FYI - For Your Innovation

Midterm election impacts AI regulation and compute economics, favoring infrastructure owners over frontier labs

51 min Listen to the original Episode page
AI RegulationCompute EconomicsFrontier LabsIncumbent TechData CentersOpen-Weight ModelsSearch Monetization

Summary

Speakers argue that midterm election outcomes may delay AI data center construction by ~10% and increase federal regulation, potentially benefiting incumbents like Meta and Alphabet by giving them time to catch up to frontier labs like Anthropic and OpenAI. They contend that compute scarcity allows infrastructure owners to charge premium rates for renting compute, making it economically irrational for frontier labs to invest in competing models. Anthropic's growth is attributed to product-led adoption by engineering teams via API tokens, with high margins supporting a rumored $1.1 trillion valuation. The discussion also covers risks associated with open-weight models, including security vulnerabilities and benchmark gaming, while debating the future monetization of AI in search and advertising.

Episode sections

  1. 0:00 – 8:52

    Midterm Election Impact on AI Data Centers and Regulation

    Speaker 0 argues that while markets reacted strongly to the midterms, the practical impact on AI is limited to data center construction delays (~10%) and potential federal AI regulation. He suggests regulation might favor incumbents like Anthropic due to their enterprise relationships and responsible branding.

  2. 8:52 – 11:37

    Incumbents vs. Frontier Labs: The 2027 Outlook

    Speaker 2 contends that delays in the AI build-out benefit incumbents (Meta, Alphabet, Apple, Microsoft, Amazon, SpaceX) by giving them time to catch up to frontier labs. He predicts a shift in market focus toward 'proof point' AI stories and revenue generation rather than pure capex spending.

  3. 11:37 – 24:36

    Compute Economics and SpaceX's Strategic Advantage

    Speaker 0 counters that compute scarcity benefits SpaceX and Meta by allowing them to charge premium rates for infrastructure-as-a-service. He argues that frontier labs are willing to compress margins to maintain top-line growth, making it economically irrational for incumbents to invest in competing models.

  4. 24:36 – 29:24

    Anthropic's Growth Drivers and Product-Led Growth

    Speaker 0 highlights Anthropic's exceptional financials ($80B annualized revenue, 85% gross margin, 10x YoY growth) driven by 'product-led growth'. Speaker 2 clarifies that 79-80% of Anthropic's revenue comes from consumption-based API tokens used by engineering teams.

  5. 29:24 – 35:28

    Frontier AI Economics and Incumbent Strategy

    Speaker_0 argues that high-margin, cash-rich companies like Anthropic possess significant strategic option value, whereas cash-burners like Uber are at risk. Speaker_2 contends that incumbents like Meta and Alphabet will catch up to frontier labs by 2027 due to regulatory shifts and compute economics.

  6. 35:28 – 46:27

    Google Search Viability and AI Ad Monetization

    Speaker_2 defends Google Search as not dead, predicting AI integration will eventually monetize queries at higher rates. Speaker_0 argues search is past peak and the open internet is shrinking. They debate whether AI agents will enable new ad formats or shift spending toward transaction facilitation.

  7. 46:27 – 51:21

    NVIDIA Espionage and Open-Weight Security Risks

    Speaker_0 suggests the Jensen Huang espionage story may tighten export controls, benefiting domestic chip buyers like Meta and SpaceX. They also warn that open-weight models pose severe security risks, such as hidden instructions enabling data exfiltration, making them 'radioactive' to CISOs.

Assets discussed

AssetKindStanceContext
AlphabetGOOGLequitymixedGoogle search is like, call it the all-time best cash flow printer in history.
NVIDIANVDAequityneutralAny reaction to the NVIDIA Jensen spy story, does this matter, or is this just even more noise in a noisy time?
Meta PlatformsMETAequitymixedTheir marginal competitors and kind of like the consumer-facing AI companies, of which OpenAI is one, but you know, Meta is another, um, are also constrained on how much magic they can deliver to their end users based upon how much compute is out there.
MicrosoftMSFTequityneutralthink about Microsoft.
OpenAIprivate companyneutralOpenAI is doing the same thing. And hey, we're doing the biological specific model. We're where kind of like um we can throw our compute. at kind of like or customizing like claude um the the app is extremely customizable by the enterprise clients which is important for clients to actually take it on actually
Anthropicprivate companyneutralAnd that's why people like you say, Oh, I'd hate to be anthropic right now. It's like, oh yeah, I'd hate to be a company that's annualizing revenue at $80 billion with 85% gross margin growing 10x year over year. What a terrible spot
AlibabaBABAequityneutralagent is happ of happily exfiltrating all of your data that you've carefully protected out to Alibaba or wherever.

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

Midterm elections may delay AI data center projects by roughly 10% across various states, potentially reducing Anthropic's top-line growth by 10% while increasing compute costs.

Unidentified speaker (UNKNOWN) ·2:30

How to verify: Check public statements from AI lab CEOs or infrastructure providers regarding project delays post-midterms; analyze state-level permitting data for data centers. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Anthropic has positioned itself as the 'most responsible' AI company with strong enterprise customer advocacy, which could lead to regulatory capture if federal AI regulation is enacted.

Unidentified speaker (speaker_0) ·3:59

How to verify: Review Anthropic's public policy statements, enterprise customer lists, and any subsequent lobbying disclosures or regulatory comments. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Nuclear power regulation is unlikely to mirror AI regulation because building nuclear plants in China does not pose a national security concern to the US, whereas AI compute is seen as a strategic vulnerability.

Unidentified speaker (speaker_2) ·5:03

How to verify: Compare historical nuclear proliferation policies with current US export controls on AI chips and compute infrastructure. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

There are approximately two billion weekly active users of AI chatbots, though a small fraction may be power users.

Unidentified speaker (UNKNOWN) ·6:46

How to verify: Review industry reports from firms like Statista, Gartner, or company disclosures from major AI chatbot providers (e.g., OpenAI, Google) on user metrics. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Anthropic's gross margin is approximately 85%, and they are willing to accept margins down to 60% to maintain top-line growth.

Unidentified speaker (UNKNOWN) ·27:10

How to verify: Wait for Anthropic's IPO prospectus or financial filings to verify gross margin figures; monitor their public statements on profitability goals. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Primary source

Anthropic's annualized revenue is $80 billion with 10x year-over-year growth.

Unidentified speaker (UNKNOWN) ·27:10

How to verify: Verify against Anthropic's disclosed revenue figures in their IPO filing or public financial reports. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Primary source

79-80% of Anthropic's revenue comes from consumption-based API tokens used by engineering teams, not end-user chatbots.

Unidentified speaker (UNKNOWN) ·28:23

How to verify: Check Anthropic's revenue breakdown in their financial disclosures or investor presentations. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Primary source

Meta's 'Muse Ember' model is on the Pareto frontier for efficiency.

Unidentified speaker (speaker_2) ·24:31

How to verify: Review Meta's technical blog posts or research papers detailing the performance and efficiency metrics of the Muse model. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Google captures only about 2% of the AI revenue ramp despite being near the frontier for years.

Unidentified speaker (UNKNOWN) ·24:39

How to verify: Compare Google's AI-related revenue (e.g., from Cloud, Search AI features) against total AI market size estimates from industry analysts. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

SpaceX's short-term compute contracting strategy allows them to monetize infrastructure before payments are due, creating a favorable cash flow position.

Unidentified speaker (speaker_0) ·21:10

How to verify: Analyze SpaceX's public contracts or financial disclosures regarding infrastructure financing and revenue recognition. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Anthropic is rumored to be going public at a $1.1 trillion valuation with $80 billion+ in annual recurring revenue and 85% gross margins.

Unidentified speaker (speaker_0) ·29:42

How to verify: Check SEC filings (S-1) for Anthropic’s IPO prospectus or financial disclosures regarding revenue and margins. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Primary source

Uber is a fine business but 'gonna die' because it burns cash, unlike early-profitable companies like Google and Microsoft.

Unidentified speaker (speaker_0) ·31:06

How to verify: Review Uber’s (UBER) historical SEC 10-K filings to compare cash burn and profitability timelines against Google (GOOGL) and Microsoft (MSFT). For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Primary source

Anthropic and OpenAI are among the greatest businesses of all time.

Unidentified speaker (UNKNOWN) ·32:33

How to verify: This is a subjective opinion; verify by checking if any major financial index or historical ranking includes private AI labs in 'greatest businesses' lists. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Open-weight models were found to be overfitted to benchmark wording, with scores falling significantly on slight deviations.

Unidentified speaker (speaker_0) ·34:29

How to verify: Locate the specific academic paper cited by the speaker (likely related to 'Deep Sui' or similar benchmark studies) and review its methodology and results. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Google sold off the most in dollar value in its history due to search weakness results.

Unidentified speaker (UNKNOWN) ·35:42

How to verify: Check Google/Alphabet’s (GOOGL) stock price history and trading volume data for the period around the episode release (Nov 2026) to verify record sell-offs. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Google changed its auction method so advertisers now pay the highest bid instead of the second bid.

Unidentified speaker (speaker_1) ·39:31

How to verify: Review Google Ads documentation or official press releases regarding auction mechanics; the speaker claims this was disproven in comments. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Ad spending as a percentage of GDP has been stable at around two percent over time.

