Latest / In The Loop / Is The AI Bubble About To Burst?

Transcript

Is The AI Bubble About To Burst? · In The Loop

• 0:00
  This week Sam Altman, perhaps AIS biggest influencer, just sat
• 0:04
  down with a journalist and answered a really interesting
• 0:07
  question whether or not we're in a bubble.
• 0:10
  And he said, and I quote my opinion is yes, in the same week
• 0:15
  that Chachi BT had a mass rebellion over users desperately
• 0:20
  wanting the drug that is Chachi BT4-O and other loved models and
• 0:25
  Sam Altman announcing billions of dollars in spending on data
• 0:28
  centres and infrastructure over the next 10 years.
• 0:30
  Today we're going to examine this contradiction and try and
• 0:33
  answer the question of whether we think AI is really in a
• 0:35
  bubble. This is in the leap with Jack
• 0:37
  Horton. I hope you enjoy the show.
• 0:51
  Let's start right at the top with what he actually said now
• 0:55
  with the journalist and The Verge.
• 0:57
  He actually said he thinks we're in a bubble.
• 0:59
  And what he said was actually quite interesting and quite
• 1:01
  clever because when he was asked are we in a phase where
• 1:05
  investors as a whole are over excited about AI?
• 1:08
  And he said his opinion is yes. And he called it insane that AI
• 1:12
  startups would say 3 people and an idea are getting massive
• 1:16
  valuations. And he warns that someone's
• 1:19
  going to get burned here and that person who loses out are
• 1:24
  going to lose in a phenomenal amount of money.
• 1:26
  Now Almond's point didn't come out of nowhere.
• 1:29
  It was just obviously days after church.
• 1:31
  GBTS work probably could be only described as a disappointing
• 1:35
  launch, or at least the reception of that launch has
• 1:37
  been disappointing. And actually you could argue
• 1:40
  that it's just damage control after maybe that big flagship
• 1:43
  product of the year didn't really go down as well as maybe
• 1:46
  they hoped. Now, his point in particular was
• 1:48
  that when bubbles happen, smart people start to get overexcited
• 1:53
  about, let's say, a kernel of truth, as he puts it.
• 1:56
  And if you look at most bubbles in history, most of the time,
• 2:00
  like the tap bubble, there was a it was a real thing.
• 2:02
  And that's that kernel of truth. We know this thing is going to
• 2:05
  reshape the world. And he thinks AI might be doing
• 2:08
  the same to people. Everyone knows it's going to
• 2:10
  reshape the world. And that's that kernel of truth
• 2:12
  that people obsess over. And therefore that's so big, but
• 2:15
  that's so badly. But by comparing it to the.com
• 2:19
  bubble, actually what he's doing is implicitly arguing that for
• 2:23
  all of the failed companies and the companies that I guess
• 2:26
  popped during that bubble, there are the Amazons.
• 2:30
  And those Amazons are the ones that will take the lion's share
• 2:33
  of the profits. And obviously he's implying that
• 2:36
  Open AI is going to be one of those Amazon companies.
• 2:39
  Now, I think a lot of this narrative, and it's been a real
• 2:41
  narrative that's emerged online in the last few days can be
• 2:45
  boiled down to this, I guess, summary and tweet by Antonio
• 2:50
  Garcia and Nick Bauman. And the reason it's such a good
• 2:54
  explanation of, I guess this narrative around the bubble is
• 2:57
  because of the historical analysis it makes.
• 3:00
  What basically they say is that during the Web 2 revolution, so
• 3:03
  when the Internet first emerged and you have the Amazons,
• 3:07
  essentially what they say is that venture capital money for
• 3:10
  very wealthy individuals were basically going towards paid ads
• 3:14
  to pump the number of monthly active users on the platforms.
