Anthropic, Cursor, FaL & Bessemer Venture Partners: The Realities of Scaling AI

The way we sell software changed. Before, if you are selling a SaaS product, the marginal cost was very little to sell to another company and/or to sell another product. But with AI, the, the new gross margins are lower because it comes with a cost to sell, sell to another person. So that means everyone has less margins. Maybe that changes over time. Models might get cheaper, and if you can have like people have stickier workflows, maybe the margins go up over time.

Hey everybody, it's Saster. Fin is the number one AI agent for resolving complex queries like refunds, transaction disputes, and technical troubleshooting, all with speed and reliability. See how Fin can deliver the highest resolution rates and highest quality customer experience at fin.ai/saster. That's fin.ai/saster. The biggest B2B and AI event of the year is back. It's the Saastr AI Summit in the SF Bay Area, aka the Saastr Annual. It'll be back in May 2026. With 36% of everyone coming CEOs, it's an incredible AI-first professional event. The very, very best S-tier folks will be there talking about sharing and learning how to scale AI in B2B in this new world. But here's the reality. The longer you wait, the higher ticket prices go up. They're really cheap in the beginning. And then, you know, just a few days before they get kind of expensive, but you've been warned. Early bird tickets are available now and I want to see you there. Once they're gone, you'll pay hundreds more. So book your spot today by going to podcast.saastrannual.com. That's podcast.saastrannual.com to get your exclusive discounts for Saastr AI SF 2026.

We will see you there.

Thank you to Saastr for having us. It is great to be here. I'm Talia Goldberg. I'm a partner at Bessemer and I'm very excited to do this panel with our friends at Anthropic, Cursor, and Fall. So before we jump in, if you guys just give like your quick bios, that'd be awesome.

Nice to meet everyone. My name is Kelly Loftus and I lead the startup sales team at Anthropic. I've had the privilege over the last year and a half working and scaling our startups team, and so very excited to be here today.

Hi everyone.

My name is Jacob. I First got my start in AI making Tabnine, uh, back in 2018. Then I worked at OpenAI as a researcher and then did SuperMaven about a year ago. And then, uh, around 8 months ago we joined Cursor. And so I've been working there on machine learning since then.

Hi everyone. My name is Gert-Jaume. I'm the CTO and co-founder of Fall. Fal is a generative media platform. We host open and closed source image and video models in our platform and expose them as easy-to-use APIs to the end users using our inference engine.

Awesome. I love having the three of you because all of these businesses are enabling this wave of AI and developers, they're always leading at every at every era of innovation. And so I, I want to start this off actually by sharing a conversation that I was having with Gorkam while we were driving up here. And Gorkam was like, gosh, UVCs, like the metrics are totally broken. Like all the metrics, all the questions we get asked, they're wrong because the businesses look different today. And we have these SaaS metrics that looked at things like CAGR and, you know, gross margin like this and and had, you know, very clear good, better, best metrics. And, you know, lo and behold, we have 3 wildly successful companies that have broken the norm. So, Gorgom, tell me to start, like, what is broken? And then maybe we can chat a little bit about what metrics actually matter for all of you.

I think the biggest difference now is companies are growing much faster, right? I think it used to be triple-double-double, or, you know, if you tripled in a year, that was considered gold standard. And now it's so many companies are breaking out, going from zero to $50 million, like really, really fast. I think that's the biggest difference, but also no one in AI really has 80%, 90% gross margins because the way we sell software changed. Before, if you are selling a SaaS product, the marginal cost was very little to sell to another company and/or to sell another product. But with AI, the new gross margins are lower because it comes with a cost to sell, sell to another person. So that means everyone has less margins. Maybe that changes over time. Models might get cheaper. And if you can have like people have stickier workflows, maybe the margins go up over time. But currently it seems like everyone has less margins than traditional SaaS, but everyone is growing like crazy. And there's like other implications of this to how people build sales teams. Seems like there's so much demand for AI that people don't need these massive sales teams and they can get things done with much leaner, leaner teams.

