February 27, 2025 • 21 mins
Karl Freund of Cambrian-AI Research joined Leonard Lee of neXt Curve to recap February 2025, another action-packed month in the world of semiconductors and accelerated and non-accelerated computing on the neXt Curve reThink Podcast series, Silicon Futures. Karl and Leonard share their thoughts on the state of NVIDIA and their Q4 25' earnings call.
This episode covers:
➡️ Karl's initial impressions from the NVIDIA Q4 25' earnings call (2:31)
➡️ The DeepSeek factor: factored in or not? (6:00)
➡️ The distilled future of AI: cheaper tokens, more models (10:03)
➡️ How will the gravity of AI computing shift and can NVIDIA shift? (11:21)
➡️ The diverging economics of AI supercomputing (16:50)
Hit both Leonard and Karl up on LinkedIn and take part in their industry and tech insights.
Check out Karl and his research at Cambrian AI Research LLC at www.cambrian-ai.com.
Please subscribe to our podcast which will be featured on the neXt Curve YouTube Channel. Check out the audio version on BuzzSprout or find us on your favorite Podcast platform.
Also, subscribe to the neXt Curve research portal at www.next-curve.com for the tech and industry insights that matter.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Karl Freund (00:06):
Next curve.
Leonard Lee (00:09):
Hi everyone.
Welcome to this next curve.
Rethink podcast episode, wherewe break down the latest tech
and industry events andhappenings into the insights
that matter.
And I'm Leonard Lee, executiveanalyst at next curve.
And in this Silicon futuristepisode, we'll be talking about,
what's happening in the month ofFebruary, In this segment, we're
(00:33):
going to be focusing probablyoverly on NVIDIA.
they just had this call, so
Karl Freund (00:40):
we just got off the call.
Leonard Lee (00:41):
yeah, so as always, I'm joined by the always
accelerated, Mr.
Carl Freund of Cambridge andhyphen.
It will say the artificial,always artificial.
No, that's.
That's kind of me.
That would be rude.
That would be evil.
I'm not that bad.
Uh, unfortunately, our, friendJim McGregor of curious research
(01:07):
isn't able to join us at leastfor this segment, but he'll join
us tomorrow and we're going tostitch all of this together to
give you.
Perspective on, the month ofFebruary, which is really turned
out to be quite another insane,month, before we get started,
please remember to like, andshare, react and comment on this
episode and also subscribe hereon YouTube and on buzzsprout to
(01:29):
listen to us on your favoritepodcast platform.
opinions and statements by myguests are their own and don't
reflect.
Mine are those of next curve andwe're doing this to provide, you
know, an open forum fordiscussion and debate on all
things related to, AI, justsemiconductors in general, but
we do a lot of AI on this,because this is really what Carl
(01:52):
loves to talk about.
So, yeah, how are you?
I'm good.
Thanks.
Really?
Little sunburn.
Oh, little sunburn.
Yeah.
That is weird.
Usually sunburns don't really gowell with colds, right?
Karl Freund (02:09):
Yeah, I got sunburn and a cold.
I think the reds from thesunburn, the nose, nose is
probably from drinking too muchalcohol in Mexico last week.
Leonard Lee (02:17):
Oh, yeah, you know, there's always a price to pay,
there's a something.
Variation of mine zoom isrevenge, but, but yeah, well,
yeah, you just got into it.
Karl Freund (02:33):
I think, you know, uh, clearly all the spots and
radical.
Prognostications about DeepSeek,certainly haven't come to pass
yet.
I don't think they will.
I still think that, cheaper isgood.
And if you can build your AImodel cheaper, you'll build more
(02:54):
AI models.
NVIDIA's revenue, didn't showany kind of slowdown whatsoever.
interestingly, you contrast thatwith AMD and, I think the gap is
widening.
I'm coming to the opinion of atleast for the next year, I can't
see anything past a year, butfor the next year, I think
(03:15):
NVIDIA's leadership is not onlysolid, but it is improving
against it's, commercialSilicon, vendor competitors.
they're probably a little moreworried about Google.
but that's just one cloud.
It's just one cloud,,NVIDIA ison all clouds.
And that's, that's a good thing.
And of course, the same would betrue for Tranium, 2, which is
(03:36):
actually looking pretty good.
but again, that's only on AWS,but yeah.
Yeah, all in all really good.
they, they increased, revenue.
They're going up to, 44 billionnext quarter.
4343.
I got it wrong.
Yeah, I bet you.
I bet you they hit 44.
what do you think?
(03:56):
We're done Not going to happen.
So 43 billions, pretty damngood.
And the growth rates compareobviously coming down the cell.
You compare it year over yearand their growth is phenomenal.
They're up tremendously.
