Latest / Strictly Business / Strictly Business Interview with Michael Power

Transcript

Strictly Business Interview with Michael Power · Strictly Business

• 0:00
  You're listening to Strictly Business Podcast with Lindsay Williams.
• 0:05
  With me is Michael Power, former investment strategist at 91, and for our
• 0:11
  podcast, he's a contributor to
• 0:13
  Daily Maverick.
• 0:15
  Last time, Michael spoke of renewables, energy that is,
• 0:17
  but now he's going to startle us by talking about the future of AI.
• 0:20
  Michael, welcome from Kenya, I understand.
• 0:22
  How's Kenya today?
• 0:24
  The sun is shining.
• 0:25
  I can see the odd cloud, but not many.
• 0:27
  It's glorious.
• 0:28
  Fantastic.
• 0:29
  I won't ask you about animals and drinks by the pool and all that sort of thing,
• 0:34
  because we want to get onto this.
• 0:36
  I just want to go back to the renewables chat very, very briefly, because in a
• 0:40
  recent Trump speech,
• 0:42
  I think it was that big ramble-a-thon he did at the United Nations,
• 0:44
  he said China just make the windmills, the wind turbines, that is, which are, of
• 0:50
  course, noisy,
• 0:52
  and they eventually rust away, and they sell them to Europe and other places,
• 0:55
  but they don't have any wind farms themselves.
• 0:58
  That's nonsense, isn't it?
• 0:59
  Complete and utter nonsense.
• 1:01
  I can't tell you how much.
• 1:03
  He famously said in July, I've never seen a wind farm in China.
• 1:07
  Can someone show me one?
• 1:10
  And the fact of the matter is in the northwest of China, where the Gobi Desert
• 1:16
  and the Taklamakan Desert are,
• 1:18
  and there's no people to annoy, no golf courses, for instance, there are vast,
• 1:23
  I don't even know what you call them farms of wind farms and also sun farms.
• 1:29
  so it's a combination of the wind and the sun in the deserts um that are basically
• 1:36
  giving china this
• 1:38
  leg up in the world of energy okay we'll come to that when it comes to ai as well
• 1:44
  and that that vision that i've got of these vast
• 1:46
  farms of wind turbines reminds me of the beginning of the film blade runner 2049 it
• 1:51
  wasn't obviously
• 1:51
  windmills but there was just miles and miles of something futuristic.
• 1:56
  Anyway, I'm getting off the point.
• 1:58
  You say the following in your article, your latest article, which has caused a
• 2:04
  lot of a comment, a bit of a stir, I might add.
• 2:06
  You say the following.
• 2:08
  We are witnessing the birth of a new paradigm, artificial intelligence as
• 2:12
  infrastructure, not product.
• 2:14
  You go on to say there is a quiet revolution unfolding, not in silicon, but
• 2:17
  in sovereignty.
• 2:19
  And this is not so much about national sovereignty, but more about consumer
• 2:22
  sovereignty.
• 2:24
  The future of AI is being decided not by which company trains the biggest model.
• 2:27
  but by who owns and above all, who uses the recipe the model provides.
• 2:33
  I'm sort of understanding now, Michael, but then after that, I get really fogged.
• 2:39
  And you admitted on LinkedIn that you were also a bit fogged with AI before you wrote
• 2:44
  this.
• 2:45
  I was intrigued and intimidated.
• 2:47
  I'm not sure I was specifically fogged.
• 2:50
  And in the time since I've written that article, I've become far more
• 2:56
  understanding of exactly what's going on.
• 2:58
  So if there's any...
• 3:00
  Doubt left in your mind or lack of clarity, I'm happy to banish it.
• 3:03
  OK, please banish by not imagining,
• 3:05
  but by actually seeing the reality that on the end of the line is somebody who didn't
• 3:12
  understand your article when you come to
• 3:14
  SLMs and all that sort of thing.
• 3:16
  But please go on.
• 3:18
  What did you discover when you delved into the murky world, murky to me anyway, the
• 3:21
  murky world of AI?
