Latest / Elon Musk Podcast / ChatGPT hits $100M in ad revenue

Transcript

ChatGPT hits $100M in ad revenue · Elon Musk Podcast

• 0:00
  Open AI has hit $100 million in annualized ad revenue with their
• 0:05
  ChatGPT pilot. Yeah.
• 0:07
  And I mean, the velocity of that number is genuinely difficult to
• 0:10
  process, especially when you add the context that this enormous
• 0:15
  revenue figure is being generated by only a really,
• 0:18
  really tiny fraction of their total user base, right?
• 0:21
  Which makes the picture incredibly intense.
• 0:23
  Exactly. And they're already actively
• 0:25
  working with over 600 advertisers.
• 0:28
  We are looking directly at how a platform built entirely to
• 0:31
  answer questions balances serving the person asking the
• 0:34
  prompt with, you know, the demands of the advertiser paying
• 0:37
  for space on the screen. So when the tool you rely on to
• 0:40
  research and make decisions gets a second customer who pays for
• 0:43
  placement, how does that alter the relationship you have with
• 0:47
  the answers it gives you? To understand the shifting
• 0:49
  dynamics of that relationship, we really have to look closely
• 0:52
  at the exact mechanics of this pilot program.
• 0:54
  Yeah, and what is technically happening behind the screen?
• 0:56
  Right, because the facts of how this pilot operates are highly
• 1:00
  specific. The ads themselves are placed at
• 1:02
  the very bottom of the answers provided by the artificial
• 1:05
  intelligence. Yes, and they're clearly labeled
• 1:08
  so you know an advertisement is actually present.
• 1:10
  And the company insists that these ad placements do not
• 1:13
  influence the actual is generated by the model.
• 1:16
  Right. And right now, this specific
• 1:19
  pilot program only targets their free users and the ChatGPT Go
• 1:23
  subscribers. It definitely helps to
• 1:25
  understand exactly how they keep the generated answer in the
• 1:27
  advertisement separated. It does so when you type a query
• 1:31
  into the interface, the underlying language model starts
• 1:34
  calculating the most probable next words to answer your
• 1:37
  prompt. It streams that text to your
• 1:40
  screen token by token, and that entire process happens in one
• 1:45
  specific part of their system. OK, completely independent of
• 1:47
  that, the visual interface you are looking at pings a separate
• 1:51
  ad server and requests A relevant banner based on the
• 1:54
  broad topic of your conversation.
• 1:55
  Oh, I see. The advertisement is simply
• 1:58
  pasted at the bottom of your screen by the user interface.
• 2:00
  So the neural network itself never saw the ad.
• 2:03
  Exactly. It did not factor the sponsor
• 2:06
  into its calculations at all. The separation is entirely at
• 2:09
  the visual layer, rather than, you know, the cognitive layer of
• 2:12
  the artificial intelligence. And the eligibility breakdown
• 2:15
  for those user tiers reveals exactly how cautious they are
• 2:19
  being with the rollout. Yeah, they are being very
• 2:21
  careful. Roughly 85% of those free and Go
• 2:25
  users in the US are eligible to see the ads.
• 2:28
  Because they fit the precise demographic and usage criteria
• 2:32
  for the pilot. Right, but fewer than 20% are
• 2:34
  actually shown an ad on a daily basis.
• 2:36
  Let me back up. Yeah.
• 2:38
  Are you saying that almost everyone in those specific tiers
• 2:40
  is sitting in the pool of potential AD recipients?
• 2:44
  Yeah, but only very small sliver of them actually encounters the
• 2:47
  ads on any given day. Exactly.
• 2:49
  Almost the entire US free and go user base is in the hopper, but
• 2:53
  the daily delivery is kept highly restricted.
• 2:56
  That restriction on delivery is a classic method for testing the
• 2:59
  waters without alienating the user base.
• 3:02
  They are basically running a highly controlled experiment to
• 3:05
  see if the presence of ads causes users to close the app or
• 3:08
  ask fewer questions or switch to a competitor.
• 3:11
  And generating a massive annualized figure like $100
• 3:15
  million from such a limited daily exposure rate creates a
• 3:19
  highly realistic path to their stated goal of generating at
• 3:24
  least $17 billion from consumers using the platform.
• 3:28
  Oh. Absolutely.
• 3:29
  When you realize the revenue potential sitting inside just a
• 3:32
  20% daily exposure rate, pulling in $100 million annualized from
• 3:37
  a fraction of a fraction of your traffic, that $17 billion target
• 3:42
  suddenly stops looking like a distant fantasy.