Leonard Lee (04:12):
Yeah,
Karl Freund (04:13):
yeah So anyway, I think jensen makes really
interesting comments on his callfirst of all, he says Blackwell
is in full production and we'llfully ramp in the next three
quarters He also made somecomments that I thought were
interesting about the impact ofai on software development And
the impact of ai on software ingeneral.
(04:35):
Yeah, and if he's right, and Ihaven't seen him be wrong for
about seven years, that wouldimply that every computer's
gotta run ai, whether it'sthrough an NPU on a CPU or on an
attached accelerator.
We'll find out.
and it may not happen next, nextearnings, but this whole, we
(04:57):
talked about last time too,right?
The whole idea of he's going tobuild the best AI PC on the
planet, and if he's right aboutsoftware, that would imply that,
his AI PC could do quite well.
Leonard Lee (05:11):
yeah, we'll see how it's positioned, further
downstream, because it is reallya workstation, right?
Karl Freund (05:18):
PC.
3000 bucks.
That's a, that's a low endworkstation.
Leonard Lee (05:22):
Yeah, it's a low end workstation, which is oddly
incredibly small.
Karl Freund (05:27):
yeah, it's the size of a deck of cards.
Leonard Lee (05:29):
Yeah.
Karl Freund (05:30):
I don't want to go back and repeat our last
broadcast, but, that's what I'mlooking forward to in May when
they start doing that.
And GTC later on in March, Iexpect we'll hear more about it.
Leonard Lee (05:41):
Oh, yeah, definitely.
And I think this was a really,interesting earnings report to
go into excuse me, GTC with,obviously, the outlook, that
they provided is very, verypositive.
But, yeah, I don't think thedeep seat thing has been
factored in yet.
I mean, this is from a quarterago.
I mean, deep seat, believe it ornot only happened.
(06:03):
What a month ago,
Karl Freund (06:05):
right?
Leonard Lee (06:05):
know, I think.
People are still figuring out,we hear a lot of talk about R1
and, there's a few things thathe said that didn't kind of
reconcile with what I'm seeing.
number 1, a lot of the rhetoricaround distillation, how that's
really bringing the cost of posttraining down.
So, you know, there'sconflicting narratives there,
(06:27):
where I think he's thinkingthere's going to be more compute
required.
For distillation, but then we'rehearing about folks who are
distilling models, at like,whatever.
Single, you know, like, doubledigit figures, right?
20 bucks, 30 bucks, you know,250 bucks.
So how does that square withwhat he's assuming and that
(06:50):
might be 1 of the interestingthings that we, investigate or
talk about and speak to thevideo folks about is how is post
training, shifting and I think alot of folks aren't putting
enough attention to whathappened with V3.
There are still a lot ofquestions.
(07:10):
That are not answered about V3.
R1 is pretty well known causethey published a lot of,
transparent.
And quite frankly, replicable,methods, in that particular
white paper.
But if you look at V3, there'sstill a lot of unknowns that
people have not addressed.
And I think 1 of the biggestchallenges is going to be to
(07:34):
figure out how they pre trainedthis model where it had this
ridiculous sparsity.
Right?
I mean.
You know, basically, it's 10times more efficient than a
llama model.
At least that's what I I've readin the research search I've
done.
You really still have to look atV3 and what are the implications
(07:56):
on pre training and how is itthat they were able to sit there
and take this and get to that.
you know, the level of sparsitywhere it's just incredibly
efficient and, you know, afriend of mine at Intel sent me
a, study that a team at Intel isdoing exactly on.
That is the efficiency.
Of the token routing within theV3, uh, MOE architecture and
(08:22):
they're examining how is it thatthis model can be so efficient
orders of magnitude moreefficient than anything else.
I don't think I have not read.
A good treatment of thisproblem, other than that paper,
and it's only the 1st step inthis investigation, or what I
would consider the 1st step inan investigation that pretty
(08:42):
much I don't think anyone hasdone.
And so I think that they're onstep 1 and they're still about 5
or 6 steps before you reallyhave a sense of what it is that
deep seek.
And quite frankly, the Chinesemodel builders have done
because.
They didn't just take stock,NVIDIA stock, open source stuff,
(09:04):
and then implement to build thismodel and there's this
misconception that they somehowjust, copied or post trained on
top of, a llama model.
That's not true.
They use the llama architecture,something similar, and they
tweak the crap out of it.
they even designed and builttheir own tokenizer.
So there are all these thingsthat have not been addressed or
(09:25):
not known.
yet we have people out theresaying that this is a nothing
burger.
And it's like, well, if youhaven't asked these questions,
you haven't gone down thesepaths to investigate what these
guys did, how do you know whatthe implications are on the
market?