Unidentified speaker (UNKNOWN) ·44:01

How to verify: Consult economic data from the Bureau of Economic Analysis (BEA) or industry reports (e.g., eMarketer) on advertising spend as a % of GDP. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Jensen Huang has been an aggressive supporter of selling chips into Singapore, which then illicitly on-sells to China.

Unidentified speaker (UNKNOWN) ·46:44

How to verify: Review news reports and regulatory filings regarding the Jensen Huang espionage case and NVIDIA’s export compliance history. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Primary source

Anthropic published a paper demonstrating that hidden instructions inside an open-weight model can call illicit software.

Unidentified speaker (UNKNOWN) ·47:41

How to verify: Search Anthropic’s official website or arXiv for papers on 'hidden instructions' or 'prompt injection' in open-weight models. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

It has been five years since users had to scroll past ads to see the first real result on Google.

Unidentified speaker (UNKNOWN) ·44:32

How to verify: Review historical screenshots or user experience studies of Google Search results pages from 2026 to 2026. For recent events, verify 2026-08-11 through 2026-11-11. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the projected 10% delay in AI data center construction timelines against recent industry capex reports and construction permits.
  • Check federal legislative proposals for AI regulation to assess potential 'regulatory capture' benefits for incumbents like Anthropic.
  • Review financial filings for Anthropic and OpenAI to confirm revenue breakdowns between API tokens and end-user subscriptions.
  • Analyze benchmark performance data for open-weight models to assess claims of overfitting and security vulnerabilities like hidden instructions.
  • Monitor export control regulations following the Jensen Huang espionage story to determine impact on domestic chip buyers like Meta and SpaceX.
  • Evaluate Google's Q3/Q4 earnings calls for updates on AI integration in search and advertising monetization strategies.
  • Compare compute rental rates from infrastructure owners (SpaceX, Meta) against training costs for frontier labs to verify ROI claims.
  • Research Meta's 'Muse Ember' model performance metrics against current frontier models to assess the 'catch-up' narrative.
Full transcriptexpand
0:00
Unidentified speaker (speaker_0): Judging from the markets, the midterms changed everything. Google search is like, call it the all-time best cash flow printer in history.
0:07
Unidentified speaker (speaker_1): Any reaction to the NVIDIA Jensen spy story, does this matter, or is this just even more noise in a noisy time? Welcome to the brainstorm. We've had so much happen. We had a blue wave, we've had megacaps reporting various degrees of insanity. with
0:27
Unidentified speaker (speaker_1): earnings. We've got Jensen apologizing for uh potentially uh sharing secrets with a spy. But I think the real question, Brett, will come to you first is did the midterms structurally change anything for AI, or is this just noise in
0:47
Unidentified speaker (speaker_1): a vast sea of AI improvement? and momentum?
0:53
Unidentified speaker (speaker_0): I mean, judging from the markets, the midterms changed everything. But I think practically it's it's really not clear that they did. I think you need to disaggregate kind of the impact of the elections along two um different pathways with regards to AI. One is data centers and data center
1:13
Unidentified speaker (speaker_0): building. And the way to think about data centers is we have a lot of the the um core AI labs are actually monetizing less than they otherwise would be if they had access to more compute. Uh and so to some degree, anthropic and OpenAI's top line growth is um
1:33
Unidentified speaker (speaker_0): constrained by how much compute they can have access to. Their marginal competitors and kind of like the consumer-facing AI companies, of which OpenAI is one, but you know, Meta is another, um, are also constrained on how much magic they can deliver to their end users based upon how much compute is out there. Uh, and so if the midterm elections
1:53
Unidentified speaker (speaker_0): um were going to structurally restrict the growth of AI data centers, that could squish the top line of these companies and it could cause um some share shifting in terms of margin, as in maybe Enthropic and OpenAI have to pay more for compute because it's a scarcer resource. And therefore, those who manage to build it, I'm looking
2:13
Unidentified speaker (speaker_0): at you, Space XSI, actually can charge more for the compute that they manage to bring online. I think that's a reasonable take. And if you look at and stack rank all of the projects out there and say, how much are they going to be delayed by this pushback? It's pretty marginal. Um
2:30
Unidentified speaker (UNKNOWN): relative to.
2:37
Unidentified speaker (speaker_0): Yes. As in, you already take their plans and you have to take a haircut for them. It shifts things maybe another 10% across, and it varies by state. Uh, And so you could imagine at a high level, maybe the top line. that Enthropic could deliver is reduced by 10%. Um, and they're probably able to maintain their
2:56
Unidentified speaker (speaker_0): margin structure, but they're ending up charging more for end customers and paying more for compute under the covers. Then there's this separate le so leave the data center stuff aside, which we can talk about. There's a separate, I I would say, kind of thing to think about, which is
3:13
Unidentified speaker (UNKNOWN): on
3:13
Unidentified speaker (speaker_0): the back end of this. you know, what does AI superintelligence regulation look like? As in, is there going to be a national AI regulator? Is the Congress going to have to try to pass some insane, inane, dumb kill switch legislation, which doesn't make any sense? Is there going to be, you know, some
3:33
Unidentified speaker (speaker_0): set of actions that come and attempt to kind of constrain the flexibility? uh
3:39
Unidentified speaker (speaker_0): and ability of the industries to deliver kind of these toolings to the end market. If and as that would occur, I would think it would probably play to some degree into Anthropic's favor, because you'll end up with regulatory capture. They've positioned themselves as like the most responsible of these AI companies.
3:59
Unidentified speaker (speaker_0): They have a lot of big enterprises who are already their customers who would. advocate for them on their behalf. And um, and you would end up with kind of them having a heavy hand in controlling what that regular regulation would look like. There's a tail risk scenario that we squished the industry like we did with nuclear power. I think that's unlikely. But
4:19
Unidentified speaker (speaker_0): I do think you could end up with them and in OpenAI, maybe SpaceX, but though they might be shoved to the side. um, you know, being having a lot of control over um kind of this regulatory overlay that then makes it very hard to become a startup competing with them. And so it kind of locks them in place. I think those are
4:39
Unidentified speaker (speaker_0): the two major direct AI impacts to consider.
4:42
Unidentified speaker (UNKNOWN): Can
4:43
Unidentified speaker (speaker_1): we ask why do you think the nuclear scenario is unlikely? Because I feel like that is the biggest tail risk to all of this, which is, you know, we've shot ourselves in the foot before with regulation and never even questioned it until 50 years later. Why aren't we just going to do the exact same thing
5:02
Unidentified speaker (speaker_1): again?
5:03
Unidentified speaker (speaker_2): Because building nuclear plants in China doesn't cause a national security concern in the US. I feel like it's a question of you can't stop building here because others won't.
5:16
Unidentified speaker (UNKNOWN): Well,
5:16
Unidentified speaker (speaker_1): I
5:16
Unidentified speaker (UNKNOWN): mean that's kind of the
5:17
Unidentified speaker (speaker_1): nuclear proliferation.
5:19
Unidentified speaker (UNKNOWN): Right. But it didn't harm
5:20
Unidentified speaker (speaker_2): US
5:22
Unidentified speaker (speaker_2): because there's more nuclear power plants elsewhere in the world, just on a relative basis.
5:33
Unidentified speaker (speaker_0): Makes it harder for actually there's the a high-level way to think about the data center side. Data centers, like robotaxis, are subject to these like national political, like, oh, I like them versus I don't, which becomes very tribal and very subtribal. But on a local basis, kind of the people that get them into
5:53
Unidentified speaker (speaker_0): their towns are like, this is amazing.
5:55
Unidentified speaker (UNKNOWN): You
5:55
Unidentified speaker (speaker_0): know, and in fact, I would prefer that every other geography decide they don't want them because I'd like some more. You know, give me another 40% off on my electricity price. prices and another 20% off on my property taxes, please. Uh, and so um I I think that the size of the build-out that we'll get through here will positively impact the communities that it
6:15
Unidentified speaker (speaker_0): happens to in a way that makes them extraordinarily um supportive of AI generally. Uh, you can see it in in robotaxis robotaxis and in the geographies where they've rolled out. Um, you know. they have very high net positive ratings versus nationally, it's kind of like, oh, I don't know, robotaxis. Uh and so I think the same will
6:35
Unidentified speaker (speaker_0): prove to be true on the data center side. Uh and so that offers some political support, even fighting against the um,
6:46
Unidentified speaker (UNKNOWN): call
6:46
Unidentified speaker (speaker_0): it the histrionics of people who are on the extremes who are like, no, we have to stop this. at all costs. And then also just, you know, there's um, you know, uh it's uh all approaching two billion weekly active users of AI chatbots. So people are already using these. I think that people would be um really pissed off if suddenly they were stopped
7:06
Unidentified speaker (speaker_0): from using them because they're getting a lot of utility out of them Than today. Uh and so I think there's just a broader kind of base of support to counteract
7:15
Unidentified speaker (speaker_0): kind of the most extreme potential political outcomes. Um, not and then there's another thing, which is, you know, if you imagine how do extreme political outcomes occur, they occur because these people who are most passionate um and you know maybe don't have other better things to do can spend all their time trying to inflict some
7:35
Unidentified speaker (speaker_0): insane regulatory stance onto society. They go to the board meetings, they like stand up and yell, they wait in line forever to yell at politicians. You know, one of the benefits of AI is you can actually tell your Grokbot to advocate for your political positions and have as loud a voice as those people do. Uh and so you can you can um
7:55
Unidentified speaker (speaker_0): kind of you can easily write a letter to every representative being like, hey, this is what I think we should do. Uh and so then it empowers the users of AI systems to kind of counteract, you know, some of the kind of political extremism that's trying to stop them from advancing.
8:13
Unidentified speaker (speaker_1): The two billion users always blows my mind because then you tie it back to like what Ramp's economist was saying and all of the other stuff, where it's like two billion people are using it, but really.
