Ramp Economist Finds AI Spend Growth Slowing Among Top Adopters Despite Rising Usage
Summary
Ara Carazian, lead economist at Ramp, presents data showing that while AI spending is growing rapidly, it is highly concentrated among the top 1% of businesses. He argues that this top tier is slowing its expenditure growth despite increasing token usage, driven by price deflation and competition from providers like OpenAI and Anthropic. The hosts discuss the risk that 70-80% of AI company revenue may come from other VC-backed firms, creating systemic correlation risks. Carazian predicts median firm AI spend will remain low, viewing AI as table stakes rather than a high-growth expense category for most enterprises.
Episode sections
- 0:24 – 1:55
Introduction and Data Scope
Host introduces Ara Carazian, lead economist at Ramp, who analyzes AI spend across 80,000+ businesses. Carazian notes that while AI spend is growing faster than any other category, it is highly concentrated among a small percentage of intense adopters.
- 1:55 – 4:46
Cracks in the AI Thesis: Top 1% Slowdown
Carazian argues that the top 1% of businesses, which drive 80% of AI spend, are slowing their expenditure despite using more tokens. This slowdown is attributed to increased competition, lower prices, and better enterprise contracts from providers like OpenAI and Anthropic.
- 4:46 – 12:57
Median vs. Top Tier Spend Dynamics
The median firm spends approximately $15 per employee per month on AI, showing steady growth. In contrast, the top 1% spends $7,000-$10,000 per employee per month but is seeing a slowdown in spend growth since September, even as token usage increases.
- 12:57 – 17:46
Competition and Market Share Shifts
The AI provider market is becoming more competitive with OpenAI, Anthropic, XAI, and Meta entering the fray. Carazian notes that Anthropic gained significant share rapidly, partly driven by consumer brand awareness, and that price cuts (e.g., OpenAI's Luna) are driving efficiency.
- 17:46 – 24:33
Revenue Quality and VC Dependency
Chamath Palihapitiya and Carazian discuss that 70-80% of AI company revenue may come from other VC-backed AI firms. This creates a correlation risk where a slowdown in VC funding or AI adoption among top spenders could cascade through the sector.
- 24:33 – 28:27
Productivity Gains and Cybersecurity
The hosts discuss whether non-tech firms can achieve significant ROI with low AI spend. Carazian disagrees with the idea that low spend yields high productivity for core business outputs, arguing that frontier models require significant spend. They also identify cybersecurity as a rising spend category due to open-weight model risks.
- 28:27 – 31:46
Future Spend Predictions
Carazian predicts median firm AI spend will remain under $1,000 per employee per month (or potentially per year) in five years, viewing AI as table stakes similar to a computer purchase. He emphasizes that headwinds currently outweigh tailwinds for rapid spend growth.
- 31:46 – 34:01
Disclaimer and Legal Notice
The segment consists entirely of standard legal disclaimers from ARC Investment Management LLC. It clarifies the separation between ARC Wolf Financial and Public entities, limits liability for damages, and states that the content is for informational purposes only and not investment advice.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| OpenAI | private company | neutral | that it's not just OpenAI's game anymore, as it was when we started tracking. |
| Anthropic | private company | neutral | Anthropic is now the most popular AI model used by businesses, that there's increasing competition from SpaceX AI, and now Meta's gonna get into the mix for enterprise, and video already has a model. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
1% of businesses drive 80% of AI spend.
Ara Carazian ·0:00
How to verify: Check Ramp's published data reports or tweets by Ara Carazian referencing the concentration of AI spend among the top 1% of their 80,000+ business dataset.
Top 1% of businesses are slowing their AI spend despite using more tokens.
Ara Carazian ·0:02
How to verify: Compare Ramp's internal data on token volume vs. spend for the top 1% segment between July and September 2026, as referenced in Carazian's posts.
Anthropic is the most popular AI model used by businesses.
Ara Carazian ·1:41
How to verify: Review Ramp's market share data for AI model usage among its 80,000+ business clients.
Median firm spends $15 per employee per month on AI.
Ara Carazian ·4:50
How to verify: Check Ramp's published charts or reports showing median AI spend per employee.
Top 1% firm spends $7,000-$10,000 per employee per month on AI.
Ara Carazian ·6:46
How to verify: Check Ramp's published charts or reports showing top 1% AI spend per employee.
AI prices have fallen by almost half over the period analyzed.
Ara Carazian ·7:48
How to verify: Analyze Ramp's data on average cost per token or total spend vs. token volume for the top 1% segment between July and September.
70-80% of revenues at AI companies are from other AI companies who've raised money from the venture industry.
Chamath Palihapitiya ·17:46
How to verify: Review financial filings or revenue breakdowns of major AI companies (e.g., Anthropic S1, OpenAI private market data) to identify customer concentration and VC backing.
OpenAI cut Luna's prices 80% to win marginal share.
Brett Hendon ·21:49
How to verify: Check OpenAI's pricing announcements or Ramp's data on spend shifts towards Luna following price changes.
Median firm AI spend will be under $1,000 per employee per month in five years.
Ara Carazian ·29:10
How to verify: Track actual median AI spend per employee in Ramp's dataset in 2031 and compare to Carazian's prediction.
ARC Investment Management LLC is an SEC registered investment advisor.
Unidentified speaker ·31:59
How to verify: Search the SEC Investment Adviser Public Disclosure database for ARC Investment Management LLC registration details.
ARC Wolf Financial and Public are separate unaffiliated entities.