Because the economics they'reintroducing, even with our one.
Are there it's disruptive,right?
(09:46):
So, I think there's still a lotof questions that need to be
answered, but, we'll see.
I think this is going to playout over the course of the next
2 quarters.
we won't see, the impacts of itfor a while.
But yeah, I think you're right.
I mean, I
Karl Freund (10:00):
also know this, Ali Baba announced a text to video.
equivalent, this morning,
Leonard Lee (10:07):
really?
Karl Freund (10:08):
Yeah, using the same kind of distillation
techniques.
Yeah, I think it's going to bepervasive.
I think it's going to help a lotof people that couldn't afford
to use a eyes and allow them touse a eye.
The question is, where's thatbalance point?
Right?
Yeah.
Devon's paradox is a generalthing.
It doesn't mean it's, If you cutthe price by 2, then the number
(10:32):
of units increased by 2.
It doesn't mean that.
Yeah.
Yeah.
It's it's a general trendobservation.
Leonard Lee (10:41):
Yeah.
And then, and definitely, invideo, or at least Jensen,
probably all of it's Jensen.
They're really banking on thiswhole.
inference time scaling, he citedthe 3 scaling observations.
They're not laws observationsand that as being 1 of them and,
how that really plays out and,factors in the economics, I
(11:03):
think, is going to be important,right?
Especially the token pricescontinue to collapse, which is
different from.
Karl Freund (11:11):
Yes,
Leonard Lee (11:13):
and people get confused.
Karl Freund (11:16):
Yeah, right.
I think we will definitely seedata center growth training
growth slow down.
There's no question in my mind.
The question I have is.
Well, okay.
How fast is Jetson going to rampup things like physical AI,
genetic AI, yes.
And IPCs, how fast will thatramp up?
(11:38):
Let's be able to take the placeof phenomenal growth they've
enjoyed for the data centertraining.
so I know we have more of thatvalid point balance point will
be, I agree.
I think we will know a lot morein the next two quarters as we
see this begin to play out.
Leonard Lee (11:51):
yeah, but already that outlook, the 43 billion is
an indication that a lot of thestuff is already baked in.
Like, he mentioned visibility ofthe pipeline.
but then, as, you know, Carl,I'm always looking at the end
markets and that situation.
I think those are factors thatsometimes are not well, let's
(12:12):
say treated in the.
AI discourse, especially thesupercomputing discourse.
But, I think people do need toconsider that when you listen on
the Microsoft call, Satya willsay, look, the data center build
outs those can be depreciateover a long period of time.
Those are investments that wecan make, that, where the
(12:33):
expenses will be, spread over, adecade or 15 years or so.
but we're going to be flexibleand we'll build out the compute
part, based on demand I think aswe look at the investments going
into the data center, not allparts are fixed.
And not all things areprioritized, in the same way.
(12:57):
when, you look at it from a riskperspective, the likes of
Microsoft, Google and others aregoing to be probably adopting a
similar kind of strategy interms of how they.
manage the risk going forwardbecause there's all risk.
It's just a matter of, what'syour mitigation or, what sort of
(13:17):
controls are you going to put inplace to respond to the either
unexpected or the inevitable.
Karl Freund (13:24):
I think that makes sense.
Leonard Lee (13:27):
Yeah, I really think you, hit the nail on the
head with, these other parts oftheir business, which I think
outside of gaming and gamingkind of took a 20 percent plus,
quarter on quarter, declineprofessional visualization, I
think was up.
It wasn't like 16%.
(13:47):
It's a relatively small market.
So I think the 1 that caught my
Karl Freund (13:51):
Automotives started to grow again.
Their automotives have beenfairly stagnant.
And, I don't think you can pinall that on Toyota.
They haven't even shipped thosevehicles yet that are using
video technology.
So it's not from Toyota.
So that was interesting to seebecause I haven't written them
off for auto, Qualcomm has amuch stronger position.
(14:13):
And much lower costs and, a verygood customer list, that is
pulling, on, on snapdragontoday.
So, that was a bit of a surpriseto me to see the automotive crow
so well.
Leonard Lee (14:26):
Yeah.
Karl Freund (14:26):
if it's sustainable.
Leonard Lee (14:28):
Yeah, exactly.
those are the segments that weneed to keep an eye out on,
especially as you mentioned,things shift away from the large
data center training.
Toward inference.
But I do think that there'sstill going to be a lot of
training.
it might be confederated, acrossthe edge, it's going to be a
different variety of training,probably more of the post fine
(14:49):
tuning type of stuff that's likethe whole life cycle of, the AI
application or the model.
Itself, because those thingsneed to be maintained, right?
yeah, it'll probably thosesegments will be a leading
indicator of how, theirbusinesses, from the data center
and diffuses.