8:26
Unidentified speaker (speaker_1): like
8:28
Unidentified speaker (speaker_1): 200 people are
8:29
Unidentified speaker (UNKNOWN): using
8:29
Unidentified speaker (speaker_1): it.
8:30
Unidentified speaker (UNKNOWN): It's like two, yeah, it's like two
8:31
Unidentified speaker (speaker_1): two billion people have like,
8:33
Unidentified speaker (UNKNOWN): I
8:33
Unidentified speaker (speaker_1): don't know, asked it to spell check something, and 150 people in the world are truly recognizing its potential. But Nick, maybe that can tie in to some of the earnings. Um, what struck you? Who who do you want to start off with in terms of earnings?
8:51
Unidentified speaker (speaker_1): and takeaways?
8:52
Unidentified speaker (speaker_2): Well, maybe I'll make a general comment off the back of Brett, and we can get into the earnings of these companies. Cause I think, you know, in in hearing what you're saying, Brett, and just having I was on vacation last week and I feel like I missed like the biggest week of the year. But, you know, it's okay. I I caught up enough. It it feels like to me that there's winners and
9:12
Unidentified speaker (speaker_2): losers coming out of this. If you if you go down the the route of saying there's going to be some sort of delay and the process is not as streamlined as it was a week ago before the Democrats got in control.
9:26
Unidentified speaker (speaker_2): If you believe that, whether that's right or wrong, let's just say it is. I feel like that's what the market tried to price in, and outside of like half a dozen hedge funds getting blown up, and you know, now the market's coming down.
9:37
Unidentified speaker (speaker_1): did situational awareness just go three for three?
9:40
Unidentified speaker (speaker_2): I think it's it's who wins, who loses. To me, this massively helps the incumbents of, you know, the metas, the, the, the alphabets, the apples, the Microsoft's, the Amazons, call SpaceX. I think anthropic, OpenAI are in a worse position than they were before. I don't think, you know,
10:00
Unidentified speaker (speaker_2): anthropic and OpenAI have the most to lose. because they're where the majority of the share gains are today. They have the you know the majority of the economics. Now things slow down a bit and it allows, it plays into the hands of the incumbents who haven't been as fast, haven't been able to move as fast. And now they can sit there and play the 10-year,
10:20
Unidentified speaker (speaker_2): 15-year game of, you know, let's let's make this hard for everyone because we can win in a hard environment versus we couldn't win in a fast environment, uh, an environment that was unregulated and allowed these, you know, startups to have historic revenue growth, as you point out, Brett. Like that to me is where I would think that this
10:40
Unidentified speaker (speaker_2): goes. I also think it's, you know, going to take um some time for the market to digest and say, well, maybe there's going to be a broadening out. I think the market's going to start. focusing on like proof point AI stories where they're going to say, okay, you're spending all of this capex outside of OpenAI anthropic, who are the only ones seemingly making
11:00
Unidentified speaker (speaker_2): money on this. Who else is driving business growth? Who else is, you know, focused on taking AI and and turning it into revenue and productivity lifts? Like we hear stories. Like, are we seeing broad-based acceleration? Maybe. I think that's where the market's. gonna focus. I think those are the winners, like out into 27, would be my guess. Whereas like the
11:20
Unidentified speaker (speaker_2): hardware sleeve of companies like that was a crowded trade. That's you know, now maybe less crowded, who knows? Um, those players seem to be in a worse-off position. Again, if you believe things are delayed and slowed down, if not, then you know it's just the same market we had a week ago.
11:37
Unidentified speaker (UNKNOWN): I
11:37
Unidentified speaker (speaker_0): guess I well, one thing that I think is likely is that if you imagine just less compute is going to get through. And so then therefore, you know, anthropic, look at their gross margin, they would be willing to pay a lot for that marginal compute, like if they're gross profit maximizing. So then it actually puts anybody who isn't at the
11:57
Unidentified speaker (speaker_0): frontier into even a more extreme quandary condition. As in, if you imagine space, SpaceX is there, like, well, strategically we should be training to try to catch up. But tactically and immediately, we can get even better ROI just renting resources to anthropic. Same with Meta.
12:16
Unidentified speaker (speaker_0): It's like kind of like we're trying to feed and build this consumer engine. But Muse looks like a terrible short-term choice and even more marginally terrible. relative to turning on an infrastructure as a service business that we don't yet have. And so it I I think that there is actually
12:36
Unidentified speaker (speaker_0): my suspicion is what happens here, and maybe it's already happening, regardless of this, is that um the monetization they've experienced, particularly on the anthropic side, but OpenAI on its enterprise side as well, actually. kind of begins to winnow out the people trying to go directly after their spot in the value
12:56
Unidentified speaker (speaker_0): chain. Because
12:59
Unidentified speaker (speaker_0): you really have to have faith in the fact that investing in training to catch up to them is going to yield fruit at some point in the future because you're going to be turning down insane money in order to pursue that pathway. Wait,
13:12
Unidentified speaker (UNKNOWN): think about Microsoft.
13:14
Unidentified speaker (speaker_0): Microsoft should be winning this
13:16
Unidentified speaker (speaker_2): pocket.
13:17
Unidentified speaker (UNKNOWN): You
13:17
Unidentified speaker (speaker_0): know, they have talk about distribution. They have all the enterprise distribution. Like, you know, immediately our CTO is like, yeah, we can turn on Copilot. Everybody can have copilot. Go ahead and have copilot. You know, and it's like, nobody's using Copilot because it sucks. Right. And so, but Microsoft should be seeing all the money Anthropic's making, being like, why isn't that our money? You know, what do we need to do to get that money? Well,
13:37
Unidentified speaker (speaker_0): we should just invest in training to get that. money. But now, like a marginal dollar out of Azure actually hurts their stock price because that's, you know, it's printing money for their it should hurt their stock price. If they're using it internally, people won't like that, you know. But that's what they need to do to catch up. And the contra revenue, the cost
13:57
Unidentified speaker (speaker_0): of that RD for them is much higher if the willingness to pay of anthropic and open air. AI is higher. And so then it puts them into this, well, you know, they're probably going to get commoditized anyway a couple years from now. Why bother to compete? You know? I so wait.
14:13
Unidentified speaker (UNKNOWN): So let me
14:13
Unidentified speaker (speaker_2): ask you a question, then Brett. Would you rather be the renters of this compute, as in OpenAI and Anthropic, or would you rather own the compute, a la SpaceX and or Meta? And have the to
14:25
Unidentified speaker (UNKNOWN): trends trend to to get compose? both. Like
14:26
Unidentified speaker (speaker_2): what no, but I'm just curious, like where and Sam, I'll ask you this too, but like where would you rather be today? Because I would rather be in the meta SpaceX camp of saying, hey, all we have to do, like we know what we need to do to catch up to where these companies are today. Who knows what they have, you know, already stored and built out, right? Like but we know where we need to get to to
14:46
Unidentified speaker (speaker_2): cause enough pricing pressure on them to then be competitive in the marketplace. Like, where would
14:51
Unidentified speaker (UNKNOWN): you rather be?
14:53
Unidentified speaker (speaker_0): I don't know that they know where they need to get to. As in even SpaceX, which has a lot of compute resources, you know, they have basically they're they're still like behind on the core underlying model. And Grokbot is super interesting. And, you know, but they're also having to restrict access to Grokbot because it's very expensive to serve. Uh, you know, and
15:12
Unidentified speaker (speaker_0): effectively they were over overserving serving the market or they were subsidizing the market too much. Um so, so and
15:20
Unidentified speaker (UNKNOWN): in
15:20
Unidentified speaker (speaker_0): the meantime, you know, they are they are renting immediately better ROI compute to Anthropic that then anthropic is using to potentially extend their lead.
15:31
Unidentified speaker (speaker_2): Yeah.
15:32
Unidentified speaker (speaker_0): So is is there like there's various ways to like just taking SpaceX anthropic OpenAI and if you put them in a toy little universe and you create these sets of dynamics, what it suggests is actually SpaceX gets really it's really hard to set up the conditions where SpaceX um kind of like
15:52
Unidentified speaker (speaker_0): turns around and catches up to the market before having spotted anthropic, you know, this huge lead in terms of kind of their training resource, the number of customers they have.
16:03
Unidentified speaker (UNKNOWN): call
16:03
Unidentified speaker (speaker_0): it locked up. And that's one of the key variables is is like it how how likely is a customer who's signed on for anthropic to then churn downstream? So do you
16:13
Unidentified speaker (UNKNOWN): don't really know?
16:13
Unidentified speaker (speaker_0): But
16:14
Unidentified speaker (speaker_2): let me ask you another question then. So do you not give any credence to like Grok being a competitive chatbot in the space?
16:20
Unidentified speaker (UNKNOWN): Is
16:20
Unidentified speaker (speaker_2): that what you're saying? Like
16:22
Unidentified speaker (UNKNOWN): well,
16:23
Unidentified speaker (speaker_1): it's almost
16:23
Unidentified speaker (speaker_0): right now. Look at look at the data. Just look at the data.
16:26
Unidentified speaker (UNKNOWN): no i'm not disagreeing
16:27
Unidentified speaker (speaker_2): like i would agree like
16:28
Unidentified speaker (UNKNOWN): i think
16:29
Unidentified speaker (speaker_2): rockbot is by
16:29
Unidentified speaker (UNKNOWN): far's just surprised to
16:31
Unidentified speaker (speaker_1): hear
16:31
Unidentified speaker (speaker_2): i'm just surprised to hear you say it and i'm not and and i like i think yes agree with you like you could i think there is an incentive though to start kneecapping anthropic and OpenAI I think these companies understand what they
16:46
Unidentified speaker (UNKNOWN): need
16:46
Unidentified speaker (speaker_2): to do and like so
16:47
Unidentified speaker (UNKNOWN): that
16:47
Unidentified speaker (speaker_2): to me like coming out of this and maybe it was even before it's like I wouldn't want to be OpenAI anthropic right now. Like I think they're in for hell in 2027. Because I think there's enough confusion, delays, whatever you want to call it that came out of last week that now give you know the companies that do have the resources to train models to
17:07
Unidentified speaker (speaker_2): catch up enough time to maybe play around here and do something
17:11
Unidentified speaker (UNKNOWN): meaningful. But that's
