaking like this. I don't want to I
don't want to dismiss it like they're they're really good
and very efficient, they're very smart engineers. But this is
these are not things that are not known, and you know,
(27:17):
these are not things that like the the other top
tier AI researchers and AI labs are not aware of
and potentially already using themselves. Right, And if we look
at what we saw last week on top of all
of this, clearly spending is going up, not down. So
Meta just just significantly increased their capex. I think on Friday,
right we had the whole Stargate announce and five hundred
(27:39):
billion dollars whatever that is, and even China a few
days ago announced a one trillion yuan which is about
one hundred and forty billion dollar US sort of like
sovereign like AI effort right, so like clearly spending is
still going up, not down. And and these models, I said,
they didn't just just dump on us. Yesterday, the Deep
Seat V three was released the end after just.
Speaker 3 (28:00):
Well, Tracey Stacey, I got to get this in. It's
too important. At UCLA. You had to survive Physics one
A a few years ago. The smartest thing I've seen
is that American AI is trying to go out to
thirty two decimal points, where the Chinese are saying, you
know what eight decimal points is?
Speaker 2 (28:19):
Okay?
Speaker 3 (28:19):
Is this a question of granularity or we're being too
perfect in AI? And the Chinese are saying, you don't
need to be too perfect.
Speaker 10 (28:28):
Now, so they use what you're trying. They use mixed precision.
They use mixed precision numbers to train this that they
used to have floating point eight eight bit numbers you
typically use floating point sixty or floating point thirty two.
It's actually really interesting though, in Vidia with their Hopper generation,
which is the H one D eight hundreds. It has
(28:49):
something called on it. It's actually called a transformer engine.
What that actually did was it enabled you to train
at eight bit precision. And in fact Blackwell, which is
their their their next generation stuff, this spending on now
actually enables four bit precision. So no, these are things
that everybody is looking at. It may it may be
(29:10):
that that deep Seek is the first ones to actually
deploy this in a very large model, but these are.
Speaker 2 (29:14):
Things they're looking at it.
Speaker 10 (29:15):
And if you actually look over time, we've seen the
numerical precisions that people are using coming in and other
the reason you do this, did you get more compute?
If I go from sixteen bit to eight in theory,
I get twice as many flops, twice as many operations
per second. But no, these are things that have been
happening all along. This is just the next step. And
they're again I don't want to dismiss them, Like I said,
they've done some really clever things and but but none
(29:37):
of this is a miracle or and none of this
I think is an unknown or a slap in the
face to the other AI labs.
Speaker 3 (29:45):
We're out of time. Don't be a stranger thing. You're
trooper to get up this early. Really really appreciate these.
In Los Angeles daily lunches at Michael's. They're closed now
because of the fire, but I'm sure when Michael's reopened,
you know Stacy will be there. Stacy redskin with burns,
definitive there.
Speaker 1 (30:05):
This is the Bloomberg Surveillance Podcast. Listen live each weekday
starting at seven am Eastern on Applecarplay and Android Auto
with the Bloomberg Business app. You can also watch us
live every weekday on YouTube and always on the Bloomberg terminal.
Speaker 3 (30:20):
The daily look at the front pages. It's brought to
you by IBKR. With the average US temperature in January
twenty twenty five be greater than thirty four degrees. Trade
your prediction with IBKR forecast Trader that yes was recently
at fourteen percent. Makes sense by yes or no anterner
dollar if you're right. Go to ibkr dot com slash
(30:43):
forecast Interactive Brokers ibkr dot com slash Forecasts.
Speaker 2 (30:50):
We dashed to the newspapers. Lisa tell me we're deep,
deep seek free.
Speaker 7 (30:55):
We are newspaper free.
Speaker 11 (30:56):
I know, trying to do something a little different. We
talked about more companies pushing workers back to the office, right,
but not everyone at the company is being treated the same.
They're saying. It's causing a lot of tension now in
the workplace because you have some people in a company,
like the top performers, the employees with those certain skill sets.
They're being given that flexibility, and so he comes in
(31:18):
once in a while because the companies don't want to
lose them to other companies. You know, it's a whole
thing of stealing each other's workers. So it's a lot
of tension in the workplace that's starting to cause attention
between the managers and the staff on top of it.
Speaker 3 (31:30):
So it's truly fluid, I mean into the warmer season
into spring.
Speaker 2 (31:35):
To me, it's a hugely fluid debate.
Speaker 7 (31:37):
Some people take Fridays off in the summertime.
Speaker 2 (31:39):
I've heard, well they do.
Speaker 7 (31:41):
Yes, I've heard we missed that memo.
Speaker 2 (31:43):
It works.
Speaker 3 (31:44):
It works out with me in PA because he takes
every Friday off, and I'm like, you know, you're in
the fam and all that.
Speaker 7 (31:50):
It works out.
Speaker 11 (31:51):
Nextll does we want to go to the airline industry
because you're seeing this shift in the pricing power now
going back to the carriers. If you've been pricing tickets
out there, notice prices for the cheapest flights up twelve
percent this month from a year earlier. That's according to Hopper.
But the airlines are saying people are gonna pay it
no matter what. And they're saying people are paying it domestic, yes,
international as well.
Speaker 8 (32:13):
And there.
Speaker 11 (32:16):
I don't know, I don't know how much of a
deal there you got, But they're saying what's shifting to
is that the discount airlines. You know, you've seen them
who usually go to these you know consumer you know,
budget conscious people. They are starting to shift to the
to the more advanced. And then you have the luxury sector.
Now you have all these luxury items being offered out
(32:36):
on the plane. So if you go on a flight,
you may have a you know, a cosmetic bag filled
with all these premium you know, lotions and perfumes. You're
getting caviars, champagne, cae.