(15:11):
as the AI workloads start to,spread across the edge, I
suppose,
Karl Freund (15:17):
right?
Leonard Lee (15:17):
yeah,
Karl Freund (15:17):
definitely.
Leonard Lee (15:18):
Yeah.
I also
Karl Freund (15:19):
noticed, I-B-M-I-B-M launched another
wave of granite models it's justevery month or two months, they
come out with a whole new way ofnew granite models, IBM clients
can take advantage of moreefficient, more advanced models.
Leonard Lee (15:35):
Yeah, I'm sure there's going to be even more
models now that everyone'sjumping on this, distillation
bandwagon, right?
They've discovered distillation,but, ultimately, I think, the
benefit is going to be.
the opportunities outside of thelarge data centers, especially
as these more robust modelcapabilities can now be brought
(15:57):
down into a smaller, modelfootprint and then, deployed,
even on a IPC, for instance,right?
Oh, that's cool.
Yeah, I mean, the edge isgetting more capable, which is
kind of cool, but and it's sortof democratizing the data center
capabilities across these,different edge environments.
(16:20):
So, I think that's probably 1 ofthe big pluses.
And can you believe it's onlyjust been a month.
Right.
It's just the month of February.
I'm almost afraid GTC around thecorner.
I know.
And then we have GTC in March.
And so, What else?
Anything else?
stand out for you on the call.
Karl Freund (16:42):
No, I kind of covered it.
Yeah.
Leonard Lee (16:46):
the only other thing that I can think about is,
like when Jensen says that theuseful life of these systems,
that they build, whether it's inthe data center or some of the
older.
systems based on older GPUs.
he mentions the fungibility.
I think that you have toreconcile that with one year
(17:08):
cadence.
Right.
Cause if you're like, what isit, the doubling the
performance, is it every year orit's more than double, right?
Karl Freund (17:19):
Yeah, I don't know if I've, I haven't seen him say
whether there's doubling ortripling or whatever, and I, I
think, he was asked thatquestion on a call and his
response was, Hey, look, it'sjust another black well with
more memory and betternetworking and he did say
improved compute cores.
(17:40):
Something to that effect, whichI'm not sure how valid that is.
We'll have to wait and see.
We'll probably hear a lot morein about three weeks.
I know.
GTC.
Yeah.
So yeah, but I think you'reright.
I think the annual Hayden's isgoing to be hard.
I think for the industry tojust, it may be possible that if
(18:03):
they, if that, that they'lljust, I don't know what the,
what people will do.
But if you got two generationsto Blackwell that are just
ramping within six months ofeach other,
Leonard Lee (18:15):
I
Karl Freund (18:16):
think it's really up to NVIDIA to correctly
position each of theseplatforms, for, specific use
cases and models.
And maybe, I don't know, maybeBlackwell becomes more of an
inference platform and, uh,Alter becomes a training.
I don't know.
I'm just guessing there'ssomething along those lines.
You got to have a lead dog inthe race and, you can't have two
(18:39):
lead dog.
Well, you could, but you'recompeting with yourself.
It's interesting that it's hardto forecast,
Leonard Lee (18:46):
yeah, but then, you know, it is interesting because
we may be at an inflectionpoint, because even based on
some of the things that we'vecited, just in the last 20
minutes or so, they're sort ofcounter bailing or contradictory
trends.
I think there's some frictionforming, right?
Whether it's what I call, sortof the token dumping dynamic
(19:08):
where prices are collapsing fortokens, the end market where
monetization is still a problem.
I just got on off of a wholeseries of calls this week.
real monetization is reallynowhere.
Insight, it's actuallydisturbing and then, and then
you have, the economics ofsupercomputing, that, may be
(19:32):
impacted by what deep seek V3,did and so how that plays out,
we still have to.
To, um, you know, see, if therewas ever a really interesting
time in this whole cycle thatwe've been witnessing with,
generative AI, this is probablyjust interesting time right now.
Karl Freund (19:54):
hang on tight.
Leonard Lee (19:55):
I know.
I know GTC is going to be crazy.
So any other last thoughts?
Karl Freund (20:00):
No, I think we pretty much covered the
waterfront
Leonard Lee (20:02):
So, yeah, I know.
And uh, Daisy took off.
So that means that we took offsomething.
Something's going on.
Yeah, I know.
I know.
It must be dinner time.
So, yeah.
With that.
Hey, Carl, thanks a lot.
That's always a pleasure.
Appreciate, yeah, you sharing.
And, before we, uh, take off,remember to like share and
subscribe to the next curve,YouTube channel and also follow
(20:25):
our research at www dot nextdash curve.
com.