• 3:23
  What matters now is not the chip.
• 3:26
  It's the ecosystem.
• 3:28
  And the ecosystem is made up by hardware and software.
• 3:33
  And of course, in the case of China,
• 3:35
  they were restricted in their access to particularly NVIDIA hardware.
• 3:41
  So they've essentially compensated by developing their software to an
• 3:47
  extraordinary degree.
• 3:49
  So today in China,
• 3:51
  the chips are built around the software and not vice versa.
• 3:58
  The software is doing the heavy lifting.
• 4:01
  I'm not trying to say that it's just software alone, it's an ecosystem.
• 4:06
  It requires both hardware and software.
• 4:09
  But those people who just look at the power of the NVIDIA chip, for instance,
• 4:14
  it's a bit like looking under the hood of a car at the horsepower of an engine
• 4:21
  without understanding
• 4:23
  the totality of how the car is put together and how it will perform on the
• 4:26
  road.
• 4:26
  So I think what's happening at the moment is people have been blinded by teraflops,
• 4:31
  which is the key measure of power, and don't understand the wider significance.
• 4:37
  And there are two statistics that have come out in the last week which make me
• 4:44
  really take notice of what's going on.
• 4:46
  The first came from the venture capital backer, Andreessen Horowitz,
• 4:50
  which is probably the most powerful venture capital backer in the world.
• 4:55
  And they say now that...
• 4:57
  80%
• 4:58
  of the presentations done to them by startups are done on
• 5:05
  Chinese software.
• 5:07
  That is a devastating indictment of the big companies in the United States.
• 5:13
  The second thing is that Qen, which is the model or family of models put out by
• 5:19
  Alibaba,
• 5:21
  recently released an update.
• 5:22
  And that update now is...
• 5:25
  top of the hugging face leaderboard,
• 5:27
  which is the way in which you basically decide how well the latest large language
• 5:33
  models
• 5:35
  are being accepted.
• 5:37
  And I should underline the point here is that the Chinese family of models,
• 5:42
  the likes of DeepSeek as well,
• 5:44
  are generally speaking outperforming most of the Mistrals and the ChatGPTs and the
• 5:51
  other models that are coming out of the West.
• 5:53
  Well, this is terribly interesting and obviously has potentially a huge impact on
• 6:00
  what
• 6:00
  we've come to know as AI.
• 6:02
  I mean, we screen and we see NVIDIA at around 175, 180, whatever it is,
• 6:06
  because of their chips that are terribly powerful and cost an absolute fortune.
• 6:12
  Are you saying that these chips might not be necessary in their current form in the
• 6:19
  future because it can
• 6:21
  be done better elsewhere and in a different format?
• 6:23
  Yes, and because of the way in which particularly the American fraternity of
• 6:30
  software has grown up,
• 6:32
  they have used what's closed weight.
• 6:32
  They've used closed weight large language models, which is what the NVIDIA is being
• 6:38
  applied to,
• 6:40
  and not open weight or not by and large open weight.
• 6:41
  There are a couple of exceptions, but the vast majority of them are closed weight.
• 6:45
  Now, this is critically important because it basically means that you end up with
• 6:51
  Large language models that run these chips that are not very flexible, whereas in
• 6:58
  China,
• 6:58
  the large language models, which are open source, can be reconfigured by the user,
• 7:04
  and they're much more flexible.
• 7:06
  Furthermore, the Chinese have introduced, it's been echoed in the West, but they
• 7:13
  haven't done it successfully yet,
• 7:15
  something called a mixture of experts.
• 7:17
  which basically allows you and I, sitting wherever we are in the world, to use one
• 7:23
  of these models,
• 7:25
  and we send in data by using it, which then gets fed into these mixture of
• 7:30
  experts,
• 7:32
  who then upgrade their knowledge as a result of that input,
• 7:36
  and then send it out in a refined form immediately thereafter to the rest of the
• 7:40
  world.
• 7:41
  This is just extraordinary.