Speaker 4 (32:50):
You have not seen.
Speaker 11 (32:53):
You haven't gotten your crystal champagne.
Speaker 2 (32:56):
What is it Hudson River, Yeah, exactly right.
Speaker 6 (32:59):
I mean it's everything's a cart though you pay for
everything now, so I mean, you know the bags that
that that the better seats so.
Speaker 3 (33:05):
We had an offspring at the airport with a lot
of luggage. Call up and say, I can't believe what
they're you know, yep, they want it for me right now,
for three extra bags or.
Speaker 7 (33:15):
Whatever, two weeks in Europe with one check on check it. Yep,
I don't take it. You don't take a bag. I
don't take the kids.
Speaker 2 (33:22):
They travels, they take too much stuff today.
Speaker 11 (33:24):
Next, lastly, okay, so you heard about costco is switching
back from coke back to coke from pepps. Really, yeah,
that's that's huge for you. I know, Paul, you know
you're the Coke.
Speaker 6 (33:36):
Pepsi coke person. But I won't die on that hill.
You will put PEPs in front of me.
Speaker 7 (33:40):
I'll be fine.
Speaker 11 (33:40):
But it's like this huge debate on.
Speaker 3 (33:42):
Someone to explain a Costco on a Saturday morning Cocono Pepsi.
Speaker 4 (33:46):
Yeah.
Speaker 11 (33:46):
People are like Ritt. They want it with their dollar
fifty hot dog. They want to have their certain soda
because they're used to it.
Speaker 2 (33:53):
Can you tell the difference?
Speaker 7 (33:55):
I can't wait.
Speaker 2 (33:56):
Wait, you haven't had mix?
Speaker 3 (34:00):
Now.
Speaker 6 (34:00):
I'm the soda person here and I actually have almost
a year and a half. I put in a ticket
every day to bring back the soda machines on the
sixth floor. It's going up to the highest level of
bloom But if you.
Speaker 11 (34:09):
Did the blind taste test, could you tell you yes,
you could?
Speaker 4 (34:11):
You can.
Speaker 2 (34:12):
Yeah.
Speaker 6 (34:13):
But again, I'm a coke person, but you put a
PEPs in front of me just as happy.
Speaker 7 (34:17):
I mean, I don't know anybody else back there coke
PEPSI I don't know who cares. I don't know.
Speaker 6 (34:23):
But the organic stuff they have on the sixth floor
has to go. If somebody's listening that can do something
here at bloom or get rids.
Speaker 7 (34:30):
They have organic soda on the sixth floor. I mean,
that's un American. It's un America exactly.
Speaker 6 (34:36):
So I'm talking to the highest people in here at
Bloomberg to get our soda.
Speaker 2 (34:39):
Machines, the newspapers. Lisa Mateo, thank you so much.
Speaker 1 (34:45):
This is the Bloomberg Surveillance podcast, available on Apple, Spotify,
and anywhere else you get your podcasts. Listen live each
weekday seven to ten am Easter and on Bloomberg dot com,
the iHeartRadio app, tune In, and the Bloomberg Business app.
You can also watch us live every weekday on YouTube
(35:05):
and always on the Bloomberg terminal
Bloomberg Audio Studios, podcasts, radio news. This is the Bloomberg
Surveillance Podcast. Catch us live weekdays at seven am Eastern
on Apple CarPlay or Android Auto with the Bloomberg Business app.
Listen on demand wherever you get your podcasts, or watch
(00:25):
us live on YouTube right now.
Speaker 2 (00:28):
In joining us.
Speaker 3 (00:29):
And you know last night when this was blowing up,
I said get me Mandeep saying.
Speaker 2 (00:32):
You'll join us in a bit here.
Speaker 3 (00:34):
But aniog Rana joins us now from Chicago with a
little different take than Mandeep. More on the cloud, more
on the systems analysis, if you will. Of these big companies,
man Anerog, thank you so much for joining.
Speaker 2 (00:46):
You don't do biholed sell.
Speaker 3 (00:48):
But would you suggest the cell side will shift opinion today.
Speaker 4 (00:54):
For the cloud providers?
Speaker 5 (00:55):
I don't think so because, as you know Paul just
mentioned about the tweet from my ACOFT CEO, he's probably
a little bit happier because for him, if the cost
curve goes down in the long run, and we don't
know if that's going to happen, you know, his adoption
rate goes up and his CAPEX goes down.
Speaker 4 (01:11):
So it's it's you know, in a roundabout way.
Speaker 5 (01:15):
Now, we don't think there is going to be an
impact on CAPEX in the near term, but in the
long run, if that just shifts, you know, it helps
their profitability. But more important to that the enterprise you know,
you could say take create on AI adoption goes up the.
Speaker 3 (01:29):
Theme this morning, Paul's been great about this is like
drunken sailors from where in Anerich you're just world class
at this. Are they drunken sailors in terms of CAPEX?
Or is there a plan? Is there a sequence? Is
there a pace to spending billions?
Speaker 5 (01:47):
So one of the things, if you would say, is
if if this AI adoption curve, which is absolutely in
infancy right now, the only thing we have seen is
in the consumer world, if enterprises start to put this
in their core app locations, then they would need a
lot more you could say, firepower to run it. Most
of the things that we have discussed is people would
(02:08):
not want to do this on their own premise. They're
bigger companies will but most of them will go to
a handful of cloud providers, and that would be Microsoft, Aws,
Google and then Oracle. So for all these companies to
increase their capacity now they are only seeing the orders
and based on that they are expanding their data center footwork.
Speaker 4 (02:28):
The question is whether do.
