My only pushback is that velocity isn’t the enemy, confusion is. Fast growth is fine as long as founders are honest about what they control and what they’re renting. The danger starts when those two blur.
Good points Nitin. When growth is unbelievably fast. The decision to adopt is typically more consumer like ("I'll try this"). So churn is higher as it's more try before you buy vs a business critical decision in the early years. So extreme velocity has some characteristics that feed risks into the system. But I think it's a great trade off to be honest 👏
Thanks, Chris, this is an excellent reminder as to why ARR is a poor indicator for real growth in the AI sector. How long do you think the market will continue to ignore the disconnect between revenue growth and actual profitability?
In all 4 layers the time horizons are different. At the application layer - patience will be shorter ie 2 to 4 Rounds of funding if the unit economics don't add up (unrelated to profitability - VC isn't about profits it's about market dominance). At the infrastructure layer it's a quite possible zero profits are made for a couple of decades - so much longer.
What this surfaces is a mismatch between how fast AI revenue can be generated and how slowly ownership, pricing power, and defensibility are actually built. The risk is founders internalizing those numbers as proof of durability before they’ve earned any real control over the stack.
That's totally my read. the demand to try something is exceptional but the depth of the commitment is very uneven and mostly shallow so time to build in genuine depth & therefore durability of the relationship/revenue is still required. Some won't make it of course or the demand goes elsewhere to repeat the cycle
ARR has always been revenue, not profit or ownership, and there’s always a cost base underneath it. What’s different here beyond AI having more variable COGS?
I’m trying to understand why that makes the metric meaningfully less real rather than just more incomplete.
Similarly with COGs. Software margins are in the 90%. AI margins that are infrastructure heavy are below 50%. Many AI vendors are subsidizing ie sucking up the costs by giving the technology away (& so investors are subsidizing the service & their patience with that model is key). It's a different business model (high vs low margin tech) and the KPIs in CaC, LTV, Burn Multiples need to reflect the true margin. So many will never pay back the CaC for example. Or they'll need to massively increase prices to do so or sell to a titan. The landscape will shift in many way.
Historically ARR is contracted revenue ie enterprises have signed up for a 1, 2, 3, + contractual length that auto renews and frequently pay annually in advance. Not all ARR is the same commercially (that's not changed per se) from one vendor to the next. What we find under the hood with AI is almost zero contractual strength in 90% of AI startups and some putting one offs in the ARR column (we've seen "grants" in the ARR column 😬). We find zero contractual lengths. We see individual signatures implying consumer like procurement vs multiple signatures implying enterprise like procurement which then infers more business critical decision making. So we see lots of AI ARR which is one off, non recurring, weak in commercial strength, PAYG and therefore huge churn. People signing to try something vs business critical. So ARR can mean many things vendor to vendor and some ARR is strong and valuable & someone else's is weak and not worth the paper it's written on & likely to be disrupted/evaporate. Maybe they disrupt their own ARR as they find their model via weak ARR.
This really captures something many founders overlook — that headline number isn’t the real engine, it’s what’s behind it that matters. According to a 2024 survey by OpenView, companies with over 70 per cent of revenue from recurring sources grow more predictably and are valued higher by investors, so understanding what is truly recurring versus pass‑through is crucial. The analogy to marketplaces’ GMV hype helps make that idea tangible without jargon. If founders focus on durable economics rather than surface growth curves, they’re much better placed for long‑term resilience. What’s one early warning sign you’d advise founders to watch for that shows their ARR is more mirage than momentum?
All about the usage. Are people using the technology in anger (ie really using it repeatedly), where & what are they using leading to what's the true value exchange btw the technology and the user/buyers. The answers in there 🎯
To be clear, this is a problem with ARR in general, since long before AI hit the mainstream. There are pass-through costs in all kinds of SaaS products, such as those that use AWS S3 or API Gateway (literally every API call can end up costing you).
But prior to AI, the pass-through tended to be either situational (e.g. payment processors taking 2.9%+) or so low that it didn't significantly impact the final numbers.
So, it feels more like AI has exposed this more clearly, because the per-call AI costs end up being quite significant.
What it ultimately reveals is that it has NEVER really made sense to look at ARR in isolation, and posts and charts and whatnot that focus on it by itself have always been empty calories.
It shouldn't really come as a surprise to realize that what ultimately matters is always profit, not revenue.
Thanks for the wisdom and expertise - this is all true it's not a brand new problem. This wave of AI (trial, error, adoption, churn, etc) has simply multiplied the challenges of what's good and what's not. Thanks for sharing 🌟
This is a key insight. The whole notion of what constitutes "Annual Recurring Revenue" is being fundamentally re-evaluated in the AI space.
I think the "Control Line" is a brilliant mental model for founders to use right now. If your growth is acceleration based on someone else's infrastructure, you're building on leased land.
Agree with this post.
My only pushback is that velocity isn’t the enemy, confusion is. Fast growth is fine as long as founders are honest about what they control and what they’re renting. The danger starts when those two blur.
