Chief Economist at Ramp: From a business data perspective, SaaS is back.

AI hasn't killed SaaS. The consumption data from the chief economist of Ramp tells the truth. This episode of the dialogue is from the "Monetary Matters" column of the MTS program under the A16Z Podcast. Hosts Jack Farley and Max Wiethe invited Ara Kharazian, the chief economist of Ramp. The reason why this episode is worth listening to carefully is not that he gave another new prediction about AI and SaaS, but that he holds a ruler that many people don't have - real business data from 50,000 enterprises with an annualized expenditure of about $100 billion, which can clearly distinguish market sentiment, investors' narratives, and enterprises' actual procurement behaviors. The core conclusions of the whole dialogue can be summarized into four points: First, the so - called "SaaSpocalypse" (SaaS doomsday theory) does not hold at least in the current enterprise expenditure data; Second, AI is indeed changing the software market, but the changes are mainly concentrated in the competition at the model layer, the concurrent use of multiple models, cost control, and the growth of the new application layer, rather than the collective collapse of traditional SaaS; Third, enterprises still mainly purchase software by seat. Although token - based billing is growing, its proportion is still extremely small; Fourth, what really deserves attention is not just who wins between OpenAI and Anthropic, but how AI is reshaping the software's infrastructure, workflow, data distribution, content production, and enterprise organization methods.

The following is the translation.

1

No "SaaS Doomsday" in Real Enterprise Expenditure Data

Host: Today we have invited Ara Kharazian, the chief economist of Ramp Economics. You have written many articles specifically studying a question that the market is most concerned about now: Have enterprises really changed their software expenditure patterns because of AI? Everyone is saying that AI will rewrite SaaS and reshape the way of software procurement. What has Ramp really seen in the real - world data?

Ara Kharazian: I think this kind of research is important because we are living in an environment where almost everyone wants to make judgments about the market trend, especially in the technology industry. There are many opinions, but few people really speak with data. And I have a very special job at Ramp: We can see the expenditure data of 50,000 enterprises with an annual expenditure of about $10 billion. So we want to answer a very simple question - what is really happening in the market? New AI companies are constantly launching products to compete with well - known SaaS companies. So, has the adoption rate of traditional SaaS declined? Has the way enterprises purchase SaaS changed? In my opinion, the so - called "SaaSpocalypse" can actually be split into two meanings: First, will enterprises switch from traditional SaaS to competitive products directly provided by model companies; Second, will enterprises abandon the traditional seat - based procurement method and switch to purchasing software based on agent capabilities and token usage. Based on the current data, neither of these two things has happened in any meaningful way. Seat - based contracts still account for about 65% to 75% of the expenditure, and fixed platform fees account for about 20% to 30%. Even for those SaaS platforms that have seriously launched token - based billing, the relevant expenditure only accounts for about 0.5%. For example, both HubSpot and Adobe have such products. As for the common saying that enterprises are abandoning established SaaS suppliers, just look at Figma: After Google launched relevant products, many people in the market were discussing whether Figma would be impacted. But we saw in the data a month ago that Figma was still one of the fastest - growing suppliers on our platform, and enterprises were still continuously purchasing it. So many judgments that "AI will immediately kill SaaS" are more of a projection of the product's future rather than a description of real - world business expenditures.

2

There Are Changes, but They Are Mainly at the Marginal Level

Host: It sounds more like "everyone is discussing it hotly, but the real - world changes are not that big." So where are the real changes? Where has there been marginal movement?

Ara Kharazian: There are indeed changes. More and more software companies are starting to offer some token - based products. For example, Adobe is a typical example because it knows that its products will be increasingly driven by AI capabilities, and these capabilities have marginal costs. Looking further ahead, if AI agents are used to operate software in the future, it will be difficult to simply charge by seat as before. So, more companies are starting to offer this new billing method, which is a fact. But if you ask whether this change has had a substantial impact on real - world procurement, the answer is still no: Currently, it only accounts for about 0.5% of the total expenditure.

Host: Does that mean that only when a company really grabs market share from others with the token model will the entire industry be forced to switch? In other words, what will be the trigger point for a full - scale transformation?

Ara Kharazian: This is a good question. All these changes will ultimately manifest as a series of competitive reactions, and these participants are not fools. They are well - capitalized, clear - minded, and highly capable companies. For example, you can look at Anthropic and Figma: Now Anthropic is also developing design - related products, but that doesn't mean Anthropic is naturally better at serving designers or understanding what designers need than Figma. Anthropic has models and a high product iteration speed, but Figma itself is an extremely mature and popular product, and it can also call the models produced and sold by Anthropic. I've never really understood the intuition behind the view that "AI will cause the collective demise of SaaS." Competition will of course occur, and Anthropic is indeed a strong new entrant in many markets, but this is completely different from "enterprises will immediately migrate to these new players on a large scale." Moreover, enterprise expenditures are inherently quite sticky.