Speaker 5 (02:29):
They need the expensive chips to do with this or
can they run this in with cheaper chips. And that's
really I think the biggest discussion going to be over
the next twelve months.
Speaker 6 (02:38):
An raglick Tom has often want to say, my head
is spinning here. Just since Friday. I've got a big,
big tech company called Meta I call it Facebook given
the IPO. We spent a lot of time thinking about
that name in the tech the ticker FB. They up
their cap x from fifty billion to sixty five billion,
a large part on AI. Now I've got this Deep
(03:00):
Seek thinking about I can do this in.
Speaker 7 (03:02):
Much lower costs. How do we square those two things?
Speaker 5 (03:06):
Yeah, but one thing is for sure, this Deep Seek
news of the paper has been out for a while,
so it does got It did get publicized over the
last few days. But I'd be very surprised if Meta
AWS and Microsoft did not know about this. It has
been around for a while, so both of them, both
Meta and Microsoft talked about their capex last week after
knowing this news, so you know they have moved.
Speaker 4 (03:29):
I mean, they're far smarter than I am when it
comes to this.
Speaker 5 (03:31):
So if they're taking up their Capex, I mean there
must be a real need for it.
Speaker 3 (03:35):
What's it mean for Coopertino there having an early morning?
They don't have lotes at Coopertino? They got something, Lisa.
What's the thing with the green powder in it?
Speaker 2 (03:45):
The machi?
Speaker 3 (03:46):
Yeah matter, they're having the machiat Coopertino some undrinkable bilge.
What are they saying at Apple today?
Speaker 5 (03:53):
They're actually smiling and saying, you guys fight it out
and tell me what's the best large language model out there.
I'll just put it in my phone and I won't
even pay for it. So they if you just look
at you know, Apple right now, that's kind of the
safety haven for people is they're not spending that capex.
If you look at that Capex line, it hasn't moved
at all in two years, and the free cash slow
just keep on coming. Their issue, honestly, is more so
(04:16):
on the you know, the growth side. They are having
challenges in China that needs to you know, smoothened out
over the next twelve months. Otherwise this is going to
be another year for them. To grow only four five percent.
That's the big issue for them. Nothing to do with AI.
Speaker 6 (04:30):
All right, Well, let's just stay on the Apple story.
It stocks down eleven percent year to date this year,
it's only a fifteen percent over the trailing twelve months,
which is radical on the performance. And for a company
like Apple, is there a bulkcase here short term?
Speaker 7 (04:45):
For Apple?
Speaker 5 (04:47):
In the short term, it's very difficult to come up
with the bulkcase because, as we talked about it, even
you know a few months ago, Apple Intelligence is not
going to be out in other parts of the world.
It's not out in China, and China is where they
are struggling honestly at this and they need to come
out and pacify investors that you know, we have a
strategy in China, we will gain market share back. Otherwise
(05:08):
this is going to be another mut the year for them.
We think the next phone, which is the one that's
coming out in September, I know that's fa I think
that's the big catalyst, but that's like, you know, that's
eight months out, nine months out.
Speaker 3 (05:19):
Right, and Agrana with us in Chicago right now. We
welcome all of you on your commute across the Nation.
For those of you in the office at home on YouTube,
thank you for being on youtubday. Just brilliant live chat
of learning a lot from people smarter than me on
the YouTube live chat this morning. Subscribe at Bloomberg Podcast.
That's the best way to get to the show out there.
(05:42):
This is a rare privilege. And for senior management listening
on radio right now. We adhere to their policy which
is Ana Agarana and man Deep Sing can never be
in the same ruct together. It's too much worse to
the to the franchise. But Paul, why don't you bring
in man Deep Sing?
Speaker 2 (05:58):
As we talked to both and Ruana and.
Speaker 6 (06:00):
Mandy man Deep Seeing covers the technology space along with
Ana rog Rana, there are two leaders globally running our
technology franchise and it is a global franchise analyst in Europe,
Asia and North America. Man Deep, But I guess what
we're hearing such Antela from Microsoft put out something that hey,
(06:20):
cost coming down is pretty natural for the technology space.
That happens, and it's a good thing because it helps adoption.
Speaker 8 (06:28):
Does that make sense to you, Well, clearly, I think
the investor expectations were different, which is why you see
this sort of stock reaction.
Speaker 7 (06:37):
So yes, we.
Speaker 8 (06:38):
Have Moore's law and we have other laws in terms
of figuring out how the cost curve may look like
over time. But clearly this is a step change in
terms of, you know, what you can do with the
existing chips that are out there, as well as what
you will need going forward. So I think this is
a big moment in terms of just that we can
(07:01):
see with these chips and the kind of scale that
is required for the next versions of these foundational models.
But there is no doubt that this makes generative AI
more accessible to software companies as well as you know,
overall users.
Speaker 6 (07:15):
Anrak, what do you think this means to you know,
I know you guys at Bloomberg Intelligence you have access
to this I d C data, which is some of
the best data on technology spending in the marketplace, and
you guys feature that in your research. Is this going
to change i DC's forecast of tech spending?
Speaker 7 (07:34):
AI spending? Do you think so?
Speaker 5 (07:37):
AI spending has been going up, Paul, We've talked about
this a few times, but it's the other non AI
spending that has actually not done so well in the
last two years and over the next I would say month,
month and a half, we will get a much clearer
picture if corporations or enterprises are moving up there over
our tech budgets or not, because right now AI spending
(07:57):
is coming at the expense of areas you know, normal
software or you know, just bording hardware upgrades or consulting
that really needs to pick up because a large portion
of the tech ecosystem is some of those companies.
Speaker 7 (08:12):
Mande.