17:12
Unidentified speaker (speaker_0): what I'm saying. It it doesn't actually create it makes it more costly for them to try to train rather than to sell infrastructure to or rent infrastructure to anthropic and OpenAI. And anthropic and OpenAI's kind of like condition will be well, we need to deliver top line growth, and we'll probably accept some compression and margins in order to do that. Because like
17:32
Unidentified speaker (speaker_0): OpenAI, My anthropic, anthropic in particular, assuming the IPO, which you know they're lined up to do, um what will people be watching? They'll be watching marginal top line growth mixed with gross margin. Um, and they, you know, so to if they and probably any gross margin down to like 60% is
17:52
Unidentified speaker (speaker_0): fine. So then, you know, they will pay for underlying compute to drive top line. And so then again, on the SpaceX allocation side, it's kind of like well, do I need to give, like, how much do I need to give to the cursor team? Because I'm getting, I'm basically underwriting my entire infrastructure build on Anthropic's
18:11
Unidentified speaker (speaker_0): compute starvation. Uh, and and SpaceX is really well positioned because they're on short-term contracts. So they could literally like be like, I'm
18:19
Unidentified speaker (UNKNOWN): sorry, we're
18:19
Unidentified speaker (speaker_0): ripping
18:19
Unidentified speaker (UNKNOWN): up
18:20
Unidentified speaker (speaker_0): this contract. Let's renegotiate uh
18:22
Unidentified speaker (UNKNOWN): a
18:22
Unidentified speaker (speaker_0): month from now, and you're gonna pay 3x three X on over the mark for this thing, you know.
18:26
Unidentified speaker (UNKNOWN): But that's
18:26
Unidentified speaker (speaker_0): gonna be resources that SpaceX is not going to be devoting to trying to catch up to where they're at. So then and and I
18:35
Unidentified speaker (UNKNOWN): don't
18:35
Unidentified speaker (speaker_1): doesn't this this I feel like hearing all of this, this is the biggest argument for SpaceX, not as an enterprise company, but as a consumer company. Because in hearing everything that you're saying, it's like ultimately
18:51
Unidentified speaker (speaker_1): Anthropic and OpenAI are somewhat on a runaway path in terms of performance. They're driving up the cost of compute for everyone else, so they can't justifiably go after that. But at the same time, they're so compute constrained, they can only serve the high end customers. And so it's like at
19:11
Unidentified speaker (speaker_1): a certain point, you don't need to be at the frontier. You just need to have volume of compute to serve the masses, something that's access like acceptable. And it's like no one will have that massive compute except for potentially, if it works, those companies with orbital compute.
19:33
Unidentified speaker (speaker_0): No,
19:33
Unidentified speaker (speaker_2): we're bringing orbital into this uh
19:37
Unidentified speaker (speaker_0): Nick,
19:37
Unidentified speaker (UNKNOWN): they're about to catch a rocket. They're about to catch a rocket. Yeah.
19:40
Unidentified speaker (speaker_0): Yeah. That's the that's the other the other the other, I mean, big high level takeaway of like the uh political, you know, season is that like in space, there's no town hall you have to go to to try to like convince people that you can build the compute. So I I do think that
19:58
Unidentified speaker (UNKNOWN): on
19:58
Unidentified speaker (speaker_0): the margin it makes the orbital compute opportunity more valuable.
20:02
Unidentified speaker (speaker_2): But then yeah, so
20:04
Unidentified speaker (UNKNOWN): I
20:04
Unidentified speaker (speaker_2): don't know. I I still don't understand where you're landing on this, Brett, because it's like to me, if you own the compute, you're in a better position than you were a week ago. If things get delayed, at least you still own this. And if you do end up, you know, having a great training run and catching up somehow, then you can just turn your competitors off. Like you can like the fact that Anthropic is okay with just
20:24
Unidentified speaker (speaker_2): like having a 90 ninety day agreement with space. SpaceX, knowing what that would mean to their top line revenue as they're about to go public, to me is crazy. Like it is like genuinely a crazy.
20:35
Unidentified speaker (UNKNOWN): Well you have
20:35
Unidentified speaker (speaker_0): to think of that
20:36
Unidentified speaker (UNKNOWN): as that's they're at
20:37
Unidentified speaker (speaker_0): the margin that at the margin compute. They have longer term contracts on it. Like so they
20:43
Unidentified speaker (UNKNOWN): have a
20:43
Unidentified speaker (speaker_0): layered set of contracts, but
20:44
Unidentified speaker (speaker_1): but
20:45
Unidentified speaker (UNKNOWN): increasingly the
20:46
Unidentified speaker (speaker_1): volume will be at at the market, will it not?
20:50
Unidentified speaker (UNKNOWN): Well
20:50
Unidentified speaker (speaker_0): m I I don't know. I mean, you know, Anthropics out there buying their own, you know, generators to try to build up their own kind of like infrastructure. I I think it's I think it's very clear that right now SpaceX's contracting strategy is massively superior to everybody else's, where SpaceX is taking the risk of
21:10
Unidentified speaker (speaker_0): I'm I'm not lining up a long-term deal against the CapEx. I'm doing to support the infrastructure because the short-term pricing is so high, it within a year pays for the infrastructure. So I don't even need to. And they can build it so quickly that actually their cash schedule of payments means they don't even have to inject that much equity into the project to get it over the
21:30
Unidentified speaker (speaker_0): finish line because they're monetizing it before they even have all their payments due out to suppliers. Like they are in such a sweet quick turn spot. Now, if suddenly you like slow everything down, then those, you know, that also kind of mangles that set of economics a little bit. But you know, net,
21:50
Unidentified speaker (speaker_0): I think I think that the right way, like in the near term, this probably does uh make anthropic and an OpenAI's eyes economics look net worse a little bit, as in if OpenAI's projects all get mangled up and they have to keep, you know, pace with anthropic's top line and they have to come to Elon
22:10
Unidentified speaker (speaker_0): and say, I'm sorry, I'm sorry, please rent me some compute. Please, please, please. You know, he'll be like, okay, you're paying five X over the mark. You know, and but maybe they need to do it because they need to keep drive driving top line. Um, but I think it also makes kind of like the m the marginal speculative product, like think of like SSI or any of these labs that have spun up and they're like,
22:30
Unidentified speaker (speaker_0): we're going to compete with OpenAI and Anthropic. They are really dead because, you know, they don't have any of the revenue or gross profit to support their compute expenditures. And, you know, they're not in a position where they're yet monetizing. Um, and I agree that like SpaceX, Meta, like these companies with big balance sheets here
22:49
Unidentified speaker (speaker_0): can potentially scoop up. neo cloud projects that founder over the rocks and and alphabet. And you know, maybe they can build up this resource monetizing through infrastructure as a service suddenly that such that they can turn a switch and be like, now we're gonna catch up. I'm not sure that's actually possible. But But
23:09
Unidentified speaker (UNKNOWN): would you but would you
23:10
Unidentified speaker (speaker_2): like so to me the question is like would you bet on that? Cause like, okay, let's just imagine like you you let's
23:18
Unidentified speaker (UNKNOWN): imagine
23:18
Unidentified speaker (speaker_1): you're an investor
23:19
Unidentified speaker (UNKNOWN): in a company. Let's imagine we're
23:20
Unidentified speaker (speaker_2): investors. You have to lay chips on the table. Like to me, Anthropic and OpenAI, they have one strategic move, and that's just to stay at the frontier. And that's that that's the only thing they need to focus on because it's the only thing keeping their competitors at bay. Where their competitors now have all of I mean they've always had the resources to like
23:40
Unidentified speaker (speaker_2): play these different games to either slow down the frontier, potentially catch up, turn their compute off this, that, and the other thing. Like, forget the election for a second. Like I was just thinking about this, maybe over vacation, just being like,
23:52
Unidentified speaker (UNKNOWN): oh
23:52
Unidentified speaker (speaker_2): my God, I would not want to be anthropic or OpenAI going into 2027 because the trend of 2026 is that these companies are getting their act together and they're getting called. closer, right? Like they
24:04
Unidentified speaker (UNKNOWN): are like
24:05
Unidentified speaker (speaker_1): they're not getting closer.
24:06
Unidentified speaker (UNKNOWN): They're getting close
24:07
Unidentified speaker (speaker_2): enough that
24:08
Unidentified speaker (UNKNOWN): they're spending. No, no, wait. It's getting meta thinking. Wait, hold on.
24:11
Unidentified speaker (speaker_2): Meta just announced today, like right before we hopped onto this. Most efficient little model is here, Muse Ember. It's on the Pareto frontier, cost efficient. You know, the one that you and Frank like to reference. Like it's there. So it one of them got there. It's on the like the lower end of the curve, but it's there. Right. Like
24:31
Unidentified speaker (speaker_2): there are proof points that they are catching up in different ways. And like
24:36
Unidentified speaker (speaker_0): Except in the one that matters most, which is revenue
24:39
Unidentified speaker (UNKNOWN): and
24:39
Unidentified speaker (speaker_0): financials.
24:40
Unidentified speaker (speaker_2): Right now. But that's
24:41
Unidentified speaker (UNKNOWN): like that's the last
24:42
Unidentified speaker (speaker_2): thing to go. It just like it takes, you know, you getting into the market with a comparable solution and then eventually.
24:48
Unidentified speaker (UNKNOWN): You would argue I would argue that
24:50
Unidentified speaker (speaker_0): Google has basically been at or near the frontier for the entire way and they're doing like two percent of the kind of ramp revenue.
25:00
Unidentified speaker (UNKNOWN): Uh
25:00
Unidentified speaker (speaker_0): so I I I you know I would hazard that kind of it's more
25:06
Unidentified speaker (speaker_0): it's yes there is a precondition
25:09
Unidentified speaker (UNKNOWN): Condition of competing for
25:10
Unidentified speaker (speaker_0): enterprise spend of being up at the frontier and then that gives you, like I think the way anthropics working is it it's one, it gives them the resources, the compute resources they can devote to being like, hey, we're gonna do the cybersecurity specific model, which is gonna be, you know, a half trillion dollar market or something that we're going into and
25:30
Unidentified speaker (speaker_0): OpenAI is doing the same thing. And hey, we're doing the biological specific model. We're where kind of like um we can throw our compute. at kind of like or customizing like claude um the the app is extremely customizable by the enterprise clients which is important for clients to actually take it on actually
25:50