I want to touch on two things you said. One, we'll talk a little bit about all of your team structures and just some tactics on how you've built up go-to-market and sales. But before we do that, you mentioned this concept of, wow, like the gross margins are lower and it's because the cost to serve each customer also, it's not, you know, effectively zero as it was with the SaaS model. It's different. There's a real cost. And that is— there's a really weird dynamic that creates, which is that your best customers in some ways become your worst customers because they're causing a lot of cost in your system. And I think as a result from— we've seen a lot of experimentation with pricing models. And you're seeing this ideas of like value-based pricing and cap, you know, maybe more usage-based pricing. Maybe share with me from Anthropic's and Cursor's viewpoint, because I know Cursor has been experimenting too with having a more usage-based model, not just the $20 or $40 a month. How are you thinking about that? And what do you think this looks like a year from now or 2 years from now?

Yeah, it's really hard to predict. I mean, if you look at where things were a year ago, one year ago there was no Sonnet 3.5, and look at where we've come since then, where just a huge number of tokens are flowing through that model. And where are things going to be in a year, two years from now? As Gorkum says, it's not like traditional software where you're— when you receive $10 from the customer, you have to spend, you know, 10 cents in AWS costs to provide that. These GPUs are expensive and they have a real footprint in electricity and heat, and that's not going to change. The way I would think about it is just, you know, you look at the value that's being delivered to the customer and then you look at the, the stack that produces that value and you look at, you know, the total value being created by this technology, which is, you know, large and quickly growing. And then you just think, how can we increase that value and how can we be part of that supply chain that takes the electricity and converts it into something that is useful to people? Yeah.

So do you think there could be like if the average software engineer— I don't know, let's make up a number. Let's say someone gets paid $150,000 a year. How much do you think you could charge for Cursor in a year from now, or the best users, the highest value users?

That's a really good question. And, you know, when, when I first started selling developer tools for, you know, a flat price of $49, it's like, well, how much does this need to increase your productivity to be worth it? It's like 0.01% and it's worth it. I think, I think when you consider it relative to a developer's salary, you know, I think many people have been accelerated more than 2x by this technology already, and the technology is only going to get better. So I think comparing it to the developer salary is the right way to look at it.

One thing we realized, I think this is true for Cursed AI and Anthropic's case too, Instead of models getting cheaper, yes, maybe running the same model got cheaper, but people trained much bigger models that are much more expensive to run now, and people expect to use the best model. So running inference in general, maybe 100x in cost. Like, this happens for us in our use case for video models. Like, yes, the image models we had a year ago, and now it's super cheap to run. But now because we have much more demanding video models, the margins even got actually lower because it's so expensive to run these models. People demand them, so the inference costs got much larger.

I have a saying at Bessemer that I started saying a year ago, which is that COGS are the new CAC, which is that you could spend a lot on COGS, but it means that you also can't then be spending a lot on CAC and customer acquisition. Because if you have low margins and really high acquisition costs, that's tricky. But the good news is all of your products, they kind of sell themselves. And so it's a, it's a little bit of a trade-off. But it is funny, I thought a year ago, you know, we're on this cost curve of, of models getting cheaper and cheaper. And so I anticipated that the margins of some of our companies would actually increase and that it was okay that they had low margins. And it turns out they've stayed really low. And they've stayed low because of exactly what you said, which is the cost to serve, the models are getting better, but, and people are using them more, but that equation hasn't held.

Advancements in software are always faster than advancements in hardware. So maybe the next generation of chips are going to make things more cost effective, but developments on the modeling side is much happening much, much faster. So it is getting more expensive to run the best models.

If—

yeah.

Awesome.

So Kelly, when you joined Anthropic, I think how many people were on the go-to-market team?

Yeah, Anthropic itself was about 250 people and there were less than 10 people on the go-to-market team.

And today, how many people are there at Anthropic?