Speaker 6 (08:12):
If I think when I think of this news today again,
I'm like everybody else. Our listeners and viewers are learning
about deep Seek and what it means. I can't help
but think about TikTok at some point. Again, I feel
like by the end of business today, we're gonna have
a story coming into Washington. There's gonna be an opinion
coming from Washington as it relates to deep Seek. Does
oppose a similar risks? Do you think to TikTok from
(08:33):
a regulator's perspective, Well.
Speaker 8 (08:35):
It's still the app is still you know, early in
terms of adoption, and look, I'm sure it's got millions
of users in East Asia and you know, China and whatnot.
But I think when it comes to AI consumers, just
look at the functionality, and you know the fact that
(08:57):
they have an app wage is accessible for free as
opposed to paying two hundred dollars a month, they'll end
up using it. So it's the same argument with TikTok.
You know they will go with the functionality. They don't
care about the privacy and.
Speaker 3 (09:09):
The data man keep saying with us and ana ug Runa,
we continue anerog, let's go to earnings. Let's like remove
ourselves from deep seek. What is the character of the
cloud mystery that you see in the earning season, Aniog.
Speaker 5 (09:24):
Yeah, we are anticipating in improvement in growth rates for
both Microsoft and AWS going in. In fact, Microsoft, we
are hoping they come out and increase their guidance for
Azure numbers for the next two quarters. They have been
struggling with supply challenges where they are not able to
fulfill the demand that's coming in in their data centers,
and if they give that signal that you know, that
(09:47):
should revive or pacify some of the fears that we
are seeing right now in the market. Same thing for
a WS AWS has been talking about a recovery in
consumption for their core business looking forward to some comments
sometimes as well from them.
Speaker 4 (10:03):
So I'm hopeful that.
Speaker 5 (10:04):
Both companies will come up with good numbers over the
next you know, ten days.
Speaker 3 (10:08):
But let's say it's not like you know, Procter and Gamble.
It's not a unit and price store hit the revenue line.
But where is the fear in the earning season analog
on the income statement. Is it a margin fear, a
revenue fear, or something I don't understand.
Speaker 5 (10:24):
I think it's going to be a backlock slash revenue
fear more than the margins, because everybody understands this is
a very you know, unique business. When you have a
fixed cost business like this, you can scale it up
in the near town, but once you reach a mature face,
you know, all the revenue comes back, or you can
distribute that over a very large platform.
Speaker 4 (10:43):
That has been the case for Cloud over the last
you know, ten years.
Speaker 5 (10:46):
So in an expansion of capacity, people don't really you know,
see that as a big issue in the near town
man Deep.
Speaker 6 (10:52):
Last Friday, Meta took their capex guidance from fifty billion,
which was consensus, which was kind of guidance, to sixty
five billion. I've never seen that before. In order of magnitude,
what's going on there.
Speaker 8 (11:06):
Well, and I guess they probably knew this deep Seek advancement,
so it's not as if, you know, they were cut
off guard with the deep Seek release, so they did
it despite the you know, the whole thing around deep Seek.
So to my mind, you know, every company is doubling
down on AI right now, even though we are talking
about pricing, compression and whatnot. But this is the moment
(11:29):
where generative AI will be incorporated in all software, all applications,
and every company wants a piece of it, so they
don't mind spending more on County in.
Speaker 3 (11:38):
The room a ROD does a public one AI. Is
there any interest that Lisa Mateo is waiting and waiting
and waiting for AI.
Speaker 5 (11:48):
See, the thing is that the consumers have already embraced
it through chat, GPT, through Gemini, through Perplexity. But when
you look at the enterprise use case, we are still
in the building phase. And that's probably this year is
going to be like next year, we should see a
much fostered adoption by companies in terms of and that
you said will show up in cloud numbers right now.
(12:09):
Even when you look for somebody like a Microsoft, most
of their AI revenue is from consumer app, which is
basically open AI is shock GPT.
Speaker 7 (12:18):
So all right, Tom, we like to talk to anog.
Speaker 2 (12:20):
Do you like Ludlow this morning?
Speaker 8 (12:22):
No?
Speaker 6 (12:22):
You just I mean we're likes will be Bloomberg Tech exactly.
Speaker 2 (12:26):
I mean, can you get the shoes with the White Souls?
Speaker 7 (12:29):
Yes, that's what David shoes.
Speaker 2 (12:33):
No, you weren't shoes of the White Souls.
Speaker 5 (12:36):
No, No, old school.
Speaker 6 (12:39):
And Tom, you know Man Deep and Hon they're pretty good.
But the research, the real investment research on this whole
Deep seek Chinese AI.
Speaker 7 (12:48):
Robert Lee. These are Bloomberg intelligence analysts in Hong Kong.
Speaker 6 (12:51):
So folks who have access to the Bloomberg terminal go
there because Man Deep and they're just he was out
with research a couple of months ago on this and
he says, hey, this stuff is there.
Speaker 2 (13:01):
Okay.
Speaker 3 (13:02):
You know Lea in Uh in Hong Kong was way
out front of this. Why are we stunned this morning?
Speaker 8 (13:09):
Well because he didn't found the table in terms of saying,
my god, the table we're negative?
Speaker 3 (13:15):
Where are we negative? Three point five percent? In YAC
right now? What I mean, let's bring it back to
what happened. Why were we surprised.
Speaker 8 (13:27):
By this, No, I don't think you are surprised because
open source has been talked about a lot. We have
been talking about how companies are looking to you know,
drive scaling uh at inference time, and how pricing was
a constraint, and so look, this is the way the
industry was headed. It's just that Deep Sea came out
with a you know, a long paper that describes what
(13:49):
they did in detail, and everyone else is kind of
closed source. So being open source has its advantages, and
that's what we're seeing with deep Seat here.