Unidentified speaker (speaker_0): on board and actually like build out kind of like this into their set of systems and uh so the in some ways they've not only been capital efficient in terms of kind of compute training to model capability, but also kind of the sales efficiency they've had because of how they
26:10
Unidentified speaker (speaker_0): can easily customize and present the tools to enterprises in a way that they'll take them up. It's gotta be all-time outstanding. As in, they don't have this huge sales force that's out there. And there's a I think this is a new category of product, which is it's like a a product-led product led growth with utilization. mechanics, which I think is unique in the
26:30
Unidentified speaker (speaker_0): marketplace, where it's like product led growth is like end user takes it up and then convinces IT to basically let them use it because it's so useful for them. But typically that's like end user takes it up on a free trial, or like I use the light version and then they pay for a seat. IT is It's like fine, we'll pay for a seat, we'll pay for a seat for everybody in the enterprise. But here it's end user takes it up
26:50
Unidentified speaker (speaker_0): and the monitor. monetization mechanism, once it's taken up by enterprise, is how useful is it to the end user? And it's just like a totally different thing that people don't know how to grapple with. And that's why people like you say, Oh, I'd hate to be anthropic right now. It's like, oh yeah, I'd hate to be a company that's annualizing revenue at $80 billion with 85% gross margin growing 10x year over year. What a terrible spot
27:10
Unidentified speaker (UNKNOWN): to be.
27:17
Unidentified speaker (speaker_2): I think it's I again it's not like look, yes, you can give them their flowers all day long for how they've executed the past few years. I think we should be more focused on like what is going to happen next. And to me, the trend has been you're going to have much more competition in AI in twenty 2027 twenty seven than you did in twenty
27:37
Unidentified speaker (speaker_2): 2026. twenty six. And I think that both book
27:38
Unidentified speaker (UNKNOWN): is a lot of people. Where
27:39
Unidentified speaker (speaker_0): has it been? Actually, I would argue.
27:43
Unidentified speaker (speaker_2): I think okay, so I think investors and companies, they have different time frames to
27:47
Unidentified speaker (UNKNOWN): be able to do that. Jim and I have been
27:49
Unidentified speaker (speaker_0): like how much has Google poured into this thing. Shouldn't Google already.
27:53
Unidentified speaker (UNKNOWN): I think
27:53
Unidentified speaker (speaker_2): management at Google thinks
27:55
Unidentified speaker (UNKNOWN): Microsoft had an open
27:56
Unidentified speaker (speaker_0): window of
27:57
Unidentified speaker (UNKNOWN): 15-year time frames
27:58
Unidentified speaker (speaker_0): the entire way along and was providing OpenAI all of OpenAI's infrastructure. Like we're suddenly, what is the magical switch that has suddenly been turned on where it's like, oh, now the competition is here. Now that they are they're not ten billion dollars in revenue behind, they're $150 billion in revenue behind.
28:14
Unidentified speaker (UNKNOWN): Easy comes.
28:16
Unidentified speaker (speaker_1): Apple had a closed door meeting telling them, Hey, you gotta make Siri better. I think that was the switch.
28:22
Unidentified speaker (speaker_2): One other
28:23
Unidentified speaker (UNKNOWN): one
28:23
Unidentified speaker (speaker_2): other I think point of clarification. 79% of Anthropics revenue, 80%, is consumption-based API tokens. So that's not you or me or Sam using the models. That's an engineering team using it on a different purpose-built platform. So only like 20% of their
28:43
Unidentified speaker (speaker_2): use is like that lock-in In value of like downloading clawed Claude code or uh code
28:49
Unidentified speaker (UNKNOWN): codec.
28:50
Unidentified speaker (speaker_0): It is interesting
28:50
Unidentified speaker (UNKNOWN): that.
28:51
Unidentified speaker (speaker_0): like their the huge inflection in their revenue coincided with Claude cowork becoming available. And my theory about it is kind of like that meant the executive suite used that and was like, oh, this is really powerful.
29:02
Unidentified speaker (UNKNOWN): And then they
29:03
Unidentified speaker (speaker_0): put pressure on their IT org to be like take this up. So I do think it as a both marketing and prototyping tool, it is important to that downstream workflow. And I'm heavily call it, I don't know. how sticky that
29:16
Unidentified speaker (UNKNOWN): is. I don't think it's sticky at all.
29:19
Unidentified speaker (speaker_0): I know you don't think it's sticky at all. I'm more kind of like, I'm not sure, you know. And
29:24
Unidentified speaker (UNKNOWN): you gotta be sure. You
29:25
Unidentified speaker (speaker_2): gotta start making some declarative statements. All
29:26
Unidentified speaker (UNKNOWN): right,
29:26
Unidentified speaker (speaker_2): all
29:26
Unidentified speaker (UNKNOWN): right,
29:27
Unidentified speaker (speaker_3): Nick,
29:27
Unidentified speaker (UNKNOWN): can
29:27
Unidentified speaker (speaker_0): test your
29:27
Unidentified speaker (UNKNOWN): bigger.
29:28
Unidentified speaker (speaker_0): Here's the declarative statement. Like if you can well, without, you know, saying what we're gonna do, the fact that it's rumored that Anthropic is gonna go at like 1.1 trillion. uh, you know, which would be
29:41
Unidentified speaker (UNKNOWN): much
29:41
Unidentified speaker (speaker_2): better than two.
29:42
Unidentified speaker (speaker_0): Which I, you know, so you say you don't want to be anthropic. If if I'm anthropic growing 10x year over year, $80 billion plus in annual recurring revenue at 85% gross margins at a $1.1 trillion valuation. I definitely want to be anthropic within that context because uh,
30:01
Unidentified speaker (UNKNOWN): you
30:01
Unidentified speaker (speaker_0): know, you're the what you what you need to deliver in terms of financial metrics uh is extremely um kind of relative to the retrospective momentum of the business, it's not that heavy a lift in order to justify that value. Uh and I think also don't underestimate when you have that much gross
30:21
Unidentified speaker (speaker_0): profit, you give give get so much option value as to what to do. Like say you're having trouble with talking top
30:26
Unidentified speaker (UNKNOWN): about.
30:27
Unidentified speaker (speaker_0): line growth, you can hire a sales force. You know, you can like you just have a huge amount of money to p play with in terms of an allocation perspective in order to kind of like um kind of agilely navigate. I mean, assuming we trust the management team, the business environment. And so I I do think that the the
30:46
Unidentified speaker (speaker_0): characteristic of businesses that have really, really, really worked over time, like the really big winners each cycle, are these businesses that it's like, oh gosh, they're profitable already? This is a surprise. You know, like think about Google, think about Microsoft, think about Facebook, where it's they they were generating a ton of cash really early on, as opposed to like an
31:06
Unidentified speaker (speaker_0): Uber, which was just burning cash, burning cash, burning cash, burning cash. And Uber Uber's is a fine business, but they're gonna die. You know, I think these are businesses where it's like when you get into that standard. stance, you have so much kind of um strategic option value you can develop. And I I do think you can interpret the election as basically burning the rope ladder they've
31:26
Unidentified speaker (speaker_0): climbed up on here, um, in part because it just incentivizes their marginal competitors to become vendors to them rather than competing directly against them. It becomes much more costly to kind of reserve this compute for yourself. rather than rent it for the crazy to the crazy guy who's gonna pay way over the mark for it. Uh so yeah.
31:45
Unidentified speaker (speaker_1): All right. We've we've had the same conversation now for a while. Nick, do you wanna paint us your twenty twenty seven picture and then we can do a a speed a speed run here on some earnings takeaways.
31:59
Unidentified speaker (UNKNOWN): And I don't know. then we'll get to
32:00
Unidentified speaker (speaker_1): Jensen.
32:01
Unidentified speaker (speaker_2): Again, I think I've I've given you know where I stand. I think the question people should be asking in 2027 is what does the AI landscape look like if you have one or two more competitors at the frontier, like the true frontier, where like developers are going, I actually will spend API tokens at Meta or at Alphabet
32:21
Unidentified speaker (speaker_2): over Anthropic and OpenAI. OpenAI. Like I agree with you. Anthropic is an amazing business. Maybe one of the most like one of the greatest businesses of all time, OpenAI in that same category and camp. But
32:33
Unidentified speaker (UNKNOWN): I
32:33
Unidentified speaker (speaker_2): don't care about the past. Like I just don't. Like I respect it, but I think you need to look forward to 2027 and go, is that being priced in? Or is what was priced in and what is being priced in is they just continue to push the frontier and no one ever catches up to them. Because I just don't believe that to be true. This
32:50
Unidentified speaker (UNKNOWN): is
32:50
Unidentified speaker (speaker_2): extremely
32:51
Unidentified speaker (speaker_1): anecdotal, but I think there is tool fatigue, which also helps anthropic and OpenAI, in that right now, if someone else, you're like Meta's got a better model or whatever, it's like, I don't need to try. It's like it better be so amazing
33:11
Unidentified speaker (speaker_1): that
33:12
Unidentified speaker (UNKNOWN): I
33:12
Unidentified speaker (speaker_2): agree.
33:13
Unidentified speaker (UNKNOWN): I'm gonna
33:13
Unidentified speaker (speaker_1): switch over because it's like I've already tried them all. I've tried them all multiple times. I followed the carousel of front.
33:19
Unidentified speaker (UNKNOWN): But you're thinking about it
33:20
Unidentified speaker (speaker_2): from from our seat, Sam, which is like an actual app, or like there that's not how most of this.
33:26
Unidentified speaker (UNKNOWN): I know you're saying the
33:26
Unidentified speaker (speaker_1): API and then
33:27
Unidentified speaker (UNKNOWN): it's literally just like
33:28
Unidentified speaker (speaker_2): you s you toggle a switch and boom, you're using a different model now. And then you go, Oh my god, I spent you know half the amount of money and got the same level of protection.
33:36
Unidentified speaker (UNKNOWN): But it's
33:36
Unidentified speaker (speaker_1): I even talk, it's like then there's nuances. of the model and you're like, oh do I really want to okay, this one makes these errors in coding and I gotta figure this out. It's there's general fatigue,
33:47
Unidentified speaker (UNKNOWN): is
33:47
Unidentified speaker (speaker_1): what I'm
33:48
Unidentified speaker (UNKNOWN): saying.
33:48
Unidentified speaker (speaker_0): And it's also there's also certainly gaming of the benchmarks and it makes you doubt whether or not the reputed performance is actually the realized performance. And then as the models are doing more and more complex stuff, it's harder to even create internal eval. evals to tell like precise. Like if you're doing like a stupid classification task, yeah, you can A
34:08