About 1,300 at Anthropic, probably 150 or so on the go-to-market team. So crazy scaling over the last year and a half.

That's crazy. And can you share with us one thing that has been really interesting to think through is like, how do you structure your go-to-market team?

When I joined, we did not have the concept of quotas. It's just hard to pick a number on what you want to actually measure each rep by. And so what I did when I started was let's just build a team around feedback and doing everything that is for the better, knowing that this team is going to scale from 10 people to hundreds. And so really focusing on that and keeping that in mind. And we still don't really have quotas. We have shadow targets. It's really hard and difficult to predict exactly what is happening. The adoption is fast. A lot of this is driven by the model intelligence, which you cannot predict over a long time period. So it's been extremely difficult, but really rallying people around the mission and being super clear about what we care about is getting feedback. On our models and continuing to work with partners to push the model capabilities forward. And that comes from a lot of the startups we work with.

No, no quotas at Anthropic.

No quotas at Anthropic today. That might change everything in this space. I feel like as soon as you say it, it's already outdated, but just shadow targets today.

Okay, so no one got, you know, a quote, no one got a commission for selling to Cursor.

No one did.

And actually, maybe just quickly, would love to hear the structure for you two as well.

I have a funny story about this actually. Beginning of this year, we were looking to hire a head of sales and like any good head of sales candidate, people were trying to negotiate a quota system because we were growing so fast and we thought, okay, maybe we had grown really fast until that point and we thought, okay, maybe doubling next year would be a good target. And we said like, okay, maybe, maybe we were going to double and that would be the OTE. And during like the interviews and during negotiations, we grew maybe 50%. So we were like almost halfway there already. And then we decided, okay, this is useless. We are not doing quotas. It's impossible to predict. One thing now we are experimenting, maybe we can do shorter term quotas, meaning maybe quarterly or monthly quotas rather than yearly. Whereas it's more predictable, you can course correct if something changes, but right now we are also not, not doing quotas. Everyone's getting on target earnings basically.

Yeah. Yeah, at Cursor, many of our first enterprise customers bought Cursor because they're developers. Came to their management and they said, we need this tool, or in many cases they were already using it, but we are growing the sales org a lot because there are a lot of companies where, you know, the developers can't necessarily make their own decisions about the tool, but they can still be substantially accelerated by this. And so we need to reach everyone.

Yeah. The, the Quota system, part of the reason it's there is, you know, to incentivize the behavior that you want. And there's always good incentives and bad incentives. So, it's a tricky system no matter what. And one of the benefits you have is that a lot of this, there's so much demand that a lot of this for all three of you is really filling the demand versus generating it. But in getting the right team members early on that have the, share the same incentives, like that is probably the prerequisite to saying we don't need a system. You know, around this because we're all aligned. And so I'm curious for each of you, are there any, any things, whether it's like small tactics or questions or ways that you've assessed people, not just for go-to-market, but even for R&D or any area that says, hey, like this is a person that would be a good fit at Cursor or a good fit at Anthropic and might actually, you know, not be a good fit at OpenAI.

On the sales side at Cursor, we have a very technical sales team. And that's partly because it's a technical product. And so it's helpful when the salesperson understands that well. But it's also because, you know, there's a lot of ways the sales process can be accelerated with software and with Cursor. And so many people in our sales org are also building tools to help the sales org using Cursor. And I think that's a trend that will continue.

Any specific examples on that of like what tools are helping automate your go-to-market team?

One, we use AI to help qualify inbound leads is one thing.

That's great.

One thing that worked really well for us, Talia, is we have a small research team now, and most of the people we hired for that research team is through our research grants program. So it's open invitation. You can basically just send us an email with a project you have in mind and we care about, let's call it efficient AI. It's either efficient fine-tuning techniques or efficient inference. And if you have a research idea around that, or maybe if you like it, it could be an adjacent area as well. And we give you compute for a couple of weeks for you to like submit a project. And we hired maybe 4 people through that research grants program and has been really useful for us. We have no strings attached, no expectations, but people have been doing great projects and we ended up hiring them.

