AI isn't the SaaS business—it's resale
Why big AI companies often invoice themselves
There’s a long list of things I find irritating. Right behind my upstairs neighbor’s parrot is restaurant menus that state some menu prices per kilo or pound.
Did I go to culinary school, or did you, Mr. Chef? Do I look like I’ve memorized the average weight of a Patagonian toothfish?
So when I ask, they usually know! “Well typically this type of fish is between 400 and 600 grams”. Why does it not say that on the menu?
Well, They don’t sell fish at fixed prices because they don’t buy it at fixed prices. Fish aren’t uniform and restaurant margins are too thin to risk losing money on a bigger-than-average turbot.
AI companies deal with the exact same problem at a much larger scale.
Why AI companies have different economics
Bad pricing can kill/harm a company in many ways (too expensive to get traction, too cheap to be sustainable). But one of them is misaligning how you charge with how you provide value (just ask Unity).
The stakes are even higher for AI companies.
Most people think AI companies work like a giant SaaS company: spend money building a product, charge customers more than it costs to run and pocket the difference.
Pure SaaS has some of the highest margins usually has margins above 80%. So AI companies should follow that practice, right?
Well, AI costs a lot more to run, which means you have to charge more. But AI models are ingredients in other products. and end users have price sensitivity (look at the backlash against OpenAI’s $200/month ChatGPT Pro plan).
AI companies’ margins are typically lower (Anthropic’s is apparently 50-55%) because they have higher variable costs.
At Lago, we see how AI companies like Together AI, Mistral and Groq charge their customers—and it’s more like a restaurant selling fish than a SaaS company.
And AI is complicated:
Nvidia buys chips from TSMC
Data centers buy Nvidia GPUs
Compute providers buy time in data centers
Cloud providers pay the compute providers for inference
Application-layer companies pay cloud providers for AI model outputs
End users pay application-layer companies to use their products.
AI isn’t SaaS, it’s reselling.
And because AI is expensive, companies lower en user cost by sharing revenue. Players agree to share revenue to lower the usage-based component of the price. This keeps prices from skyrocketing for startups (because fewer players to add their margin) and incentivizes everyone to facilitate a good end user experience.
Revenue sharing sounds simple enough, but is actually a billing nightmare. It might sound like an easy handshake agreement (“You’ll get 20% of what we make from each customer we send you. But implementing it for AI specifically is annoying.
Why it’s so hard to bill your partners in AI
Paying out partners is easy for seat-based pricing: Pay your partner a percentage of the customers they sent you, which is either going to be a one-time yearly amount or a uniform monthly charge.
That’s not the same for usage-based pricing, where you have the pleasure of:
Calculating what you owe each partner manually
Telling each partner individually what you owe them
Waiting for an invoice and then recording that spend
…on top of needing your own billing system!
Speaking of which, it means your billing system is no longer the single source of truth because invoices are managed elsewhere.
You end up running manual calculations, chasing invoices, and double-checking tools to keep track of everything. One mistake and you mess up a partner relationship.
Sharing AI revenue without a dedicated tool is horrible for both finance an engineering teams. So we built that dedicated tool (included in Lago).
It’s called partner billing, which is essentially the capability to invoice yourself.
The way this works is like this:
There’s a new subtype of the customer object which can be linked to an add-on, subscription or invoice. An example of this could be a cloud provider partnering with OpenAI to host the proprietary GPT models.
The partner would be OpenAI. Any usage of GPT models could then be compensated to OpenAI with a given percentage.
The types of things you rev-share for are arbitrary. You can link a partner to specific subscriptions (for acquisition partners and affiliate fees). Or to specific add-ons or usage (for products you resell). Or for specific invoices (for one-off partnerships).
Lago then issues an invoice addressed to you as coming from your partner. That way you have all of your billing in one tool and maintain a single source of truth, no matter how complex your billing gets.
Billing unblocked
Billing often blocks shipping new features, pricing and monetization.
It’s precisely because niche things like partner billing take up time after the feature has long been built. Companies usually solve this in one of two ways:
They build an in-house billing system that soon requires an entire engineering team to maintain.
They use a billing vendor and need to do a ton of custom computation before they send anything to their billing system, which effectively demotes their billing system to a PDF generator.
But it shouldn’t work that way. Getting paid and paying your vendors should be effortless, not a major blocker .
So if you have rev share partners and want to stop worrying about stuff that’s not worth your time, check out Lago.