Unidentified speaker ·32:03
How to verify: Review the corporate structure and entity filings for ARC Wolf Financial and ARC Public to confirm lack of affiliation.
The podcast content does not constitute a provision of services or products by ARC.
Unidentified speaker ·32:46
How to verify: Review ARC's Terms of Service and Privacy Policy regarding podcast content usage.
Statements regarding companies are strictly beliefs and points of view, not endorsements.
Unidentified speaker ·32:59
How to verify: Compare the podcast statements against any official press releases or endorsements from the mentioned companies.
Historical results are not indications of future results.
Unidentified speaker ·33:14
How to verify: Analyze historical performance data of discussed assets against subsequent market performance to test correlation.
Statements may include forward-looking statements based on current views and assumptions.
Unidentified speaker ·33:17
How to verify: Compare forward-looking statements made in the podcast with actual subsequent market events and company reports.
Information was obtained from sources ARC believes to be reliable, but accuracy is not guaranteed.
Unidentified speaker ·33:50
How to verify: Cross-reference specific data points mentioned in the podcast with primary sources cited or available public records.
Do your own research
- Verify Ramp's methodology for analyzing spend across 80,000+ businesses to understand data scope and potential biases.
- Review recent earnings reports and investor presentations from OpenAI, Anthropic, and Meta for evidence of price cuts and market share shifts.
- Analyze revenue breakdowns of major AI companies to assess the percentage of revenue derived from other VC-backed firms.
- Check cybersecurity spending trends and regulatory filings related to open-weight model risks.
- Compare AI productivity claims in non-tech sectors against actual implementation case studies and ROI data.
- Monitor VC funding rounds and burn rates of AI startups to assess systemic correlation risks mentioned by the speakers.
Full transcriptexpand
- 0:00
- Unidentified speaker: 1% of businesses drive 80% of that spend.
- 0:02
- Unidentified speaker: That top 1% of businesses, those are the really intense adopters of AI, are already, it seems, starting to slow their spend on AI.
- 0:10
- Unidentified speaker: And it's not that they're using it less, right?
- 0:13
- Unidentified speaker: They actually are using it more based on all the indicators we have.
- 0:15
- Unidentified speaker: They're using more tokens.
- 0:17
- Unidentified speaker: If these companies, if you can't raise the next up round, then your AI spend stops.
- 0:23
- Unidentified speaker: Welcome to the brainstorm.
- 0:24
- Unidentified speaker: We've got an amazing guest today, the lead economist from Ramp, Ara Carazian.
- 0:30
- Unidentified speaker: Ara, thank you so much for joining us.
- 0:33
- Unidentified speaker: You've been putting out tons of great tweets.
- 0:35
- Unidentified speaker: You're close to the source of uh seeing what's coming through from people spending on AI.
- 0:42
- Unidentified speaker: What are you seeing out there?
- 0:44
- Unidentified speaker: I don't know if you looked at the anthropic S1.
- 0:46
- Unidentified speaker: You know, how are you thinking about the entire industry right now?
- 0:50
- Unidentified speaker: Yeah, thanks for having me.
- 0:52
- Unidentified speaker: The also the amended S1 just came out as at the time we're recording, just maybe less than an hour ago.
- 0:57
- Unidentified speaker: Um I mean, it's a privileged data set to have access to.
- 1:01
- Unidentified speaker: I mean, it's all of our research comes from RAMP's data set.
- 1:05
- Unidentified speaker: So 80,000 plus businesses, 275 billion uh plus every year in business spend.
- 1:11
- Unidentified speaker: And so it's a great look at where businesses are spending and where they are pulling back.
- 1:16
- Unidentified speaker: It also helps us to look very granularly at who is actually investing in AI specifically.
- 1:22
- Unidentified speaker: And so over the past two years that we've been tracking AI spend, we have noted that AI spend has grown faster than any other spend category we've tracked, that it's also been spread out across multiple providers, that it's not just OpenAI's game anymore, as it was when we started tracking.
- 1:41
- Unidentified speaker: Anthropic is now the most popular AI model used by businesses, that there's increasing competition from SpaceX AI, and now Meta's gonna get into the mix for enterprise, and video already has a model.
- 1:55
- Unidentified speaker: But now my last series of posts has been called Cracks in the AI thesis.
- 2:00
- Unidentified speaker: Because although we have observed this rapid run-up in AI spend, and note AI spend does continue to grow in our data set.
- 2:07
- Unidentified speaker: If you have this uh preconceived notion that AI spend is going to continue to rise, you should maybe check that assumption with respect to the data that we've already reported, which is that one AI spend is highly concentrated, 1% of businesses drive 80% of that spend.
- 2:25
- Unidentified speaker: That top 1% of businesses, those those are the really intense adopters of AI, are already, it seems, starting to slow their spend on AI.
- 2:35
- Unidentified speaker: And it's not that they're using it less, right?
- 2:37
- Unidentified speaker: They actually are using it more based on all the indicators we have.
- 2:40
- Unidentified speaker: They're using more tokens.
- 2:41
- Unidentified speaker: But it's a combination of factors.
- 2:43
- Unidentified speaker: One, AI is getting cheaper because there's so much competition between OpenAI and Anthropic and stating price cuts and also introducing more efficient models.
- 2:52
- Unidentified speaker: They're getting better deals on that spend because OpenAI and Anthropic are now offering contracts for large enterprises that introduce new discounts.
- 3:01
- Unidentified speaker: And there's also now pressure from uh you know the the sort of emergence of new and highly performant models at the standard and light tiers.