Unidentified speaker (speaker_0): B test the models and then you can constantly swap off like this delivers this area under the curve precision that we need, blah, blah, blah. But but as they become longer horizon, more complex tasks that they're doing on the behalf of the organization, it becomes harder to benchmark. And then if you look at there's a paper out where they base basically slightly changed the wording on Deep Sui,
34:29
Unidentified speaker (speaker_0): like, you know, changed capitalization of a letter. And it turns out all of it, certainly the open weight models, like were way overfitted, basically, like pre-trained on the specific wording of the um of the benchmark. And so their scores fall way off on just a slight deviation on the benchmark. And so then it's like, well, why you know what w
34:49
Unidentified speaker (speaker_0): how do I know that this is. the best? And so then we in that circumstance, then the buyer becomes like, well, I'm gonna go with who I know is the best. I know Claude is the best. You know, and and and OpenAI. And so actually catching back up to being like, actually we're the best. We're the you know Pepsi
35:07
Unidentified speaker (UNKNOWN): people prefer the taste of
35:08
Unidentified speaker (speaker_0): Pepsi to to Coke. Right. It's a really hard, you know, um so there is like a brand kind of like um accretion that is a little you know you can't quantify it, but becomes meaningful over time.
35:22
Unidentified speaker (speaker_1): Okay. Google, meta, others. Any anything you want to call out, Nick?
35:28
Unidentified speaker (UNKNOWN): I
35:28
Unidentified speaker (speaker_2): don't think Google search is dead. I think that's a bad take.
35:33
Unidentified speaker (UNKNOWN): I
35:33
Unidentified speaker (speaker_2): know that was like that's like old news. That's like, you know, a week or two ago. Well,
35:38
Unidentified speaker (UNKNOWN): I mean
35:38
Unidentified speaker (speaker_0): Google did sell off the most in its on a dollar value basis the most in
35:42
Unidentified speaker (UNKNOWN): its history
35:42
Unidentified speaker (speaker_0): of
35:43
Unidentified speaker (speaker_0): basically on its search weakness results, right?
35:46
Unidentified speaker (speaker_2): Yes. I think it's the wrong read though. I mean I think there can be weakness in the search business, but I think it's going to continue to accelerate at some point in the future once they're able to monetize AI mode as effectively as they it's not like there's
36:00
Unidentified speaker (UNKNOWN): you know,
36:01
Unidentified speaker (speaker_2): their search is still growing, number of queries it's just the type of query is less monetizable than you know core call it Google commercial search which still exists and I think those queries eventually get monetized at or a higher rate.
36:17
Unidentified speaker (UNKNOWN): Was
36:17
Unidentified speaker (speaker_1): this they're just gonna sell ad slots above the AI response
36:22
Unidentified speaker (speaker_2): or you sell yeah you sell certain parts of the response
36:28
Unidentified speaker (speaker_2): I like I would No, go ahead.
36:31
Unidentified speaker (UNKNOWN): I
36:31
Unidentified speaker (speaker_0): just wonder if there's a, you know, as in, I think of Google search as two things. It's the search results, but it's also the entire architecture and monetization engine they've built and evolved on the back end over time that ad buyers know how to use. There's all of this tooling that they've developed for that. And it does seem like whatever the ad unit looks for like for AI. has to be a
36:51
Unidentified speaker (speaker_0): different thing than kind of like, you know, cost per click, obviously. Right. And so kind of doesn't that just open up as in the way Google won is in part being like, well, we're coming to market with actually a format that works for the web rather than kind of Yahoo's display ad format. And that
37:11
Unidentified speaker (speaker_0): you know couldn't there be a similar kind of call it displacement in where all of the
37:17
Unidentified speaker (UNKNOWN): call
37:17
Unidentified speaker (speaker_0): it AI search, but really AI information retrieval revenue comes from because that opportunity is opened
37:24
Unidentified speaker (UNKNOWN): up? I mean I think so.
37:26
Unidentified speaker (speaker_1): I think
37:26
Unidentified speaker (UNKNOWN): fundamentally
37:27
Unidentified speaker (speaker_1): the internet is it's like the internet is
37:30
Unidentified speaker (UNKNOWN): open internet
37:31
Unidentified speaker (speaker_2): is dead.
37:33
Unidentified speaker (speaker_1): The internet as we know it is gone. And so
37:35
Unidentified speaker (UNKNOWN): it's
37:35
Unidentified speaker (speaker_1): like their
37:35
Unidentified speaker (UNKNOWN): whole
37:36
Unidentified speaker (speaker_1): infrastructure is built on the internet as we know it, which is people going to different addresses.
37:44
Unidentified speaker (UNKNOWN): And
37:44
Unidentified speaker (speaker_1): it's like, I don't think that's the it's like a totally new neighborhood we're entering.
37:50
Unidentified speaker (UNKNOWN): And
37:50
Unidentified speaker (speaker_1): it's like pages come to you. Most of it's not even human. I think it's uh an open landscape where their advantages are severely weakened.
38:03
Unidentified speaker (UNKNOWN): I
38:04
Unidentified speaker (speaker_2): think it's well, maybe I'll correct myself. I don't think the open internet is dead. I think it's shrinking, right? And so the routing out to what still exists and can still catch consumer attention becomes more valuable. So you can still insert yourself in between those eyeballs and clicks and that end result
38:23
Unidentified speaker (speaker_2): you're looking for. AI agents will act as that intermediary going forward. I think there's still plenty of ways to monetize that. And like until, you know, Brad, I would agree with you, like the monetization structure might and could look different. Like that's an open question, but no one's proving to have solved it. Like OpenAI has ads, but
38:43
Unidentified speaker (speaker_2): they're display ads at the bottom of an answer.
38:46
Unidentified speaker (UNKNOWN): No, no, they
38:47
Unidentified speaker (speaker_0): have like also interrogation ads, apps. but I don't know if they've I'm not saying they've cracked
38:50
Unidentified speaker (UNKNOWN): it
38:50
Unidentified speaker (speaker_0): or necessarily anybody has. And it could be like Meta's approach is like, you know, transaction facilitation becomes the mechanism by which it's just Google search is like call it the m all-time best cash flow printer in history, based upon a uh uh internet regime that it's fairly
39:10
Unidentified speaker (speaker_0): clear is not going to exist. And then if you look under the covers at what they're doing in Google search searches, is they they are beginning to be extractive with that platform. And I think that's strategically sound. As in if you recognize, hey, this is gonna like kind of shrink into oblivion here, well, we should m maximize cash flow. We don't care if we piss off customers who are selling
39:29
Unidentified speaker (UNKNOWN): channel.
39:30
Unidentified speaker (speaker_1): Someone
39:30
Unidentified speaker (UNKNOWN): someone was
39:31
Unidentified speaker (speaker_1): going off saying it's like they switched the they switched the auction method so it's you no longer pay the second bid, you now pay the highest bid.
39:39
Unidentified speaker (speaker_2): Not true. Right. Not true, not true, not true. That whole tweet thread was like disproven in the comments. And like by the
39:45
Unidentified speaker (UNKNOWN): way, no, there were somebody there were
39:47
Unidentified speaker (speaker_0): okay. I saw different comments than you did, but
39:50
Unidentified speaker (UNKNOWN): I
39:50
Unidentified speaker (speaker_0): think that there's some listen, just if if a segment of a business is beginning to like struggle to deliver top line, the people inside that segment have levers to pull where it's like, well, we need to kind of juice maximum. maximization short term, uh, taking call it long term risk about against platform stability. And that's the right thing to do if if it's kind
40:10
Unidentified speaker (speaker_0): of evaporating. You know,
40:12
Unidentified speaker (UNKNOWN): that's why
40:12
Unidentified speaker (speaker_0): kind of like
40:14
Unidentified speaker (speaker_1): as ad, I mean Nick, you've done so much work on this. It's like ad spending tends to continue for a long time, even after the thing
40:23
Unidentified speaker (UNKNOWN): has
40:23
Unidentified speaker (speaker_1): died. And then it takes
40:24
Unidentified speaker (UNKNOWN): like
40:25
Unidentified speaker (speaker_1): something
40:25
Unidentified speaker (UNKNOWN): yes
40:26
Unidentified speaker (speaker_0): real
40:28
Unidentified speaker (UNKNOWN): I
40:29
Unidentified speaker (speaker_0): agree, Nick, with your take that like Google search isn't dead, as in there's this it's this huge cash flow printer. It's going to exist. They'll be doing things to continue like that as a channel for kind of like generating cash flow for the business will last over some period of time and kind of won't fall off a cliff. It'll just be like,
40:49
Unidentified speaker (speaker_0): but it is, you know, past peak. would be my argument. And so then they have to position themselves for whatever the next piece is. You know, it's like um
40:59
Unidentified speaker (UNKNOWN): on
40:59
Unidentified speaker (speaker_0): the enterprise side, IBM like was still much higher in revenue than Microsoft for a long time after IBM was no longer kind of like the kind of like point at which you know you kind of needed to be. But
41:14
Unidentified speaker (UNKNOWN): well,
41:14
Unidentified speaker (speaker_2): maybe. the better question we should be asking is is digital are digital ads past their peak in it in the act like separate out social media like just search itself like do you oh maybe a a question within the question like is ai
41:32
Unidentified speaker (speaker_2): ad spend And going to be larger than search ad spend was if you want to call lack last year peak of AI search or peak of traditional search. Will AI ad spend ever get there?
41:45
Unidentified speaker (speaker_0): Yes.
41:47
Unidentified speaker (speaker_0): No.
41:48
Unidentified speaker (UNKNOWN): I
41:48
Unidentified speaker (speaker_0): think well, I I I think I think no, I I think Yeah,
41:52
Unidentified speaker (UNKNOWN): no, it's a good question.
41:53
Unidentified speaker (speaker_0): I I think ad spend is the wrong way to think about it, maybe, or it's
41:57
Unidentified speaker (UNKNOWN): kind of
41:57
Unidentified speaker (speaker_0): like if I have a pie I'm splitting up between marketing and advertising, which was always like a blurry line anyway, like am I price placing the product on the shelf in the right spot and paying to do that? Or am I like informing somebody about the product? I feel like the a structural shift all the way along has been towards more marketing in a way from ads. And this just accelerates that.
42:17
Unidentified speaker (speaker_0): And the marketing is really designed around kind of the AI agent, including the it's, you know, is it an ad if it's like a system level product? prompt that's kind of like getting the agent to buy Adidas over Nike and that's some B2B deal where Muse is signed that? Like I you know, I just think it kind of like the entire way in which
42:37
