Tokens Are the New Customer Acquisition Cost

That meant you could afford to be pretty generous. You could give somebody a 30-day free trial and let them properly explore the product, or have a free tier that somebody used for months before eventually deciding they were getting enough value to upgrade. In some cases they might have been using the thing for years before becoming a paying customer. The expensive bit wasn't really letting people use the product; it was getting them there in the first place.

So a lot of SaaS economics became about customer acquisition. You spent money on sales, marketing, content, events and, increasingly, large amounts of money on Google and Meta to get people through the front door. Once they arrived, you could let them kick the tyres for quite a while without worrying too much about what that was costing you.

AI changes those dynamics because suddenly the usage itself has a meaningful cost. With LLMs we talk about tokens, while image, video and other generative tools have slightly different economics, but the basic principle is the same. Every time somebody asks the product to do something, somebody somewhere is paying for the compute.

I think this creates a particularly awkward problem for early-stage AI companies because a lot of these products take time to understand. Take one of the many new AI design or coding tools. You sign up and get a small number of free prompts . Maybe you get lucky, type exactly the right thing and the product produces something amazing. You immediately think, "Wow, this is great," and reach for your credit card.

That's rarely how it works for me though. Normally I'll spend the first few prompts trying to work out what the product actually wants from me. I'll phrase something badly, it'll misunderstand me, I'll try again, get something vaguely interesting and start tweaking it. Then, just as I'm beginning to understand how the thing works, I run out of credits and get asked for £20 or £30.

At that point I'm not really being asked to pay for a product I like. I'm being asked to pay £20 or £30 to find out whether I might like it, and unless the first few attempts were particularly impressive, I generally don't bother.

I come across this all the time with new AI design tools. Five or six generations are usually enough to get a rough sense of what the product does, but nowhere near enough to understand whether it could become genuinely useful, let alone something I'd want to build into my workflow.

So maybe five prompts isn't enough. Maybe I need 20 or 50. Maybe I need a few prompts every day for a couple of weeks. Or a couple of months? Maybe the company effectively needs to subsidise me while I learn how to use the product and bake it into my workflow. The problem is that every additional chance you give me to reach that aha moment costs you money.

That's why I increasingly think of tokens as a form of CAC. In the old SaaS world, you might have been perfectly happy paying Google £30 to acquire somebody because, once they arrived, letting them spend the next month exploring your software cost almost nothing. With an AI product, you still have that original acquisition cost, but you may now need to spend another (potentially much larger) chunk of money on compute in order to give somebody enough experience of the product that they're willing to convert.

Strictly speaking you might not count that as CAC, but economically it's doing much the same job. You're deliberately spending money on somebody who isn't yet a customer in the hope that they'll eventually become one.

Once you start looking at compute this way, there's an interesting modelling problem. Let's say you've decided you're happy to spend $20 of compute on each prospective customer. You can then look at what percentage of those people convert. What happens if you increase that allowance to $50? Does conversion go up enough to justify the extra spend? What about $100?

Presumably there's a sweet spot somewhere. Giving somebody twice as much compute might tripple their likelihood of converting, in which case great. Or you might discover that you've doubled the cost for almost no additional conversion. At that point you're effectively doing the same CAC and payback calculations SaaS companies have always done, except part of the acquisition budget is now being spent through the product itself.

I suspect it will get more sophisticated than simply giving everybody the same allowance. SaaS companies have been doing versions of this for years. Somebody registering with an email address from a large company might be treated very differently from somebody arriving with a random Gmail account, because one looks considerably more likely to turn into a valuable customer.

AI products will probably need to learn similar signals. Maybe somebody who uploads real company data, connects a work account or comes back three days in a row is worth giving another $50 of compute to. Somebody firing off hundreds of random prompts with no obvious sign of purchase intent probably isn't. You start asking not just how much free compute you're willing to give away, but who it's worth giving it to.

Freeloading has always been part of freemium. You might have had 10% of customers paying for a product that 90% used for free, and that could still be a perfectly decent business because those free users barely cost you anything. They helped with word of mouth, some eventually converted, and the paying customers comfortably covered the cost of serving everyone else.

That equation becomes much harder when the free users are burning real money every time they press a button. A freemium model where 90% of the audience never pays looks quite different if that 90% is also consuming expensive compute every day.