- 3:11
- Unidentified speaker: So that's also pushing down some of the spend.
- 3:14
- Unidentified speaker: Can I ask a few questions about this?
- 3:16
- Unidentified speaker: Because I think it's you are the basis for that statement is basically looking at, well, one, the overall data set, let me give you my perception of kind of what the sample is and tell me where I'm wrong.
- 3:29
- Unidentified speaker: But I uh I expect that RAMP has basically like a higher share of call it more aggressive companies, more VC companies, probably more California, certainly more US.
- 3:42
- Unidentified speaker: Uh, and so then that would explain some of the over-indexing to AI spend generally amongst your data set relative to like this what the census measures as well, right?
- 3:52
- Unidentified speaker: Um and and is there, is it does it are there other kind of like biases in the sample?
- 3:58
- Unidentified speaker: And I don't mean bias in a in a bad way.
- 4:00
- Unidentified speaker: I just mean just to understand kind of like the scope of the data.
- 4:03
- Unidentified speaker: Yeah, yeah, yeah.
- 4:04
- Unidentified speaker: Well, you know, you have a point, right?
- 4:07
- Unidentified speaker: I think ramp data absolutely does skew toward tech forward businesses that are likely to adopt AI in the first place and also likely to adopt it pretty intensely.
- 4:16
- Unidentified speaker: Yeah.
- 4:17
- Unidentified speaker: At the same time, you know, it's the way that private sector data sets typically deal with this is by breaking out data set by like size and sector.
- 4:26
- Unidentified speaker: We do that, but I actually think that's a less helpful way of breaking out AI spend, which truly is highly concentrated and highly skewed.
- 4:34
- Unidentified speaker: I think the most important measure that we publish is showing AI spend at the median, the top 10%, and the top 1% level.
- 4:42
- Unidentified speaker: Yeah.
- 4:43
- Unidentified speaker: And the median.
- 4:44
- Unidentified speaker: That's exactly the chart I'm looking at.
- 4:45
- Unidentified speaker: Okay.
- 4:46
- Unidentified speaker: Yeah, yeah, yeah.
- 4:46
- Unidentified speaker: So the median firm on our platform.
- 4:48
- Unidentified speaker: 15 bucks, right.
- 4:50
- Unidentified speaker: That spends on AI only spends about $15 per employee per month.
- 4:53
- Unidentified speaker: So that's like a chat subscription for every employee, and maybe someone else has a coding agent.
- 5:01
- Unidentified speaker: And that's just the median AI spender.
- 5:04
- Unidentified speaker: So I think one of the really important ways of tracking this market is to first of all do away with some of the, you know, uh uh sort of balancing metrics that we've gotten used to.
- 5:16
- Unidentified speaker: They're just not useful nowadays.
- 5:19
- Unidentified speaker: At least in this market where AI spend is so concentrated and also correlated within that concentration, what you really want to look at is intensity of spend.
- 5:26
- Unidentified speaker: Something that is by itself becoming increasingly difficult to define, by the way.
- 5:31
- Unidentified speaker: Have you have confusion?
- 5:33
- Unidentified speaker: It's like how how quickly, or is there an accelerating pace at which uh spend is making its way to the median type company, or not really?
- 5:47
- Unidentified speaker: The the leaders are continuing to pull away from the rest of the pack?
- 5:52
- Unidentified speaker: Well, the median firm is continues to grow its spend on AI.
- 5:57
- Unidentified speaker: I mean, just a month ago it was by maybe about $12 per employee per month, now it's $15, right?
- 6:02
- Unidentified speaker: Okay, so there is some growth happening there.
- 6:05
- Unidentified speaker: Meanwhile, the top 1% firm is starting to slow its spend, the top 10% firm also continues to grow its spend.
- 6:11
- Unidentified speaker: All to say that we will continue to see greater proliferation across the median firm.
- 6:18
- Unidentified speaker: You know, we'll continue to see them use AI more intensely.
- 6:21
- Unidentified speaker: My predictions there have not changed.
- 6:24
- Unidentified speaker: Uh where it matters is how far is it going to go and how is it going to end.
- 6:31
- Unidentified speaker: Now, the top 1% firm, remember a few months ago, even you know, we were talking about how the top 1% firm should be spending as much on AI as it does per employee on a software engineer's salary.
- 6:42
- Unidentified speaker: So that's already not happening at the top 1% firm.
- 6:46
- Unidentified speaker: Top 1% firm spends upwards of 7 to 10K per employee per month, and it's starting to slow.
- 6:52
- Unidentified speaker: Can I can I just clarify or question kind of that assertion it's starting to slow is based on the last two months, I'm assuming, within the data set?
- 7:02
- Unidentified speaker: As in year over year, clearly top 1% is up something like four or five X year over year.
- 7:09
- Unidentified speaker: But you're saying that, well, measured on September versus July.
- 7:15
- Unidentified speaker: I guess it's not two months, it's uh whatever that works too.
- 7:18
- Unidentified speaker: Yeah, September versus July, September, and our earliest data from October shows a slowdown from the hypergrowth that we observed over the rest of the beginning and middle half of this year.
- 7:27
- Unidentified speaker: And meanwhile, you know, we've you know, people have written to me on Twitter being like, oh, it's seasonal.
- 7:32
- Unidentified speaker: And it's like, well, it's not really seasonal, because if you look at token volumes over that time, they've continued to increase.
- 7:39
- Unidentified speaker: So there are important market dynamics here, including that over that same period, the price of AI has fallen by almost half.