Of course, the other side cannot be ignored: This is still one of the most dynamic software procurement markets we've ever seen, especially in the field of AI - related software. New players may replace old players on a monthly basis. Anthropic has just surpassed OpenAI in Ramp's data and become the most widely used model by enterprises; Cursor has also replaced GitHub Copilot. So, this market is definitely worth continuous tracking. But my core judgment remains: It's too early to conclude that "SaaS doomsday" has already happened, and such conclusions are usually not based on real - world enterprise behaviors.

3

Enterprises Haven't Abandoned SaaS

Host: That is to say, your conclusion can be summarized as: The SaaSpocalypse is not reflected in the data?

Ara Kharazian: Quantitatively speaking, yes. The two meanings of the SaaSpocalypse - "the software purchase method will be completely rewritten" and "traditional SaaS companies will be quickly killed" - are not supported by real - world business expenditures. At least so far, it has not changed the way enterprises purchase in any meaningful way, nor has it eliminated the companies that are most often cited as examples in the market. You can even look at Perplexity. Many people describe it as a more vulnerable pure AI player that is even more likely to be directly crushed by model companies. Compared with application - layer companies like Figma, it seems more dangerous. But in Ramp's data, Perplexity is also one of the fastest - growing suppliers because it still offers product forms that model companies have not directly competed in.

Host: Does the token - based billing launched by those SaaS companies replace the original seat - based billing, or is it just adding AI functions to the existing software packages? In a sense, part of the expenditure may flow back to these so - called "competitors."

Ara Kharazian: Currently, it seems more like an addition rather than a replacement. It is indeed adding new AI capabilities to the existing product packages. Users may pay by token for these, but it is usually still superimposed on the seat - based contracts. However, this part of the expenditure is really too small. We are talking about companies that have seriously launched such pricing, but on these platforms, the token - related expenditure still accounts for less than 1%. So it's hard to judge whether it has become the profit - growth engine for these companies.

4

Anthropic Surpasses OpenAI, but More Importantly, Enterprises Are Moving towards "Multi - Model"

Host: Everyone basically knows that Anthropic has surpassed OpenAI. Apart from this most obvious change, what other trends in the expenditure on other models have not been fully recognized by the outside world?

Ara Kharazian: Actually, when I wrote the article about Anthropic, I had some reservations. We started publishing an index last year called the Ramp AI Index to track the adoption rates of Anthropic and OpenAI by enterprises. For a long time, OpenAI has been leading, and now Anthropic has taken the lead for the first time. But over the past period, I've increasingly felt that people often misinterpret my research results as "OpenAI is finished" or "Anthropic has won," which is neither my intention nor the way I think the market will actually evolve. More and more enterprises are using multiple models in a deployed way, either by subscribing for employees to use or embedding them in their AI - native products. Enterprises that adopted AI earlier are most likely to use multiple models simultaneously and will continue to add more AI suppliers over time. If you take these early adopters as a signal of the future market direction, and our data often develops in this way, then it can be inferred that the average enterprise in the future will probably access more than one model at the same time.

On the other hand, enterprises are also becoming more cost - conscious. For enterprises that spend a large amount on tokens, especially those using models in high - intensity scenarios such as agent coding, their token costs have increased by 13 times in the past year. Even so, this part of the expenditure currently only accounts for about 2% of the total non - salary expenditure of enterprises, so the absolute value is not large. But if you linearly extrapolate the growth rate of 13 times, it is obviously an unsustainable cost curve that most enterprises cannot and should not bear. For this reason, enterprises that use AI most intensively have increasingly migrated a part of their AI expenditure to platforms like OpenRouter, hoping to switch between multiple models and use cheaper open - source models as much as possible. Even so, the proportion of this part is still very small at present: In the AI expenditure on our platform, the expenditure directly through OpenRouter only accounts for about 3%, but it is growing faster among enterprises that are most motivated to find low - cost solutions. They will route the workload to cheaper models for appropriate tasks. So, I think the main pressure on model companies comes from two aspects: cost and competition; and this competition even partly comes from themselves - from their cheaper model versions and open - source alternatives.

