AI transcript
0:00:01 Are we in an AI bubble?
0:00:04 I do not believe we’re in an AI bubble today.
0:00:05 I was, depending on how you look at it,
0:00:08 the privilege and the misfortune of being a tech investor
0:00:09 during the year 2000 bubble,
0:00:11 which was really a telecom bubble.
0:00:13 And I think it’s really helpful to compare and contrast today
0:00:14 to the year 2000.
0:00:18 The year 2000 internet bubble or telecom bubble
0:00:20 was defined by something called dark fiber.
0:00:25 At the peak, 97% of the fiber that had been laid was dark.
0:00:26 Contrast that with today.
0:00:29 There are no dark GPUs.
0:00:32 Every major technology cycle raises the same question.
0:00:34 Is it real or are we in a bubble?
0:00:37 Today, you’ll hear a conversation from runtime
0:00:40 between Gavin Baker, managing director and CIO
0:00:42 of Atreides Management and David George,
0:00:44 general partner at A16Z
0:00:47 about how AI is reshaping the global economy.
0:00:49 From capital allocation and infrastructure spending
0:00:51 to business models and margins.
0:00:53 It’s a detailed data-driven look
0:00:55 at where we actually are in the AI cycle
0:00:57 and what’s likely to happen next.
0:00:58 Let’s get into it.
0:01:03 And that brings us to our opening fireside chat.
0:01:06 We’re going to start with a taboo question
0:01:07 right out of the gate.
0:01:08 Are you ready for it?
0:01:12 If AI is the biggest trend in the world right now,
0:01:14 where is the evidence for it?
0:01:16 Why is it only just beginning to show up in the economy?
0:01:19 And as Andre Carpathie asked,
0:01:20 are agents really just ghosts?
0:01:24 To kick this off and to help us answer this question,
0:01:27 please join us in welcoming Gavin Baker,
0:01:29 managing partner and CIO of Atreides.
0:01:31 Now, some of you may know Gavin
0:01:33 as that really thoughtful guy on Twitter.
0:01:36 Anytime some big piece of AI news comes out,
0:01:39 I know more than a few people who count on Gavin
0:01:41 to explain what the F is really going on.
0:01:44 So, a huge thank you to Gavin for being with us today.
0:01:46 Joining him is our very own David George,
0:01:48 general partner at A16Z.
0:02:08 Who knows what that music was from?
0:02:11 Glad they got our pump-up music right.
0:02:12 Yes.
0:02:16 Battlestar Galactica, the original 1977 one.
0:02:19 In case we have to all fight Cylons in a few years.
0:02:22 Yeah, it could segue into the topic, I guess.
0:02:24 So, thank you for being here.
0:02:25 I always love talking to you.
0:02:26 Same.
0:02:28 Really grateful to you for inviting me.
0:02:30 Grateful to your colleagues for having me here.
0:02:32 I really look forward to the next two days.
0:02:34 I think I’m going to learn a lot, so thank you.
0:02:34 Yeah.
0:02:35 Okay, all right.
0:02:38 So, the big topic is AI bubble,
0:02:40 kind of macro view of things.
0:02:43 So, maybe just to start with a couple stats
0:02:44 to set the stage,
0:02:45 and then I want to get your take on where we’re at.
0:02:47 So, we have about a trillion dollars
0:02:49 of data centers in the U.S.
0:02:52 The plan is to add three to four trillion dollars
0:02:53 in the next five years.
0:02:55 Over the past three years,
0:02:59 we have already built out in data center capacity
0:03:00 a larger amount of dollars
0:03:03 than the entire U.S. interstate highway system,
0:03:04 which took 40 years,
0:03:06 just in terms of dollars,
0:03:07 and that’s inflation adjusted.
0:03:09 Open AI alone, I think,
0:03:12 has more than a trillion dollars of deals
0:03:14 set up that they’ve committed to,
0:03:15 and we can talk about that.
0:03:17 But, at the same time,
0:03:19 so those are all like big numbers on infrastructure,
0:03:20 and they’re scary,
0:03:21 and they say, oh, bubble.
0:03:24 And Google released a stat recently
0:03:28 that they have seen a 150X increase
0:03:30 in the amount of tokens processed
0:03:31 in the last 17 months.
0:03:33 So, on the one hand,
0:03:35 you’ve got this crazy, scary-sounding build-out.
0:03:36 On the other hand,
0:03:39 you actually have a bunch of usage that’s happening.
0:03:41 So, are we in an AI bubble?
0:03:44 I do not believe we’re in an AI bubble today.
0:03:46 I had, depending on how you look at it,
0:03:47 the privilege and the misfortune
0:03:48 of being a tech investor
0:03:50 during the year 2000 bubble,
0:03:51 which was really a telecom bubble.
0:03:52 And I think it’s really helpful
0:03:54 to compare and contrast today
0:03:56 to the year 2000.
0:03:58 First, I think Cisco peaked
0:04:01 at 150 or 180 times trailing earnings.
0:04:02 NVIDIA’s at more like 40 times.
0:04:04 So, valuations are very different.
0:04:06 Most important, however,
0:04:09 is that the year 2000 internet bubble
0:04:10 or telecom bubble
0:04:12 was defined by something called dark fiber.
0:04:15 And if you’re a veteran of the year 2000,
0:04:17 you’ll know what that was.
0:04:19 But dark fiber was literally fiber
0:04:20 that was laid down in the ground
0:04:23 and not lit up.
0:04:24 Fiber is useless
0:04:26 unless you have the optics
0:04:27 and switches and routers
0:04:28 that you need on either side.
0:04:30 So, I vividly remember
0:04:31 companies like Level 3
0:04:32 or Global Crossing
0:04:33 or WorldCom
0:04:34 would come in and they say,
0:04:36 we laid 200,000 miles
0:04:37 of dark fiber this quarter.
0:04:38 This is so amazing.
0:04:40 The internet’s going to be so big.
0:04:42 We can’t wait to light these up.
0:04:44 At the peak of the bubble,
0:04:47 97% of the fiber
0:04:48 that had been laid
0:04:49 in America
0:04:50 was dark.
0:04:52 Contrast that with today.
0:04:54 There are no dark GPUs.
0:04:56 All you have to do
0:04:58 is read any technical paper
0:04:59 and that one of the biggest problems
0:05:00 in a trading run
0:05:02 is that GPUs are melting.
0:05:05 And there’s a very simple way
0:05:06 to kind of cut to the heart
0:05:06 of all of this.
0:05:08 It is a return on investing capital
0:05:10 of the biggest spenders
0:05:11 on GPUs
0:05:12 who are all public.
0:05:15 And those companies,
0:05:16 since they ramped up CapEx,
0:05:17 have seen,
0:05:19 call it a 10-point increase
0:05:20 in their ROICs.
0:05:21 So, thus far,
0:05:22 the ROI on all the spending
0:05:23 has been really positive.
0:05:25 It’s an interesting
0:05:26 and open debate
0:05:27 about whether or not
0:05:27 it will continue
0:05:28 to be positive.
0:05:30 With the quantum of spend
0:05:30 we’re going to have
0:05:31 on Blackwell,
0:05:32 I personally think it will.
0:05:34 But there’s no debate
0:05:35 that thus far
0:05:36 the ROI on AI
0:05:38 has been really positive
0:05:40 and valuation-wise
0:05:41 we’re just not in a bubble.
0:05:43 I couldn’t agree more.
0:05:44 The other thing
0:05:45 that I would say is
0:05:46 you can contrast
0:05:48 the actual adoption
0:05:49 and usage
0:05:50 of the technology
0:05:51 from then, right?
0:05:52 The internet was actually
0:05:53 really hard
0:05:54 because you had to build
0:05:55 a two-sided network.
0:05:56 Like, you had to build
0:05:56 websites
0:05:57 and then you had to get users
0:05:59 and it’s much more difficult.
0:06:01 In the case of the AI tool,
0:06:02 all you have to do
0:06:03 is kind of light them up
0:06:04 via API
0:06:05 or turn on your website
0:06:06 ChatGPT
0:06:07 and everybody has access
0:06:08 to them, right?
0:06:09 Built on top of cloud computing
0:06:10 on top of the internet
0:06:11 and you can get to
0:06:12 instant distribution
0:06:13 a billion people right away.
0:06:14 Absolutely.
0:06:16 So, the other thing
0:06:16 is the counterparty.
0:06:17 So, you mentioned this.
0:06:18 They happen to be
0:06:19 the best companies
0:06:21 in the history of the world, right?
0:06:22 I think collectively
0:06:23 the people who are
0:06:24 coming out of pocket
0:06:25 the writing checks
0:06:26 for this CapEx
0:06:28 I think they collectively
0:06:29 generate like
0:06:30 $300 billion
0:06:31 of free cash flow
0:06:31 a year.
0:06:32 Is that right?
0:06:34 Some directionally?
0:06:34 Round numbers.
0:06:35 Yeah, and they have
0:06:36 $500 billion
0:06:37 of cash on the balance sheet.
0:06:39 So, whenever people are like
0:06:39 oh my god, it’s a bubble
0:06:40 or is it going to pop?
0:06:41 I’m like
0:06:42 I think it’s kind of fine.
0:06:43 I mean, it costs like
0:06:45 $40 or $50 billion
0:06:46 to light up
0:06:47 one gigawatt.
0:06:48 Yeah, if you’re
0:06:49 on NVIDIA chips.
0:06:50 On NVIDIA chips.
0:06:51 Yeah.
0:06:51 Yeah.
0:06:52 So, you know,
0:06:53 there’s kind of like
0:06:54 an $800 billion buffer
0:06:56 growing $300 billion
0:06:56 every year.
0:06:58 Yeah, I mean
0:06:59 free cash flow
0:06:59 at some of them
0:07:01 has begun
0:07:03 to maybe, you know.
0:07:04 Well, this goes to your point
0:07:05 on return on invested capital.
0:07:06 Yes, we should see that next year.
0:07:07 It might creep down a little bit.
0:07:08 Yeah, a little bit of a mismatch
0:07:09 in the build out.
0:07:10 But, you know,
0:07:11 Larry Page apparently
0:07:11 internally said
0:07:13 I’m happy to go bankrupt
0:07:14 rather than lose this race.
0:07:15 And I think that is
0:07:16 the mentality for sure
0:07:17 at Google
0:07:18 and perhaps Meta.
0:07:20 It’s just seen as existential
0:07:22 and you have to win.
0:07:24 Okay, so lots
0:07:25 has been written
0:07:27 about these round-tripping deals.
0:07:29 So, because round-tripping
0:07:30 is a very scary concept
0:07:32 from the internet build-out.
0:07:33 That was a big problem.
0:07:34 What do you make of it here?
0:07:36 It is objectively happening.
0:07:37 Money is fungible.
0:07:38 So, NVIDIA,
0:07:39 if they sign a deal
0:07:40 with OpenAI,
0:07:40 they can say,
0:07:42 hey, you can’t use our money
0:07:43 to buy our chips,
0:07:44 but money is fungible.
0:07:45 But it’s happening
0:07:46 at a very small scale.
0:07:48 Yes.
0:07:49 Yeah.
0:07:50 And I think…
0:07:50 I know this is like
0:07:51 a crypto or blockchain.
0:07:52 Yeah, exactly.
0:07:54 And I think
0:07:55 what is driving this
0:07:57 isn’t the need
0:07:59 to finance GPU
0:08:01 or data center purchases,
0:08:02 but it’s actually
0:08:03 competitive dynamics.
0:08:05 So, NVIDIA’s
0:08:05 biggest competitor,
0:08:06 it’s not AMD,
0:08:08 it’s not Broadcom,
0:08:10 it’s certainly not Marvell,
0:08:11 it’s not Intel,
0:08:13 it’s Google.
0:08:14 And more specifically,
0:08:16 it is Google
0:08:17 because Google
0:08:19 owns the TPU chip.
0:08:20 And this is by far,
0:08:22 maybe perhaps today,
0:08:24 the only alternative
0:08:25 to NVIDIA for training
0:08:27 and maybe the best
0:08:28 inference alternative.