- 7:48
- Unidentified speaker: So you could actually double your usage of AI and just use the models as is and still see token spend flatline or slow down because of those market dynamics.
- 8:00
- Unidentified speaker: I try to base most of our work for that reason, you know, in multiple metrics that tend to be moving in the same direction.
- 8:07
- Unidentified speaker: And right now, I do think there are multiple metrics moving in a direction that suggests not that AI spend is going to, you know, crash, but that if you're a top 1% firm, there are now measures that allow you to control that spend and also just allow you to sort of use cheaper models because they're getting cheaper over time and gain from those models at a lower price than you could have a year ago.
- 8:35
- Unidentified speaker: And that's what we're observing in the top 1%.
- 8:38
- Unidentified speaker: And so, Brett, when you when you hear this, how do you think about the way that you look at the future?
- 8:44
- Unidentified speaker: And is it just that we're still so early on and the median, I mean, the gap between the median firm and the top 10%, top 1% is so massive, and that as this diffuses, necessarily everyone needs to move up to that higher point.
- 9:02
- Unidentified speaker: Is your car about to drive away?
- 9:04
- Unidentified speaker: Yeah, sorry.
- 9:05
- Unidentified speaker: I had a phone call interrupting.
- 9:07
- Unidentified speaker: Um I think, well, I think one, you can't read it independent of the dynamics between open air and anthropic in particular, because they are um they're both anthropic is preparing itself for um going public and needs to continue to deliver aggressive top line growth, which you know, if you talk to them leading up to the IPO, they'll say we are being more aggressive on pricing and kind of like we basically invested in pricing to drive token volume.
- 9:38
- Unidentified speaker: And then over time, that'll yield a um demand elasticity response.
- 9:43
- Unidentified speaker: And that's what they've seen over time.
- 9:45
- Unidentified speaker: If you look at the ramp data on Opus 5.5, um, that's clearly taking a ton of share.
- 9:51
- Unidentified speaker: Um, you know, and I don't doubt what you're saying, Aura, that they're going and trying to sign longer-term deals.
- 9:56
- Unidentified speaker: If you think of like the doubt plaguing the overall market, it is will OpenAI and Anthropic be good for the bill?
- 10:04
- Unidentified speaker: Basically, all these companies are lining up to supply hundreds of billions of dollars in infrastructure that they are promising to rent and use.
- 10:11
- Unidentified speaker: And so then therefore they have to continue to deliver revenue growth in order to afford and pay that set of bills.
- 10:18
- Unidentified speaker: Uh, and so um it's in the interest.
- 10:21
- Unidentified speaker: I mean, I know OpenAI had posted about it a few months ago, and it doesn't surprise me if Anthropic's doing the same.
- 10:28
- Unidentified speaker: It's in their interest to try to line up longer-term, more secure revenue against contracts so that they can lay that against the liabilities they have to assuage that concern.
- 10:39
- Unidentified speaker: Um, I do wonder if, like, if you think about naturally, like the state of the way the technology should work, is there's a set of really aggressive companies trying to experiment with if we use as much of this stuff as possible, how quickly can we accelerate?
- 10:53
- Unidentified speaker: Can we outcompete?
- 10:54
- Unidentified speaker: Some of them will fail, some of them won't.
- 10:56
- Unidentified speaker: Like a 1% portion of kind of your sample is necessarily a smaller part of the sample.
- 11:02
- Unidentified speaker: And then there's a set of call it like, you know, median enterprises who are like a bigger blend who are figuring things out and trying to take things on in a call it more measured or kind of like a uh uh uh slightly carefuler way.
- 11:19
- Unidentified speaker: Necessarily that's actually the over time, you know, whatever remains on the back end of the singularity here, I suspect it's gonna be more AI aggressive folks, but it's gonna be the median of the sample, not the top 1% spenders, unless uh that 1% becomes as big as the entire rest of the median.
- 11:38
- Unidentified speaker: So I do think that there has to be a diffusion that carry takes place here.
- 11:42
- Unidentified speaker: And I'd just be careful about like a month-to-month change or call it a uh three-month change, isn't does not mean that kind of like even amongst those top 1%, the the you know, they've they're gonna stop growing spend.
- 11:59
- Unidentified speaker: I would be extraordinary.
- 12:00
- Unidentified speaker: And I don't know if you disagree with this, but if you go forward six months from now, do you honestly believe that the top 1% are gonna be like, no, we're good.
- 12:09
- Unidentified speaker: Basically, like $9,000 per employee per month is like the optimum spend here because it's it's very contra at least what the Frontier Labs are doing, which are the fastest revenue growth growing companies of all time.
- 12:21
- Unidentified speaker: I don't necessarily think AI spend is going to drop or that it's gonna completely flatline, but I do think the market should take note.
- 12:28
- Unidentified speaker: Maybe the market actually is taking note.
- 12:29
- Unidentified speaker: Say people in tech should probably take note of the fact that there are several dynamics that are headwinds to rising AI spend in ways that, you know, when I propose these thoughts in many tech circles, they're they're seen as uh it's like they're seeing the data for the first time.
- 12:47
- Unidentified speaker: It's not just the fact that we're observing a slowdown in spent, right?
- 12:50
- Unidentified speaker: There are dynamics that are driving that slowdown that are only going to get stronger over time.
- 12:57
- Unidentified speaker: Competition in the markets, which will drive down the price of AI further and further, because it's not just between open I and enthropic anymore.
- 13:04
- Unidentified speaker: Now, again, wait, I have a question about that, actually.