Also follow Carl Freund'sresearch at Cambrian hyphen, AI
research at, Cambrian hyphen.
ai.com.
Karl Freund (20:41):
Dot
Leonard Lee (20:41):
Yeah.
Karl Freund (20:41):
In ai.com.
Leonard Lee (20:42):
Yeah.
And, follow us and, check outour Silicon Futures, show,
current past, and, stay tuned.
Yeah, stay tuned and check usout for the.
tech and industry insights inthe semiconductor industry and
all this AI stuff that matters,and we'll see you next time.
(21:02):
Thanks a lot.
Karl Freund (21:04):
Take care.
Next curve.
Leonard Lee (00:09):
Hi everyone.
Welcome to this next curve.
Rethink podcast episode, wherewe break down the latest tech
and industry events andhappenings into the insights
that matter.
And I'm Leonard Lee, executiveanalyst at next curve.
And in this Silicon futuristepisode, we'll be talking about,
what's happening in the month ofFebruary, In this segment, we're
(00:33):
going to be focusing probablyoverly on NVIDIA.
they just had this call, so
Karl Freund (00:40):
we just got off the call.
Leonard Lee (00:41):
yeah, so as always, I'm joined by the always
accelerated, Mr.
Carl Freund of Cambridge andhyphen.
It will say the artificial,always artificial.
No, that's.
That's kind of me.
That would be rude.
That would be evil.
I'm not that bad.
Uh, unfortunately, our, friendJim McGregor of curious research
(01:07):
isn't able to join us at leastfor this segment, but he'll join
us tomorrow and we're going tostitch all of this together to
give you.
Perspective on, the month ofFebruary, which is really turned
out to be quite another insane,month, before we get started,
please remember to like, andshare, react and comment on this
episode and also subscribe hereon YouTube and on buzzsprout to
(01:29):
listen to us on your favoritepodcast platform.
opinions and statements by myguests are their own and don't
reflect.
Mine are those of next curve andwe're doing this to provide, you
know, an open forum fordiscussion and debate on all
things related to, AI, justsemiconductors in general, but
we do a lot of AI on this,because this is really what Carl
(01:52):
loves to talk about.
So, yeah, how are you?
I'm good.
Thanks.
Really?
Little sunburn.
Oh, little sunburn.
Yeah.
That is weird.
Usually sunburns don't really gowell with colds, right?
Karl Freund (02:09):
Yeah, I got sunburn and a cold.
I think the reds from thesunburn, the nose, nose is
probably from drinking too muchalcohol in Mexico last week.
Leonard Lee (02:17):
Oh, yeah, you know, there's always a price to pay,
there's a something.
Variation of mine zoom isrevenge, but, but yeah, well,
yeah, you just got into it.
Karl Freund (02:33):
I think, you know, uh, clearly all the spots and
radical.
Prognostications about DeepSeek,certainly haven't come to pass
yet.
I don't think they will.
I still think that, cheaper isgood.
And if you can build your AImodel cheaper, you'll build more
(02:54):
AI models.
NVIDIA's revenue, didn't showany kind of slowdown whatsoever.
interestingly, you contrast thatwith AMD and, I think the gap is
widening.
I'm coming to the opinion of atleast for the next year, I can't
see anything past a year, butfor the next year, I think
(03:15):
NVIDIA's leadership is not onlysolid, but it is improving
against it's, commercialSilicon, vendor competitors.
they're probably a little moreworried about Google.
but that's just one cloud.
It's just one cloud,,NVIDIA ison all clouds.
And that's, that's a good thing.
And of course, the same would betrue for Tranium, 2, which is
(03:36):
actually looking pretty good.
but again, that's only on AWS,but yeah.
Yeah, all in all really good.
they, they increased, revenue.
They're going up to, 44 billionnext quarter.
4343.
I got it wrong.
Yeah, I bet you.
I bet you they hit 44.
what do you think?
(03:56):
We're done Not going to happen.
So 43 billions, pretty damngood.
And the growth rates compareobviously coming down the cell.
You compare it year over yearand their growth is phenomenal.
They're up tremendously.
Leonard Lee (04:12):
Yeah,
Karl Freund (04:13):
yeah So anyway, I think jensen makes really
interesting comments on his callfirst of all, he says Blackwell
is in full production and we'llfully ramp in the next three
quarters He also made somecomments that I thought were
interesting about the impact ofai on software development And
the impact of ai on software ingeneral.
(04:35):
Yeah, and if he's right, and Ihaven't seen him be wrong for
about seven years, that wouldimply that every computer's
gotta run ai, whether it'sthrough an NPU on a CPU or on an
attached accelerator.
We'll find out.
and it may not happen next, nextearnings, but this whole, we
(04:57):
talked about last time too,right?