0:08:30 And Google’s
0:08:31 a problematic competitor
0:08:32 because they also own
0:08:33 a company called DeepMind
0:08:34 and they have
0:08:36 a product called Gemini.
0:08:37 And I think you could argue
0:08:37 that they’re
0:08:39 the leading AI company today.
0:08:40 I think they’ve taken
0:08:41 15 or 20 points
0:08:42 of traffic share
0:08:43 in the last two
0:08:44 or three months
0:08:45 and that’s just traffic
0:08:45 to Gemini.
0:08:47 it does not include
0:08:48 search overviews.
0:08:49 I suspect
0:08:50 on a actual
0:08:51 traffic basis,
0:08:52 Google is bigger
0:08:53 than OpenAI,
0:08:54 Anthropic,
0:08:55 anyone today.
0:08:58 And that business
0:08:58 is going to run
0:08:59 on TPUs.
0:09:00 And then we have
0:09:01 three other labs
0:09:02 that are relevant today.
0:09:02 There’s Anthropic
0:09:04 and that’s an Amazon
0:09:05 and Google captive.
0:09:05 You know,
0:09:06 Anthropic is really
0:09:07 going to run
0:09:07 on TPUs
0:09:08 and Traniums.
0:09:09 And so you’re left
0:09:10 with XAI
0:09:11 and OpenAI
0:09:12 at the forefront.
0:09:13 And if Google
0:09:15 is going to
0:09:16 a lab like Anthropic
0:09:16 and saying,
0:09:17 I’m going to help
0:09:18 you fundraise
0:09:20 and give you chips
0:09:21 for competitive reasons,
0:09:22 it’s very hard
0:09:23 for NVIDIA
0:09:24 not to respond.
0:09:25 And as Jensen said,
0:09:25 he thinks it’s going
0:09:26 to be a good investment.
0:09:27 So I think
0:09:28 the round-tripping
0:09:29 concerns
0:09:32 are pretty overblown.
0:09:33 I mean,
0:09:34 what NVIDIA really needs
0:09:35 is they need
0:09:36 Meta to get
0:09:36 their act together
0:09:38 or another
0:09:38 American open-source
0:09:39 player to emerge
0:09:41 or maybe some sort
0:09:42 of detente
0:09:42 with China
0:09:43 and AI.
0:09:45 When people ask me
0:09:46 about NVIDIA
0:09:46 and all the moves
0:09:47 and the round-tripping,
0:09:48 my reaction is
0:09:49 everything they’ve done
0:09:50 is completely rational.
0:09:52 100% rational.
0:09:53 Yeah, long-term.
0:09:54 Yeah, long-term.
0:09:54 Sure, some of the things
0:09:55 they do may not have
0:09:56 as high of a return
0:09:57 on capital as other things,
0:09:58 but strategically,
0:09:59 I think they’re all
0:10:00 kind of the right moves.
0:10:01 Jensen’s one of the
0:10:01 two best CEOs
0:10:02 along with Elon
0:10:04 I have ever known,
0:10:05 and I think he’s
0:10:06 playing a strong hand
0:10:07 really well.
0:10:07 Yeah.
0:10:08 All right,
0:10:09 so you started getting
0:10:10 into the model companies.
0:10:11 Let’s just talk about
0:10:13 the model.
0:10:14 So we can come back
0:10:15 to chips and memory
0:10:16 and networking
0:10:16 because I want to get
0:10:17 your take on that,
0:10:18 but since we’re
0:10:19 on the model side,
0:10:20 what do you think
0:10:21 happens with market structure?
0:10:22 Who wins where?
0:10:24 Who are you most
0:10:24 optimistic about?
0:10:25 Where do you have concerns?
0:10:28 So I think humility is an important
0:10:30 virtue for an investor.
0:10:32 And I’m just,
0:10:33 if we’re going to make
0:10:35 an analogy and say
0:10:37 that ChatGPT is to AI,
0:10:39 has Netscape Navigator
0:10:40 was to the internet.
0:10:43 At this point in the internet boom,
0:10:45 Google had not been founded.
0:10:48 Mark Zuckerberg was in middle school.
0:10:50 Travis Kalanick was in kindergarten.
0:10:53 So it’s just very early.
0:10:54 So I think it’s important
0:10:55 to be humble
0:10:56 about making
0:10:57 high confidence predictions
0:10:59 at the application layer.
0:11:00 It’s one reason I think
0:11:01 the infrastructure layer
0:11:02 is often maybe
0:11:03 a safe place to be
0:11:04 at the beginning
0:11:05 of one of these
0:11:06 new technology waves.
0:11:07 Well, actually talk about
0:11:08 the role they play
0:11:09 at the infrastructure layer
0:11:10 because there’s a piece of them
0:11:11 that obviously they serve
0:11:12 as an infrastructure layer
0:11:13 powering other
0:11:14 application providers
0:11:15 and then they also have
0:11:16 their own applications.
0:11:16 So I think
0:11:18 I would draw a distinction.
0:11:18 Yeah.
0:11:19 I mean, that’s most true
0:11:20 of Google,
0:11:21 but I think it’s hard
0:11:22 to have high conviction
0:11:24 other than to observe
0:11:25 the internet was
0:11:27 a very disruptive innovation.
0:11:28 I think there’s
0:11:29 reasonable arguments
0:11:30 that AI could be
0:11:32 a sustaining innovation
0:11:34 because the raw ingredients
0:11:35 of kind of data,
0:11:38 the capital to buy compute
0:11:39 and distribution,
0:11:40 which is what you need,
0:11:42 all of today’s
0:11:43 biggest tech companies
0:11:44 have all of those in spades.
0:11:46 So as long as they execute well,
0:11:48 hire good people
0:11:50 and have a sound strategy,
0:11:52 like I think you could see
0:11:53 it be a sustaining innovation
0:11:54 for a lot of members
0:11:55 of the Mag 7.
0:11:56 On the other hand,
0:11:57 I do think it’s existential
0:11:59 and if you don’t execute,
0:12:00 you know,
0:12:03 IBM might be a good fate.
0:12:04 Yeah.
0:12:04 Yeah.
0:12:05 Yeah.
0:12:06 That’s tough.
0:12:07 Yeah.
0:12:08 Data distribution,
0:12:09 compute,
0:12:10 dollars,
0:12:11 talent.
0:12:12 Yeah.
0:12:14 They have every right to win.
0:12:14 Yeah.
0:12:15 They have every right to win.
0:12:16 And it seems now
0:12:17 more than before
0:12:18 they’re taking it quite seriously.
0:12:20 Yeah.
0:12:21 Maybe Google in particular,
0:12:23 but obviously Meta’s making
0:12:24 the dramatic moves
0:12:24 they’re making too.
0:12:25 No,
0:12:25 to me,
0:12:27 ChatGPT was Pearl Harbor
0:12:28 for Google
0:12:29 and we’re going to see
0:12:30 how they responded
0:12:31 and they’re slowly
0:12:32 starting to respond.
0:12:32 Yeah.
0:12:34 And then what do you think,
0:12:35 what’s your forecast
0:12:37 for that sort of
0:12:38 the platform piece
0:12:38 of their business,
0:12:39 the infrastructure piece?
0:12:41 What do you think,
0:12:42 how do you think it shakes out
0:12:43 in terms of like
0:12:44 business model,
0:12:45 market structure?
0:12:46 So do you think
0:12:47 they end up
0:12:48 as high margin businesses
0:12:50 like the clouds
0:12:52 or like aircraft manufacturers
0:12:53 or do you think
0:12:54 they end up
0:12:55 very competitive
0:12:56 in low margin businesses
0:12:57 like airlines?
0:12:59 I don’t think
0:13:00 they’ll be airlines,
0:13:01 but you can,
0:13:02 anybody can just look
0:13:03 at the P&L,
0:13:04 you know,
0:13:05 of a SaaS company
0:13:08 circa 2021 and 2022
0:13:08 and you see,
0:13:09 you know,
0:13:11 80, 90% gross margins
0:13:14 and the nature of AI
0:13:15 because of scaling laws,
0:13:16 Richard Sutton’s
0:13:17 the better,
0:13:17 the better listen.
0:13:20 They’re just more
0:13:21 compute intensive.
0:13:23 So their gross margins
0:13:24 are structurally
0:13:25 going to be lower
0:13:26 but that doesn’t mean
0:13:27 they can’t be
0:13:27 great businesses.
0:13:28 I just,
0:13:29 I think it’s going
0:13:30 to be a long time
0:13:31 before we see
0:13:33 a truly kind of,
0:13:34 you know,
0:13:35 an AI lab,
0:13:36 a frontier lab
0:13:37 with gross margins
0:13:39 anywhere near SaaS
0:13:41 or internet era margins.
0:13:41 Now,
0:13:42 their OPEX
0:13:43 can be a lot lower
0:13:45 and,
0:13:45 you know,
0:13:46 maybe that’s how
0:13:47 you square it
0:13:48 but just the gross margins
0:13:49 are fundamentally different
0:13:51 and until scaling laws change
0:13:52 and the importance
0:13:53 of test time compute,
0:13:54 things like that,
0:13:55 change,
0:13:56 which I don’t see
0:13:56 happening,
0:13:58 they are going to be
0:13:58 lower margin.
0:13:59 Yeah.
0:14:00 Okay,
0:14:01 so let’s talk about
0:14:02 application layer.
0:14:04 So you just,
0:14:05 you just kind of got into
0:14:05 it a little bit
0:14:06 with the SaaS businesses
0:14:08 and I don’t know
0:14:08 if you’ve waded
0:14:09 into this fight
0:14:10 on Twitter
0:14:11 but it’s sort of,
0:14:11 you know,
0:14:12 the like,
0:14:12 you know,
0:14:13 every few months
0:14:14 it comes up
0:14:14 and it’s like,
0:14:16 SaaS is terrible
0:14:16 and it’s dead
0:14:16 and,
0:14:17 you know,
0:14:17 it’s all going
0:14:18 to go away
0:14:18 and then,
0:14:19 you know,
0:14:21 with Andre’s
0:14:23 Duarkesh interview
0:14:23 he just did,
0:14:24 it’s,
0:14:24 you know,
0:14:25 like the market’s reacting
0:14:26 positively to it
0:14:27 and it’s like
0:14:29 a whipsaw reaction.
0:14:29 So what do you think
0:14:31 happens with SaaS
0:14:31 and software?
0:14:32 You know,
0:14:32 I think I,
0:14:33 you know,
0:14:34 first said probably
0:14:34 in early 24
0:14:35 that I thought
0:14:37 all of application SaaS
0:14:37 might be a zero
0:14:39 different than
0:14:41 infrastructure SaaS.
0:14:43 I would say
0:14:44 I have a more nuanced
0:14:45 view now
0:14:47 and I think
0:14:48 there could be
0:14:49 some really big
0:14:50 application SaaS winners
0:14:52 especially if you serve
0:14:53 like a more fragmented
0:14:54 SMB customer base.
0:14:56 You know,
0:14:57 Google is making it
0:14:57 really easy
0:14:58 if you’re a customer
0:14:58 of theirs
0:15:00 to use your data
0:15:01 and essentially
0:15:02 make any SaaS app
0:15:02 you want
0:15:03 and then your data
0:15:04 isn’t shared
0:15:04 with anyone else.
0:15:07 But the critical mistake
0:15:08 that I think
0:15:09 a lot of retailers
0:15:09 made
0:15:12 in dealing with Amazon
0:15:13 is they looked
0:15:14 at Amazon’s margins
0:15:15 and they said
0:15:16 we don’t want
0:15:17 to be in that business.
0:15:19 And
0:15:20 that was obviously
0:15:20 a terrible mistake
0:15:21 and here we are
0:15:23 25 years later
0:15:23 and, you know,
0:15:24 Amazon has really
0:15:26 healthy retail margins.
0:15:27 And
0:15:28 I worry
0:15:29 that application
0:15:30 SaaS companies
0:15:31 are trying
0:15:32 to preserve
0:15:34 their existing
0:15:34 gross margin
0:15:35 structures
0:15:37 because they believe
0:15:38 that if their
0:15:38 gross margins
0:15:39 go down,
0:15:40 their stocks
0:15:41 will go down.