- 13:07
- Unidentified speaker: Like you assert you asserted that, but like in the in the ramp data, it's basically XAI is taking a little share here and um and what 95% of the market is open AI and enthropic or something like that.
- 13:22
- Unidentified speaker: Yeah.
- 13:23
- Unidentified speaker: I mean, so that's a pretty healthy oligopoly.
- 13:27
- Unidentified speaker: I don't know that there's a now the the statement that meta is gonna take share, is that based on data that you're seeing in October or something, or is that more of an assertion from kind of how they're approaching the market?
- 13:40
- Unidentified speaker: We don't we don't see a significant share for Meta at all today.
- 13:44
- Unidentified speaker: I mean, it just sort of came onto the uh market just recently for enterprise.
- 13:48
- Unidentified speaker: But what I'm saying is that I don't think people should underrate how competitive this market is.
- 13:54
- Unidentified speaker: And within a matter of months, we can and have already seen new entrants take more than half the market.
- 14:03
- Unidentified speaker: I mean, that's kind of what Anthropic did.
- 14:04
- Unidentified speaker: It's a rise in market share happened over the span of like a year to become both the most popular model and the largest uh in terms of business spend.
- 14:14
- Unidentified speaker: And so we shouldn't underrate that.
- 14:16
- Unidentified speaker: And I wouldn't underrate Meta either, because part of what's so interesting about this market being a market for what is essentially B2B spend is that it's actually developed in many ways uh as a sort of lagging indicator of how the consumer market has developed.
- 14:34
- Unidentified speaker: OpenAI's initial rise came out of its standard default status as being the consumer product in ChatGPT.
- 14:42
- Unidentified speaker: Anthropics rise, yes, driven by coding agents.
- 14:46
- Unidentified speaker: But if you really look and line up when it started to rise rapidly, it actually coincides with a few other things.
- 14:54
- Unidentified speaker: One, it's introduction of Claude Cowork, which followed coding agents and cloud code.
- 14:58
- Unidentified speaker: And then two, uh, its pivot to consumer, Super Bowl ads, the Department of Defense labeling it as supply chain risk, and then Katy Perry, for some reason, posted about how she's using anthropic.
- 15:12
- Unidentified speaker: That was the season in which we saw Anthropic's run up.
- 15:16
- Unidentified speaker: And so I wouldn't underrate the extent to which consumer strength as a product can actually drive enterprise adoption and spend in this market, because with Grok, you know, with XAI's recent models, too, we're noting a recent rise, slow but meaningful, that also coincides with the introduction of Grok bot.
- 15:36
- Unidentified speaker: Yeah.
- 15:36
- Unidentified speaker: So I agree that like consumers factors matter.
- 15:40
- Unidentified speaker: Sure.
- 15:41
- Unidentified speaker: The consumer and the enterprise are entangled.
- 15:43
- Unidentified speaker: I would just I don't think that like you can look at you know Cloud's consumer adoption, and that's clearly not a line leading indicator for how they've penetrated the market.
- 15:53
- Unidentified speaker: As in my view of the that they are going to customers that typically wouldn't take on call it startup enterprise B2B, and they're able to deliver them this product that is magic, but that they also feel safe with and that is heavily customizable to the enterprise.
- 16:11
- Unidentified speaker: And they can do that customization very quickly because of I mean they have their AI magic in order to help them do it.
- 16:18
- Unidentified speaker: Um and so I think that they have hit, you know, just a sweet spot in being able to deliver what enterprises want.
- 16:26
- Unidentified speaker: And yes, the awareness of a Super Bowl ad helps, you know, that that stuff helps, but I I don't know that I think, you know, the contrast, if if we really were to believe that kind of like the consumer adoption were led naturally to the enterprise adoption, then you would say open AI is best positioned, and we've lost our guest, but and and maybe disagree hard disagree.
- 16:54
- Unidentified speaker: Well, and and like Google would have a much bigger play.
- 16:57
- Unidentified speaker: I don't know.
- 16:58
- Unidentified speaker: Sam, Sam, what do you think about kind of AI adoption here?
- 17:02
- Unidentified speaker: Should should should we be taking Aura at his word that oh, actually at the leading edge, the top one percent spend is moderately slowing and therefore spend can slow dramatically.
- 17:16
- Unidentified speaker: Well, I think spend, of course, spend can slow, but I don't oh here we we've got aura back.
- 17:22
- Unidentified speaker: Aura, welcome back.
- 17:25
- Unidentified speaker: You know, if if if that was a commercial break, great.
- 17:27
- Unidentified speaker: Yeah, yeah.
- 17:28
- Unidentified speaker: If it was fundamentally true, then Gemini in theory should be doing better than it is in enterprise.
- 17:35
- Unidentified speaker: Um and what what I wanted to touch on as well here is Tramoth has a point that he's made, and I'll just read a tweet.
- 17:46
- Unidentified speaker: He's like, I suspect you know, 70 to 80% of revenues at AI companies are from other AI companies who've raised money from the venture industry.
- 17:54
- Unidentified speaker: It's not a circular revenue problem, but it's a revenue quality problem.
- 17:58
- Unidentified speaker: In the data that you're looking at, are these top 1% spenders, are they unprofitable VC-backed companies that are doing this?
- 18:09
- Unidentified speaker: Or do you think they're more grounded and can sustain that type of spend?
- 18:16
- Unidentified speaker: Well, I I I agree with Chamath's point here.
- 18:20
- Unidentified speaker: I do think he's actually right about it, that it's not a circular revenue problem, but it is a revenue quality problem.