The whole idea of he's going tobuild the best AI PC on the
planet, and if he's right aboutsoftware, that would imply that,
his AI PC could do quite well.
Leonard Lee (05:11):
yeah, we'll see how it's positioned, further
downstream, because it is reallya workstation, right?
Karl Freund (05:18):
PC.
3000 bucks.
That's a, that's a low endworkstation.
Leonard Lee (05:22):
Yeah, it's a low end workstation, which is oddly
incredibly small.
Karl Freund (05:27):
yeah, it's the size of a deck of cards.
Leonard Lee (05:29):
Yeah.
Karl Freund (05:30):
I don't want to go back and repeat our last
broadcast, but, that's what I'mlooking forward to in May when
they start doing that.
And GTC later on in March, Iexpect we'll hear more about it.
Leonard Lee (05:41):
Oh, yeah, definitely.
And I think this was a really,interesting earnings report to
go into excuse me, GTC with,obviously, the outlook, that
they provided is very, verypositive.
But, yeah, I don't think thedeep seat thing has been
factored in yet.
I mean, this is from a quarterago.
I mean, deep seat, believe it ornot only happened.
(06:03):
What a month ago,
Karl Freund (06:05):
right?
Leonard Lee (06:05):
know, I think.
People are still figuring out,we hear a lot of talk about R1
and, there's a few things thathe said that didn't kind of
reconcile with what I'm seeing.
number 1, a lot of the rhetoricaround distillation, how that's
really bringing the cost of posttraining down.
So, you know, there'sconflicting narratives there,
(06:27):
where I think he's thinkingthere's going to be more compute
required.
For distillation, but then we'rehearing about folks who are
distilling models, at like,whatever.
Single, you know, like, doubledigit figures, right?
20 bucks, 30 bucks, you know,250 bucks.
So how does that square withwhat he's assuming and that
(06:50):
might be 1 of the interestingthings that we, investigate or
talk about and speak to thevideo folks about is how is post
training, shifting and I think alot of folks aren't putting
enough attention to whathappened with V3.
There are still a lot ofquestions.
(07:10):
That are not answered about V3.
R1 is pretty well known causethey published a lot of,
transparent.
And quite frankly, replicable,methods, in that particular
white paper.
But if you look at V3, there'sstill a lot of unknowns that
people have not addressed.
And I think 1 of the biggestchallenges is going to be to
(07:34):
figure out how they pre trainedthis model where it had this
ridiculous sparsity.
Right?
I mean.
You know, basically, it's 10times more efficient than a
llama model.
At least that's what I I've readin the research search I've
done.
You really still have to look atV3 and what are the implications
(07:56):
on pre training and how is itthat they were able to sit there
and take this and get to that.
you know, the level of sparsitywhere it's just incredibly
efficient and, you know, afriend of mine at Intel sent me
a, study that a team at Intel isdoing exactly on.
That is the efficiency.
Of the token routing within theV3, uh, MOE architecture and
(08:22):
they're examining how is it thatthis model can be so efficient
orders of magnitude moreefficient than anything else.
I don't think I have not read.
A good treatment of thisproblem, other than that paper,
and it's only the 1st step inthis investigation, or what I
would consider the 1st step inan investigation that pretty
(08:42):
much I don't think anyone hasdone.
And so I think that they're onstep 1 and they're still about 5
or 6 steps before you reallyhave a sense of what it is that
deep seek.
And quite frankly, the Chinesemodel builders have done
because.
They didn't just take stock,NVIDIA stock, open source stuff,
(09:04):
and then implement to build thismodel and there's this
misconception that they somehowjust, copied or post trained on
top of, a llama model.
That's not true.
They use the llama architecture,something similar, and they
tweak the crap out of it.
they even designed and builttheir own tokenizer.
So there are all these thingsthat have not been addressed or
(09:25):
not known.
yet we have people out theresaying that this is a nothing
burger.
And it's like, well, if youhaven't asked these questions,
you haven't gone down thesepaths to investigate what these
guys did, how do you know whatthe implications are on the
market?
Because the economics they'reintroducing, even with our one.
Are there it's disruptive,right?
(09:46):
So, I think there's still a lotof questions that need to be
answered, but, we'll see.
I think this is going to playout over the course of the next
2 quarters.
we won't see, the impacts of itfor a while.
But yeah, I think you're right.
I mean, I
Karl Freund (10:00):
also know this, Ali Baba announced a text to video.
equivalent, this morning,
Leonard Lee (10:07):
really?
Karl Freund (10:08):
Yeah, using the same kind of distillation
techniques.
Yeah, I think it's going to bepervasive.