0:15:42 It is
0:15:43 definitionally impossible
0:15:44 given what we just
0:15:45 discussed
0:15:46 to succeed
0:15:46 in AI
0:15:47 without gross
0:15:47 margin pressure
0:15:49 and I do not
0:15:49 know
0:15:50 why they have
0:15:51 concerns
0:15:52 because we have
0:15:52 an existence
0:15:53 proof
0:15:54 that a software
0:15:55 company can
0:15:56 deal well
0:15:56 with declining
0:15:57 margins
0:15:58 and Microsoft
0:15:59 and Adobe
0:15:59 to the whole
0:16:00 AI thing
0:16:00 came along.
0:16:01 It used to be
0:16:02 that companies
0:16:02 were scared
0:16:03 to go from
0:16:03 on-premise
0:16:04 to the cloud
0:16:05 because margins
0:16:05 were lower.
0:16:06 Cloud margins
0:16:08 are lower.
0:16:08 They’re still good.
0:16:10 And Microsoft,
0:16:11 they transitioned
0:16:13 from on-premise
0:16:14 perpetual licenses
0:16:15 with maintenance
0:16:16 to a cloud model
0:16:18 and it was a
0:16:18 pretty good stock
0:16:19 for 10 years.
0:16:20 So I don’t,
0:16:21 if you’re an
0:16:22 application SaaS
0:16:22 company,
0:16:23 like what I would
0:16:24 just say is
0:16:25 don’t be scared
0:16:26 and look at
0:16:27 declining gross
0:16:27 margins
0:16:28 kind of has a
0:16:29 mark of success
0:16:30 rather than
0:16:31 a badge of shame
0:16:32 or something
0:16:32 to be feared.
0:16:33 It’s actually so
0:16:34 funny you say
0:16:34 that because
0:16:35 whenever we have
0:16:35 these discussions
0:16:36 about companies,
0:16:38 basically every
0:16:38 company that comes
0:16:39 to present to us
0:16:39 is like we’re
0:16:40 an AI company
0:16:42 and we always
0:16:42 look at the gross
0:16:43 margins and it’s
0:16:44 become like a
0:16:45 badge of honor
0:16:45 for them to
0:16:46 actually have
0:16:46 low gross
0:16:47 margins
0:16:47 because we’re
0:16:47 like oh my
0:16:47 God,
0:16:47 people are
0:16:48 actually using
0:16:49 your AI stuff.
0:16:50 But if you
0:16:50 show up and
0:16:50 you’re like I’m
0:16:51 an AI company
0:16:51 and it’s like I
0:16:53 got 82% gross
0:16:53 margins,
0:16:54 you’re like I
0:16:54 don’t think
0:16:55 anybody’s really
0:16:55 using it.
0:16:58 It’s interesting.
0:16:58 If you’re one
0:16:59 of these public
0:16:59 companies,
0:16:59 would you rather
0:17:00 have like 10
0:17:01 bucks of revenue
0:17:01 with 90% gross
0:17:02 margins or 50
0:17:03 bucks of revenue
0:17:04 with 60% gross
0:17:04 margins?
0:17:05 Not hard.
0:17:05 It’s not that
0:17:06 complicated.
0:17:07 It’s hard to do
0:17:07 in the public
0:17:08 market.
0:17:08 It’s hard to do
0:17:09 in publics,
0:17:09 but if you
0:17:10 communicate it,
0:17:10 you draw parallels
0:17:11 to the cloud
0:17:11 transition.
0:17:13 I mean I’m
0:17:14 an investor and
0:17:14 I would be
0:17:14 excited about
0:17:16 it and I
0:17:16 don’t think I’m
0:17:17 alone in the
0:17:17 world.
0:17:18 And then the
0:17:19 big advantage
0:17:20 these legacy
0:17:21 application SaaS
0:17:21 companies have
0:17:23 is they do
0:17:24 have these
0:17:24 really profitable
0:17:26 existing businesses
0:17:27 and so you
0:17:28 can run
0:17:28 your new
0:17:29 AI products
0:17:30 at breakeven
0:17:34 and catch up
0:17:34 to the leaders,
0:17:35 et cetera,
0:17:35 et cetera.
0:17:36 And I’m just
0:17:37 surprised more
0:17:37 people have not
0:17:38 done that.
0:17:39 Like why are
0:17:39 none of the
0:17:40 public coding
0:17:41 companies even
0:17:42 trying to compete
0:17:42 with Cursor?
0:17:44 And the reality
0:17:44 is Cursor now
0:17:45 they have a
0:17:46 trillion tokens
0:17:49 and there will
0:17:49 be a point
0:17:50 where they have
0:17:51 enough coding
0:17:52 tokens that it’s
0:17:53 tough to catch
0:17:53 them.
0:17:54 But I think
0:17:55 today if you’re
0:17:55 a public coding
0:17:56 company and
0:17:57 you said I’m
0:17:57 going to lean
0:17:58 in, I’m going
0:17:58 to run it
0:17:59 breakeven, I have
0:17:59 an existing
0:18:00 business, I’m
0:18:01 going to attach
0:18:01 it to
0:18:02 everything, hey
0:18:03 you have a
0:18:04 chance and
0:18:05 you know the
0:18:06 prize is clearly
0:18:07 really big.
0:18:07 I see Martin
0:18:08 is skeptical.
0:18:08 Martin
0:18:09 said you have
0:18:10 a chance.
0:18:10 I said a chance.
0:18:11 I said a chance.
0:18:12 It’s like a
0:18:12 dumb and dumber
0:18:13 you’re telling
0:18:13 me there’s a
0:18:14 chance not like a
0:18:14 real chance.
0:18:15 You’re telling
0:18:16 me there’s a
0:18:16 chance.
0:18:20 Yeah exactly.
0:18:20 I totally agree.
0:18:21 Yeah we actually
0:18:22 saw I mean you
0:18:22 know we see it
0:18:24 you know we may
0:18:25 if we you know
0:18:26 Figma for
0:18:28 example like when
0:18:28 they went out they
0:18:29 are extremely high
0:18:30 gross margin and
0:18:30 they’re like hey
0:18:31 we’re going to you
0:18:32 know pretty
0:18:33 aggressively distribute
0:18:34 our AI tools and
0:18:34 our gross margins
0:18:34 are going to go
0:18:36 down and you know
0:18:36 investors asked a
0:18:37 few clarifying
0:18:38 questions and then
0:18:38 they were like oh
0:18:39 that actually would
0:18:39 be a good thing.
0:18:40 And so I’m
0:18:40 surprised more
0:18:41 people in the
0:18:41 public markets
0:18:42 aren’t doing it.
0:18:43 worked out okay for
0:18:43 them.
0:18:44 It’s working out
0:18:45 well.
0:18:46 Long game to
0:18:46 play.
0:18:47 What about on the
0:18:48 consumer side of the
0:18:49 application layer?
0:18:51 So obviously Google
0:18:52 was the portal to the
0:18:53 internet is kind of
0:18:53 still is the portal to
0:18:54 the internet and the
0:18:55 whole business model
0:18:57 was predicated upon
0:18:59 taking some intent and
0:19:00 directing you to
0:19:01 someone else’s website
0:19:03 where they would do
0:19:03 stuff with you.
0:19:06 It’s kind of not
0:19:06 going to be that way.
0:19:07 It already is not that
0:19:09 way with AI.
0:19:10 although I tried the
0:19:11 browser today and I
0:19:12 tried to do some
0:19:13 pretty basic shopping
0:19:14 stuff and it’s you
0:19:15 know still still some
0:19:16 work to do but I
0:19:17 think it will get
0:19:17 there.
0:19:18 So what do you
0:19:19 actually think happens
0:19:21 with the sort of
0:19:22 market structure of
0:19:22 the consumer internet
0:19:23 companies?
0:19:24 Do they get
0:19:26 subsumed into a
0:19:28 component of a
0:19:29 chatbot interface or
0:19:30 do you think it’s
0:19:30 something else?
0:19:33 So one humility
0:19:34 hard to say.
0:19:36 Two I would just
0:19:39 say I think the AI
0:19:40 companies that have
0:19:41 launched these AI
0:19:42 browsers may come to
0:19:43 regret it because
0:19:43 there’s something
0:19:45 called Chrome that
0:19:46 has whatever it is
0:19:49 five billion users and
0:19:51 if you’re Google you
0:19:52 know you can just go
0:19:52 look and what
0:19:52 happened with Google
0:19:54 Buzz they are very
0:19:55 cautious you know
0:19:56 there’s you know
0:19:57 they’re currently in
0:19:57 litigation with the
0:20:01 government and they
0:20:02 could easily do this
0:20:04 and probably do it
0:20:05 even better but they
0:20:06 didn’t want to be
0:20:07 first.
0:20:09 so now you have
0:20:10 two AI native
0:20:11 companies with their
0:20:12 own browsers let
0:20:13 them run for three
0:20:15 to six months get a
0:20:16 little head start and
0:20:18 then wow here we are
0:20:19 we had to do this and
0:20:21 I don’t know how that’s
0:20:23 going to work maybe
0:20:24 for the companies other
0:20:25 than Google who don’t
0:20:26 own Chrome.
0:20:30 I guess data and
0:20:30 distribution is pretty
0:20:31 powerful in that.
0:20:32 Yeah hindsight’s 20-20
0:20:35 and the one thing I
0:20:36 would say is I do
0:20:37 think it’s tough to
0:20:37 bet against the
0:20:39 companies with large
0:20:40 existing user bases
0:20:44 today and I also
0:20:45 think reasoning has
0:20:46 fundamentally changed
0:20:47 the economics of these
0:20:49 frontier models you
0:20:51 know pre-reasoning I
0:20:52 often said if you are a
0:20:54 frontier model without
0:20:56 access to unique
0:20:59 valuable data and
0:21:00 internet scale
0:21:01 distribution you’re the
0:21:02 fastest depreciating
0:21:03 asset in history.