- 18:25
- Unidentified speaker: If you look at that top 1% of firms on our platform that spend on AI, yes, they are by and large VC-backed tech companies.
- 18:33
- Unidentified speaker: So that's all fine as long as AI spend continues to grow and that they continue to grow in terms of their own sort of customer bases.
- 18:42
- Unidentified speaker: But that concentration problem is also a correlation problem in that if the AI trade starts to slow down, all of these companies will simultaneously slow down their AI spend.
- 18:56
- Unidentified speaker: And that could lead to some cascading impact across the rest of the sector.
- 19:01
- Unidentified speaker: Well, it's almost a nice idea.
- 19:03
- Unidentified speaker: If these companies, if you can't raise the next up round, then your AI spend stops.
- 19:12
- Unidentified speaker: Yeah, or you slow it down, or you you come under more cost conscious pressures.
- 19:16
- Unidentified speaker: You know, and we do, by the way, see that across some of the you know, there are, you know, RAMP, for example, too, is easily a top 1% business in terms of AI spend.
- 19:26
- Unidentified speaker: We have money coming in as a profitable company.
- 19:28
- Unidentified speaker: And yet, as you become advanced in AI adoption, you also become more advanced in controlling that spend.
- 19:36
- Unidentified speaker: You implement controls against uh usage of frontier models on automated tasks, you implement better defaults, you start to use routers to optimize certain tasks.
- 19:46
- Unidentified speaker: You know, all of those are natural business behaviors.
- 19:50
- Unidentified speaker: And, you know, we haven't observed them necessarily so far from that top 1% business because those top 1% businesses are also fast-growing businesses that have money coming in.
- 19:58
- Unidentified speaker: They don't have to make that many compromises.
- 20:01
- Unidentified speaker: Yeah, but at a certain point, they do, and we are starting to see that across some of them.
- 20:07
- Unidentified speaker: I that's one of the main drivers of slowing spend at the top 1%, just using AI more, just using cheaper models over time.
- 20:13
- Unidentified speaker: My gut read.
- 20:14
- Unidentified speaker: But you know, it's uh right.
- 20:16
- Unidentified speaker: I was just gonna say, it's like my gut read of this all of this is it's like I agree.
- 20:23
- Unidentified speaker: It's like three months do not make a trend.
- 20:27
- Unidentified speaker: And it seems it's like everything has happened so fast from the AI side of things, from the venture side of things, and from company side of things.
- 20:36
- Unidentified speaker: To your point, Aura, it's like people just need to people are saying, woo, it's like we're coming up on year end too.
- 20:42
- Unidentified speaker: It's like, whoa, what just happened?
- 20:45
- Unidentified speaker: And I do think it's like the V, right?
- 20:47
- Unidentified speaker: I do, it's like people are looking around and they're like, What, we're talking about like hundred billion dollars, like it's ten dollars all of a sudden.
- 20:54
- Unidentified speaker: Like we should reconsider and just make sure we're going in the right direction here.
- 20:59
- Unidentified speaker: On the VC side, Brett, you're probably tied in closer than I am.
- 21:02
- Unidentified speaker: But it's like I've my sense is people are like, wait, can it's like how many hundred billion dollar companies can there be that are you know two years old?
- 21:16
- Unidentified speaker: Um some some you know hesitation on that front.
- 21:20
- Unidentified speaker: But Brett, you you were reacting.
- 21:22
- Unidentified speaker: Go ahead.
- 21:23
- Unidentified speaker: Well, I I I just think or my perception is the beginnings of like people evaluating spend more happened in the second quarter, third quarter.
- 21:33
- Unidentified speaker: And maybe this is the back end of that, where it's kind of like anthropic was uh extraordinarily expensive, but everybody had gotten addicted to anthropic and and using those models.
- 21:43
- Unidentified speaker: And so then open AI looked on a per task cheaper, and so people were evaluating shifting to open AI.
- 21:49
- Unidentified speaker: OpenAI was providing Luna, which is a big um spin driver in your data set, which is very efficient on it, and and they cut Luna's prices 80% to then win marginal share.
- 22:00
- Unidentified speaker: Um, and like I think it's interesting that 6.1 doesn't show up in your data set.
- 22:05
- Unidentified speaker: I feel like that's a very efficient set of models.
- 22:08
- Unidentified speaker: And then Opus 5.5 is also like call it more efficient than what Enthropic had been delivered in the past delivering in the past.
- 22:15
- Unidentified speaker: So I, you know, I do think that you're absolutely right.
- 22:19
- Unidentified speaker: If you have performance per cost having like every two months or whatever one, that means you can keep up your same set of activities and half your costs every two months, you know.
- 22:30
- Unidentified speaker: And um, so the I guess the open kind of business discovery question is do kind of like these companies that are approaching things in an AI-peeled way, you know, do they really win share from kind of like this other set of businesses?
- 22:47
- Unidentified speaker: And there is going to be some friction in that necessarily, you know, like you're fighting against a better product, does not a business make.
- 22:55
- Unidentified speaker: You have to cross it with like winning the battle against status quo.
- 22:59
- Unidentified speaker: Uh, and that can have time lags in it.
- 23:02
- Unidentified speaker: I I think it's hard to think the model the model companies do need to drive more adoption, more intense adoption from that median and top 10% firm.
- 23:13
- Unidentified speaker: Fundamentally, they do need to do that.
- 23:15
- Unidentified speaker: At the same time, today, the most commercially advanced form of AI is coding agents, which are clearly productivity enhancing, at least for the tech sector and for that job category that uses them actively.