• 3:44
  The math really crystallizes when you look at an observation
• 3:47
  from commentator Peter Gosteff. Oh yeah, his point was
• 3:49
  fascinating. He.
• 3:50
  Pointed out that you need a log scale to even look at the
• 3:53
  numbers sensibly. Right, because the difference in
• 3:55
  size is just too extreme for a normal chart.
• 3:57
  Exactly. Think about trying to graph the
• 3:59
  height of a blade of grass next to the height of Mount Everest
• 4:02
  on a standard piece of paper. The grass is completely
• 4:05
  invisible. Yeah, so a log scale changes the
• 4:08
  rules of the graph so that every step up the side multiplies the
• 4:11
  value by 10. Allowing you to fit extremely
• 4:14
  disparate numbers onto the same page.
• 4:16
  Right. You only use a log scale when
• 4:18
  something is growing so aggressively that a normal chart
• 4:21
  physically breaks. And this pilot program has
• 4:24
  reached a 6th of the New York Times revenue in almost no time
• 4:28
  at all. Think about that.
• 4:29
  You have an established media entity with a century of
• 4:32
  history, and a tiny pilot program spun up by a technology
• 4:36
  company is already pulling in a sixth of their revenue.
• 4:39
  Generating that much cash entirely through a pilot program
• 4:42
  that most users rarely even see completely changes how we
• 4:46
  calculate the financial ceiling of these platforms.
• 4:48
  Shifting focus directly to the perspective of the users and the
• 4:51
  builders relying on this technology technology, the
• 4:54
  gravity of that ad revenue becomes impossible to ignore.
• 4:57
  It really does, observer Leonardo Jacques prop.
• 5:00
  A really interesting point regarding this exact dynamic.
• 5:03
  Right about the tool people use for drafting and product
• 5:06
  decisions. Exactly.
• 5:07
  That tool now has a paying customer sitting right there at
• 5:11
  the exact moment a decision is being made.
• 5:14
  You are typing a highly specific query about a marketing strategy
• 5:18
  or a software choice for your business.
• 5:21
  And an advertiser has purchased the right to be visually present
• 5:25
  at the exact second you hit enter.
• 5:27
  It fundamentally changes the psychology of the interaction.
• 5:30
  It really does. Imagine you have a highly
• 5:32
  trusted librarian for years. You ask her for book
• 5:36
  recommendations on gardening and she consistently hands you the
• 5:40
  absolute best, most objective materials available.
• 5:43
  Her only goal is to help you learn about gardening, right?
• 5:46
  Then one day you ask for a recommendation.
• 5:49
  She hands you a book, and a glossy flyer for a major
• 5:52
  hardware store falls out of the pages.
• 5:54
  Wow. She assures you that she still
• 5:56
  picked the best book and the flyer has nothing to do with her
• 5:59
  recommendation. Even if she is telling the
• 6:01
  complete truth, your brain immediately starts calculating
• 6:04
  the incentives. You start wondering if the
• 6:05
  hardware store is paying her. And if that payment will
• 6:08
  eventually influence which books she keeps on the shelf it.
• 6:12
  Perfectly mirrors the historical pattern of Internet platforms.
• 6:15
  Yeah, search engines and social feeds all started neutral,
• 6:18
  chronological or organic. They built immense trust by
• 6:22
  giving you exactly what you asked for in the exact order it
• 6:26
  appeared, with 0 interference. The pure utility was the entire
• 6:30
  draw. But once the ad business gets
• 6:32
  big enough to matter, the product quietly reorganizes
• 6:36
  around it. $100 million creates its own financial gravity.
• 6:40
  The engineering resources, the user interface updates and the
• 6:43
  strategic goals of the company slowly begin to align with the
• 6:47
  needs of the people paying the largest bills.
• 6:49
  The early days of web search felt like pure magic.
• 6:52
  You typed a phrase and the algorithm objectively ranked the
• 6:56
  most relevant information on the entire Internet.
• 6:58
  Then came the lightly shaded boxes at the top of the results.
• 7:01
  Yeah, then the shopping carousels.
• 7:03
  Eventually, the actual organic results were pushed entirely
• 7:07
  below the fold and the entire top half of your screen became
• 7:10
  purchased real estate. The primary function of the tool
• 7:13
  shifted from organizing the world's information to capturing
• 7:17
  intent and selling it to the highest bidder.
• 7:19
  The AI you use is still good, but it is no longer just yours.
• 7:23
  And for anyone building products on top of ChatGPT or integrating
• 7:27
  its recommendations into a daily workflow, the ground just
• 7:31
  shifted beneath them. It limits the reality of a
• 7:33
  purely neutral artificial intelligence assistant.