Speaker 6 (13:58):
So the baniers to entry here on ROD just in general,
on AI are the banished entry low for AI.
Speaker 5 (14:08):
I think that's exactly what this paper that Mandep is
alluding to is talking about. That you and I can
go out and build some of this thing without much capital.
Now the question is how are you going to distribute it?
And which is why I go back to the cloud
providers that those are the guys who are going to
work with enterprises to embed some of this technology. If
you look at somebody like an AWS, they work with
all these providers, they work with Meta Islama model or
(14:31):
Anthropics cloud models. So it's really up to the customer
which model they choose or sometimes they would write the
reroute the inquiry based on what's better for that particular question.
Speaker 4 (14:43):
So it's really the distribution.
Speaker 5 (14:45):
What I'm saying is is the value here which takes
us back to the cloud providers and also do Apple.
Speaker 3 (14:50):
This has been fun erragran and your trooper to get
up this early in Chicago aner Agroana with Bloomberg Intelligence
at Man Deep saying, can we say on behalf of
all of us?
Speaker 2 (14:59):
Mandep, thank you much.
Speaker 3 (15:01):
Okay, we'll try not to do it three times tomorrow,
maybe once, or maybe we'll see.
Speaker 2 (15:05):
A Wednesday or whatever.
Speaker 3 (15:06):
Man Deep sing and enter Agrana. Is what the show
is about, just world class on tech analogy.
Speaker 1 (15:18):
You're listening to the Bloomberg Surveillance Podcast. Catch us live
weekday afternoons from seven to ten am Eastern Listen on
Applecarplay and Android Otto with the Bloomberg Business app, or
watch us live on YouTube.
Speaker 2 (15:31):
We Get Lucky.
Speaker 3 (15:31):
Joe Wisenthal has done so much for Bloomberg, I should say,
for business journalism, and of course it's work here at
Bloomberg with Tracy Alloway. But one of the great things
is he did Paul, he you know, I'm like a
I E I E I O. Joe's actually doing it,
he said, no, but but nobody. He looked at the
summer tour of Light Sweet Crude and said, I can't
(15:54):
do the Light Sweet Crude analysis without AI. You three
days ago you nailed deep Seek. You're using it at home. Yeah,
what's the experience of deep Sea?
Speaker 9 (16:07):
So I'll just see the main There are two main
takeaways for me. One is I have never experienced any
technology in the history of any tech, computers, apps, whatever,
that have zero switching.
Speaker 7 (16:19):
Costs the way AI chatbots have.
Speaker 9 (16:21):
I used Chad GBT for a long time. Claude came out,
the one from Anthropic, I was like, oh, this works Gemini.
I used Gemini now that for some reason, that one
hasn't sucking as much then I. But then deep Sea
came out and I said, or I saw the chap
out last week online and I was like, I'm gonna
try this out. It took me thirty seconds to register,
and I was like, this is just as good as
(16:42):
what I'm getting from Chad. You know, I ask at
random things during the day and.
Speaker 2 (16:47):
It's just zerop radio. It's all, you know.
Speaker 9 (16:52):
I asked a random historical questions or linguistic I'm interested
in Lenco, like.
Speaker 2 (16:55):
Why did Texas lose to Ohio? No, I can't. That's
too painful. I don't ask that.
Speaker 9 (17:00):
But then the other thing is I made a twenty
twenty five one of my New Year's resolutions. Yep, I've
never I said I'm going to try and code something
with AI one of my goals. Yeah, And so I
started like building a little software. But I just asking
to create this code, and uh, I was able to
plug my app that I'm building on my home. It's
a little linguistics app into the Deepseek API, and it's
(17:24):
like a fraction of the cost of what I use
for the open AI API. So, like I just played
around with it. I was like, Wow, here's this thing
that seems just as good as changchipyt. Sorry, here's this thing.
Speaker 2 (17:35):
That seems just good changing formosa.
Speaker 6 (17:39):
So yes, I'm not surprised that, Like, all right, the
dumb question there, Yeah, are we not using Google anymore
because of this?
Speaker 10 (17:45):
No?
Speaker 9 (17:45):
I still well, look it is you know, like I
still do. I still use Google all the time. I
still use Wikipedia all the time. They're just different.
Speaker 7 (17:53):
But what's your initial go to, oh, I need to
know about this.
Speaker 9 (17:56):
I think it really depends on the context. I say,
I really like using AI for things like understanding a
term within a specific context or something like that or something.
But you know, like you don't know if it's true,
So you know, I always do a backup. I see
something and then I say, oh, look at Google is
like is this actually true?
Speaker 3 (18:16):
Morgan Brown, Dropbox Imperial w He's out with the smartest
note I've seen where he's talking about decimal points. I
don't want to go into that. You and Tracy at
audience talk a lot more to techy people about this.
What's their message on the future of American AI versus
(18:36):
these shock threats.
Speaker 9 (18:39):
I mean, you know what, here's what my my take,
which is that they're already I mean, we saw the
Trump announcement last week with Stargate right half a trillion dollar.
Speaker 3 (18:49):
I thought that was I thought that was the movie
with what's happening now?
Speaker 9 (18:54):
I think because of the arrival of this very compelling
Chinese model, is it just cements the notion that there
is a global arms race allah the bomb And I
don't know if achieving AGI or whatever people talk about
is a breakthrough on par or is strategic as a
nuclear bomb. But now everyone is just going to be
we have to beat them, we have to beat them,
and there's gonna be so much money flowing to tech
(19:16):
companies now in this idea that there is a global race, Joe.
Speaker 6 (19:19):
I feel like by the end of business today, the
top star on the Bloomberg terournal will be US government
looking at deep Seek as it's looking at TikTok TikTok.