0:21:04 I think reasoning really
0:21:06 changed that because the
0:21:07 way RL works during
0:21:08 post-training having a
0:21:10 big user base now kind
0:21:12 of unlocks that flywheel
0:21:13 that was at the center of
0:21:14 every great consumer
0:21:16 internet company where
0:21:17 you have a good product
0:21:19 you get a lot of users
0:21:20 the users make the
0:21:22 algorithm better the
0:21:23 algorithm makes the
0:21:24 product better and it
0:21:26 just spins and that
0:21:28 it’s not quite spinning
0:21:30 yet in AI but you can
0:21:31 squint and see it and
0:21:32 so I think that
0:21:33 fundamentally changes
0:21:34 the economics for
0:21:38 anthropic for XAI for
0:21:41 open AI but I mean
0:21:43 Mark Zuckerberg’s trying
0:21:45 hard yeah we’ll see
0:21:47 yeah yeah yeah a lot of
0:21:48 smart people in there
0:21:50 now yeah for sure I
0:21:51 think the worry is and
0:21:52 I think this is another
0:21:54 interesting thing is if
0:21:56 you don’t like in a
0:21:57 strange way the Chinese
0:21:58 open source model
0:22:00 ecosystem is a godsend
0:22:02 to any American company
0:22:03 that’s trying to catch
0:22:04 those four leading labs
0:22:06 because the problem is
0:22:07 if you don’t have
0:22:10 Gemini 2.5 Pro or a
0:22:11 later checkpoint of it
0:22:13 or a later checkpoint of
0:22:14 Grok that we don’t see
0:22:15 or a later GPT
0:22:18 checkpoint training the
0:22:19 next model you’re at a
0:22:20 big disadvantage oh by
0:22:21 the way one thing I
0:22:22 just want to say that
0:22:23 drives me crazy is all
0:22:24 these people who say
0:22:25 that GPT-5 is the end
0:22:28 of scaling loss GPT-5 is
0:22:30 a smaller model it was
0:22:31 not designed to be
0:22:32 better it was designed to
0:22:33 be more economical for
0:22:35 open AI and Microsoft to
0:22:38 run any reference to
0:22:40 GPT-5 and scaling laws is
0:22:44 crazy yeah sorry rant rant
0:22:45 over we get the pedestal
0:22:46 up here if you want
0:22:47 yeah exactly shaking your
0:22:48 hand yeah that’d be good
0:22:50 yeah that’d be good uh do
0:22:51 you want to talk about
0:22:53 chips sure so okay I know
0:22:56 you love NVIDIA talk about
0:22:57 you know your view of
0:23:01 NVIDIA AMD TPUs ASICs and
0:23:02 how do you think sort of
0:23:04 market structure shakes out
0:23:06 their you know competitive
0:23:07 advantage that the various
0:23:11 players have yeah I think
0:23:12 it goes I think it is
0:23:15 really um it’s a fight
0:23:18 between NVIDIA and um the
0:23:20 Google TPU and that’s
0:23:20 something that I don’t
0:23:21 think is broadly
0:23:23 appreciated is the extent
0:23:25 which Broadcom and AMD are
0:23:26 effectively going to market
0:23:28 together NVIDIA is no
0:23:29 longer just a semiconductor
0:23:30 company as I’m sure you’ll
0:23:31 hear from Jensen tomorrow
0:23:33 you know not it was a
0:23:34 semiconductor company then
0:23:35 a software company with
0:23:36 CUDA now systems company
0:23:37 with these rack level
0:23:39 solutions and now
0:23:39 arguably you know a
0:23:40 data setter level
0:23:42 uh company with the you
0:23:43 know level of architecting
0:23:44 they’re doing with scale
0:23:46 up scale across and um
0:23:48 scale out scale across
0:23:51 networking um so the
0:23:53 networking the fabric the
0:23:54 software it’s all
0:23:55 important and what
0:23:56 Broadcom is saying to
0:23:57 companies like Meta is
0:23:59 hey we will build you a
0:24:01 fabric that can
0:24:02 theoretically compete with
0:24:04 NVIDIA’s fabric which is
0:24:06 a mixture of NVLink and
0:24:07 either InfiniBand or
0:24:08 Ethernet um it will build
0:24:09 it on the Ethernet it’s
0:24:10 going to be an open
0:24:12 standard and hey we’ll
0:24:13 we’ll make you your
0:24:15 version of of TPU which
0:24:15 by the way took Google
0:24:17 three generations to get
0:24:18 working and you know what
0:24:19 if your ASIC isn’t good
0:24:21 you can just plug AMD
0:24:24 right in um but I
0:24:26 personally believe most of
0:24:26 those ASICs are going to
0:24:29 fail um particularly if
0:24:30 it’s in the fullness of
0:24:31 time like over a period of
0:24:32 time or in the fullness of
0:24:35 time in the next three
0:24:36 years I think you’ll see a
0:24:38 bunch of high profile um
0:24:39 ASIC programs canceled
0:24:41 especially if Google um
0:24:42 starts selling TPUs
0:24:44 externally which has been
0:24:46 all over X and then you
0:24:47 know they you know who
0:24:48 knows exactly how that
0:24:49 would work because if
0:24:51 you’re an anthropic it was
0:24:52 just rumored anthropic wants
0:24:52 to buy tens of billions of
0:24:54 TPUs if you’re anthropic
0:24:55 maybe you don’t want Google
0:24:56 seeing your secret sauce but
0:24:58 there’s ways around that so
0:25:00 I think this is really a
0:25:01 battle between Google and
0:25:03 it’s TPU enabled by
0:25:04 Broadcom for now and
0:25:06 Google can take the TPU
0:25:07 away from Broadcom whenever
0:25:09 they want yeah now they
0:25:10 can’t do the Ethernet
0:25:11 networking that Broadcom is
0:25:13 is doing uh but they
0:25:15 control the TPU um so it’s
0:25:16 really Google and the TPU
0:25:20 verse um NVIDIA you know
0:25:21 with with you know Amazon
0:25:23 like that’s a very talented
0:25:25 team arguing the most
0:25:26 talented silicon team at any
0:25:27 hyperscaler the Annapurna
0:25:29 team like I think the
0:25:30 Tranium 3 will probably be a
0:25:31 much better chip than the
0:25:33 Tranium 2 it took Google
0:25:34 three generations to get the
0:25:36 TPU right um and then AMD
0:25:38 will you know will always be
0:25:39 kind of the second source and
0:25:40 you need a second source
0:25:43 all right exciting uh what do
0:25:45 you think happens okay so I
0:25:46 want to go back um to
0:25:48 business models so one of the
0:25:49 big things that is widely
0:25:51 discussed is like you know
0:25:53 source of disruption and most
0:25:55 of the CEOs in this room are
0:25:56 CEOs of startups who are
0:25:58 trying to go beat some
0:26:00 incumbent or find you know
0:26:01 some new market
0:26:03 opportunity and the most
0:26:04 ripe opportunities tend to
0:26:05 come when you have a big
0:26:07 platform shift that is also
0:26:08 accompanied with a business
0:26:11 model shift um and so there
0:26:12 are a couple of areas where I
0:26:14 can see it I feel like in an
0:26:16 obvious way so you know we’re
0:26:17 investors in Decagon customer
0:26:19 support like you can pretty
0:26:21 easily see a business model that
0:26:23 is priced on the resolution of
0:26:24 a task because it’s so
0:26:27 measurable um you can see you
0:26:29 know like encoding like a lot of
0:26:30 the business model has now
0:26:32 shifted to consumption and you
0:26:33 know obviously especially for
0:26:34 developer facing things like
0:26:37 that’s comfortable um and pretty
0:26:39 well known what about the rest
0:26:40 of the industry because I feel
0:26:42 like there’s sort of this hand
0:26:44 wave thing that is going on
0:26:45 which is like we’re gonna go get
0:26:47 all of services but it’s like
0:26:48 okay so how do you actually go do
0:26:50 that it’s gonna be pretty hard so
0:26:52 do you have any prediction on how
0:26:54 that plays out well I think what
0:26:56 you’re seeing in customer service
0:26:57 which is kind of like an easy
0:27:00 first example um we have a lot of
0:27:02 textual data the LLMs are good at
0:27:04 text you can kind of you know
0:27:07 probably really easily run some
0:27:08 RL to make sure that they you know
0:27:11 get a good verified reward you
0:27:12 know verified reward being happy
0:27:14 customer first call resolution or
0:27:17 whatever it is um and but I do
0:27:19 think you will see that played out
0:27:20 like humans were fundamentally
0:27:22 played for out paid paid based on
0:27:25 outcomes and a lot of AI will be
0:27:27 augmenting humans but probably
0:27:29 also replacing some humans and
0:27:30 that will involve being paid um
0:27:31 paid for outcomes you know going
0:27:33 back to the consumer business
0:27:34 model you know everybody’s
0:27:35 talking about affiliate fees and
0:27:37 for sure I’m gonna have you know
0:27:40 my own AI it will be a version of
0:27:42 Grok um because we’re both ex AI
0:27:43 shareholders it will be a version
0:27:45 of Grok that knows me and it likes
0:27:47 me um and you know when I when I
0:27:49 want to you know the next time I
0:27:50 want to go on vacation it will
0:27:52 know the hotels that I like to go
0:27:55 to and it’ll say hey three hotels I
0:27:57 have Gavin you know I have Gavin
0:27:59 coming who’s got the best price of
0:28:02 the best room um it’s gonna
0:28:03 massively upgrade the gifts that you
0:28:04 give to Becky
0:28:08 yes Becky Becky’s in the audience
0:28:09 she really appreciated your dumb and
0:28:10 dumber reference I’ll have you know
0:28:15 um but um yeah and then there will
0:28:16 probably be some sort of affiliate fee
0:28:18 and again that’s just being paid for
0:28:20 an outcome and kind of closing that
0:28:23 loop which will be probably a little
0:28:24 bit of a business model degradation
0:28:26 because the great why did Google
0:28:29 never start a marketplace because
0:28:32 people overvalue systematically their
0:28:34 ability once they’ve acquired a
0:28:37 customer through Google to keep it
0:28:39 as an organic customer so they
0:28:41 systematically overpay and they
0:28:42 continue doing that that’s why Google
0:28:45 never went to outcomes our marketplace
0:28:47 because advertising leads to the
0:28:49 advertisers systematically overpaying so
0:28:52 that inefficiency will be squeezed out
0:28:54 but yeah it will go to outcomes and you
0:28:56 know I think Elon tweeted today that you
0:28:57 know work would become optional you know
0:28:59 like instead of buying your vegetables
0:29:03 um you know at a at a supermarket you
0:29:05 can grow your own garden if you want now
0:29:07 who knows how long it takes us to get
0:29:11 there but I that doesn’t sound wildly
0:29:13 implausible to me for how powerful this
0:29:15 technology is and I was just struck
0:29:16 Karpathy you know whatever two days
0:29:18 ago you know is being painted as like a
0:29:21 skeptic for saying AGI is 10 years away
0:29:22 are you kidding
0:29:26 10 years yeah that’s wild yeah sign me up
0:29:28 while we’re short of timelines please
0:29:30 yeah well so no that’s awesome while
0:29:32 we’re on the topic of like very exciting
0:29:36 futuristic things robotics do you have you
0:29:38 on yeah very real and it’s gonna be
0:29:40 Tesla versus the Chinese in the same way
0:29:43 it’s Tesla versus the Chinese and in
0:29:46 cars electric cars yeah yeah I would just
0:29:48 say cars not electric cars yeah cars yeah
0:29:51 do you have a sense of timeline I mean
0:29:53 you can you can all watch the Optimus
0:29:57 videos every roboticist I know is
0:29:59 extremely impressed you know there’s
0:30:01 there’s a giant debate is it gonna be
0:30:02 humanoids are not humanoids I think that
0:30:04 debate is over because humanoids can
0:30:06 kind of learn you know from watching
0:30:08 YouTube videos and then it’s easier for
0:30:11 a human being you know to put on a suit
0:30:13 and show the robot how to do it I mean
0:30:15 it’s kind of crazy to watch the video of
0:30:18 all you know the 50 Optimus robots doing
0:30:21 50 different tasks and then it’s very
0:30:23 simple you know did you did you put the
0:30:25 glass in the dishwasher correctly or not
0:30:27 this is so fun Gavin I always love
0:30:28 chatting with you let’s give a hand to
0:30:30 Gavin thank you David thank you
0:30:35 all right next up we have a very
0:30:37 exciting panel on building out real
0:30:40 world infrastructure but first give us a
0:30:41 few minutes we got to do a quick
0:30:45 stage change here so thank you thanks
0:30:45 everybody
0:30:51 thanks for listening to this episode of
0:30:53 the a16z podcast if you like this
0:30:55 episode be sure to like comment subscribe
0:30:58 leave us a rating or review and share it
0:31:00 with your friends and family for more
0:31:03 episodes go to YouTube Apple Podcasts and
0:31:07 Spotify follow us on x a16z and subscribe to
0:31:10 our sub stack at a16z.substack.com thanks
0:31:12 again for listening and I’ll see you in the
0:31:15 next episode as a reminder the content here
0:31:17 is for informational purposes only should not
0:31:19 be taken as legal business tax or investment
0:31:22 advice or be used to evaluate any investment
0:31:24 or security and is not directed at any
0:31:27 investors or potential investors in any a16z
0:31:29 fund please note that a16z and its
0:31:31 affiliates may also maintain investments in
0:31:32 the companies discussed in this podcast
0:31:35 for more details including a link to our
0:31:39 investments please see a16z.com forward slash
0:31:40 disclosures
0:31:40 you
0:31:41 you
0:00:04 I do not believe we’re in an AI bubble today.
0:00:05 I was, depending on how you look at it,
0:00:08 the privilege and the misfortune of being a tech investor
0:00:09 during the year 2000 bubble,
0:00:11 which was really a telecom bubble.
0:00:13 And I think it’s really helpful to compare and contrast today
0:00:14 to the year 2000.
0:00:18 The year 2000 internet bubble or telecom bubble
0:00:20 was defined by something called dark fiber.
0:00:25 At the peak, 97% of the fiber that had been laid was dark.