5

Most Jobs Don't Require Frontier Models, and the Market Hasn't Caught Up with This Reality

Host: This actually sounds like a bearish judgment on the total AI expenditure. Is it possible that enterprises are just gradually realizing that many tasks don't require frontier models at all? In the past, people just defaulted to "throwing more computing power at the problem," but in the future, what really matters is to break down the tasks correctly, design the processes well, and then decide which level of model to use.

Ara Kharazian: I completely agree. You don't need a frontier model all the time. In many cases, it may even be worse for what you want to do because it is expensive and slow. Even in my daily work, I often default to routing tasks to the most expensive model, but in fact, I often only need it to use lighter and faster models like Haiku or Sonnet when necessary. However, you can't expect ordinary employees to make such judgments. The reason I'm relatively familiar with these choices is that my job requires me to continuously track the AI market, know how new models are changing, and be more familiar with the best practices in use. But most office employees should not be required to master these things, and the market has not really productized this demand.

The problem is that OpenAI and Anthropic themselves don't have much incentive to provide a routing product that can automatically help you reduce AI expenditure because they make money from token revenue. More specifically, about 80% of the commercial revenue of Anthropic and OpenAI comes from token - based billing, so they naturally encourage you to use more. In contrast, large companies like Google have many ways to make money and don't necessarily rely on you spending more on tokens.

Host: Is there any plug - in or software in the market that can help enterprises automatically judge "which model this task should be assigned to"?

Ara Kharazian: Yes, Cursor is doing this. In fact, I think model companies will eventually be forced to respond to this demand under competitive pressure, especially Anthropic. More and more people have encountered its computing power limitations, so from its own interests, it should first make its products more efficient so that more users can use it more reasonably. But most likely, the first step will be taken by other companies, and then force OpenAI and Anthropic to follow. From this perspective, this is also a typical bullish logic for Cursor: It can compete in the developer experience, provide cheaper and more suitable model choices, and better routing, which are things that OpenAI and Anthropic are not willing to do actively without competitive pressure.

6

The "Bustle" of DeepSeek and the "Silence" of Gemini: Statistical Differences behind Adoption Data

Host: In your chart, Anthropic and OpenAI are in fierce competition, Google is behind, and DeepSeek is almost invisible at the bottom. But a few months ago, we were still hearing that many VC - backed technology startups were using DeepSeek because it was cheap. So why is it hardly visible on the chart?

Ara Kharazian: I don't think the statement that "80% of VC - backed startups are using DeepSeek" is true. When DeepSeek first came out, we did look at the relevant data. About a year ago, maybe a little more, we saw a spike in the adoption rate, but it never exceeded 1% of the number of enterprises on the platform. The reason we still keep it in the index is more to provide a reference background - although some people say that it may be time to remove it at this stage. However, DeepSeek is not the only open - source or low - cost model in the market. Next, we will expand the Ramp AI Index to track the adoption of more cheap models. Even if the model itself is open - source, you usually still have to pay for its actual deployment in the cloud, which is what platforms like OpenRouter provide.

DeepSeek also has a rather tricky problem: It may be cheap, but even if you deploy it locally, many enterprises are not willing to use it from a security perspective. Especially if you are developing products for enterprises or consumers, this concern will be stronger, and there are already many other options in the market. So, I don't think DeepSeek will be widely adopted soon.

Host: What about Google's Gemini? The line for it on the chart is not very high either. Do you think the recent news will change its trend?

Ara Kharazian: Google is actually underestimated, but there is a premise in the research design: We only track "paid adoption." Many enterprises are indeed using Google's AI, but it is integrated for free through Google Workspace - that's how we use it. So the question becomes: How do you define "what counts as AI adoption," and to what intensity does this adoption need to reach to bring about a real productivity boost? From an economic research perspective, many researchers are actually skeptical about the idea that "simply subscribing to a chat service can significantly improve enterprise productivity." Maybe we need to see a more comprehensive and in - depth use than the built - in functions of Google Workspace before we can talk about a productivity leap. But in any case, Google's distribution advantage is objective because almost all enterprises using Google Workspace are naturally exposed to Gemini.

Host: Speaking of the Ramp AI Index, just as the public market speculates about "which company will be the next to enter the S&P 500," people are also curious: Besides Anthropic and OpenAI, who is qualified to be the "top model player" that you will focus on tracking next?

Ara Kharazian: To be honest, from the perspective of measuring economic productivity, what really matters is not "who is first, who is second, and who is third." Adoption will ultimately lead to the co - existence of multiple players. In the next stage of our research, the focus is not just on whether enterprises have adopted, but on the "intensity" of adoption - what kind of enterprises are using AI particularly well and how they got there. This is not easy because intuitively you might say "look at who spends the most money," but spending money doesn't