That's a genius tactic.

Works.

I really like that.

Yeah, I like the topic on how are people using AI within their day-to-day. I'm curious on your end, or what's the coolest use cases within all of our companies that people are using AI in their day-to-day to help speed up different workflows or augment them?

We are trying a lot of different sales tools to automate email marketing or automate, okay, you check the pricing page, now you get an email, like things like that. Maybe you don't even need AI to do those, but a lot of the next generation tools are all AI enabled. So we probably subscribe to like 3, 4 of them trying to see if any of them works better, but it's still on the experimentation phase. And I know some companies can do this really well.

We're really excited about Background Agent, but with this feature, you can give tasks to AI that it will complete asynchronously, and you can have multiple running at once, and then you can check in on the progress and easily— for example, you know, if it's 90% right, 10% is off, you can very easily drop into what the AI has been doing and bring it into your editor and correct it. So it has this property that you can very easily steer and fix any mistakes, which I think is really important. Because these models are really, really smart and they're getting smarter. But they do still make mistakes sometimes. And sometimes, you know, they don't really perfectly understand what you wanted. And so being able to correct it and be in control is really important.

Yeah, it's very cool.

Across both the technical and non-technical org, one of the favorite use cases is the Slack channel that we spun up. And what the Slack channel does is employees can go in, ask questions. Claude is on the back end and uses to go search over our internal knowledge bases and then retrieve and answer the employee's question. It's been extremely useful for productivity, so folks can get a really great in-depth answer really quickly that they would have otherwise had to ask a busy manager or busy engineer. And so it's been extremely great for time to onboard and just across, especially across time zones as well.

Were there any really key decisions, you know, all of your companies like you've totally broken the norms of revenue per FTE, like in such wild ways. It totally breaks the brain across, you know, all 3 of these companies. So obviously you've had a lot of success, but there's definitely in every startup journey, inevitably there's these challenging points in the road. I'm sure there will continue to be. But were there any key decisions that you can think of that you go back to that you were like, gosh, like this was a really critical thing and we got it right or we got it wrong that might be interesting for the group.

So, when we first started the company, we were more of a data infrastructure company and we had like a serverless Python runtime. We decided, okay, we want to focus on image and video. First of all, because we had a technical advantage in the beginning, some of our earliest customers were were using image models at the time. So we decided to double down on that technical advantage. But also everyone pretended, oh, all AI models is the same market, you know, same companies will be running all the models. But also we identified early that the buyers of these models are going to be very different. Therefore, this is going to be a completely different market. So branded, we positioned the company as a generative media platform. It's actually a term we came up with, or it was being used, but we kind of owned it for ourselves. So that positioning was also really, really important in our company journey and helped us with marketing, helped us with getting like really big people. And now we are trying to associate our brand more with generative media. And I think it's working very well.

Yeah, I think that's that focus. From what I've seen so far has been so critical. And meanwhile, you know, a lot of people talk about AI and they're talking about, you know, broader LLMs and otherwise. And so it feels really tempting. And you see a lot of other companies trying to be everything to everyone. Okay, we only have a couple minutes left, but there's an elephant in the room, which is that Cursor and Anthropic have such an incredible symbiotic relationship. I think it's been publicly reported that Cursor is at least one of Anthropic's, you know, largest customers. Anthropic has really done an incredible job in the code vertical as well and also has cloud code. But it's interesting because Anthropic, you have these partners on the infrastructure side like AWS, you have the application side like Cursor. How do you balance competition versus collaboration.