0:00:26 Contrast that with today.
0:00:29 There are no dark GPUs.
0:00:32 Every major technology cycle raises the same question.
0:00:34 Is it real or are we in a bubble?
0:00:37 Today, you’ll hear a conversation from runtime
0:00:40 between Gavin Baker, managing director and CIO
0:00:42 of Atreides Management and David George,
0:00:44 general partner at A16Z
0:00:47 about how AI is reshaping the global economy.
0:00:49 From capital allocation and infrastructure spending
0:00:51 to business models and margins.
0:00:53 It’s a detailed data-driven look
0:00:55 at where we actually are in the AI cycle
0:00:57 and what’s likely to happen next.
0:00:58 Let’s get into it.
0:01:03 And that brings us to our opening fireside chat.
0:01:06 We’re going to start with a taboo question
0:01:07 right out of the gate.
0:01:08 Are you ready for it?
0:01:12 If AI is the biggest trend in the world right now,
0:01:14 where is the evidence for it?
0:01:16 Why is it only just beginning to show up in the economy?
0:01:19 And as Andre Carpathie asked,
0:01:20 are agents really just ghosts?
0:01:24 To kick this off and to help us answer this question,
0:01:27 please join us in welcoming Gavin Baker,
0:01:29 managing partner and CIO of Atreides.
0:01:31 Now, some of you may know Gavin
0:01:33 as that really thoughtful guy on Twitter.
0:01:36 Anytime some big piece of AI news comes out,
0:01:39 I know more than a few people who count on Gavin
0:01:41 to explain what the F is really going on.
0:01:44 So, a huge thank you to Gavin for being with us today.
0:01:46 Joining him is our very own David George,
0:01:48 general partner at A16Z.
0:02:08 Who knows what that music was from?
0:02:11 Glad they got our pump-up music right.
0:02:12 Yes.
0:02:16 Battlestar Galactica, the original 1977 one.
0:02:19 In case we have to all fight Cylons in a few years.
0:02:22 Yeah, it could segue into the topic, I guess.
0:02:24 So, thank you for being here.
0:02:25 I always love talking to you.
0:02:26 Same.
0:02:28 Really grateful to you for inviting me.
0:02:30 Grateful to your colleagues for having me here.
0:02:32 I really look forward to the next two days.
0:02:34 I think I’m going to learn a lot, so thank you.
0:02:34 Yeah.
0:02:35 Okay, all right.
0:02:38 So, the big topic is AI bubble,
0:02:40 kind of macro view of things.
0:02:43 So, maybe just to start with a couple stats
0:02:44 to set the stage,
0:02:45 and then I want to get your take on where we’re at.
0:02:47 So, we have about a trillion dollars
0:02:49 of data centers in the U.S.
0:02:52 The plan is to add three to four trillion dollars
0:02:53 in the next five years.
0:02:55 Over the past three years,
0:02:59 we have already built out in data center capacity
0:03:00 a larger amount of dollars
0:03:03 than the entire U.S. interstate highway system,
0:03:04 which took 40 years,
0:03:06 just in terms of dollars,
0:03:07 and that’s inflation adjusted.
0:03:09 Open AI alone, I think,
0:03:12 has more than a trillion dollars of deals
0:03:14 set up that they’ve committed to,
0:03:15 and we can talk about that.
0:03:17 But, at the same time,
0:03:19 so those are all like big numbers on infrastructure,
0:03:20 and they’re scary,
0:03:21 and they say, oh, bubble.
0:03:24 And Google released a stat recently
0:03:28 that they have seen a 150X increase
0:03:30 in the amount of tokens processed
0:03:31 in the last 17 months.
0:03:33 So, on the one hand,
0:03:35 you’ve got this crazy, scary-sounding build-out.
0:03:36 On the other hand,
0:03:39 you actually have a bunch of usage that’s happening.
0:03:41 So, are we in an AI bubble?
0:03:44 I do not believe we’re in an AI bubble today.
0:03:46 I had, depending on how you look at it,
0:03:47 the privilege and the misfortune
0:03:48 of being a tech investor
0:03:50 during the year 2000 bubble,
0:03:51 which was really a telecom bubble.
0:03:52 And I think it’s really helpful
0:03:54 to compare and contrast today
0:03:56 to the year 2000.
0:03:58 First, I think Cisco peaked
0:04:01 at 150 or 180 times trailing earnings.
0:04:02 NVIDIA’s at more like 40 times.
0:04:04 So, valuations are very different.
0:04:06 Most important, however,
0:04:09 is that the year 2000 internet bubble
0:04:10 or telecom bubble
0:04:12 was defined by something called dark fiber.
0:04:15 And if you’re a veteran of the year 2000,
0:04:17 you’ll know what that was.
0:04:19 But dark fiber was literally fiber
0:04:20 that was laid down in the ground
0:04:23 and not lit up.
0:04:24 Fiber is useless
0:04:26 unless you have the optics
0:04:27 and switches and routers
0:04:28 that you need on either side.
0:04:30 So, I vividly remember
0:04:31 companies like Level 3
0:04:32 or Global Crossing
0:04:33 or WorldCom
0:04:34 would come in and they say,
0:04:36 we laid 200,000 miles
0:04:37 of dark fiber this quarter.
0:04:38 This is so amazing.
0:04:40 The internet’s going to be so big.
0:04:42 We can’t wait to light these up.
0:04:44 At the peak of the bubble,
0:04:47 97% of the fiber
0:04:48 that had been laid
0:04:49 in America
0:04:50 was dark.
0:04:52 Contrast that with today.
0:04:54 There are no dark GPUs.
0:04:56 All you have to do
0:04:58 is read any technical paper
0:04:59 and that one of the biggest problems
0:05:00 in a trading run
0:05:02 is that GPUs are melting.
0:05:05 And there’s a very simple way
0:05:06 to kind of cut to the heart
0:05:06 of all of this.
0:05:08 It is a return on investing capital
0:05:10 of the biggest spenders
0:05:11 on GPUs
0:05:12 who are all public.
0:05:15 And those companies,
0:05:16 since they ramped up CapEx,
0:05:17 have seen,
0:05:19 call it a 10-point increase
0:05:20 in their ROICs.
0:05:21 So, thus far,
0:05:22 the ROI on all the spending
0:05:23 has been really positive.
0:05:25 It’s an interesting
0:05:26 and open debate
0:05:27 about whether or not
0:05:27 it will continue
0:05:28 to be positive.
0:05:30 With the quantum of spend
0:05:30 we’re going to have
0:05:31 on Blackwell,
0:05:32 I personally think it will.
0:05:34 But there’s no debate
0:05:35 that thus far
0:05:36 the ROI on AI
0:05:38 has been really positive
0:05:40 and valuation-wise
0:05:41 we’re just not in a bubble.
0:05:43 I couldn’t agree more.
0:05:44 The other thing
0:05:45 that I would say is
0:05:46 you can contrast
0:05:48 the actual adoption
0:05:49 and usage
0:05:50 of the technology
0:05:51 from then, right?
0:05:52 The internet was actually
0:05:53 really hard
0:05:54 because you had to build
0:05:55 a two-sided network.
0:05:56 Like, you had to build
0:05:56 websites
0:05:57 and then you had to get users
0:05:59 and it’s much more difficult.
0:06:01 In the case of the AI tool,
0:06:02 all you have to do
0:06:03 is kind of light them up
0:06:04 via API
0:06:05 or turn on your website
0:06:06 ChatGPT
0:06:07 and everybody has access
0:06:08 to them, right?
0:06:09 Built on top of cloud computing
0:06:10 on top of the internet
0:06:11 and you can get to
0:06:12 instant distribution
0:06:13 a billion people right away.
0:06:14 Absolutely.
0:06:16 So, the other thing
0:06:16 is the counterparty.
0:06:17 So, you mentioned this.
0:06:18 They happen to be
0:06:19 the best companies
0:06:21 in the history of the world, right?
0:06:22 I think collectively
0:06:23 the people who are
0:06:24 coming out of pocket
0:06:25 the writing checks
0:06:26 for this CapEx
0:06:28 I think they collectively
0:06:29 generate like
0:06:30 $300 billion
0:06:31 of free cash flow
0:06:31 a year.
0:06:32 Is that right?
0:06:34 Some directionally?
0:06:34 Round numbers.
0:06:35 Yeah, and they have
0:06:36 $500 billion
0:06:37 of cash on the balance sheet.
0:06:39 So, whenever people are like
0:06:39 oh my god, it’s a bubble
0:06:40 or is it going to pop?
0:06:41 I’m like
0:06:42 I think it’s kind of fine.
0:06:43 I mean, it costs like
0:06:45 $40 or $50 billion
0:06:46 to light up
0:06:47 one gigawatt.
0:06:48 Yeah, if you’re
0:06:49 on NVIDIA chips.
0:06:50 On NVIDIA chips.
0:06:51 Yeah.
0:06:51 Yeah.
0:06:52 So, you know,
0:06:53 there’s kind of like
0:06:54 an $800 billion buffer
0:06:56 growing $300 billion
0:06:56 every year.
0:06:58 Yeah, I mean
0:06:59 free cash flow
0:06:59 at some of them
0:07:01 has begun
0:07:03 to maybe, you know.
0:07:04 Well, this goes to your point
0:07:05 on return on invested capital.
0:07:06 Yes, we should see that next year.
0:07:07 It might creep down a little bit.
0:07:08 Yeah, a little bit of a mismatch
0:07:09 in the build out.
0:07:10 But, you know,
0:07:11 Larry Page apparently
0:07:11 internally said
0:07:13 I’m happy to go bankrupt
0:07:14 rather than lose this race.
0:07:15 And I think that is
0:07:16 the mentality for sure
0:07:17 at Google
0:07:18 and perhaps Meta.
0:07:20 It’s just seen as existential
0:07:22 and you have to win.
0:07:24 Okay, so lots
0:07:25 has been written
0:07:27 about these round-tripping deals.
0:07:29 So, because round-tripping
0:07:30 is a very scary concept
0:07:32 from the internet build-out.
0:07:33 That was a big problem.
0:07:34 What do you make of it here?
0:07:36 It is objectively happening.
0:07:37 Money is fungible.
0:07:38 So, NVIDIA,
0:07:39 if they sign a deal
0:07:40 with OpenAI,
0:07:40 they can say,
0:07:42 hey, you can’t use our money
0:07:43 to buy our chips,
0:07:44 but money is fungible.
0:07:45 But it’s happening
0:07:46 at a very small scale.
0:07:48 Yes.
0:07:49 Yeah.
0:07:50 And I think…
0:07:50 I know this is like
0:07:51 a crypto or blockchain.
0:07:52 Yeah, exactly.
0:07:54 And I think
0:07:55 what is driving this
0:07:57 isn’t the need
0:07:59 to finance GPU
0:08:01 or data center purchases,
0:08:02 but it’s actually
0:08:03 competitive dynamics.
0:08:05 So, NVIDIA’s
0:08:05 biggest competitor,
0:08:06 it’s not AMD,
0:08:08 it’s not Broadcom,
0:08:10 it’s certainly not Marvell,
0:08:11 it’s not Intel,
0:08:13 it’s Google.
0:08:14 And more specifically,
0:08:16 it is Google
0:08:17 because Google
0:08:19 owns the TPU chip.
0:08:20 And this is by far,
0:08:22 maybe perhaps today,
0:08:24 the only alternative
0:08:25 to NVIDIA for training
0:08:27 and maybe the best
0:08:28 inference alternative.
0:08:30 And Google’s
0:08:31 a problematic competitor
0:08:32 because they also own
0:08:33 a company called DeepMind
0:08:34 and they have
0:08:36 a product called Gemini.
0:08:37 And I think you could argue
0:08:37 that they’re
0:08:39 the leading AI company today.
0:08:40 I think they’ve taken
0:08:41 15 or 20 points
0:08:42 of traffic share
0:08:43 in the last two
0:08:44 or three months
0:08:45 and that’s just traffic
0:08:45 to Gemini.
0:08:47 it does not include
0:08:48 search overviews.
