Cal AI Capped Creator CPMs at Its Own RPM, and Most Brands Never Do That Math
Cal AI, the calorie-tracking app MyFitnessPal acquired in March, generated more than $30 million in revenue in two years largely by paying small creators with tight communities to post, founder Zach Yadegari said at TechCrunch Disrupt. The part worth copying is the contract: payouts were tied to guaranteed reach, and creator CPM was never allowed to climb above the app's revenue per thousand views.
Most coverage of Cal AI fixates on the founder's age (he's 19) and the exit. Fair enough, it's a good story. But the useful bit for anyone running creator budgets is buried in the middle of TechCrunch's write-up of his Disrupt session, and it's a lot less glamorous than "teen builds viral app." It's a pricing rule. A pretty strict one.
A growth formula that was boring on purpose
The origin is small. One early influencer post brought in roughly 1,000 new users over two days, and after that, Yadegari said, the company recruited creators "as fast as we could." He described the results as "a clear formula" that was "pretty linear, but extremely fast." Each post pulled a predictable number of viewers, and a predictable slice of them converted.
The deals were negotiated around reach, with a floor baked in. His paraphrase of the pitch to creators: "we'll pay you this, but we need to get us at least this amount of use." A lot of those creators, in his words, were "broke college kids" who "just want cash." The selection filter was creators with "very tight communities," not big follower counts.
The acquisition coverage back in March described an AI calorie app that soared to over 15 million downloads in under two years. Accurate. Also accurate: a spreadsheet, a minimum-view clause, and a lot of students who needed rent money.
Quick note on the numbers, because they don't line up neatly. TechCrunch now says "over $30 million in revenue" across two years (cumulative revenue), while the March acquisition coverage used the same $30 million figure as annual revenue. A FunnelFox recap of co-founder Jake Castillo's AIM 2026 talk says $50M+ in annual recurring revenue, which is an annualized run-rate, not money in the bank. Different metrics, different dates, different people counting. I'd anchor on the $30M figure and treat the rest as directional.
The one rule: CPM never beats RPM
Castillo spelled out the logic on RevenueCat's Sub Club podcast. Once you know your revenue per thousand views (RPM), "you can't pay like a CPM that is higher than that RPM." Creators push back on that sometimes. His position, roughly, is that there isn't an alternative if you want the channel to be profitable.
What made the RPM easy to estimate for Cal AI is that a creator post showed up as a visible spike on an otherwise flat download chart. You could literally see which video did what. That matters more than it sounds, and I'll come back to it.
The FunnelFox recap adds some operating numbers, all secondhand from the talk, so take them with a bit of salt: 300+ creator partnerships within 18 months, CPMs in the $0.50 to $3.00 range, and an early benchmark of roughly 10 installs per 1,000 views (installs, not paid subscribers). At peak they were onboarding about 20 creators a week, with contracts generated in one click. Paid ads, by comparison, returned roughly $1.20 to $1.50 per dollar spent.
And then there's the line I'd tape to the monitor:
Break-even is set by the average video. The profit is in the outlier.
Castillo's example on the podcast: one creator sat at about 10,000 to 12,000 views per video for two months, then hit 7 million views in month three, and stayed above 500,000 views per video afterward. If you'd judged that creator in month one, you would have cut him. A flat fee with a view floor protects you on the average and hands you the upside on the outlier. A pure per-view deal hands the upside to the creator instead.
Why this works for a $30 app and maybe not for your $400 blender
This is where I get a bit skeptical about people copying it wholesale. Cal AI was priced around $30 a year, according to the RevenueCat episode, and co-founder Blake Anderson framed the low price as optimizing for mindshare over profit. That's an impulse download. Someone watches a video, taps a link, scans their lunch. The conversion happens inside the same session, which is exactly why the spike was visible.
Most brands don't sell like that. A considered purchase spreads conversions over days or weeks, across devices, through search and email and a friend's recommendation. The spike flattens out and your RPM estimate gets fuzzy, which makes the "never pay CPM above RPM" rule harder to enforce. We covered this in why promo codes miss most influencer sales, and honestly, it's the unglamorous problem underneath most creator programs.
Think of it like a freight contract. You pay per load with a guaranteed minimum, and you stop calling the trucker whose loads keep showing up half empty. That only works if you can weigh the truck. Cal AI could weigh every truck. A lot of DTC brands are guessing.
Comments, not followers
The selection piece is probably the most portable part. Castillo told RevenueCat that he reads the comments on a niche's biggest creators looking for people asking for something the app solves. "Views can lie. Followers can lie, but comments cannot." The FunnelFox recap describes the same filter: comments like "thank you for this" were the signal, generic reactions weren't.
It also lines up with what we found in the micro vs macro sales data: smaller creators tend to win on sales even when they lose on raw views. And paying a college student directly skips the layer we wrote about in creator agency fees, where a big chunk of budget never reaches the person actually posting.
One honest counterpoint from the same podcast: Cal AI spent "thousands and thousands of dollars" on a couple of UFC creators and got no download spike at all. Big names in an adjacent niche didn't translate. So even the playbook's authors had misses.
The disclosure catch that case studies skip
Something that bugged me a little while reading the RevenueCat notes. Cal AI gave on-screen captions naming the app only to creators whose integrations were subtle, because captions cost views on videos where the promo was already obvious. Smart creative testing. It's also a habit that can drift toward ads designed not to look like ads.
The FTC's Disclosures 101 guidance for influencers is pretty clear that creators can't assume followers know about a paid relationship. Pair that with a roster of cash-motivated students, sometimes 50 to 70 of them posting on a single day (Castillo's number), and I think the compliance risk scales faster than the views do. Budget for the clear disclosure and accept the small view hit. It's cheaper than the alternative.
Running the Cal AI contract on your next 10 creators
If you want to try this without a 300-creator operation, here's how I'd set it up:
- Calculate your RPM first. Take the last 90 days of creator posts and divide attributed revenue by total views, in thousands. If you can't see spikes, combine link clicks, codes, and a post-purchase "where did you hear about us" survey. It won't be precise. It just has to be honest.
- Cap CPM at about 60-70% of that RPM. Castillo's rule is CPM below RPM. I'd leave extra margin because attribution on anything that isn't an impulse app is going to undercount some posts and overcount others.
- Write a flat fee with a view floor. Something like $600 for at least 200,000 views within 14 days, which works out to a $3 CPM at the floor. Below the floor, the creator owes a make-good post. Above it, you keep the upside.
- Score renewals after 30 days. Cal AI uses an internal renewal score that weighs secondary signals before re-signing. A spreadsheet with CPM, conversions, and comment quality is enough to start.
- Test 10 creators, not 2. The outlier math only works with volume. Two creators is a coin flip, and a coin flip tells you nothing about a channel.
For app marketers, FunnelFox's roughly 10 installs per 1,000 views is a useful reference point. Personally, if a creator delivers well under a third of that after two posts, I'd look at creator fit before blaming the price.
Radio figured this out decades ago
My prediction: by the end of 2027, at least a third of mid-market creator contracts will include some form of view floor or make-good clause, because CFOs are going to keep asking why creator spend has no guaranteed delivery when nearly every other media line does.
And the funny part is that none of this is particularly new. Broadcast buyers have negotiated audience guarantees and make-goods for a very long time. Cal AI just applied the idea to students posting from their dorms, and it apparently took a team without agency habits to notice it was missing. Yadegari has already moved on to an AI agent startup, Persona, per TechCrunch. The contract structure is still sitting there for whoever wants it.