• 3:18
  And then crypto essentially use a very similar mechanism, but
• 3:22
  raising money from retail investors and people who are
• 3:25
  desperate to make some money. And what they were doing was
• 3:28
  spending money on user rewards to then try and get more people
• 3:31
  using the platform, again, inflating usage with artificial
• 3:36
  access to cash and AI. What these people are claiming
• 3:39
  is actually using venture capital money like Web 2 did.
• 3:43
  And what they're arguing is that in the AI industry right now,
• 3:47
  what's happening is this venture capital money being used again
• 3:51
  to subsidize compute costs. So costs to pay API providers
• 3:55
  such as open AI or data centers and infrastructure, which as a
• 3:59
  result inflates the number of people using it for the amount
• 4:03
  they're paying. And what they say is that
• 4:05
  eventually this is just going to blow up at some point because
• 4:08
  the venture capitalists, whether they stop putting money in,
• 4:10
  which means that nobody can afford to get usage, and because
• 4:13
  they never delivered real value, the entire industry just kind of
• 4:17
  collapses like a souffle. So in other words, every
• 4:19
  technology boom basically is being driven by external
• 4:23
  capital, by investors that subsidizes much of the costs,
• 4:26
  which helps create artificial demand that then evaporates when
• 4:30
  eventually the money runs out. And he thinks that following
• 4:33
  this playbook, that AI is just therefore in a bubble, and
• 4:36
  eventually it's going to follow the exact same path.
• 4:38
  Now, if you've been following this podcast, you might guess
• 4:41
  that I don't agree with the analogies and the comparisons
• 4:44
  they're making. But I guess before I give my
• 4:46
  personal opinion, I'll go through some of the arguments
• 4:48
  and the data supporting this opinion.
• 5:01
  So TechCrunch did an investigation into the cost of
• 5:04
  these, for example, AI coding assistance is just one example,
• 5:08
  and the cost of compute, so paying for API services and all
• 5:12
  the infrastructure that sits behind everything.
• 5:14
  And so they did an investigation and they actually found that
• 5:17
  margins on many of these products are either neutral or
• 5:20
  negative. So that means they're losing
• 5:22
  money on almost every single user interaction.
• 5:24
  And if the user always wants the best model, because why would
• 5:27
  you not, then each competing company is forced to always have
• 5:31
  the best model or the newest thing, which just cost more and
• 5:34
  more money. And when the gross margins are
• 5:36
  sometimes negative, then that screams like a big problem.
• 5:40
  And The Information reported that Replits are another big AI
• 5:43
  coding platform. Its revenue and gross margins
• 5:46
  fluctuated between or its gross margins, shall I say, fluctuated
• 5:49
  between 36% and -14% which really showcases that pain and
• 5:55
  relying on 3rd party AI models. And for context, as a SAS
• 5:59
  companies are technology providers selling software as a
• 6:02
  service for a subscription or a license fee.
• 6:04
  You're looking for eighty 8590% gross margins.
• 6:08
  So this data makes it I guess easy to maybe assume or think
• 6:11
  that we must be in a bubble because clearly venture capital
• 6:15
  money is just subsidizing these big companies and they're all
• 6:18
  going to go bust eventually. Well, personally, I don't buy
• 6:22
  that. There are many, many reasons why
• 6:34
  and I'm going to go through some of them.
• 6:35
  And one of the first ones is that this data, what it didn't
• 6:38
  reveal is actually the speed in which these companies are trying
• 6:41
  to solve these problems. Because naturally clever
• 6:44
  successful companies just adapt. They make something work.
• 6:48
  They've got a very good incentive to fix this problem.
• 6:51
  A great example of this is Replit.
• 6:53
  So one of the companies that I talked about just a minute ago,
• 6:56
  they actually replaced their previous pricing model to a
• 7:00
  usage based pricing model and that increased the gross margins
• 7:03
  to 23%. So from -14 to 23%.
• 7:07
  Again, still very small numbers, but big and clever companies
• 7:11
  have the incentive to figure it out.
• 7:13
  Another reason I don't buy this historical analysis is that
• 7:17
  there's just this huge infrastructure build out right
• 7:19
  now, which is extremely long term.