0:08:49 I suspect
0:08:50 on a actual
0:08:51 traffic basis,
0:08:52 Google is bigger
0:08:53 than OpenAI,
0:08:54 Anthropic,
0:08:55 anyone today.
0:08:58 And that business
0:08:58 is going to run
0:08:59 on TPUs.
0:09:00 And then we have
0:09:01 three other labs
0:09:02 that are relevant today.
0:09:02 There’s Anthropic
0:09:04 and that’s an Amazon
0:09:05 and Google captive.
0:09:05 You know,
0:09:06 Anthropic is really
0:09:07 going to run
0:09:07 on TPUs
0:09:08 and Traniums.
0:09:09 And so you’re left
0:09:10 with XAI
0:09:11 and OpenAI
0:09:12 at the forefront.
0:09:13 And if Google
0:09:15 is going to
0:09:16 a lab like Anthropic
0:09:16 and saying,
0:09:17 I’m going to help
0:09:18 you fundraise
0:09:20 and give you chips
0:09:21 for competitive reasons,
0:09:22 it’s very hard
0:09:23 for NVIDIA
0:09:24 not to respond.
0:09:25 And as Jensen said,
0:09:25 he thinks it’s going
0:09:26 to be a good investment.
0:09:27 So I think
0:09:28 the round-tripping
0:09:29 concerns
0:09:32 are pretty overblown.
0:09:33 I mean,
0:09:34 what NVIDIA really needs
0:09:35 is they need
0:09:36 Meta to get
0:09:36 their act together
0:09:38 or another
0:09:38 American open-source
0:09:39 player to emerge
0:09:41 or maybe some sort
0:09:42 of detente
0:09:42 with China
0:09:43 and AI.
0:09:45 When people ask me
0:09:46 about NVIDIA
0:09:46 and all the moves
0:09:47 and the round-tripping,
0:09:48 my reaction is
0:09:49 everything they’ve done
0:09:50 is completely rational.
0:09:52 100% rational.
0:09:53 Yeah, long-term.
0:09:54 Yeah, long-term.
0:09:54 Sure, some of the things
0:09:55 they do may not have
0:09:56 as high of a return
0:09:57 on capital as other things,
0:09:58 but strategically,
0:09:59 I think they’re all
0:10:00 kind of the right moves.
0:10:01 Jensen’s one of the
0:10:01 two best CEOs
0:10:02 along with Elon
0:10:04 I have ever known,
0:10:05 and I think he’s
0:10:06 playing a strong hand
0:10:07 really well.
0:10:07 Yeah.
0:10:08 All right,
0:10:09 so you started getting
0:10:10 into the model companies.
0:10:11 Let’s just talk about
0:10:13 the model.
0:10:14 So we can come back
0:10:15 to chips and memory
0:10:16 and networking
0:10:16 because I want to get
0:10:17 your take on that,
0:10:18 but since we’re
0:10:19 on the model side,
0:10:20 what do you think
0:10:21 happens with market structure?
0:10:22 Who wins where?
0:10:24 Who are you most
0:10:24 optimistic about?
0:10:25 Where do you have concerns?
0:10:28 So I think humility is an important
0:10:30 virtue for an investor.
0:10:32 And I’m just,
0:10:33 if we’re going to make
0:10:35 an analogy and say
0:10:37 that ChatGPT is to AI,
0:10:39 has Netscape Navigator
0:10:40 was to the internet.
0:10:43 At this point in the internet boom,
0:10:45 Google had not been founded.
0:10:48 Mark Zuckerberg was in middle school.
0:10:50 Travis Kalanick was in kindergarten.
0:10:53 So it’s just very early.
0:10:54 So I think it’s important
0:10:55 to be humble
0:10:56 about making
0:10:57 high confidence predictions
0:10:59 at the application layer.
0:11:00 It’s one reason I think
0:11:01 the infrastructure layer
0:11:02 is often maybe
0:11:03 a safe place to be
0:11:04 at the beginning
0:11:05 of one of these
0:11:06 new technology waves.
0:11:07 Well, actually talk about
0:11:08 the role they play
0:11:09 at the infrastructure layer
0:11:10 because there’s a piece of them
0:11:11 that obviously they serve
0:11:12 as an infrastructure layer
0:11:13 powering other
0:11:14 application providers
0:11:15 and then they also have
0:11:16 their own applications.
0:11:16 So I think
0:11:18 I would draw a distinction.
0:11:18 Yeah.
0:11:19 I mean, that’s most true
0:11:20 of Google,
0:11:21 but I think it’s hard
0:11:22 to have high conviction
0:11:24 other than to observe
0:11:25 the internet was
0:11:27 a very disruptive innovation.
0:11:28 I think there’s
0:11:29 reasonable arguments
0:11:30 that AI could be
0:11:32 a sustaining innovation
0:11:34 because the raw ingredients
0:11:35 of kind of data,
0:11:38 the capital to buy compute
0:11:39 and distribution,
0:11:40 which is what you need,
0:11:42 all of today’s
0:11:43 biggest tech companies
0:11:44 have all of those in spades.
0:11:46 So as long as they execute well,
0:11:48 hire good people
0:11:50 and have a sound strategy,
0:11:52 like I think you could see
0:11:53 it be a sustaining innovation
0:11:54 for a lot of members
0:11:55 of the Mag 7.
0:11:56 On the other hand,
0:11:57 I do think it’s existential
0:11:59 and if you don’t execute,
0:12:00 you know,
0:12:03 IBM might be a good fate.
0:12:04 Yeah.
0:12:04 Yeah.
0:12:05 Yeah.
0:12:06 That’s tough.
0:12:07 Yeah.
0:12:08 Data distribution,
0:12:09 compute,
0:12:10 dollars,
0:12:11 talent.
0:12:12 Yeah.
0:12:14 They have every right to win.
0:12:14 Yeah.
0:12:15 They have every right to win.
0:12:16 And it seems now
0:12:17 more than before
0:12:18 they’re taking it quite seriously.
0:12:20 Yeah.
0:12:21 Maybe Google in particular,
0:12:23 but obviously Meta’s making
0:12:24 the dramatic moves
0:12:24 they’re making too.
0:12:25 No,
0:12:25 to me,
0:12:27 ChatGPT was Pearl Harbor
0:12:28 for Google
0:12:29 and we’re going to see
0:12:30 how they responded
0:12:31 and they’re slowly
0:12:32 starting to respond.
0:12:32 Yeah.
0:12:34 And then what do you think,
0:12:35 what’s your forecast
0:12:37 for that sort of
0:12:38 the platform piece
0:12:38 of their business,
0:12:39 the infrastructure piece?
0:12:41 What do you think,
0:12:42 how do you think it shakes out
0:12:43 in terms of like
0:12:44 business model,
0:12:45 market structure?
0:12:46 So do you think
0:12:47 they end up
0:12:48 as high margin businesses
0:12:50 like the clouds
0:12:52 or like aircraft manufacturers
0:12:53 or do you think
0:12:54 they end up
0:12:55 very competitive
0:12:56 in low margin businesses
0:12:57 like airlines?
0:12:59 I don’t think
0:13:00 they’ll be airlines,
0:13:01 but you can,
0:13:02 anybody can just look
0:13:03 at the P&L,
0:13:04 you know,
0:13:05 of a SaaS company
0:13:08 circa 2021 and 2022
0:13:08 and you see,
0:13:09 you know,
0:13:11 80, 90% gross margins
0:13:14 and the nature of AI
0:13:15 because of scaling laws,
0:13:16 Richard Sutton’s
0:13:17 the better,
0:13:17 the better listen.
0:13:20 They’re just more
0:13:21 compute intensive.
0:13:23 So their gross margins
0:13:24 are structurally
0:13:25 going to be lower
0:13:26 but that doesn’t mean
0:13:27 they can’t be
0:13:27 great businesses.
0:13:28 I just,
0:13:29 I think it’s going
0:13:30 to be a long time
0:13:31 before we see
0:13:33 a truly kind of,
0:13:34 you know,
0:13:35 an AI lab,
0:13:36 a frontier lab
0:13:37 with gross margins
0:13:39 anywhere near SaaS
0:13:41 or internet era margins.
0:13:41 Now,
0:13:42 their OPEX
0:13:43 can be a lot lower
0:13:45 and,
0:13:45 you know,
0:13:46 maybe that’s how
0:13:47 you square it
0:13:48 but just the gross margins
0:13:49 are fundamentally different
0:13:51 and until scaling laws change
0:13:52 and the importance
0:13:53 of test time compute,
0:13:54 things like that,
0:13:55 change,
0:13:56 which I don’t see
0:13:56 happening,
0:13:58 they are going to be
0:13:58 lower margin.
0:13:59 Yeah.
0:14:00 Okay,
0:14:01 so let’s talk about
0:14:02 application layer.
0:14:04 So you just,
0:14:05 you just kind of got into
0:14:05 it a little bit
0:14:06 with the SaaS businesses
0:14:08 and I don’t know
0:14:08 if you’ve waded
0:14:09 into this fight
0:14:10 on Twitter
0:14:11 but it’s sort of,
0:14:11 you know,
0:14:12 the like,
0:14:12 you know,
0:14:13 every few months
0:14:14 it comes up
0:14:14 and it’s like,
0:14:16 SaaS is terrible
0:14:16 and it’s dead
0:14:16 and,
0:14:17 you know,
0:14:17 it’s all going
0:14:18 to go away
0:14:18 and then,
0:14:19 you know,
0:14:21 with Andre’s
0:14:23 Duarkesh interview
0:14:23 he just did,
0:14:24 it’s,
0:14:24 you know,
0:14:25 like the market’s reacting
0:14:26 positively to it
0:14:27 and it’s like
0:14:29 a whipsaw reaction.
0:14:29 So what do you think
0:14:31 happens with SaaS
0:14:31 and software?
0:14:32 You know,
0:14:32 I think I,
0:14:33 you know,
0:14:34 first said probably
0:14:34 in early 24
0:14:35 that I thought
0:14:37 all of application SaaS
0:14:37 might be a zero
0:14:39 different than
0:14:41 infrastructure SaaS.
0:14:43 I would say
0:14:44 I have a more nuanced
0:14:45 view now
0:14:47 and I think
0:14:48 there could be
0:14:49 some really big
0:14:50 application SaaS winners
0:14:52 especially if you serve
0:14:53 like a more fragmented
0:14:54 SMB customer base.
0:14:56 You know,
0:14:57 Google is making it
0:14:57 really easy
0:14:58 if you’re a customer
0:14:58 of theirs
0:15:00 to use your data
0:15:01 and essentially
0:15:02 make any SaaS app
0:15:02 you want
0:15:03 and then your data
0:15:04 isn’t shared
0:15:04 with anyone else.
0:15:07 But the critical mistake
0:15:08 that I think
0:15:09 a lot of retailers
0:15:09 made
0:15:12 in dealing with Amazon
0:15:13 is they looked
0:15:14 at Amazon’s margins
0:15:15 and they said
0:15:16 we don’t want
0:15:17 to be in that business.
0:15:19 And
0:15:20 that was obviously
0:15:20 a terrible mistake
0:15:21 and here we are
0:15:23 25 years later
0:15:23 and, you know,
0:15:24 Amazon has really
0:15:26 healthy retail margins.
0:15:27 And
0:15:28 I worry
0:15:29 that application
0:15:30 SaaS companies
0:15:31 are trying
0:15:32 to preserve
0:15:34 their existing
0:15:34 gross margin
0:15:35 structures
0:15:37 because they believe
0:15:38 that if their
0:15:38 gross margins
0:15:39 go down,
0:15:40 their stocks
0:15:41 will go down.
0:15:42 It is
0:15:43 definitionally impossible
0:15:44 given what we just
0:15:45 discussed
0:15:46 to succeed
0:15:46 in AI
0:15:47 without gross
0:15:47 margin pressure
0:15:49 and I do not
0:15:49 know
0:15:50 why they have
0:15:51 concerns
0:15:52 because we have
0:15:52 an existence
0:15:53 proof
0:15:54 that a software
0:15:55 company can
0:15:56 deal well
0:15:56 with declining
0:15:57 margins
0:15:58 and Microsoft
0:15:59 and Adobe
0:15:59 to the whole
0:16:00 AI thing
0:16:00 came along.