We want to partner with companies like Cursor to drive the models forward and push the capabilities of what is actually possible with these models. And so that's how we think about our partnerships there. And then when you look at some of the tools like Cloud Code versus IDEs like Cursor, developers use these in complementary ways. And how can we actually continue to build products that developers want and will use alongside each other that continue to push this space forward. At the end of the day, we want to build really strong models to advance forward areas like coding and development. And so that is really what we're focused on. And one of the big things we care a lot about is feedback and partnership. So Cursor has given us feedback on our models in the coding area and had access to our models before we released them and been able to actually give us feedback to meaningfully improve the user experience on both of our ends.

Yeah, I think if you look at the core mission of the two companies, you know, Anthropic's mission is to build artificial general intelligence that is aligned with human values and Cursor's mission— or I don't know what— I don't want to be on the record as, you know, our one mission. We're broadly interested in making tools that are as useful as possible to software developers. And I think what we want is to give people the best model. And whenever the models get better, we're very happy because it means Cursor becomes more valuable to our users. And, you know, we've been really happy that Claude-Sana is so good at code, and we're excited for that to continue.

Jacob, you said one thing earlier, which was that when we were talking about usage-based pricing and just value-based pricing, you said a lot of people are getting 2x productivity, you know, plus with, with Cursor. What do you think I should be seeing from our portfolio companies? Should we be seeing 2 times more products developed? Should we be seeing, you know, half the headcount? Like, what do you think is the right rule of thumb? Today?

I think, I think these tools, there are many people who previously didn't write code, and now these models have allowed them to create great products and tools. And there are many other people who are already very experienced had writing code, but these products can accelerate them by, you know, helping with the stuff that you maybe didn't want to do before, or just, um, predicting your next edit and saving you some time. And we want to serve both of those demographics. And so I think to pick a single acceleration factor is difficult because it depends where you were previously. And I think compared to many other companies in the space, we very much care about making something that we want to use and that an experienced software developer really enjoys using. So I wouldn't, I wouldn't venture a specific number, but I would say that we really want to be— to make tools that, that help everyone regardless of your experience level and that you should be seeing substantial gains no matter where you are.

Yeah, it's a hard— it's hard to measure that productivity, but I imagine at least seeing the continued usage and adoption and the retention piece of it is really critical. Awesome. Well, one last question to bring it totally full circle with that, and then we'll take it to questions from the audience. When speaking of retention and then how all the metrics for AI companies are totally different and the VC frameworks are broken, what metric do you and each of your respective businesses— like, what do you care about most?

We care about big logos we bring to the platform. That's definitely something we pay attention to. One really interesting thing that's happening is lots of more AI-native companies and newer companies are actually spending more than like bigger enterprises who are maybe not super sure about putting things into production, but we want more of them. So whenever the models are good enough for them to say yes to a couple of the products that are currently being built, we are there for them. So big logos is one thing we care about. We obviously care about revenue and, but we want to make sure there's, it's coming from at least 30 to 35 different companies rather than being concentrated at the top. And retention matters a lot. Like we care about churn all the time. And we are, we are like doing all, everything we can to make sure people on the platform are happy, they are growing with us. And if they have spend elsewhere, we want to bring that spend to the platform as well. That's why we are hiring customer success managers, making sure we get a bigger share of their, let's call it, generic to media spend.

Wallet share.

A lot of share.

Yeah. I think we care the most about making a product that we personally want to use. And obviously we care about revenue. We care about users as well. I think, you know, revenue is an indicator that maybe lags behind users, and users is an indicator that lags behind of, you know, the fundamental, like, quality of the product and the tool. And so I think the thing we care about most of all is whether we personally want to use it or use a new feature in our, like, day-to-day life.

I know you work on the tab completion model. Is there a dashboard, like, number of tabs pressed on the platform that you look at every day?

That's a great question. We definitely track that, but, you know, if you make the suggestion longer, like, maybe it results in fewer suggestions accepted, but, like, more value Yeah, it's something we think about a lot.

Awesome. Well, thank you all for being the building blocks that is paving the way for so many incredible startups and companies across all industries. And so it's awesome to have all three of you. It was great to be here. And thank you to our panelists. Prost of all!