I think it's going to help a lotof people that couldn't afford
to use a eyes and allow them touse a eye.
The question is, where's thatbalance point?
Right?
Yeah.
Devon's paradox is a generalthing.
It doesn't mean it's, If you cutthe price by 2, then the number
(10:32):
of units increased by 2.
It doesn't mean that.
Yeah.
Yeah.
It's it's a general trendobservation.
Leonard Lee (10:41):
Yeah.
And then, and definitely, invideo, or at least Jensen,
probably all of it's Jensen.
They're really banking on thiswhole.
inference time scaling, he citedthe 3 scaling observations.
They're not laws observationsand that as being 1 of them and,
how that really plays out and,factors in the economics, I
(11:03):
think, is going to be important,right?
Especially the token pricescontinue to collapse, which is
different from.
Karl Freund (11:11):
Yes,
Leonard Lee (11:13):
and people get confused.
Karl Freund (11:16):
Yeah, right.
I think we will definitely seedata center growth training
growth slow down.
There's no question in my mind.
The question I have is.
Well, okay.
How fast is Jetson going to rampup things like physical AI,
genetic AI, yes.
And IPCs, how fast will thatramp up?
(11:38):
Let's be able to take the placeof phenomenal growth they've
enjoyed for the data centertraining.
so I know we have more of thatvalid point balance point will
be, I agree.
I think we will know a lot morein the next two quarters as we
see this begin to play out.
Leonard Lee (11:51):
yeah, but already that outlook, the 43 billion is
an indication that a lot of thestuff is already baked in.
Like, he mentioned visibility ofthe pipeline.
but then, as, you know, Carl,I'm always looking at the end
markets and that situation.
I think those are factors thatsometimes are not well, let's
(12:12):
say treated in the.
AI discourse, especially thesupercomputing discourse.
But, I think people do need toconsider that when you listen on
the Microsoft call, Satya willsay, look, the data center build
outs those can be depreciateover a long period of time.
Those are investments that wecan make, that, where the
(12:33):
expenses will be, spread over, adecade or 15 years or so.
but we're going to be flexibleand we'll build out the compute
part, based on demand I think aswe look at the investments going
into the data center, not allparts are fixed.
And not all things areprioritized, in the same way.
(12:57):
when, you look at it from a riskperspective, the likes of
Microsoft, Google and others aregoing to be probably adopting a
similar kind of strategy interms of how they.
manage the risk going forwardbecause there's all risk.
It's just a matter of, what'syour mitigation or, what sort of
(13:17):
controls are you going to put inplace to respond to the either
unexpected or the inevitable.
Karl Freund (13:24):
I think that makes sense.
Leonard Lee (13:27):
Yeah, I really think you, hit the nail on the
head with, these other parts oftheir business, which I think
outside of gaming and gamingkind of took a 20 percent plus,
quarter on quarter, declineprofessional visualization, I
think was up.
It wasn't like 16%.
(13:47):
It's a relatively small market.
So I think the 1 that caught my
Karl Freund (13:51):
Automotives started to grow again.
Their automotives have beenfairly stagnant.
And, I don't think you can pinall that on Toyota.
They haven't even shipped thosevehicles yet that are using
video technology.
So it's not from Toyota.
So that was interesting to seebecause I haven't written them
off for auto, Qualcomm has amuch stronger position.
(14:13):
And much lower costs and, a verygood customer list, that is
pulling, on, on snapdragontoday.
So, that was a bit of a surpriseto me to see the automotive crow
so well.
Leonard Lee (14:26):
Yeah.
Karl Freund (14:26):
if it's sustainable.
Leonard Lee (14:28):
Yeah, exactly.
those are the segments that weneed to keep an eye out on,
especially as you mentioned,things shift away from the large
data center training.
Toward inference.
But I do think that there'sstill going to be a lot of
training.
it might be confederated, acrossthe edge, it's going to be a
different variety of training,probably more of the post fine
(14:49):
tuning type of stuff that's likethe whole life cycle of, the AI
application or the model.
Itself, because those thingsneed to be maintained, right?
yeah, it'll probably thosesegments will be a leading
indicator of how, theirbusinesses, from the data center
and diffuses.
(15:11):
as the AI workloads start to,spread across the edge, I
suppose,
Karl Freund (15:17):
right?
Leonard Lee (15:17):
yeah,
Karl Freund (15:17):
definitely.
Leonard Lee (15:18):
Yeah.
I also
Karl Freund (15:19):
noticed, I-B-M-I-B-M launched another
wave of granite models it's justevery month or two months, they
come out with a whole new way ofnew granite models, IBM clients
can take advantage of moreefficient, more advanced models.