- 23:31
- Unidentified speaker: But it's hard to for me to see a world in which the median business uh uh gets as much value out of a coding agent as the typical software engineer at a tech company.
- 23:44
- Unidentified speaker: Well, and so I don't know.
- 23:46
- Unidentified speaker: I mean if your business is not that they won't get value out of them, but I just don't see I don't think we can expect the median firm to get to those levels.
- 23:54
- Unidentified speaker: And I do think most of the predictions around AI are around well, we don't know what the models are gonna do yet.
- 24:00
- Unidentified speaker: New products are gonna be released that are gonna be the next coding agent revolution, and that's all fine and good.
- 24:07
- Unidentified speaker: But that's why my post is called Cracks in the AI thesis is that hey, if you do have some of these notions about what's happening today, I will provide some data to suggest that you should maybe ground them those assumptions a little bit because they're not happening to the extent that people think they are.
- 24:28
- Unidentified speaker: Well, imagine I have a wait, imagine I have a thousand firms out there.
- 24:33
- Unidentified speaker: They're just out there existing in the world.
- 24:35
- Unidentified speaker: These are non-tech firms, they don't have uh I am my connection's unstable.
- 24:40
- Unidentified speaker: Can you hear me?
- 24:41
- Unidentified speaker: Hey, you're good.
- 24:41
- Unidentified speaker: Keep going.
- 24:42
- Unidentified speaker: Yep.
- 24:43
- Unidentified speaker: Okay.
- 24:43
- Unidentified speaker: Imagine there's a thousand firms out there.
- 24:45
- Unidentified speaker: They don't have software developers in these firms, these are non-coding firms, and then I take, I don't know, call it 10 of them, and I absolutely cripple them with like a cyber attack from open weight models, right?
- 24:58
- Unidentified speaker: Like the the the even though they're not tech firms, they do have like a lot of tech exposure that they will need to pay to basically protect.
- 25:09
- Unidentified speaker: Um, I think that cybersecurity is going to be a spin category that everybody's gonna have to spend on to because everybody's infrastructure is its own unique kind of like set of nested like stuff that they've uh you know acquired or paid for that is uh needs its own essentially security force in an age where any open channel is a recipe for infiltration.
- 25:34
- Unidentified speaker: So I was gonna say, yeah, I mean I think that will be a rising spend category too, right?
- 25:39
- Unidentified speaker: I just, you know, I um I still part of the the message that of of AI's proliferation, at least in 2026, though, and in the last couple months, is that you act you can actually get pretty significant productivity gains with a relatively low amount of spend.
- 25:57
- Unidentified speaker: You know, particularly if you are a business outside of the tech sector, you know, you don't need to spend that much on AI to get a pretty good ROI, even if you want to AI pill your small business or your doctor's office.
- 26:10
- Unidentified speaker: See, I think I disagree with that.
- 26:12
- Unidentified speaker: As in having used these models really aggressively and having dialed back and up the spend, it's they're still really, they're really powerful and they're incredibly still not very good.
- 26:24
- Unidentified speaker: As in, give it, give it, give it, give it a task that would take you half an hour and really try to have it do it and deliver the business product outcome.
- 26:33
- Unidentified speaker: And it's just not good.
- 26:34
- Unidentified speaker: Like for like fighting bureaucracy and that kind of stuff, I think they're reasonable because they kind of like can, but but for actual business product outputs, at least core to what we try to do, it requires a lot of hand holding.
- 26:47
- Unidentified speaker: And the degree to which you can loosen on the reins is very, very correlated with kind of like the power of the model you're sending after the problem and the amount you're spending on it.
- 26:56
- Unidentified speaker: And so I I think it the at least at a if you if you distinguish between like exploration function for business and exploitation function for business, I think exploration and head-to-head competition, the the marginal like strategic necessity will be to spend on the spend on a model on the frontier and actually increase spend over time, regardless of your job category, I think.
- 27:22
- Unidentified speaker: Well, that does is this a huge opportunity for private equity.
- 27:27
- Unidentified speaker: It's like that's what all I hear here is it's like the gap between 1% and the median is like well, yeah.
- 27:36
- Unidentified speaker: If you can effectively forget spending as much as the top 1%, if you can go into a median firm and effectively spend like a 10% firm is spending, then it's like you should be able to outcompete the median company in your industry here.
- 27:54
- Unidentified speaker: It's like that certainly seems the case.
- 27:56
- Unidentified speaker: And then the other my optimistic read of the spend slowdown too is it's like the models are moving, the capability of the models are moving faster than even the 1% of companies.
- 28:09
- Unidentified speaker: It's like there's more power being released by the comp by the technology than even the leading edge companies can figure out how to spend enough to keep up with it.
- 28:21
- Unidentified speaker: And so all of this makes me extremely bullish on on the technology.
- 28:25
- Unidentified speaker: R we're we are wrapping up.
- 28:27
- Unidentified speaker: I want to ask you okay, five years from now, or if you want to do three years, whatever, whatever medium time frame you feel comfortable with, what do you think the spend of the top 1% and of the median company will be on AI?
- 28:47
- Unidentified speaker: It's hard to say the actual level because uh the definition of what counts as AI spend also continues to change over time.
- 28:55
- Unidentified speaker: Like you kind of have to draw the line at some point.
- 28:58
- Unidentified speaker: You know, what do you see as you can?
- 29:01
- Unidentified speaker: Is kind of gonna be AI versions of spend at some point.
- 29:05
- Unidentified speaker: Um I'll say for the median firm, it's probably gonna be easier for me.