• 7:42
  It's a sort of a circle that goes round and round between consumer...
• 7:47
  And the models, which is why I say this is leading to the paramountcy of the
• 7:53
  consumer.
• 7:55
  You say, and from a market point of view, and this is a very, my podcasts tend to
• 8:00
  lean towards market analysis.
• 8:02
  You say Wall Street still bets on the incumbent triumvirate closed weight LLMs,
• 8:05
  NVIDIA chips and hyperscale clouds with the tens and hundreds of billions that are
• 8:11
  being invested
• 8:13
  in these closed weight systems of which you speak.
• 8:16
  is that all going to come to naught?
• 8:19
  Is it going to be all a giant waste of money?
• 8:21
  If I was the man from Mars coming down at this with a fresh perspective,
• 8:26
  the thing about the Chinese models is while they do need data centers in a very
• 8:31
  select way,
• 8:33
  they don't need them anything like as much as they do in the United States,
• 8:37
  because they have something called distributed
• 8:41
  AI, which means that sitting in my phone,
• 8:45
  There is the capacity to do small amounts of computing that is part and parcel of
• 8:51
  their wider, dare I say it,
• 8:53
  data center.
• 8:55
  That means that when I do something or say something or work on my AI on my phone,
• 9:00
  I am essentially taking the weight of the need to have data centers and cloud,
• 9:07
  and thereby massively reducing the cost.
• 9:10
  Not altogether, they still need data centers, but nothing like on the scale.
• 9:15
  that the system of America, the American system, which is, as I say,
• 9:19
  using closed-weight software with the likes of the very powerful chips from
• 9:24
  NVIDIA, but nevertheless,
• 9:26
  it's not something that is distributed.
• 9:28
  The Chinese model is essentially increasingly distributed.
• 9:32
  It's all over the world.
• 9:33
  It's embedded in a factory, not just in China, but anywhere where there's a degree
• 9:39
  of automation.
• 9:41
  You can have that software embedded in your factory.
• 9:43
  And your factory increasingly will be able to essentially detect a problem,
• 9:49
  send that problem back to the center that immediately will be looked at and a
• 9:56
  solution is immediately dispatched.
• 9:58
  And it's happening in, you know.
• 10:00
  pretty close to nanoseconds.
• 10:02
  And that's something about this is that it's a self-upgrading system to a far
• 10:07
  greater degree than the
• 10:09
  closed-weight systems of the United States, which is why it is so powerful,
• 10:13
  which is why I would say that, you know,
• 10:15
  in Hugging Face now, it's Chinese large-language models like QAN and
• 10:20
  DeepSeek that are winning the race.
• 10:22
  So what you're saying is it's all really about adaptability and being nimble,
• 10:25
  essentially.
• 10:27
  Absolutely.
• 10:29
  Yeah.
• 10:31
  Absolutely.
• 10:33
  This is Darwin in motion.
• 10:35
  This is, you know, the survival of the fittest by not so who's the strongest or
• 10:35
  has the biggest chip,
• 10:37
  but the one who's most adaptable.
• 10:38
  Yes.
• 10:39
  And of course, adaptability comes down to some of the examples that you give.
• 10:43
  You say, what if the next trillion dollar AI company isn't built in Silicon Valley?
• 10:49
  You put that intriguing question to us and you say, what if it's built in Hyderabad?
• 10:53
  This is one of your examples.
• 10:55
  where a startup fine-tunes a 7 billion parameter model on local dialects serving
• 11:02
  millions who speak
• 11:04
  Tamil, Telugu or Bengali.
• 11:05
  So, again, that's the adaptability and ability to be nimble that I spoke about.
• 11:11
  And this could change everything, Michael.
• 11:14
  This could change everything.
• 11:15
  And I think there are one or two people on whom this whole idea is beginning to dawn.
• 11:21
  The narrative...
• 11:22
  in the United States in particular is so strong because it's so caught up,
• 11:27
  particularly in the tech boom in the stock market,
• 11:31
  that you're not really allowed to say anything against it.