Speaker 7 (19:29):
Well, you know what, so this is interesting.
Speaker 9 (19:30):
The other thing, and thanks for reminding me of this,
which is that deep Seek is a piece of open,
open source software. It is it's unbannable. Now, there is
an app that theoretically could be banned from the app store,
but I think it'll be high. But it's a piece
of open source software.
Speaker 3 (19:44):
I have a friend.
Speaker 9 (19:45):
You know, you can run it on your own computer
if you have the And so this is fundamentally different
from TikTok in the sense that anyone can run this
piece of code, just download it, and so it's fundamentally unbannable.
Speaker 3 (19:56):
You've been on fire, Adam Tooms this weekend. Very nice,
come your euro dollar three part series. You have Howard
Marx on Good morning, mister Mark, thank you so much
for your support over the years. What's your message from
Howard Marx about this?
Speaker 9 (20:11):
Odd He doesn't think we're in an AYE bubble yet,
he said, you know, he like doesn't zuberans he doesn't see,
which I thought was really interesting because to me, it
feels like exuberance. But the idea of me questioning Howard
marks on questions like what is exuberance really feels like
someone who has experienced multiple cycles over the years, right,
I don't think I would go with my take versus take.
Speaker 2 (20:31):
Okay, did you go to NAM were you out?
Speaker 4 (20:32):
And well?
Speaker 3 (20:35):
Well, I mean the endorsements like sweet crudez. Should I
go with the Gratch, pink Rander Penguin?
Speaker 2 (20:42):
I mean, that's what's got the fishman pick you got it?
You got it? I mean it's five forty nine and Sweetwater.
I'll tell missus that's like you got to it's free
and it's pink.
Speaker 7 (20:54):
Is that done with the Jeff?
Speaker 4 (20:56):
Oh?
Speaker 3 (20:56):
Yeah, they did, Joe bid But I'm sorry, Wretches on fire,
Retches on fire, that's the real break. Fender bought Gretch
and you know what's light sweet crew?
Speaker 2 (21:07):
Doing you playing.
Speaker 9 (21:08):
We have a show this Wednesday the eleventh Street Bar
in Manhattan. It's free, eight o'clock East Village. Everyone should
come out. We're gonna play some new songs.
Speaker 2 (21:18):
Are you really for?
Speaker 3 (21:19):
And I like, you know, I'm in bed at a thirty?
Does it really start at eight? Or are you doing
the rock star?
Speaker 9 (21:25):
Do we have an opening act at eight? We're going
to be on just after nine? But you know after
come out one night. It's one night, yeah.
Speaker 7 (21:32):
But the next morning have one.
Speaker 9 (21:35):
You have one off day the next day.
Speaker 2 (21:37):
Okay. So you got Howard Marks coming up online out today.
Check it out?
Speaker 3 (21:41):
Okay, Joe Wisenthal, thank you so much and really folks
followed Joe the Stalwart and uh on Twitter, and also
followed Joe on LinkedIn where he's really going step by
step through his experience with AI, which I think is
really cool.
Speaker 1 (22:00):
This is the Bloomberg Surveillance Podcast. Listen live each weekday
starting at seven am Eastern on Apple Corplay and Android
Auto with the Bloomberg Business app. You can also listen
live on Amazon Alexa from our flagship New York station,
Just Say Alexa play Bloomberg eleven thirty.
Speaker 3 (22:18):
Stacey Raskin is double chemical engineering definitive out of UCLA
and mit he wandered back to the Avenue of the
Stars where he is a star for Bernstein. What you
need to know is in the old days, you had
a black covered book from Sanford Bernstein and you knew
good morning Brad Hints that you were dealing with the
(22:40):
best Stacy honored that you could join us this morning.
Will you change by hold sell on the American tech
companies because of this Deep Seak announcement, I mean.
Speaker 2 (22:52):
Lok, so we're not change.
Speaker 10 (22:54):
We haven't changed anything. We get killed ourselves over the
weekend to go through some of this stuff and put
some views in places like I had three kind of
broad takeaways please from yeah, from from from the work.
But number one is I think there's been a headline.
But I think this is why the circus. There was
a headline that kind of came out all is that
(23:16):
was kind of like, hey, deep Sea duplicated open AI
for five million dollars, and then went so I do
not think they duplicated open AI for five million dollars.
Number two, Look, I think the models that they built
are fantastic. They're really really good. But I don't think
they're miracles. We can talk about why and and number
three finally, like, I think the panic and there's clearly
(23:37):
a full blown on panic going on over the weekend
and this morning, I think it's overblown. I think, you know,
I do not think that deep seek is the doomsday
for for AI infrastructure.
Speaker 2 (23:48):
For those of you are for those of you are
on radio.
Speaker 3 (23:51):
On YouTube, you can see a space shuttle behind a
gentleman from m I T I.
Speaker 10 (23:56):
I didn't even know that I lived.
Speaker 3 (23:58):
That Stacey, and we were certain it worked until the
heat panels didn't work in January of nineteen eighty six.
Do we have a certitude that the stuff is going
to work for Nvidia or Microsoft? Are the other fourteen
names you follow?
Speaker 10 (24:15):
I cover ten names Microsoft and some of the others
are covered by my college cover semicaracters. When you talk
about work, like, I don't understand what you mean. But
the question here is these models look really efficient, and
the resources that were required to train them, you know,
look lower than what we've seen from some other models
in the past. And then even you look at where
(24:35):
they're actually pricing the stuff, and they're pricing and they're
pricing it at levels that are quite a bit lower
where say open ay makes their models available on if
you want to go use them. And so the question is,
given all of that that you need less, you need
fewer GPUs, that's that's the big one, and fewer infrastructure
buildouts and everything else. And so I mean, just just
to talk about it's like, what have they done. They've
(24:58):
got two flavors of models. They've got something called V three,
which is sort of like a chat model. It uses
something called a mixture of experts that makes it very efficient.