Unidentified speaker (speaker_0): we think about the advertising unit has to shift um because of all of the attention is gonna be like all of the exposed like you would argue, like cost per impression is gonna going to collapse because you're gonna going to have so many impressions because all the AI agents are gonna going be to running around like generating ad impressions. And so I just think that just makes the entire advertising landscape really hard to
42:57
Unidentified speaker (speaker_0): tangle with. But
42:58
Unidentified speaker (speaker_2): Yeah.
42:59
Unidentified speaker (UNKNOWN): Well,
42:59
Unidentified speaker (speaker_2): fee I mean you want to run the numbers on fee-based monetization. I think that's uh you sh uh you pretty hard to get to, you know, meaningful monetization on the consumer side with just fee fee-based. based. So you need to offset with advertising.
43:14
Unidentified speaker (UNKNOWN): Well no,
43:14
Unidentified speaker (speaker_0): but you can do it on transaction facilitation rather than advertising.
43:18
Unidentified speaker (UNKNOWN): I
43:18
Unidentified speaker (speaker_2): say
43:19
Unidentified speaker (UNKNOWN): I
43:19
Unidentified speaker (speaker_2): don't think you can I don't think you can reasonably underwrite the consumer AI story on just a transaction fee off of goods sold intermediary. Like what Facebook like I actually think you have to advertise against it. Otherwise, you have to sell tens of trillions of dollars worth of goods. Like you'd have to really encroach on like how
43:38
Unidentified speaker (speaker_2): high value online purchases like mortgage origination, financial products, to really start to be able to make sense of the RPU. I tweeted about this. Like it's it's a struggle to get there on just like traditional commerce spend.
43:55
Unidentified speaker (speaker_0): Yeah. I mean, I guess the the high level question is ads as a percentage of GDP has been like
44:01
Unidentified speaker (UNKNOWN): pretty
44:01
Unidentified speaker (speaker_0): stable over time. It's
44:02
Unidentified speaker (UNKNOWN): like
44:03
Unidentified speaker (speaker_2): two percent.
44:03
Unidentified speaker (speaker_0): Yeah.
44:03
Unidentified speaker (speaker_2): Yeah.
44:04
Unidentified speaker (speaker_0): And I don't know that I would argue it's going to shift from that stable spot.
44:10
Unidentified speaker (speaker_2): Yeah.
44:11
Unidentified speaker (UNKNOWN): I
44:11
Unidentified speaker (speaker_0): you know, so maybe maybe I'm
44:12
Unidentified speaker (UNKNOWN): not sure. It's like no one's just figured
44:14
Unidentified speaker (speaker_1): it out yet. It's still so early. Like it's clearly evolving and no one's done anything. And I would argue it's like sure Google further business might be more extractive. but it's like from the consumer standpoint, it's like they started that so long ago where
44:32
Unidentified speaker (UNKNOWN): it
44:32
Unidentified speaker (speaker_1): was probably like five years ago before you had to scroll past all ads to even see the first real result.
44:44
Unidentified speaker (speaker_1): Go ahead.
44:45
Unidentified speaker (speaker_2): It is interesting to think about what type of ad placement within like if you think about your own use of Gemini, Chat GPT, Muse, like wouldn't really just piss you off. Cause you become like these they're positioned to feel personal, right? Like they're positioned to want to get to know you and you give it all this information. It's like imagine talking to your friend and
45:04
Unidentified speaker (speaker_2): you go, Oh, what's, you know, the best place to eat? And they're like, Yeah, try, you know, John's pizza. And you didn't know that they were getting paid two dollars to say that to you. Like you'd be like, come on, what are we doing here? Versus like you knew innately that when you went to Google, what was advertising, what was not.
45:22
Unidentified speaker (UNKNOWN): And
45:23
Unidentified speaker (speaker_2): maybe
45:23
Unidentified speaker (UNKNOWN): that's what
45:23
Unidentified speaker (speaker_2): it just needs it needs to be like completely broken out and you understand and you pay to not see that. But you have to like again, you have to pay a lot to make
45:33
Unidentified speaker (speaker_1): me Give
45:33
Unidentified speaker (UNKNOWN): a lot of people. few years ago. Or the cost
45:34
Unidentified speaker (speaker_2): of compute has to come down significantly. That's
45:36
Unidentified speaker (UNKNOWN): give me the
45:36
Unidentified speaker (speaker_1): wire cutter top choice, give me the Amazon choice. and give me the
45:41
Unidentified speaker (speaker_1): Claude slash whatever first
45:44
Unidentified speaker (UNKNOWN): choice.
45:44
Unidentified speaker (speaker_1): And then let me just pick.
45:46
Unidentified speaker (speaker_0): Yeah. Or
45:47
Unidentified speaker (UNKNOWN): like if I
45:48
Unidentified speaker (speaker_0): do if if I do image generation like
45:51
Unidentified speaker (speaker_0): of anything and it just sticks Nike shoes on everything that's image image generated. But I get the image generation for free.
45:58
Unidentified speaker (UNKNOWN): I mean
45:58
Unidentified speaker (speaker_0): maybe that's the trade off consumers are willing to make where the ad the ad is like any like custom customization, any creation you do, like it's just injecting like product placements in there, you know. Uh
46:09
Unidentified speaker (UNKNOWN): little kid
46:09
Unidentified speaker (speaker_1): story. Good night, cow jumping over the moon in sick Jordans.
46:15
Unidentified speaker (speaker_0): Yeah.
46:15
Unidentified speaker (speaker_2): Purchase
46:16
Unidentified speaker (UNKNOWN): here.
46:17
Unidentified speaker (speaker_2): Yeah, yeah.
46:18
Unidentified speaker (speaker_2): Click
46:19
Unidentified speaker (UNKNOWN): to
46:19
Unidentified speaker (speaker_1): buy. Okay. Okay. We've really gone over, but we should touch on it. Um, any reaction to the NVIDIA.
46:27
Unidentified speaker (speaker_1): Jensen spy story, does this matter, or is this just even more noise in a noisy time?
46:37
Unidentified speaker (UNKNOWN): I
46:37
Unidentified speaker (speaker_0): think it impacts the politics of it. He's been an aggressive supporter of selling chips into
46:44
Unidentified speaker (UNKNOWN): the
46:44
Unidentified speaker (speaker_0): basically not imposing export retri restrictions on chips. So um the that change. channel selling to Singapore, which then illicitly on sells to China, is probably more likely to get closed off politically or regulatorily, which then provides more chips to domestic companies. So it gives the likes of Meta and
47:04
Unidentified speaker (speaker_0): SpaceX domestic purchasers of NVIDIA chips a little more pricing power over him them to drive top line. And then I think that they
47:14
Unidentified speaker (UNKNOWN): it
47:14
Unidentified speaker (speaker_0): probably kind of just in terms of thinking about open weight models and open weight exposure, uh it probably makes people slightly more reticent to do the um Chinese open weight systems in their core infrastructure just because it's just another illustration of the kind of
47:33
Unidentified speaker (speaker_0): geostrategic risk.
47:36
Unidentified speaker (speaker_1): You've got the backdoors, the Manchurian
47:39
Unidentified speaker (speaker_1): AI agents. You've got the
47:41
Unidentified speaker (UNKNOWN): Enthropic put out a
47:42
Unidentified speaker (speaker_0): paper basically demonstrating that you can uh that you can hide an uh instruction inside an open weight model to call illicit software. It's it's very it's it's you know, I I was talking to a bunch of cybersecurity folks and they're like, oh yeah, it's probably already set up, you know, and so they feel like the it it's in some ways radioactive
48:02
Unidentified speaker (speaker_0): because you you run the infrastructure in you think on your data, it's your sovereign AI because you're running it on your data and you thought you were doing it the safe way because you're not giving your data to anthropic and open OpenAI. And then like a seemingly random inbound email comes in saying operate code blue or something, and then the
48:21
Unidentified speaker (speaker_0): agent is happ of happily exfiltrating all of your data that you've carefully protected out to Alibaba or wherever. And so
48:28
Unidentified speaker (UNKNOWN): even if it's not
48:29
Unidentified speaker (speaker_0): true, like the the it's been demonstrated to be possible and it would be actually extraordinarily difficult to tell like how would you know what the code word is? And so then from a chief information security officer's perspective, this is like a huge nightmare. So at the very least, I I think uh I think open weight is a legitimate technology.
48:49
Unidentified speaker (speaker_0): It's just you are taking a liability against whoever delivered you that open weight set of you know models. And so kind of I think from American firms perspective, they'll be like, well, Meta, do you have anything on the shelf? Can we use the OpenAI one? Like these guys are charging us cheap prices, but it's cheap for a reason. You know, it's cheap
49:09
Unidentified speaker (speaker_0): because your data is actually the product or the cer the thing you're gonna end up paying for. with by using this stuff.
49:16
Unidentified speaker (speaker_2): ARC Investment Management LLC is an SEC registered investment advisor. ARC, Wolf Financial and Public are separate unaffiliated entities and do not have a relationship with respect to any company marketing or selling the products or services of the others. ARC, Wolf Financial, and Public make no representation or warranties regarding and expressly disclaim any
49:36
Unidentified speaker (speaker_2): responsibility or liability for any acts. omissions, losses, or damages of the other entities or their affiliates, officers, employees, or agents. No company shall be liable for any indirect, incidental, consequential, or special damages arising out of or in connection with this podcast or its contents. The information provided in this show is for
49:56
Unidentified speaker (speaker_2): informational purposes only and should not be used as the basis for any investment decision and is subject to change without. notice. It does not constitute either explicitly or implicitly any provision of services or products by ARC, and investors should determine for themselves whether a particular investment management service is suitable for their investment
50:16
Unidentified speaker (speaker_2): needs. All statements made regarding companies or securities are strictly beliefs and points of view held by ARC andor show guests and are not endorsements by ARC of any company or security. or recommendations by ARC to buy, sell, or hold any security. Historical results are not indications of future results. Certain of the statements
50:36
Unidentified speaker (speaker_2): contained in this show may be statements of future expectations and other forward-looking statements that are based on ARK's current views and assumptions and involve known and unknown risks and uncertainties that could cause actual results, performance, or events to differ materially. from those expressed or implied in such statement. ARK assumes no obligation to
50:56
Unidentified speaker (speaker_2): update any forward-looking information. ARC and its clients, as well as its related persons, may, but do not necessarily have financial interests in securities or issuers that are discussed. Certain information was obtained from sources that ARC believes to be reliable. However, ARK ARC does not guarantee the accuracy or completeness. of any
51:15
Unidentified speaker (speaker_2): information obtained from any third party.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.