• 11:35
  But the reality is that if 80% of the startups presenting to Andreessen's whole
• 11:41
  widths are using
• 11:43
  Chinese large language models, that's a pretty stunning indictment.
• 11:48
  This is the most powerful venture capital company in the world.
• 11:51
  And all the new prospects they're looking at, I say all 80% of them, are using
• 11:58
  Chinese software.
• 12:00
  Tell us about the power implications, because America is not running out of
• 12:05
  power, but it certainly needs to develop more for the future.
• 12:07
  What about China?
• 12:09
  China's way ahead.
• 12:11
  Massively ahead.
• 12:13
  And of course, the two systems that we're talking about, the two ecosystems here,
• 12:14
  the Chinese one is far less power hungry, because so much of the power...
• 12:19
  usage is distributed, whereas the U.S.
• 12:22
  system is much more power hungry because it's concentrated, particularly with
• 12:27
  things like the data centers.
• 12:29
  So, yes, I mean, China has been running away with particularly renewable energy,
• 12:35
  as my previous article showed,
• 12:37
  and that has built in absolutely no shortage whatsoever when it comes to the
• 12:43
  availability
• 12:45
  of power for China's AI ecosystem.
• 12:46
  Quite the opposite in the United States, where...
• 12:48
  You know, electricity prices are rising dramatically.
• 12:52
  They're up 10% this year, and there are shortages breaking out everywhere.
• 12:56
  And it's potentially, electricity prices going to become quite a hefty political
• 13:03
  issue in the coming
• 13:05
  elections.
• 13:07
  So I think that what we've got here is that the United States has neglected
• 13:09
  power, which in a strange sort of way,
• 13:11
  they tend to need more given their particular model than the Chinese.
• 13:14
  Whereas the Chinese have gone hell for leather and produced Lots of wind farms
• 13:21
  and solar panel farms.
• 13:23
  And also they're now putting in massive hydro projects.
• 13:24
  And that means they've got all the power they need and more.
• 13:27
  So it's, again, the man from Mars looking at this situation from a fresh perspective
• 13:34
  would
• 13:36
  start scratching his head.
• 13:38
  The Aladdin's cave of AI isn't locked, you say.
• 13:40
  It never was.
• 13:42
  The world just thought it was.
• 13:44
  I would say it's the big titans made us believe it was the titans in America.
• 13:45
  Well, that's probably right.
• 13:47
  I mean, that's a better refinement to that phrase, but it doesn't run off the tongue
• 13:52
  quite so quickly.
• 13:54
  But yes, absolutely.
• 13:56
  I think we have been, dare I say it, conned.
• 13:57
  But the reality is, as I say, and perversely, because of the hardware
• 14:03
  restrictions that were placed on China,
• 14:05
  they took another route.
• 14:07
  And that route was software heavy.
• 14:09
  Now, there has been a precedent in all of this, of course, and that was Linux.
• 14:12
  And beyond that, Android, which is built on Linux.
• 14:16
  And it was Linux that really upset IBM at the end of the 1990s. Obviously,
• 14:23
  IBM has reconfigured its offering and is now very heavy into data centers and cloud
• 14:28
  and the like,
• 14:30
  out of the frying pan into the fire, I think so.
• 14:32
  But the reality is that 100% of all supercomputers in the world today are not
• 14:39
  run on Microsoft, they're run on Linux.
• 14:41
  And our cell phones are much more geared towards...
• 14:42
  Android, which as I say is a Linux spinoff, and like the Chinese software
• 14:48
  packages,
• 14:50
  it's open source.
• 14:52
  And the result is that you can have these precedents, the Android and before that
• 14:58
  the Linux precedent,
• 15:00
  that give you some indication of where things might be going.
• 15:02
  Who knows about this?
• 15:03
  If I know about it, and you, you're a clever chap, and you created this article.
• 15:09
  I mean, has it caused a bit of a stir is what I'm trying to say.
• 15:12
  Has Silicon Valley contacted you?