It also creates the potential for an arms race. Maybe your modelling tells you that spending $50 on a prospective user works really well, only for your competitor to decide that they're willing to spend $100. Then another well-funded competitor decides $200 is perfectly acceptable because they're more interested in grabbing market share than making the unit economics work today.

We've seen this movie before. For years companies bid against each other for Google keywords, and the price of acquiring a customer gradually rose until the winners were often the companies with either the best economics or the deepest pockets. You could have a great product and still lose because somebody else could afford to pay twice as much for the same customer.

I can see free compute heading in exactly the same direction. One AI product gives you five prompts, another gives you 50, while a heavily funded competitor lets you use the thing almost without limits until you've built it into your workflow. At that point it isn't necessarily the best product that wins. The company with the most capital can effectively buy itself more opportunities to convince you.

This is where the big AI platforms have a massive advantage. If you're OpenAI, Anthropic or Google, you've got enormous amounts of capital behind you, access to huge amounts of compute and millions of paying customers already helping fund the system. Compared with a seed-stage startup, your ability to subsidise usage can look almost unlimited.

So you can afford to be much more aggressive. You can give people far more product than they're actually paying for, get it embedded into their daily habits, grab market share and worry about improving the economics later. It's essentially a form of penetration pricing, except instead of simply discounting the sticker price you're also subsidising the actual usage.

We've already seen a fairly extreme version of this with ChatGPT. In January 2025, Sam Altman said OpenAI was actually losing money on its $200-a-month Pro plan because customers were using more of the service than the company had expected. That's a pretty difficult thing for a startup to compete with. You're agonising over whether you can afford to give somebody another $20 worth of compute in the hope they'll become a customer, while one of the largest players in the market is prepared to subsidise people who are already paying it $200 a month.

It also makes the normal SaaS payback model much messier. Traditionally you might spend £100 acquiring a customer and know that, assuming they stayed around long enough, you'd slowly earn that money back through the margin on their subscription. With AI you can spend money getting somebody to subscribe and then continue subsidising their usage after they've converted. Your most engaged customers can also be some of your most expensive customers to serve.

If you've raised billions, you can live with that for quite a long time. You can effectively overpay for customers today because you think owning the market or becoming the default habit is worth more in the long run. If you're a small startup, it's a much tougher problem.

Maybe your accelerator or VC has negotiated $250k of model credits for you, which sounds like a huge amount of free compute when you start out. But most of that may already be getting burned on development and testing, and you know that eventually it runs out. So there's a perfectly understandable tendency to ration what you're willing to spend on prospective customers.

Financially that makes complete sense, but it can also be self-defeating because the amount of free usage you're willing to fund may simply be less than the amount somebody needs before the product becomes compelling.

I wonder whether this is also one reason we're seeing plenty of AI startups gravitate towards B2B rather than trying to build large consumer freemium businesses. A traditional enterprise sale is expensive in a different way, but you can pitch somebody, demo the product and then run a controlled pilot with a relatively small number of users. You're not necessarily funding millions of anonymous people to play with the product in the hope that a small percentage eventually convert.

I've seen this with a number of companies I work with. Rather than opening the doors and letting everybody pile in, they go after a handful of companies, show them what the product can do, run a pilot and try to turn that into a proper contract. Once you've got meaningful revenue coming in, maybe you can afford to become more generous with self-serve usage later.

So I think the acquisition question for AI companies is becoming rather different. It's still "how much can we afford to spend acquiring a customer?", but now part of the answer sits inside the product. How much compute do we need to give somebody before they genuinely experience the value? Which users are worth spending more on? And how does the conversion lift compare with the additional cost?

For early-stage companies, getting somebody to value quickly therefore isn't just good product design. It has a direct bearing on the economics of the business. If one product needs $100 of inference before somebody gets the magic and another gets there after $20, that's a pretty meaningful advantage before you've even started talking about model costs or subscription pricing.

The danger is ending up somewhere in the middle: paying somebody else's model costs, unable to afford much free usage, and putting up a paywall before customers have seen enough to want to pay. Meanwhile a better-capitalised competitor can simply keep feeding them compute until they get there.

For the last generation of startups, a big part of the acquisition game was figuring out how much you could afford to pay Google for a customer. For this generation, it may increasingly be about how much compute you can afford to burn before somebody becomes one.

And, just as with Google Ads, the uncomfortable possibility is that the company with the deepest pockets gets to set the price.