- 29:10
- Unidentified speaker: The median firm per employee under a thousand per employee per month in five years.
- 29:20
- Unidentified speaker: Um and um not that they won't use AI extremely effectively, I actually think it's gonna be pretty table stakes.
- 29:28
- Unidentified speaker: I think for the median firm today, being able to use AI is an absolute edge.
- 29:33
- Unidentified speaker: It's like being the restaurant in your neighborhood who used Google Ads for the first time.
- 29:38
- Unidentified speaker: You are gonna get more customers in the door, and your competitors who can't get customers in the door because they're not advertising in the modern era may end up going out of business.
- 29:48
- Unidentified speaker: We're gonna see the same exact thing here.
- 29:50
- Unidentified speaker: At the same time, it's then gonna become pretty table stakes.
- 29:53
- Unidentified speaker: Yes, that will drive more adoption.
- 29:56
- Unidentified speaker: Might drive actually some more spend as well, but it also means the median firm can get pretty significant ROI without that much spend.
- 30:05
- Unidentified speaker: So just just to so $1,000 a month times call it a billion knowledge workers globally, that's against uh $12 trillion in spend in 20.
- 30:14
- Unidentified speaker: So this is from the AI bear.
- 30:16
- Unidentified speaker: So I think we're in good shape.
- 30:21
- Unidentified speaker: Um I wouldn't wouldn't say a thousand, I mean, because I said under a thousand per employee per month.
- 30:25
- Unidentified speaker: So I'm giving myself a lot of room.
- 30:28
- Unidentified speaker: Uh but at the same time, I would say it's you know it's not limited to you know knowledge workers, right?
- 30:34
- Unidentified speaker: Like I do think there will be applications outside of that.
- 30:38
- Unidentified speaker: I just think that there's gonna be a large range.
- 30:43
- Unidentified speaker: But yeah, I'm definitely going under a thousand is what I mean.
- 30:46
- Unidentified speaker: I realize that sounds like either extremely bullish or just a massive range.
- 30:50
- Unidentified speaker: All to say that like in the circles that I inhabit sometimes, like there are predictions that the median firm will spend, you know, tens of thousands per employee per month.
- 31:01
- Unidentified speaker: And I'm trying to benchmark it around like a purchase of a computer per person, you know, like a thousand dollars for a computer.
- 31:09
- Unidentified speaker: Yeah.
- 31:09
- Unidentified speaker: Well, if we go.
- 31:10
- Unidentified speaker: You know what per month is probably wrong for that.
- 31:12
- Unidentified speaker: You don't buy a computer every month.
- 31:14
- Unidentified speaker: Yeah.
- 31:14
- Unidentified speaker: Unless you're very uh Yeah, I'm gonna get back to you guys when I adjust my understanding of like months and years.
- 31:21
- Unidentified speaker: Um so maybe a thousand per year.
- 31:25
- Unidentified speaker: I don't know, it's hard to say.
- 31:26
- Unidentified speaker: I just so you think anywhere from 10x to 100x?
- 31:30
- Unidentified speaker: From 10x to 100x spend from where it is today?
- 31:34
- Unidentified speaker: I should probably just avoid giving levels because I don't like giving people some sense of where I think the market's gonna land so much as hey, there are some tailwinds and some headwinds, and today I think there are more headwinds to that growth than not.
- 31:46
- Unidentified speaker: All right.
- 31:46
- Unidentified speaker: Yeah.
- 31:48
- Unidentified speaker: Thank you so much for joining us.
- 31:50
- Unidentified speaker: Um, we love seeing all the research you put out there.
- 31:54
- Unidentified speaker: And so look forward to reading more as you get more data and publish it.
- 31:59
- Unidentified speaker: ARC Investment Management LLC is an SEC registered investment advisor.
- 32:03
- Unidentified speaker: 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.
- 32:12
- Unidentified speaker: ARC, Wolf Financial and Public make no representation or warranties regarding and expressly disclaim any responsibility or liability for any acts, omissions, losses, or damages of the other entities or their affiliates, officers, employees, or agents.
- 32:27
- Unidentified speaker: No company shall be liable for any indirect, incidental, consequential, or special damages arising out of or in connection with this podcast or its content.
- 32:36
- Unidentified speaker: The information provided in this show is for informational purposes only and should not be used as the basis for any investment decision, and it's subject to change without notice.
- 32:46
- Unidentified speaker: 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 needs.
- 32:59
- Unidentified speaker: All statements made regarding companies or securities are strictly beliefs and points of view held by ARC and or show guests, and are not endorsements by ARK of any company or security or recommendations by ARC to buy, sell, or hold any security.
- 33:14
- Unidentified speaker: Historical results are not indications of future results.
- 33:17
- Unidentified speaker: Certain of the statements contained in this show may be statements of future expectations and other forward-looking statements that are based on ARC'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 statements.
- 33:36
- Unidentified speaker: ARK assumes no obligation to update any forward-looking information.
- 33:41
- Unidentified speaker: ARK and its clients, as well as its related persons, may, but do not necessarily have financial interest in securities or issuers that are discussed.
- 33:50
- Unidentified speaker: Certain information was obtained from sources that ARC believes to be reliable.
- 33:54
- Unidentified speaker: However, ARC does not guarantee the accuracy or completeness of any information obtained from any third party.
More from 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.
Did The Midterms Just Break The AI Trade? | The Brainstorm 157
Episode discussion: Midterm election impacts AI regulation and compute economics, favoring infrastructure owners over frontier labs
Speakers’ claims: 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.