• 15:15
  I can't really, out of respect for the people that I'm in contact with,
• 15:19
  I'm going to have to be anonymous about the sorts of people that I've had feedback
• 15:22
  from.
• 15:24
  Suffice it to say, I've been invited to the Bahamas next February to speak on this
• 15:29
  subject at a Brains Trust.
• 15:31
  and suffice it to say, I'm actually in.
• 15:31
  in touch with one of the chief engineers of the big boys in Silicon Valley.
• 15:35
  And he's just scratching his head and not able really to answer what I'm saying.
• 15:42
  He eventually dismissed it very cryptically as saying, it's a fantastic
• 15:46
  vision.
• 15:48
  It just cuts across the typical left brain approach that we in Silicon Valley have
• 15:53
  towards where we
• 15:55
  stand.
• 15:57
  Being a right brain person myself, that probably ties up.
• 16:00
  So you're going to be an AI consultant to the big boys, Michael, from now on,
• 16:03
  traveling around to exotic places and imparting your wisdom.
• 16:08
  No, I doubt it.
• 16:09
  I just will be a commentator.
• 16:10
  But I will say this, that I think there is a reckoning in the not too distant future
• 16:17
  that's based on the
• 16:19
  idea of recognizing that the Chinese ecosystems, and I cannot repeat that word
• 16:24
  more and more,
• 16:26
  more often, ecosystems are winning and they're beating.
• 16:29
  I mean, as I say, the horsepower of NVIDIA with cars that are much better suited to
• 16:35
  the roads that
• 16:35
  they're forced to drive on.
• 16:36
  What are you tackling next, Michael?
• 16:39
  Look, I'm actually going to, and you'll love this, address what does all of this
• 16:46
  mean for the stock market?
• 16:48
  Because I think that that's the unfinished question in this.
• 16:50
  And by the way, if you want to sort of amuse yourself or perhaps shock yourself,
• 16:56
  run up.
• 16:58
  Up two charts comparing the performance in September of NVIDIA versus Alibaba.
• 17:02
  OK, I'm doing that.
• 17:04
  I'm writing a note immediately.
• 17:06
  And I suppose if I short sell ahead of any talk you give, it's not insider trading,
• 17:12
  is it?
• 17:14
  It's good it's out there.
• 17:16
  Not at all.
• 17:18
  Not at all.
• 17:20
  But seriously, I mean, these things have happened before.
• 17:22
  Suddenly everybody thought that nothing could be, no monopoly could be penetrated.
• 17:23
  And suddenly a few cracks appeared and that was it.
• 17:26
  It was gone.
• 17:27
  And there's one thing I haven't mentioned in all of this, which is the real killer,
• 17:31
  I suppose.
• 17:33
  In the end, it is the killer.
• 17:35
  It's that Chinese software, LLMs, large language models, are free to the user.
• 17:41
  There's no cost to the user.
• 17:45
  Okay.
• 17:46
  And this will, of course, evolve by the time you get to the Caribbean in February.
• 17:51
  it will be a completely different story that you'll be telling based on this one,
• 17:56
  of course.
• 17:56
  But Michael, it's fascinating stuff.
• 17:58
  Thanks so much for your time.
• 17:59
  Even I am starting to understand this now.
• 18:01
  Michael Power is a contributor to Daily Maverick, formerly investment strategist
• 18:07
  at 91 in Cape Town.
• 18:09
  The views and opinions expressed in these podcasts are those of Lindsay Williams and
• 18:14
  various contributors and do not reflect the policy,
• 18:16
  position or opinion of any other agency, organisation, employer or company
• 18:21
  associated with the use of the
• 18:23
  with StrictlyBusinessPodcast.com.
• 18:24
  Assumptions made on the analyses are not reflective of the position of any other
• 18:31
  entity other than the speaker or the author.
• 18:33
  And since we are critically thinking human beings, these views are always subject to
• 18:37
  change, revision,
• 18:39
  and rethinking at any time.
• 18:39
  Please do not hold us to them in perpetuity.