0:16:01 It used to be
0:16:02 that companies
0:16:02 were scared
0:16:03 to go from
0:16:03 on-premise
0:16:04 to the cloud
0:16:05 because margins
0:16:05 were lower.
0:16:06 Cloud margins
0:16:08 are lower.
0:16:08 They’re still good.
0:16:10 And Microsoft,
0:16:11 they transitioned
0:16:13 from on-premise
0:16:14 perpetual licenses
0:16:15 with maintenance
0:16:16 to a cloud model
0:16:18 and it was a
0:16:18 pretty good stock
0:16:19 for 10 years.
0:16:20 So I don’t,
0:16:21 if you’re an
0:16:22 application SaaS
0:16:22 company,
0:16:23 like what I would
0:16:24 just say is
0:16:25 don’t be scared
0:16:26 and look at
0:16:27 declining gross
0:16:27 margins
0:16:28 kind of has a
0:16:29 mark of success
0:16:30 rather than
0:16:31 a badge of shame
0:16:32 or something
0:16:32 to be feared.
0:16:33 It’s actually so
0:16:34 funny you say
0:16:34 that because
0:16:35 whenever we have
0:16:35 these discussions
0:16:36 about companies,
0:16:38 basically every
0:16:38 company that comes
0:16:39 to present to us
0:16:39 is like we’re
0:16:40 an AI company
0:16:42 and we always
0:16:42 look at the gross
0:16:43 margins and it’s
0:16:44 become like a
0:16:45 badge of honor
0:16:45 for them to
0:16:46 actually have
0:16:46 low gross
0:16:47 margins
0:16:47 because we’re
0:16:47 like oh my
0:16:47 God,
0:16:47 people are
0:16:48 actually using
0:16:49 your AI stuff.
0:16:50 But if you
0:16:50 show up and
0:16:50 you’re like I’m
0:16:51 an AI company
0:16:51 and it’s like I
0:16:53 got 82% gross
0:16:53 margins,
0:16:54 you’re like I
0:16:54 don’t think
0:16:55 anybody’s really
0:16:55 using it.
0:16:58 It’s interesting.
0:16:58 If you’re one
0:16:59 of these public
0:16:59 companies,
0:16:59 would you rather
0:17:00 have like 10
0:17:01 bucks of revenue
0:17:01 with 90% gross
0:17:02 margins or 50
0:17:03 bucks of revenue
0:17:04 with 60% gross
0:17:04 margins?
0:17:05 Not hard.
0:17:05 It’s not that
0:17:06 complicated.
0:17:07 It’s hard to do
0:17:07 in the public
0:17:08 market.
0:17:08 It’s hard to do
0:17:09 in publics,
0:17:09 but if you
0:17:10 communicate it,
0:17:10 you draw parallels
0:17:11 to the cloud
0:17:11 transition.
0:17:13 I mean I’m
0:17:14 an investor and
0:17:14 I would be
0:17:14 excited about
0:17:16 it and I
0:17:16 don’t think I’m
0:17:17 alone in the
0:17:17 world.
0:17:18 And then the
0:17:19 big advantage
0:17:20 these legacy
0:17:21 application SaaS
0:17:21 companies have
0:17:23 is they do
0:17:24 have these
0:17:24 really profitable
0:17:26 existing businesses
0:17:27 and so you
0:17:28 can run
0:17:28 your new
0:17:29 AI products
0:17:30 at breakeven
0:17:34 and catch up
0:17:34 to the leaders,
0:17:35 et cetera,
0:17:35 et cetera.
0:17:36 And I’m just
0:17:37 surprised more
0:17:37 people have not
0:17:38 done that.
0:17:39 Like why are
0:17:39 none of the
0:17:40 public coding
0:17:41 companies even
0:17:42 trying to compete
0:17:42 with Cursor?
0:17:44 And the reality
0:17:44 is Cursor now
0:17:45 they have a
0:17:46 trillion tokens
0:17:49 and there will
0:17:49 be a point
0:17:50 where they have
0:17:51 enough coding
0:17:52 tokens that it’s
0:17:53 tough to catch
0:17:53 them.
0:17:54 But I think
0:17:55 today if you’re
0:17:55 a public coding
0:17:56 company and
0:17:57 you said I’m
0:17:57 going to lean
0:17:58 in, I’m going
0:17:58 to run it
0:17:59 breakeven, I have
0:17:59 an existing
0:18:00 business, I’m
0:18:01 going to attach
0:18:01 it to
0:18:02 everything, hey
0:18:03 you have a
0:18:04 chance and
0:18:05 you know the
0:18:06 prize is clearly
0:18:07 really big.
0:18:07 I see Martin
0:18:08 is skeptical.
0:18:08 Martin
0:18:09 said you have
0:18:10 a chance.
0:18:10 I said a chance.
0:18:11 I said a chance.
0:18:12 It’s like a
0:18:12 dumb and dumber
0:18:13 you’re telling
0:18:13 me there’s a
0:18:14 chance not like a
0:18:14 real chance.
0:18:15 You’re telling
0:18:16 me there’s a
0:18:16 chance.
0:18:20 Yeah exactly.
0:18:20 I totally agree.
0:18:21 Yeah we actually
0:18:22 saw I mean you
0:18:22 know we see it
0:18:24 you know we may
0:18:25 if we you know
0:18:26 Figma for
0:18:28 example like when
0:18:28 they went out they
0:18:29 are extremely high
0:18:30 gross margin and
0:18:30 they’re like hey
0:18:31 we’re going to you
0:18:32 know pretty
0:18:33 aggressively distribute
0:18:34 our AI tools and
0:18:34 our gross margins
0:18:34 are going to go
0:18:36 down and you know
0:18:36 investors asked a
0:18:37 few clarifying
0:18:38 questions and then
0:18:38 they were like oh
0:18:39 that actually would
0:18:39 be a good thing.
0:18:40 And so I’m
0:18:40 surprised more
0:18:41 people in the
0:18:41 public markets
0:18:42 aren’t doing it.
0:18:43 worked out okay for
0:18:43 them.
0:18:44 It’s working out
0:18:45 well.
0:18:46 Long game to
0:18:46 play.
0:18:47 What about on the
0:18:48 consumer side of the
0:18:49 application layer?
0:18:51 So obviously Google
0:18:52 was the portal to the
0:18:53 internet is kind of
0:18:53 still is the portal to
0:18:54 the internet and the
0:18:55 whole business model
0:18:57 was predicated upon
0:18:59 taking some intent and
0:19:00 directing you to
0:19:01 someone else’s website
0:19:03 where they would do
0:19:03 stuff with you.
0:19:06 It’s kind of not
0:19:06 going to be that way.
0:19:07 It already is not that
0:19:09 way with AI.
0:19:10 although I tried the
0:19:11 browser today and I
0:19:12 tried to do some
0:19:13 pretty basic shopping
0:19:14 stuff and it’s you
0:19:15 know still still some
0:19:16 work to do but I
0:19:17 think it will get
0:19:17 there.
0:19:18 So what do you
0:19:19 actually think happens
0:19:21 with the sort of
0:19:22 market structure of
0:19:22 the consumer internet
0:19:23 companies?
0:19:24 Do they get
0:19:26 subsumed into a
0:19:28 component of a
0:19:29 chatbot interface or
0:19:30 do you think it’s
0:19:30 something else?
0:19:33 So one humility
0:19:34 hard to say.
0:19:36 Two I would just
0:19:39 say I think the AI
0:19:40 companies that have
0:19:41 launched these AI
0:19:42 browsers may come to
0:19:43 regret it because
0:19:43 there’s something
0:19:45 called Chrome that
0:19:46 has whatever it is
0:19:49 five billion users and
0:19:51 if you’re Google you
0:19:52 know you can just go
0:19:52 look and what
0:19:52 happened with Google
0:19:54 Buzz they are very
0:19:55 cautious you know
0:19:56 there’s you know
0:19:57 they’re currently in
0:19:57 litigation with the
0:20:01 government and they
0:20:02 could easily do this
0:20:04 and probably do it
0:20:05 even better but they
0:20:06 didn’t want to be
0:20:07 first.
0:20:09 so now you have
0:20:10 two AI native
0:20:11 companies with their
0:20:12 own browsers let
0:20:13 them run for three
0:20:15 to six months get a
0:20:16 little head start and
0:20:18 then wow here we are
0:20:19 we had to do this and
0:20:21 I don’t know how that’s
0:20:23 going to work maybe
0:20:24 for the companies other
0:20:25 than Google who don’t
0:20:26 own Chrome.
0:20:30 I guess data and
0:20:30 distribution is pretty
0:20:31 powerful in that.
0:20:32 Yeah hindsight’s 20-20
0:20:35 and the one thing I
0:20:36 would say is I do
0:20:37 think it’s tough to
0:20:37 bet against the
0:20:39 companies with large
0:20:40 existing user bases
0:20:44 today and I also
0:20:45 think reasoning has
0:20:46 fundamentally changed
0:20:47 the economics of these
0:20:49 frontier models you
0:20:51 know pre-reasoning I
0:20:52 often said if you are a
0:20:54 frontier model without
0:20:56 access to unique
0:20:59 valuable data and
0:21:00 internet scale
0:21:01 distribution you’re the
0:21:02 fastest depreciating
0:21:03 asset in history.