Good points Nitin. When growth is unbelievably fast. The decision to adopt is typically more consumer like ("I'll try this"). So churn is higher as it's more try before you buy vs a business critical decision in the early years. So extreme velocity has some characteristics that feed risks into the system. But I think it's a great trade off to be honest 👏
Thanks, Chris, this is an excellent reminder as to why ARR is a poor indicator for real growth in the AI sector. How long do you think the market will continue to ignore the disconnect between revenue growth and actual profitability?
In all 4 layers the time horizons are different. At the application layer - patience will be shorter ie 2 to 4 Rounds of funding if the unit economics don't add up (unrelated to profitability - VC isn't about profits it's about market dominance). At the infrastructure layer it's a quite possible zero profits are made for a couple of decades - so much longer.
What this surfaces is a mismatch between how fast AI revenue can be generated and how slowly ownership, pricing power, and defensibility are actually built. The risk is founders internalizing those numbers as proof of durability before they’ve earned any real control over the stack.
That's totally my read. the demand to try something is exceptional but the depth of the commitment is very uneven and mostly shallow so time to build in genuine depth & therefore durability of the relationship/revenue is still required. Some won't make it of course or the demand goes elsewhere to repeat the cycle
Hey Chris.
Help me understand this distinction a bit more.
ARR has always been revenue, not profit or ownership, and there’s always a cost base underneath it. What’s different here beyond AI having more variable COGS?
I’m trying to understand why that makes the metric meaningfully less real rather than just more incomplete.
Similarly with COGs. Software margins are in the 90%. AI margins that are infrastructure heavy are below 50%. Many AI vendors are subsidizing ie sucking up the costs by giving the technology away (& so investors are subsidizing the service & their patience with that model is key). It's a different business model (high vs low margin tech) and the KPIs in CaC, LTV, Burn Multiples need to reflect the true margin. So many will never pay back the CaC for example. Or they'll need to massively increase prices to do so or sell to a titan. The landscape will shift in many way.
Historically ARR is contracted revenue ie enterprises have signed up for a 1, 2, 3, + contractual length that auto renews and frequently pay annually in advance. Not all ARR is the same commercially (that's not changed per se) from one vendor to the next. What we find under the hood with AI is almost zero contractual strength in 90% of AI startups and some putting one offs in the ARR column (we've seen "grants" in the ARR column 😬). We find zero contractual lengths. We see individual signatures implying consumer like procurement vs multiple signatures implying enterprise like procurement which then infers more business critical decision making. So we see lots of AI ARR which is one off, non recurring, weak in commercial strength, PAYG and therefore huge churn. People signing to try something vs business critical. So ARR can mean many things vendor to vendor and some ARR is strong and valuable & someone else's is weak and not worth the paper it's written on & likely to be disrupted/evaporate. Maybe they disrupt their own ARR as they find their model via weak ARR.
This really captures something many founders overlook — that headline number isn’t the real engine, it’s what’s behind it that matters. According to a 2024 survey by OpenView, companies with over 70 per cent of revenue from recurring sources grow more predictably and are valued higher by investors, so understanding what is truly recurring versus pass‑through is crucial. The analogy to marketplaces’ GMV hype helps make that idea tangible without jargon. If founders focus on durable economics rather than surface growth curves, they’re much better placed for long‑term resilience. What’s one early warning sign you’d advise founders to watch for that shows their ARR is more mirage than momentum?
All about the usage. Are people using the technology in anger (ie really using it repeatedly), where & what are they using leading to what's the true value exchange btw the technology and the user/buyers. The answers in there 🎯
Excellent post.
To be clear, this is a problem with ARR in general, since long before AI hit the mainstream. There are pass-through costs in all kinds of SaaS products, such as those that use AWS S3 or API Gateway (literally every API call can end up costing you).
But prior to AI, the pass-through tended to be either situational (e.g. payment processors taking 2.9%+) or so low that it didn't significantly impact the final numbers.
So, it feels more like AI has exposed this more clearly, because the per-call AI costs end up being quite significant.
What it ultimately reveals is that it has NEVER really made sense to look at ARR in isolation, and posts and charts and whatnot that focus on it by itself have always been empty calories.
It shouldn't really come as a surprise to realize that what ultimately matters is always profit, not revenue.
Thanks for the wisdom and expertise - this is all true it's not a brand new problem. This wave of AI (trial, error, adoption, churn, etc) has simply multiplied the challenges of what's good and what's not. Thanks for sharing 🌟
This is a key insight. The whole notion of what constitutes "Annual Recurring Revenue" is being fundamentally re-evaluated in the AI space.
I think the "Control Line" is a brilliant mental model for founders to use right now. If your growth is acceleration based on someone else's infrastructure, you're building on leased land.
Some "ARR" in AI startups is actually "I've simply paid so I can try this" and the challenge is it's quite a high % 😬
Chris, that Lego stack is a perfect visual summary of your main point.
I'll keep that in mind as more and more people in educator/coaching world start hyping "their own OS" tools.
Or when clients ask about the real cost of having one.
Glad it was helpful. Thanks for checking in 🌟