Ara Karazian | The a16z Show Summary

Participants: Ara Karazian (Lead Economist, Ramp), Jack Farley, Max Wheathey

1. Key Themes

SaaSpocalypse Is a Narrative, Not a Data Story

The most dominant theme of the episode is that the widely circulated thesis that AI will destroy SaaS companies and transform how software is bought is simply not backed by actual business spending data. Seat-based and flat platform fee contracts still dominate overwhelmingly.

"Seat-based contracts are still about 65, 75% spend, flat platform fees about 20, 30%. Even amongst the companies that have offered some kind of token-based pricing for their SaaS tools, we're only seeing uptake about half a percent of spend on those platforms." [00:02:55]

"SaaSpocalypse as a pronouncement has come way too soon and is typically not informed by actual business behavior." [00:06:42]

The AI Model Layer Is Commoditizing Fast — Cost Pressure Is Real

Businesses that are heavy AI spenders are already routing around expensive frontier models, and the economics of staying loyal to a single model provider are increasingly unsustainable. Token costs for high-intensity spenders grew 13x in one year, and routing tools like OpenRouter are quietly gaining ground.

"For the typical business that spends on tokens...token costs for that business have increased 13x just over the last year...you project out that 13x and you get to an extremely unsustainable spend path." [00:10:24]

"OpenAI and Anthropic have no incentive to offer an auto-routing product that allows you to lower your AI spend. Because they make money on tokens. Anthropic and OpenAI, 80% of their business revenue is token-based." [00:13:21]

The Real AI Growth Is in the Ecosystem Layer, Not the Model Layer

The most interesting and fastest growing companies in this AI cycle are not the model providers themselves, but the infrastructure, workflow, and application layers built around them — categories that the model companies structurally cannot compete in.

"A lot of the growth that's happening amongst vendors...a lot of the really compelling growth is not necessarily with the model companies themselves...AEO as a category, so answer engine optimization...huge growth. I mean, that sector did not exist. The companies that are making that software are not the same companies that were selling the SEO software." [00:22:52]

"Not a product offered directly by the model companies, and arguably the model companies could never offer that product because Anthropic can offer it for Anthropic, but they wouldn't be able to offer it across all of the different models in any effective way." [00:23:44]

2. Contrarian Perspectives

Model Company "AI Will Replace All Jobs" Messaging Is TAM Posting, Not Conviction

Most people take frontier AI labs' proclamations about replacing labor at face value. But the actual motivation is likely investor relations — maximizing perceived total addressable market to justify trillion-dollar valuations — not a genuine belief held by economists or even the companies themselves.

"I would love to see some research about why the model companies keep saying that it's going to destroy all jobs because I don't see why that is helpful to them... that's not even the position of most economists." [00:20:34]

"They're trying to justify a $2 trillion valuation, a $10 trillion valuation. That's what they're shooting towards. And so if you want to justify that, well, you know, the entire labor market is a pretty big total addressable market." [00:21:59] — Jack Farley

DeepSeek's Supposed Dominance Among Startups Was Largely Myth

The narrative that 80% of VC-backed startups were using DeepSeek was never supported by actual spend data and DeepSeek never cracked even 1% of firms on Ramp's platform.

"I don't think 80% of venture-backed companies are using DeepSeek. We did — look, when DeepSeek first came onto the scene...we did see a spike in adoption for DeepSeek. But it never hit even 1% of firms on the platform." [00:15:07]

"No one wants to use DeepSeek for security purposes. Even if you're locally hosting it, there's just this perception around it, particularly if you're building a product that appeals to businesses or consumers." [00:16:00]

Google's AI Adoption Is Massively Underreported Because It's Free

Google appears to be losing the model adoption race when looking at paid spend data, but this is structurally misleading — the majority of Google AI usage flows through Google Workspace at no additional cost, making it invisible in spend trackers like Ramp's index.

"We only track paid adoption. So there are a lot of firms that are using Google, but they're using it through Google Workspace, which integrates Gemini for free...Google's definitely underrated. Certainly has a distributional advantage in that it's at all businesses that use Google Workspace." [00:16:31]

Companies Adopting AI Most Aggressively Are Adding Headcount, Not Cutting It

The common narrative is that AI adoption leads to layoffs. The data suggests the opposite at fast-growing firms — AI adopters tend to be high-growth companies with more work to do, not companies shrinking their workforce.

"Firms that adopt AI are typically quite fast growing and they have a lot of opportunity ahead of them...my instinct is that the firms that are adopting AI particularly well probably have a lot of work for people to do as well." [00:19:58]

Being Cautious on AI Adoption Is Becoming a Deliberate Competitive Positioning Strategy

Rather than racing to adopt AI, at least one major firm (Deloitte) is explicitly positioning conservative AI adoption as a feature — a differentiator to clients who are nervous about AI risk in sensitive domains like audit and finance.

"It was fascinating because I actually wasn't sure until I got to the Deloitte part why they were doing this advertisement for the other accounting firms. And then I realized, oh, this is an explicit effort by Deloitte to position itself as not anti-AI, but certainly a little bit more conservative about implementation of AI." [00:30:30]

3. Companies Identified

Ramp Corporate spend management platform. Mentioned as the source of a unique, large-scale data set covering 50,000 businesses and $100 billion in annual spend, enabling real empirical research on AI and SaaS adoption trends.

"We see the spend data from 50,000 businesses, $100 billion of annual spend." [00:02:00]

Anthropic Frontier AI model company. Mentioned as having just overtaken OpenAI as the most popular model used by businesses on Ramp's platform — a significant milestone signaling model market share is highly fluid.