0:21:04 I think reasoning really
0:21:06 changed that because the
0:21:07 way RL works during
0:21:08 post-training having a
0:21:10 big user base now kind
0:21:12 of unlocks that flywheel
0:21:13 that was at the center of
0:21:14 every great consumer
0:21:16 internet company where
0:21:17 you have a good product
0:21:19 you get a lot of users
0:21:20 the users make the
0:21:22 algorithm better the
0:21:23 algorithm makes the
0:21:24 product better and it
0:21:26 just spins and that
0:21:28 it’s not quite spinning
0:21:30 yet in AI but you can
0:21:31 squint and see it and
0:21:32 so I think that
0:21:33 fundamentally changes
0:21:34 the economics for
0:21:38 anthropic for XAI for
0:21:41 open AI but I mean
0:21:43 Mark Zuckerberg’s trying
0:21:45 hard yeah we’ll see
0:21:47 yeah yeah yeah a lot of
0:21:48 smart people in there
0:21:50 now yeah for sure I
0:21:51 think the worry is and
0:21:52 I think this is another
0:21:54 interesting thing is if
0:21:56 you don’t like in a
0:21:57 strange way the Chinese
0:21:58 open source model
0:22:00 ecosystem is a godsend
0:22:02 to any American company
0:22:03 that’s trying to catch
0:22:04 those four leading labs
0:22:06 because the problem is
0:22:07 if you don’t have
0:22:10 Gemini 2.5 Pro or a
0:22:11 later checkpoint of it
0:22:13 or a later checkpoint of
0:22:14 Grok that we don’t see
0:22:15 or a later GPT
0:22:18 checkpoint training the
0:22:19 next model you’re at a
0:22:20 big disadvantage oh by
0:22:21 the way one thing I
0:22:22 just want to say that
0:22:23 drives me crazy is all
0:22:24 these people who say
0:22:25 that GPT-5 is the end
0:22:28 of scaling loss GPT-5 is
0:22:30 a smaller model it was
0:22:31 not designed to be
0:22:32 better it was designed to
0:22:33 be more economical for
0:22:35 open AI and Microsoft to
0:22:38 run any reference to
0:22:40 GPT-5 and scaling laws is
0:22:44 crazy yeah sorry rant rant
0:22:45 over we get the pedestal
0:22:46 up here if you want
0:22:47 yeah exactly shaking your
0:22:48 hand yeah that’d be good
0:22:50 yeah that’d be good uh do
0:22:51 you want to talk about
0:22:53 chips sure so okay I know
0:22:56 you love NVIDIA talk about
0:22:57 you know your view of
0:23:01 NVIDIA AMD TPUs ASICs and
0:23:02 how do you think sort of
0:23:04 market structure shakes out
0:23:06 their you know competitive
0:23:07 advantage that the various
0:23:11 players have yeah I think
0:23:12 it goes I think it is
0:23:15 really um it’s a fight
0:23:18 between NVIDIA and um the
0:23:20 Google TPU and that’s
0:23:20 something that I don’t
0:23:21 think is broadly
0:23:23 appreciated is the extent
0:23:25 which Broadcom and AMD are
0:23:26 effectively going to market
0:23:28 together NVIDIA is no
0:23:29 longer just a semiconductor
0:23:30 company as I’m sure you’ll
0:23:31 hear from Jensen tomorrow
0:23:33 you know not it was a
0:23:34 semiconductor company then
0:23:35 a software company with
0:23:36 CUDA now systems company
0:23:37 with these rack level
0:23:39 solutions and now
0:23:39 arguably you know a
0:23:40 data setter level
0:23:42 uh company with the you
0:23:43 know level of architecting
0:23:44 they’re doing with scale
0:23:46 up scale across and um
0:23:48 scale out scale across
0:23:51 networking um so the
0:23:53 networking the fabric the
0:23:54 software it’s all
0:23:55 important and what
0:23:56 Broadcom is saying to
0:23:57 companies like Meta is
0:23:59 hey we will build you a
0:24:01 fabric that can
0:24:02 theoretically compete with
0:24:04 NVIDIA’s fabric which is
0:24:06 a mixture of NVLink and
0:24:07 either InfiniBand or
0:24:08 Ethernet um it will build
0:24:09 it on the Ethernet it’s
0:24:10 going to be an open
0:24:12 standard and hey we’ll
0:24:13 we’ll make you your
0:24:15 version of of TPU which
0:24:15 by the way took Google
0:24:17 three generations to get
0:24:18 working and you know what
0:24:19 if your ASIC isn’t good
0:24:21 you can just plug AMD
0:24:24 right in um but I
0:24:26 personally believe most of
0:24:26 those ASICs are going to
0:24:29 fail um particularly if
0:24:30 it’s in the fullness of
0:24:31 time like over a period of
0:24:32 time or in the fullness of
0:24:35 time in the next three
0:24:36 years I think you’ll see a
0:24:38 bunch of high profile um
0:24:39 ASIC programs canceled
0:24:41 especially if Google um
0:24:42 starts selling TPUs
0:24:44 externally which has been
0:24:46 all over X and then you
0:24:47 know they you know who
0:24:48 knows exactly how that
0:24:49 would work because if
0:24:51 you’re an anthropic it was
0:24:52 just rumored anthropic wants
0:24:52 to buy tens of billions of
0:24:54 TPUs if you’re anthropic
0:24:55 maybe you don’t want Google
0:24:56 seeing your secret sauce but
0:24:58 there’s ways around that so
0:25:00 I think this is really a
0:25:01 battle between Google and
0:25:03 it’s TPU enabled by
0:25:04 Broadcom for now and
0:25:06 Google can take the TPU
0:25:07 away from Broadcom whenever
0:25:09 they want yeah now they
0:25:10 can’t do the Ethernet
0:25:11 networking that Broadcom is
0:25:13 is doing uh but they
0:25:15 control the TPU um so it’s
0:25:16 really Google and the TPU
0:25:20 verse um NVIDIA you know
0:25:21 with with you know Amazon
0:25:23 like that’s a very talented
0:25:25 team arguing the most
0:25:26 talented silicon team at any
0:25:27 hyperscaler the Annapurna
0:25:29 team like I think the
0:25:30 Tranium 3 will probably be a
0:25:31 much better chip than the
0:25:33 Tranium 2 it took Google
0:25:34 three generations to get the
0:25:36 TPU right um and then AMD
0:25:38 will you know will always be
0:25:39 kind of the second source and
0:25:40 you need a second source
0:25:43 all right exciting uh what do
0:25:45 you think happens okay so I
0:25:46 want to go back um to
0:25:48 business models so one of the
0:25:49 big things that is widely
0:25:51 discussed is like you know
0:25:53 source of disruption and most
0:25:55 of the CEOs in this room are
0:25:56 CEOs of startups who are
0:25:58 trying to go beat some
0:26:00 incumbent or find you know
0:26:01 some new market
0:26:03 opportunity and the most
0:26:04 ripe opportunities tend to
0:26:05 come when you have a big
0:26:07 platform shift that is also
0:26:08 accompanied with a business
0:26:11 model shift um and so there
0:26:12 are a couple of areas where I
0:26:14 can see it I feel like in an
0:26:16 obvious way so you know we’re
0:26:17 investors in Decagon customer
0:26:19 support like you can pretty
0:26:21 easily see a business model that
0:26:23 is priced on the resolution of
0:26:24 a task because it’s so
0:26:27 measurable um you can see you
0:26:29 know like encoding like a lot of
0:26:30 the business model has now
0:26:32 shifted to consumption and you
0:26:33 know obviously especially for
0:26:34 developer facing things like
0:26:37 that’s comfortable um and pretty
0:26:39 well known what about the rest
0:26:40 of the industry because I feel
0:26:42 like there’s sort of this hand
0:26:44 wave thing that is going on
0:26:45 which is like we’re gonna go get
0:26:47 all of services but it’s like
0:26:48 okay so how do you actually go do
0:26:50 that it’s gonna be pretty hard so
0:26:52 do you have any prediction on how
0:26:54 that plays out well I think what
0:26:56 you’re seeing in customer service
0:26:57 which is kind of like an easy
0:27:00 first example um we have a lot of
0:27:02 textual data the LLMs are good at
0:27:04 text you can kind of you know
0:27:07 probably really easily run some
0:27:08 RL to make sure that they you know
0:27:11 get a good verified reward you
0:27:12 know verified reward being happy
0:27:14 customer first call resolution or
0:27:17 whatever it is um and but I do
0:27:19 think you will see that played out
0:27:20 like humans were fundamentally
0:27:22 played for out paid paid based on
0:27:25 outcomes and a lot of AI will be
0:27:27 augmenting humans but probably
0:27:29 also replacing some humans and
0:27:30 that will involve being paid um
0:27:31 paid for outcomes you know going
0:27:33 back to the consumer business
0:27:34 model you know everybody’s
0:27:35 talking about affiliate fees and
0:27:37 for sure I’m gonna have you know
0:27:40 my own AI it will be a version of
0:27:42 Grok um because we’re both ex AI
0:27:43 shareholders it will be a version
0:27:45 of Grok that knows me and it likes
0:27:47 me um and you know when I when I
0:27:49 want to you know the next time I
0:27:50 want to go on vacation it will
0:27:52 know the hotels that I like to go
0:27:55 to and it’ll say hey three hotels I
0:27:57 have Gavin you know I have Gavin
0:27:59 coming who’s got the best price of
0:28:02 the best room um it’s gonna
0:28:03 massively upgrade the gifts that you
0:28:04 give to Becky
0:28:08 yes Becky Becky’s in the audience
0:28:09 she really appreciated your dumb and
0:28:10 dumber reference I’ll have you know
0:28:15 um but um yeah and then there will
0:28:16 probably be some sort of affiliate fee
0:28:18 and again that’s just being paid for
0:28:20 an outcome and kind of closing that
0:28:23 loop which will be probably a little
0:28:24 bit of a business model degradation
0:28:26 because the great why did Google
0:28:29 never start a marketplace because
0:28:32 people overvalue systematically their
0:28:34 ability once they’ve acquired a
0:28:37 customer through Google to keep it
0:28:39 as an organic customer so they
0:28:41 systematically overpay and they
0:28:42 continue doing that that’s why Google
0:28:45 never went to outcomes our marketplace
0:28:47 because advertising leads to the
0:28:49 advertisers systematically overpaying so
0:28:52 that inefficiency will be squeezed out
0:28:54 but yeah it will go to outcomes and you
0:28:56 know I think Elon tweeted today that you
0:28:57 know work would become optional you know
0:28:59 like instead of buying your vegetables
0:29:03 um you know at a at a supermarket you
0:29:05 can grow your own garden if you want now
0:29:07 who knows how long it takes us to get
0:29:11 there but I that doesn’t sound wildly
0:29:13 implausible to me for how powerful this
0:29:15 technology is and I was just struck
0:29:16 Karpathy you know whatever two days
0:29:18 ago you know is being painted as like a
0:29:21 skeptic for saying AGI is 10 years away
0:29:22 are you kidding
0:29:26 10 years yeah that’s wild yeah sign me up
0:29:28 while we’re short of timelines please
0:29:30 yeah well so no that’s awesome while
0:29:32 we’re on the topic of like very exciting
0:29:36 futuristic things robotics do you have you
0:29:38 on yeah very real and it’s gonna be
0:29:40 Tesla versus the Chinese in the same way
0:29:43 it’s Tesla versus the Chinese and in
0:29:46 cars electric cars yeah yeah I would just
0:29:48 say cars not electric cars yeah cars yeah
0:29:51 do you have a sense of timeline I mean
0:29:53 you can you can all watch the Optimus
0:29:57 videos every roboticist I know is
0:29:59 extremely impressed you know there’s
0:30:01 there’s a giant debate is it gonna be
0:30:02 humanoids are not humanoids I think that
0:30:04 debate is over because humanoids can
0:30:06 kind of learn you know from watching
0:30:08 YouTube videos and then it’s easier for
0:30:11 a human being you know to put on a suit
0:30:13 and show the robot how to do it I mean
0:30:15 it’s kind of crazy to watch the video of
0:30:18 all you know the 50 Optimus robots doing
0:30:21 50 different tasks and then it’s very
0:30:23 simple you know did you did you put the
0:30:25 glass in the dishwasher correctly or not
0:30:27 this is so fun Gavin I always love
0:30:28 chatting with you let’s give a hand to
0:30:30 Gavin thank you David thank you
0:30:35 all right next up we have a very
0:30:37 exciting panel on building out real
0:30:40 world infrastructure but first give us a
0:30:41 few minutes we got to do a quick
0:30:45 stage change here so thank you thanks
0:30:45 everybody
0:30:51 thanks for listening to this episode of
0:30:53 the a16z podcast if you like this
0:30:55 episode be sure to like comment subscribe
0:30:58 leave us a rating or review and share it
0:31:00 with your friends and family for more
0:31:03 episodes go to YouTube Apple Podcasts and
0:31:07 Spotify follow us on x a16z and subscribe to
0:31:10 our sub stack at a16z.substack.com thanks
0:31:12 again for listening and I’ll see you in the
0:31:15 next episode as a reminder the content here
0:31:17 is for informational purposes only should not
0:31:19 be taken as legal business tax or investment
0:31:22 advice or be used to evaluate any investment
0:31:24 or security and is not directed at any
0:31:27 investors or potential investors in any a16z
0:31:29 fund please note that a16z and its
0:31:31 affiliates may also maintain investments in
0:31:32 the companies discussed in this podcast
0:31:35 for more details including a link to our
0:31:39 investments please see a16z.com forward slash
0:31:40 disclosures
0:31:40 you
0:31:41 you
In this conversation from a16z’s Runtime conference, Gavin Baker, Managing Partner and CIO of Atreides Management, joins David George, General Partner at a16z, to unpack the macro view of AI: the trillion-dollar data center buildout, the new economics of GPUs, and what this boom means for investors, founders, and the global economy.
Resources:
Follow Gavin on X: https://x.com/GavinSBaker
Follow Atreides Management on X: https://x.com/atreidesmgmt
Follow David on X: https://x.com/DavidGeorge83
Stay Updated:
If you enjoyed this episode, be sure to like, subscribe, and share with your friends!
Find a16z on X: https://x.com/a16z
Find a16z on LinkedIn: https://www.linkedin.com/company/a16z
Listen to the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX
Listen to the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711
Follow our host: https://x.com/eriktorenberg
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Stay Updated:
Find a16z on X
Find a16z on LinkedIn
Listen to the a16z Podcast on Spotify
Listen to the a16z Podcast on Apple Podcasts
Follow our host: https://twitter.com/eriktorenberg
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
Leave a Reply
You must be logged in to post a comment.