• 7:36
  If your workflow relies heavily on this platform for research,
• 7:40
  drafting, or coding, the presence of a lucrative
• 7:43
  advertising model proves that the best time to think about
• 7:46
  platform dependency is always before the platform finds its
• 7:50
  primary business model. Because once the business model
• 7:53
  is locked in and generating hundreds of millions of dollars,
• 7:56
  you are subject to the gravity of their advertisers.
• 7:59
  If you built your entire company's internal tools around
• 8:01
  a clean, objective text generation API, you have to
• 8:05
  seriously consider how an ad supported consumer interface
• 8:08
  might eventually bleed into the underlying logic of the models
• 8:11
  you depend on. The industry fallout from this
• 8:13
  rapid monetization strategy has created intense friction and
• 8:16
  public rivalry. Yeah.
• 8:18
  Competitor Anthropic ridiculed this specific ad push.
• 8:22
  They went as far as making it the entire focus of their Super
• 8:25
  Bowl campaign. Which is wild?
• 8:28
  Choosing the most expensive advertising stage in the world
• 8:31
  to attack the concept of an ad supported artificial
• 8:34
  intelligence. And industry voices like John
• 8:36
  Batel view Open a Eyes revenue announcement through a really
• 8:40
  critical lens. He called the announcement a
• 8:42
  particularly thirsty leak. Right.
• 8:45
  And in the dynamic of technology companies, a thirsty leak
• 8:48
  usually happens when a company feels immense pressure to
• 8:52
  justify by its massive valuation.
• 8:54
  So in his view, releasing these specific numbers was meant to
• 8:58
  aggressively declare they are succeeding with this new
• 9:01
  initiative amidst heavy internal pressure to generate cash.
• 9:04
  The cost of running these enormous language models
• 9:07
  requires vast amounts of computing power, and he argues
• 9:10
  they have a perilous path to navigate.
• 9:12
  Navigating between achieving product market fit, maintaining
• 9:15
  simple utility, and handling the pressure to pay for the server
• 9:18
  costs. I actually look at the
• 9:19
  motivation behind releasing those numbers completely
• 9:21
  differently. Oh, really?
• 9:22
  Yeah, I think it functions instead as a highly necessary
• 9:26
  signal to marketers. OK, how so?
• 9:28
  Well, Open AI claims there has been 0 impact on privacy related
• 9:32
  trust metrics since the ads rolled out, right?
• 9:35
  That is exactly the data point those 600 advertisers need to
• 9:38
  hear to keep spending their budgets.
• 9:40
  Oh, I see. If a chief marketing officer is
• 9:43
  going to authorize putting their brand next to an artificial
• 9:46
  intelligence prompt, they need absolute assurance that users
• 9:50
  are not fleeing the platform over data privacy fears.
• 9:53
  Because advertisers are inherently risk averse, they
• 9:56
  worry that the language model might say something
• 9:58
  controversial right next to their company logo.
• 10:00
  So by publicly stating that user trust remains stable and revenue
• 10:04
  is accelerating, Open AI is building a safety blanket for
• 10:07
  the next wave of corporate ad buyers.
• 10:09
  Exactly. And the momentum of the rollout
• 10:11
  is continuing entirely regardless of the public
• 10:14
  criticism. The company's exploring
• 10:16
  additional testing in Canada, Australia and New Zealand.
• 10:19
  They are taking the technical framework established in the US
• 10:22
  pilot and pushing it into international markets.
• 10:24
  To see how different privacy laws and consumer behaviors
• 10:28
  react to the presence of sponsored content inside an
• 10:31
  artificial intelligence interface.
• 10:34
  This public rivalry really limits their ability to run this
• 10:37
  monetization quietly. Every single move they make is
• 10:41
  instantly broadcast, analyzed, and often mocked by competitors
• 10:45
  on massive stages. But it opens up a very clear
• 10:48
  dividing line in the market between ad supported models and
• 10:52
  pure subscription models. It forces you to choose exactly
• 10:55
  how you want to pay for your artificial intelligence.
• 10:58
  Either with your wallet or with your attention.
• 11:00
  So Open AI proved they can generate extraordinary ad
• 11:03
  revenue from a fraction of their users without immediately
• 11:06
  breaking consumer trust. But history shows that kind of
• 11:09
  financial gravity inevitably reshapes how a platform
• 11:12
  operates. As this advertising model
• 11:14
  expands, will an artificial intelligence train to be as
• 11:17
  helpful as possible eventually redefine the word helpful to
• 11:20
  include buying a sponsor's product?
• 11:23
  If you're not subscribed yet, take a second and hit follow on
• 11:25
  whatever app you're using. It helps us keep making this.
• 11:27
  We appreciate you being here.