• 7:21
  Because the historical comparison to say, Web 2
• 7:24
  suggests that these external subsidies of venture capital
• 7:28
  money, for example, props up completely unsustainable usage,
• 7:32
  which completely vanishes when funding stops coming.
• 7:35
  But I don't really see that. For example, infrastructure
• 7:38
  investments look nothing like what you see in a big bubble
• 7:42
  because the scale is just unprecedented and it's long term
• 7:45
  focused. A great example of this is
• 7:47
  Equinox who has just secured a GW of nuclear energy.
• 7:51
  These are 20 year power purchase agreements for factories that
• 7:56
  are still in development. To me, this just shows that AI
• 7:58
  is reaching such an industrious scale it's now being treated
• 8:02
  like aluminium or steel mills in power planning.
• 8:06
  And Bloomberg just announced that U.S. data center
• 8:08
  construction is actually on pace to now take over office building
• 8:13
  in America. So that is now more data centers
• 8:16
  being built out than there are general office buildings being
• 8:20
  built. In a world still adjusting to
• 8:21
  remote work, this slack in office construction is now being
• 8:25
  filled by server farms to basically just fuel AI.
• 8:29
  Another really big thing to me is that adoption is through the
• 8:33
  roof. So when adoption metrics are
• 8:36
  disconnected from real value creation.
• 8:38
  So Web 2's monthly active users or cryptos total value on the
• 8:42
  chain, then I'd worry about a bubble.
• 8:45
  But AI presents a different dynamic because each interaction
• 8:48
  with an AI system is representing usually some form
• 8:51
  of work being performed. So usage data just as of this
• 8:55
  month. So in August 2025.
• 8:57
  It's staggering in both scale and what it shows us about
• 9:00
  value. We're literally a point where
• 9:03
  ChatGPT weekly users have hit 700 million, which is up from
• 9:07
  400 million in just February, and ChatGPT is on track to have
• 9:11
  more conversations per day than all human beings combined.
• 9:15
  And even more telling is token processing, because Google
• 9:18
  announced that it's now processing 980 trillion tokens
• 9:22
  monthly, which is double the level of just two months
• 9:25
  earlier, which represents 104% growth in usage just in two
• 9:29
  months alone. Unlike previous bubbles, each
• 9:31
  token represents a specific number of words and is actually
• 9:36
  a request for AI to perform actual work or something, often
• 9:40
  quite helpful, whether that's writing code or analyzing data
• 9:43
  or drafting documents. And the ChatGPT 5 backlash, to
• 9:47
  me, validates real usage and value.
• 9:49
  Users weren't suddenly abandoning ChatGPT, and AI tools
• 9:53
  because they weren't happy with GBT 5.
• 9:55
  They were actually demanding online in a movement that they
• 9:59
  bring back the old models that people used to live.
• 10:01
  There are literal Reddit posts online about how people feel
• 10:04
  like they've just lost a friend because the GPT 4-O model has
• 10:08
  been killed. This is the behavior of people
• 10:10
  who have integrated AI so deeply into their lives that they
• 10:13
  couldn't imagine not having this thing in their life anymore.
• 10:16
  So this bubble thesis also assumes that the subsidized
• 10:20
  costs remain super, super high, making there for a eventual
• 10:24
  pricing correction almost inevitable.
• 10:27
  But I don't agree with that either because what we're seeing
• 10:30
  is pricing improvements happening faster than ever.
• 10:33
  So for example, in the past year open air has GPT 4 with an 8000
• 10:38
  context window had cost $30 per million input tokens and then
• 10:42
  $60.00 per million output tokens.