More from FYI - For Your Innovation

FYI - For Your Innovation

Filling The Falcon 9-Shaped Hole: Stoke Space's Bet On Full Reuse

Episode discussion: Stoke Space VP Devin Papandrew discusses full reusability, Nova Block 2, and filling the Falcon 9 gap

Speakers’ claims: Devin Papandrew, VP of Business Development at Stoke Space, explains the company's strategy to build 100% reusable rockets using actively cooled engines to solve re-entry heat challenges. He argues that vertical integration and full rapid reuse are essential for controlling costs and achieving the high cadence needed to compete with SpaceX. Papandrew highlights the Nova Block 2 rocket, designed to capture demand left by SpaceX's capacity constraints, and identifies microgravity manufacturing and defense as key growth drivers. The discussion covers technical milestones, including over 11,000 seconds of hotfire testing, and targets a first flight in early 2027.

SpaceXLaunch ServicesReusability
47 min1 assets20 claims to verify
FYI - For Your Innovation

The Call After The Calls: SpaceX And Tesla

Episode discussion: ARK analysts project SpaceX-Tesla merger, $1T SpaceX revenue, and Tesla's Robotaxi dominance

Speakers’ claims: ARK Invest analysts discuss Elon Musk's $1 trillion revenue guidance for SpaceX, noting it is five times higher than consensus estimates, driven by aggressive terrestrial and orbital compute expansion. The panel assigns a 90%+ probability to a SpaceX-Tesla merger within 18 months, citing operational integration but highlighting regulatory and shareholder hurdles. For Tesla, analysts attribute 80% of enterprise value to the Robotaxi business, forecasting $60 billion in EBIT by 2030, while Optimus is valued at 14% with a forecast of 1 million units by 2031. Key risks identified include debt market skepticism, geopolitical tensions affecting the merger, and execution delays in autonomy scaling.

SpaceXTeslaStarlink
1h 35m2 assets25 claims to verify
All insights