Leonard Lee (15:35):
Yeah, I'm sure there's going to be even more
models now that everyone'sjumping on this, distillation
bandwagon, right?
They've discovered distillation,but, ultimately, I think, the
benefit is going to be.
the opportunities outside of thelarge data centers, especially
as these more robust modelcapabilities can now be brought
(15:57):
down into a smaller, modelfootprint and then, deployed,
even on a IPC, for instance,right?
Oh, that's cool.
Yeah, I mean, the edge isgetting more capable, which is
kind of cool, but and it's sortof democratizing the data center
capabilities across these,different edge environments.
(16:20):
So, I think that's probably 1 ofthe big pluses.
And can you believe it's onlyjust been a month.
Right.
It's just the month of February.
I'm almost afraid GTC around thecorner.
I know.
And then we have GTC in March.
And so, What else?
Anything else?
stand out for you on the call.
Karl Freund (16:42):
No, I kind of covered it.
Yeah.
Leonard Lee (16:46):
the only other thing that I can think about is,
like when Jensen says that theuseful life of these systems,
that they build, whether it's inthe data center or some of the
older.
systems based on older GPUs.
he mentions the fungibility.
I think that you have toreconcile that with one year
(17:08):
cadence.
Right.
Cause if you're like, what isit, the doubling the
performance, is it every year orit's more than double, right?
Karl Freund (17:19):
Yeah, I don't know if I've, I haven't seen him say
whether there's doubling ortripling or whatever, and I, I
think, he was asked thatquestion on a call and his
response was, Hey, look, it'sjust another black well with
more memory and betternetworking and he did say
improved compute cores.
(17:40):
Something to that effect, whichI'm not sure how valid that is.
We'll have to wait and see.
We'll probably hear a lot morein about three weeks.
I know.
GTC.
Yeah.
So yeah, but I think you'reright.
I think the annual Hayden's isgoing to be hard.
I think for the industry tojust, it may be possible that if
(18:03):
they, if that, that they'lljust, I don't know what the,
what people will do.
But if you got two generationsto Blackwell that are just
ramping within six months ofeach other,
Leonard Lee (18:15):
I
Karl Freund (18:16):
think it's really up to NVIDIA to correctly
position each of theseplatforms, for, specific use
cases and models.
And maybe, I don't know, maybeBlackwell becomes more of an
inference platform and, uh,Alter becomes a training.
I don't know.
I'm just guessing there'ssomething along those lines.
You got to have a lead dog inthe race and, you can't have two
(18:39):
lead dog.
Well, you could, but you'recompeting with yourself.
It's interesting that it's hardto forecast,
Leonard Lee (18:46):
yeah, but then, you know, it is interesting because
we may be at an inflectionpoint, because even based on
some of the things that we'vecited, just in the last 20
minutes or so, they're sort ofcounter bailing or contradictory
trends.
I think there's some frictionforming, right?
Whether it's what I call, sortof the token dumping dynamic
(19:08):
where prices are collapsing fortokens, the end market where
monetization is still a problem.
I just got on off of a wholeseries of calls this week.
real monetization is reallynowhere.
Insight, it's actuallydisturbing and then, and then
you have, the economics ofsupercomputing, that, may be
(19:32):
impacted by what deep seek V3,did and so how that plays out,
we still have to.
To, um, you know, see, if therewas ever a really interesting
time in this whole cycle thatwe've been witnessing with,
generative AI, this is probablyjust interesting time right now.
Karl Freund (19:54):
hang on tight.
Leonard Lee (19:55):
I know.
I know GTC is going to be crazy.
So any other last thoughts?
Karl Freund (20:00):
No, I think we pretty much covered the
waterfront
Leonard Lee (20:02):
So, yeah, I know.
And uh, Daisy took off.
So that means that we took offsomething.
Something's going on.
Yeah, I know.
I know.
It must be dinner time.
So, yeah.
With that.
Hey, Carl, thanks a lot.
That's always a pleasure.
Appreciate, yeah, you sharing.
And, before we, uh, take off,remember to like share and
subscribe to the next curve,YouTube channel and also follow
(20:25):
our research at www dot nextdash curve.
com.
Also follow Carl Freund'sresearch at Cambrian hyphen, AI
research at, Cambrian hyphen.
ai.com.
Karl Freund (20:41):
Dot
Leonard Lee (20:41):
Yeah.
Karl Freund (20:41):
In ai.com.
Leonard Lee (20:42):
Yeah.
And, follow us and, check outour Silicon Futures, show,
current past, and, stay tuned.
Yeah, stay tuned and check usout for the.
tech and industry insights inthe semiconductor industry and
all this AI stuff that matters,and we'll see you next time.
(21:02):
Thanks a lot.
Karl Freund (21:04):
Take care.
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