• 10:45
  Whereas GPT 4 turbo just a year later costs 50% less and then
• 10:50
  67% less than that. And ChatGPT 5 is coming in even
• 10:55
  cheaper than that. So this pricing model validates
• 10:58
  underlying demand. When Replit shifted to compute
• 11:01
  time billing so usage based Millet bidding margins
• 11:04
  immediately increased. When cursor added usage charges
• 11:08
  to their product, customers started paying more money rather
• 11:11
  than churning. I'm not saying that this startup
• 11:13
  world is going to be easy for many of these companies in the
• 11:16
  space, but there's evidence of real value creation, not
• 11:21
  artificial demand propped up by extremely low prices.
• 11:24
  And finally, to me, this ignores a very important discussion
• 11:26
  because many of these metrics are talking about consumer
• 11:29
  dynamics, whereas most sustainable businesses are going
• 11:32
  to be built on enterprise integration.
• 11:34
  So this survey found the integration complexity was the
• 11:36
  number one barrier to enterprise adoption, ranking even above
• 11:40
  cost. It's expensive, it's time
• 11:42
  consuming, it's really specific to individual businesses, but
• 11:46
  it's where the recurring revenue gets built.
• 11:59
  So let's come back to where I think this could be heading.
• 12:02
  So scenario 1 is that we have, I guess, small pop of the bubble
• 12:06
  within a very large build out of infrastructure.
• 12:10
  So we might see some correction of the markets over this next
• 12:13
  period. So maybe there's going to be a
• 12:15
  drop in the value people are placing on equity by maybe 30 to
• 12:20
  50%. So you're not going to get these
• 12:22
  hugely inflated valuations really quickly, which may be
• 12:28
  results in many start-ups being bought by other companies and
• 12:32
  then resulting in less venture funding going into the space
• 12:35
  potentially. But infrastructure spending will
• 12:36
  just continue because major tech companies see this as absolutely
• 12:41
  essential no matter what the short term sentiment is.
• 12:46
  And to track this, it'd be quite simple.
• 12:47
  We track GPU utilization rates, maybe data center construction,
• 12:51
  build outs over the world, despite the fact that stock
• 12:55
  markets might go up and down. So let's go to scenario 2 of
• 12:57
  what could happen, which is I guess a very soft landing and
• 13:01
  massive scale up. So eventually markets slowly
• 13:04
  adjust what they expect to see. Revenue growth from many of
• 13:07
  these companies starts to catch up from the amount they've spent
• 13:12
  and gross margins for all these companies again start to go up,
• 13:16
  not down. In this scenario, enterprise
• 13:19
  adoption will just accelerate as integration challenges get
• 13:22
  solved. Both scenarios assume again that
• 13:25
  infrastructure build out continues.
• 13:27
  And I guess the question really to me is just how bumpy the
• 13:30
  financial ride becomes. So I guess to conclude, this
• 13:33
  discussion definitely captures something real about the current
• 13:37
  AI market because we are seeing subsidized usage, we're seeing
• 13:41
  massive valuations and really unsustainable unit economics and
• 13:47
  gross margins for many of these companies.
• 13:50
  So the application layer absolutely shows bubble
• 13:53
  characteristics. But at the same time, there's a
• 13:56
  lot of evidence saying the absolute opposite, the immediate
• 13:59
  utility that's driving massive adoption, the infrastructure
• 14:02
  commitment spanning decades, the rapidly reducing costs of all
• 14:06
  these incredible services. And most importantly for the
• 14:09
  success of these products, the deep psychological integration
• 14:13
  that ChatGPT really showcased. And this, to me, points to
• 14:17
  something that's very durable, more durable than a typical
• 14:20
  bubble. So the next six months are going
• 14:22
  to test these narratives. And I might turn out to be
• 14:26
  completely wrong, but if infrastructure spending
• 14:29
  maintains its current trajectory despite these corrections that
• 14:33
  we'll probably experience, and if enterprise adoption continues
• 14:37
  to grow, and if users keep demonstrating a love and in some
• 14:43
  very bad cases, a dependency on these AI tools, we'll know that
• 14:47
  we're witnessing a build out disguised as a bubble.
• 14:51
  Anyway, I hope you enjoyed today's episode and I'll see you
• 14:54
  next week.