And they've got a reasoning model called one where they
used reinforcement learning and some other techniques from the V
three base model to build rivals to say to open
a eyes one reasoning model. If you look at the
(25:23):
why they're so efficient in terms of the training that
they did on and the compute resources they used to
train the V three models, it's very efficient. They compared
it to say like LAMA four or five B, and
it took you know, roughly ten percent of the computer
resource to train it then metas a LAMA four or
five B did. However, the structure of the model, would
that make sense? It uses something called a mixture of experts.
(25:46):
That's the whole point. Other mixture of experts, models that
are out there, you know, take anywhere from you know,
a six to one seventh as mount compute resources.
Speaker 2 (25:55):
This is a good one.
Speaker 10 (25:55):
It was like one tent or even even better than that.
But I don't think it's a miracle. And you got
to remember, like remember I'm a semiconductor analyst. Like for
fifty years, costs and semiconductors got cut in half, like
for every every cheers for fifty years. That was not
a bad thing for semiconductor man, that was a good thing.
We all want adoption. To get adoption, we need lower costs.
(26:19):
This is something called Jevans paradox, right, like increasing efficiency
drives actually more demand, more net demand, not less demand.
I'm a firm believer in that. I have a strong
suspicion that if we can free up more compute resources,
it will get it.
Speaker 2 (26:32):
Will be used.
Speaker 10 (26:33):
I don't think we're anywhere close to the end of
compute needs for artificial intelligence. We're gonna need this just
given I mean.
Speaker 2 (26:39):
The costs have been going up. Right next year, we
need this.
Speaker 3 (26:43):
But anytime anybody coast William Stanley, Jefvins on the show,
they get to come back.
Speaker 7 (26:47):
They get to continue.
Speaker 6 (26:48):
Stacey, what do you expect to hear from the semiconductor
companies that you cover in terms of their response to
this news, whether it's and video or others.
Speaker 10 (26:56):
Yeah, well, I mean, so we were in the middle
of earning season right now, so we'll see as they
come out, right, but we forget the semiconductor comings from it.
We've already seen, like like some of the other like responses,
and you have to remember everything that that that we've
seen deep seaking like this. I don't want to I
don't want to dismiss it like they're they're really good
and very efficient, they're very smart engineers. But this is
these are not things that are not known, and you know,
(27:17):
these are not things that like the the other top
tier AI researchers and AI labs are not aware of
and potentially already using themselves. Right, And if we look
at what we saw last week on top of all
of this, clearly spending is going up, not down. So
Meta just just significantly increased their capex. I think on Friday,
right we had the whole Stargate announce and five hundred
(27:39):
billion dollars whatever that is, and even China a few
days ago announced a one trillion yuan which is about
one hundred and forty billion dollar US sort of like
sovereign like AI effort right, so like clearly spending is
still going up, not down. And and these models, I said,
they didn't just just dump on us. Yesterday, the Deep
Seat V three was released the end after just.
Speaker 3 (28:00):
Well, Tracey Stacey, I got to get this in. It's
too important. At UCLA. You had to survive Physics one
A a few years ago. The smartest thing I've seen
is that American AI is trying to go out to
thirty two decimal points, where the Chinese are saying, you
know what eight decimal points is?
Speaker 2 (28:19):
Okay?
Speaker 3 (28:19):
Is this a question of granularity or we're being too
perfect in AI? And the Chinese are saying, you don't
need to be too perfect.
Speaker 10 (28:28):
Now, so they use what you're trying. They use mixed precision.
They use mixed precision numbers to train this that they
used to have floating point eight eight bit numbers you
typically use floating point sixty or floating point thirty two.
It's actually really interesting though, in Vidia with their Hopper generation,
which is the H one D eight hundreds. It has
(28:49):
something called on it. It's actually called a transformer engine.
What that actually did was it enabled you to train
at eight bit precision. And in fact Blackwell, which is
their their their next generation stuff, this spending on now
actually enables four bit precision. So no, these are things
that everybody is looking at. It may it may be
(29:10):
that that deep Seek is the first ones to actually
deploy this in a very large model, but these are.
Speaker 2 (29:14):
Things they're looking at it.
Speaker 10 (29:15):
And if you actually look over time, we've seen the
numerical precisions that people are using coming in and other
the reason you do this, did you get more compute?
If I go from sixteen bit to eight in theory,
I get twice as many flops, twice as many operations
per second. But no, these are things that have been
happening all along. This is just the next step. And
they're again I don't want to dismiss them, Like I said,
they've done some really clever things and but but none
(29:37):
of this is a miracle or and none of this
I think is an unknown or a slap in the
face to the other AI labs.
Speaker 3 (29:45):
We're out of time. Don't be a stranger thing. You're
trooper to get up this early. Really really appreciate these.
In Los Angeles daily lunches at Michael's. They're closed now
because of the fire, but I'm sure when Michael's reopened,
you know Stacy will be there. Stacy redskin with burns,
definitive there.
Speaker 1 (30:05):
This is the Bloomberg Surveillance Podcast. Listen live each weekday
starting at seven am Eastern on Applecarplay and Android Auto
with the Bloomberg Business app. You can also watch us
live every weekday on YouTube and always on the Bloomberg terminal.