"Anthropic just did that with OpenAI, now the most popular model used by businesses, according to Ramp data." [00:00:00]

Cursor AI-powered coding assistant. Mentioned as a prime example of a product-layer company that displaced a major incumbent (GitHub Copilot), and as a strategic acquisition target for XAI precisely because it drives model adoption through superior developer experience.

"Cursor did that with GitHub Copilot...that's the bull case for something like Cursor. Which can compete on the developer experience in that way. In a way that OpenAI and Anthropic, in the absence of Cursor, are not incentivized to do." [00:14:13]

OpenRouter Model routing infrastructure platform. Mentioned as a growing beneficiary of enterprise cost pressure — high-intensity AI spenders are increasingly routing through OpenRouter to access cheaper and open-source models.

"It's those firms specifically that are increasingly shifting some of their AI spend over to platforms like OpenRouter...3% of spend on our platform goes directly through an OpenRouter versus an Anthropic." [00:10:53]

Profound New AEO (Answer Engine Optimization) software company. Flagged as a fast-growing emerging vendor in a category that did not exist previously, helping companies track and optimize their visibility in AI model responses.

"Profound is a new one growing extremely quickly...not a product offered directly by the model companies, and arguably the model companies could never offer that product." [00:23:18]

Figma Collaborative design platform. Mentioned as a counter-example to SaaSpocalypse — despite a new Anthropic design product launching in direct competition, Figma was one of the fastest growing vendors on Ramp's platform.

"Figma was one of the fastest growing vendors on our platform, that businesses were continuing to buy from it." [00:03:53]

Perplexity AI-native search company. Highlighted as a fast-growing vendor on Ramp's platform, growing despite being in the AI model-adjacent space, specifically because it is serving use cases the model companies haven't competed on yet.

"Perplexity is also one of the fastest growing vendors on Ramp. Seeing a lot of uptakes specifically because it's offering products that the model companies have not competed on yet." [00:07:35]

Attio AI-native CRM company based in London. Called out as a fast-growing challenger to Salesforce, representing a new class of AI-native application layer companies growing independently of model company performance.

"Attio is this extremely fast growing London-based CRM...Salesforce is 80% of the market...but Attio is like the AI native alternative." [00:24:38]

4. People Identified

Ara Karazian Lead Economist at Ramp. One of a very small number of people with access to large-scale, real-world business spend data on AI and SaaS, giving him a uniquely empirical vantage point in a discourse dominated by speculation. Publishes research at econlab.substack.com and on Twitter/X @AraKarazian.

"I have this unique job at Ramp where we see the spend data from 50,000 businesses, $100 billion of annual spend." [00:02:00]

5. Operating Insights

Don't Build Your AI Stack Around a Single Frontier Model

The data from early AI adopters — who tend to be the most sophisticated operators — shows a clear pattern of multi-model usage. Building dependency on a single model creates both cost and availability risk as token costs scale.

"Firms that were early adopters are the most likely to use multiple models and add more AI vendors over time. So if you want to consider the early adopters a sign of where the rest of the markets are going to go...you can assume that the average business is also most likely going to have a couple onboarded models." [00:09:34]

Auto-Routing Is an Underbuilt Internal Capability — Build It or Buy It Now

The market hasn't yet solved the problem of intelligently routing tasks to the right (often cheaper) model. Operators who implement routing logic now — whether through tools like OpenRouter, Cursor-style intelligent routing, or internal logic — will get dramatically better unit economics than those defaulting every task to the most expensive frontier model.

"I find myself in my day-to-day oftentimes automatically routing things to the most expensive model, even though it's slower. And I would actually prefer it just use Haiku or Sonnet when it's necessary. But you can't expect the typical worker to make that decision." [00:12:26]

Track Your Company's Visibility in AI Models — It's the New SEO

AEO (Answer Engine Optimization) is an emerging and fast-growing category. Companies that are not yet tracking whether and how AI models recommend them are leaving a new acquisition and brand channel entirely unmanaged.

"AEO as a category...is the software that firms use to track their performance in AI models and whether or not they're being recommended. Huge growth. I mean, that sector did not exist." [00:23:18]

6. Overlooked Insights

XAI's Acquisition of Cursor Is About Model Distribution, Not Code

This was mentioned very briefly and in passing, but is potentially one of the most strategically significant moves in the AI industry. XAI has massive compute but weak model adoption among businesses. Cursor is the fastest-growing developer tool with deep enterprise and developer penetration. The acquisition is essentially a distribution play — buying a captive, high-intent user base to drive model adoption in exactly the segment where Anthropic is currently winning.

"If their acquisition of Cursor goes through...I think there was a good reason for that acquisition as well. I mean, when you look at what XAI really had available to it, which is this incredible compute power...Acquiring something like Cursor to actually drive some kind of model adoption, I think will be a really good move for XAI." [00:38:31]

This reframes the entire competitive dynamic: the battle for AI model market share may ultimately be won not through model quality, but through distribution via beloved developer and productivity tools. Whoever owns the interface owns the adoption.

Data Aggregators Face a Structural Threat That Nobody Is Talking About

Ara briefly mentioned that Ramp now provides its data for free through Anthropic and OpenAI connectors. This is a quiet but potentially devastating development for the entire paid data industry. If high-quality, real-world business data becomes freely accessible via AI interfaces, the pricing power of legacy data vendors like Bloomberg, Refinitiv, and similar platforms could erode significantly — particularly for commoditized data sets.

"I would be happy if a lot of data aggregators lost market share because of both the free data that Ramp provides and then the ease of access that Anthropic provides to produce data. A lot of that market is unfortunately behind closed doors. The methodologies are not always clear. And in general, I think information is better when more firms and consumers have access to it." [00:26:38]