Speaker 3 (30:20):
The daily look at the front pages. It's brought to
you by IBKR. With the average US temperature in January
twenty twenty five be greater than thirty four degrees. Trade
your prediction with IBKR forecast Trader that yes was recently
at fourteen percent. Makes sense by yes or no anterner
dollar if you're right. Go to ibkr dot com slash
(30:43):
forecast Interactive Brokers ibkr dot com slash Forecasts.
Speaker 2 (30:50):
We dashed to the newspapers. Lisa tell me we're deep,
deep seek free.
Speaker 7 (30:55):
We are newspaper free.
Speaker 11 (30:56):
I know, trying to do something a little different. We
talked about more companies pushing workers back to the office, right,
but not everyone at the company is being treated the same.
They're saying. It's causing a lot of tension now in
the workplace because you have some people in a company,
like the top performers, the employees with those certain skill sets.
They're being given that flexibility, and so he comes in
(31:18):
once in a while because the companies don't want to
lose them to other companies. You know, it's a whole
thing of stealing each other's workers. So it's a lot
of tension in the workplace that's starting to cause attention
between the managers and the staff on top of it.
Speaker 3 (31:30):
So it's truly fluid, I mean into the warmer season
into spring.
Speaker 2 (31:35):
To me, it's a hugely fluid debate.
Speaker 7 (31:37):
Some people take Fridays off in the summertime.
Speaker 2 (31:39):
I've heard, well they do.
Speaker 7 (31:41):
Yes, I've heard we missed that memo.
Speaker 2 (31:43):
It works.
Speaker 3 (31:44):
It works out with me in PA because he takes
every Friday off, and I'm like, you know, you're in
the fam and all that.
Speaker 7 (31:50):
It works out.
Speaker 11 (31:51):
Nextll does we want to go to the airline industry
because you're seeing this shift in the pricing power now
going back to the carriers. If you've been pricing tickets
out there, notice prices for the cheapest flights up twelve
percent this month from a year earlier. That's according to Hopper.
But the airlines are saying people are gonna pay it
no matter what. And they're saying people are paying it domestic, yes,
international as well.
Speaker 8 (32:13):
And there.
Speaker 11 (32:16):
I don't know, I don't know how much of a
deal there you got, But they're saying what's shifting to
is that the discount airlines. You know, you've seen them
who usually go to these you know consumer you know,
budget conscious people. They are starting to shift to the
to the more advanced. And then you have the luxury sector.
Now you have all these luxury items being offered out
(32:36):
on the plane. So if you go on a flight,
you may have a you know, a cosmetic bag filled
with all these premium you know, lotions and perfumes. You're
getting caviars, champagne, cae.
Speaker 4 (32:50):
You have not seen.
Speaker 11 (32:53):
You haven't gotten your crystal champagne.
Speaker 2 (32:56):
What is it Hudson River, Yeah, exactly right.
Speaker 6 (32:59):
I mean it's everything's a cart though you pay for
everything now, so I mean, you know the bags that
that that the better seats so.
Speaker 3 (33:05):
We had an offspring at the airport with a lot
of luggage. Call up and say, I can't believe what
they're you know, yep, they want it for me right now,
for three extra bags or.
Speaker 7 (33:15):
Whatever, two weeks in Europe with one check on check it. Yep,
I don't take it. You don't take a bag. I
don't take the kids.
Speaker 2 (33:22):
They travels, they take too much stuff today.
Speaker 11 (33:24):
Next, lastly, okay, so you heard about costco is switching
back from coke back to coke from pepps. Really, yeah,
that's that's huge for you. I know, Paul, you know
you're the Coke.
Speaker 6 (33:36):
Pepsi coke person. But I won't die on that hill.
You will put PEPs in front of me.
Speaker 7 (33:40):
I'll be fine.
Speaker 11 (33:40):
But it's like this huge debate on.
Speaker 3 (33:42):
Someone to explain a Costco on a Saturday morning Cocono Pepsi.
Speaker 4 (33:46):
Yeah.
Speaker 11 (33:46):
People are like Ritt. They want it with their dollar
fifty hot dog. They want to have their certain soda
because they're used to it.
Speaker 2 (33:53):
Can you tell the difference?
Speaker 7 (33:55):
I can't wait.
Speaker 2 (33:56):
Wait, you haven't had mix?
Speaker 3 (34:00):
Now.
Speaker 6 (34:00):
I'm the soda person here and I actually have almost
a year and a half. I put in a ticket
every day to bring back the soda machines on the
sixth floor. It's going up to the highest level of
bloom But if you.
Speaker 11 (34:09):
Did the blind taste test, could you tell you yes,
you could?
Speaker 4 (34:11):
You can.
Speaker 2 (34:12):
Yeah.
Speaker 6 (34:13):
But again, I'm a coke person, but you put a
PEPs in front of me just as happy.
Speaker 7 (34:17):
I mean, I don't know anybody else back there coke
PEPSI I don't know who cares. I don't know.
Speaker 6 (34:23):
But the organic stuff they have on the sixth floor
has to go. If somebody's listening that can do something
here at bloom or get rids.
Speaker 7 (34:30):
They have organic soda on the sixth floor. I mean,
that's un American. It's un America exactly.
Speaker 6 (34:36):
So I'm talking to the highest people in here at
Bloomberg to get our soda.
Speaker 2 (34:39):
Machines, the newspapers. Lisa Mateo, thank you so much.
Speaker 1 (34:45):
This is the Bloomberg Surveillance podcast, available on Apple, Spotify,
and anywhere else you get your podcasts. Listen live each
weekday seven to ten am Easter and on Bloomberg dot com,
the iHeartRadio app, tune In, and the Bloomberg Business app.
You can also watch us live every weekday on YouTube
(35:05):
and always on the Bloomberg terminal
Popular Podcasts
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
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.