The Hidden Cost of Betting on One AI Provider

Most enterprise teams are stressed about the same thing. Which AI provider do we bet on? Is it OpenAI, Anthropic, Gemini? Divanch thinks the answer is all of them.

Divanch is a solo developer in the Canopy program at Founders Inc. right now building an overlay, an open-source AI workspace that pulls together every Frontier model, every integration, every AI feature in one place. He built it across web, desktop, mobile, and a Chrome extension by himself with the assistance of our Composio SDK. What I love about this conversation is Devanch came to AI from a completely different direction than most founders.

Biomedical engineering at John Hopkins, then hardware, then VR, and now this. And his whole thesis across everything he's built is that an interface is what limits how fast a technology actually spreads. We get into vendor lockin, why the cost of not switching compounds over time, and the three things he tell every founder not to get wrong when building with agents. Hope you enjoy this conversation.

Hi Dian. So I want to ask you the first question which is what are you building at layer norm and yeah tell me all about it. Hi Julia. Uh I am building overlay which is an all-in-one AI workspace of source.

So if you have if you use chat GPT or cloud it is basically one place to get all your best models all your personal context and every AI feature that you want to use like agents automations and have just one place to run it and it's also like cheaper and open source so it can be extended upon it can be customized to your needs or your enterprises needs and uh yeah that's that's kind of what it's like a oneliner of sum uh oneline summary of of overlay and so what does day one look like for your users what do they do they can't do before. Yeah, it's so I initially built overlay for power users who sort of like have multiple AI subscriptions let it be like JBD or claude or perplexity or Gemini Gro and they were like managing multiple subscriptions and it was supposed to be like a subscription aggregator but very soon I realized that like the the problem is not being able to access all of the best models. It is usually these these interfaces that each of these providers has. They have one feature that the others don't and that's that one feature is very important to one user.

So it would become like the vision sort of expanded into making a workspace that had all the best features from all the best providers and all the models and it would also end up being cheaper because many of the models many of the SOTA models like state-of-the-art are extremely expensive and we're also getting like these open source models that are just as just as intelligent put on hyperscalers for cheaper inference and faster inference and they would not people would not even know of them uh like even even here like as as as technical optimist missed a place as founders Inc. 6 fix and many like many people did not even hear about it. To make any person AI native, it is it is important to know of the best models and to also know what they can do with AI and overlay is intending to be that place uh that they could just like they could just go to and become AI native and even sort of looks like being able to work across models make notes that are then inferred into their chats extending extending their capabilities using like integrations from Gmail Slack notion and uh being able to run automations. Wow, that's really cool.

And how have your past experiences set you up for for success for to be building this? Yeah, it's funny because um and and I I I I get teased about this sometimes, but I kind of take it as a compliment. I did my uh engineering uh from from Johns Hopkins and I I did biomedical engineering from Johns Hopkins University in Baltimore and it happens to be the number one uh university in the world to do biomedical engineering and I basically have worked with every single layer of interfaces. I've worked on on in ear EEG uh earbuds to to get EEG from from your ear canals.

That's that's PCI plus wet lab plus hardware. Uh, I've also worked on concussion games in VR for like for concussion recovery, VR games for concussion recovery. Uh, so that's software plus hardware. And then I'm I'm I've now graduated to just software.

So I built an open-source clone of Whisper Flow last year in my summer crowd surfing in San Francisco. And now I've graduated to becoming like the the open-source interface layer for AI and making that my my full-time thing. Usually I've heard people go in the opposite direction. and they go from from B2B SAS to consumer SAS to hardware to like wet lab or like biological stuff.

I've gone the exact opposite direction. But the one the one thing that that sort of ties them all together is that they are all interface layer, all interface based. So, I've I've always been obsessed with like making interfaces because that is usually what the end consumer interacts with and the and the goodness of the interface is almost the limiting factor of the ability to disseminate a technology into the world and and how how positive an opinion a person has on a technology. Before the iPhone came about, those Blackberries and like those those keypad to uh tools were just like they were just they were just that tools.

they became ecosystems after after the iPhone sort of extended them into into the interface that they've become today. So, I've I've kind of been obsessed with interfaces and that's sort of primed me to make an open source interface for AI. Awesome story. And so, here's what I keep hearing from enterprises.

They're scared to go all in on any one provide AI provider. Is this a real problem or is it mostly just anxiety? What are your thoughts? It is extremely real.

I will tell you this because I was talking to an enterprise, a very large enterprise, I think 3,000 people only 3 days ago. They they felt this acutely. The the head of the enterprise actually had contacts at Anthropic and OpenAI that could give them subsidized API pricing. But what they didn't realize is that what they sort of realize is that if if they were to lock themselves into this one model provider and their interfaces and their APIs and the way they do things are not they are not going to be they're not going to be well positioned if the state-of-the-art changes and if the best model changes.

7 was was not the best uh coding model? 5 was not the best at reasoning? What if an open source model happened to be better at either of those things? What if Gemini wasn't the best at image understanding?

Then you have this problem of vendor lockin that becomes a true constraint for for for how your how your company is sort of adopting AI. And that is why I'm so bullish on like aggregator businesses. Overlay essentially is an aggregator business where it aggregates the best models. It aggregates uh your personal context from from all from all the places that you hold your personal context all your integrations um and aggregated of all the AI features that that an enterprise could need.

So I'm super bullish on that. It's a very real need and the enterprises who aim to be AI native are acutely aware of this. If someone is already deep in say OpenAI's ecosystem, why would they even bother switching to something modular? I think my best argument for this and and I think in how many how many other customer interviews I've had this has been validated to the tea is that the one-time cost of switching to an aggregator the cost of switching to a one time the onetime cost of switching to an aggregator decreases greatly over time like the marginal cost goes to zero almost entirely because once you once once you've switched to an aggregator you can also use the model that you were previously locked into if if it happens to be very good at something else but then you have so many other options to be able to use for anything and everything.

It is a onetime cost that is worth bearing that only divides itself with time. But the marginal cost of not switching, the opportunity cost of not switching only compounds as the capability capabilities of AI uh increase in other models or some other some other interfaces become better or some other or some other features become better. 6 gives me or the image understanding that Gemini gives me or the the the unbiasedness that Grock gives me. Switching once would of course bear some cost with it.

But in the asmtote it would just diminish to zero. If I were to not switch by losing out on these many features and of course they build up other companies are not going to slow down. uh this is the worst AI is ever going to be. You're almost you're almost guaranteeing that your that your AI that your company would not be as AI native as it could be uh because you're sort of locking in um and paying compounding marginal costs for not switching.

And where do you see this going in 2 years? What is winning architecture for enterprise agents look like? Yeah, I I have very strong opinions on this and and overlay is almost built entirely on those strong opinions which is it would have to be the winning architecture would have to be extensible uh it could be extended upon by the enterprise itself for their own needs. It could it should be customizable.

So the existing uh solutions could be uh uh could be customized which which is largely taken care of by by virtue of being open source in my opinion and it should also be an aggregator uh that aggregates uh the best models aggregates the best services and you could build that in house but if there are existing solutions that you could white label and customize and extend you have no reason to. So this is not only a winning solution for a company that wants to adopt AI, it is also a winning solution for for a company that wants to develop AI because their their their development costs have been cut in have been cut by a good margin. Their their operation costs have been cut by a good margin to maintain that architecture. Aggregating customizable extensible inter extensible interfaces or architectures will win in the next next two three two to three years in my opinion.

And what what are those architectures like what are those products that companies could use instead? Yeah, I I I think across every single across every single um feature that at least I have I have built uh in overlay I have looked at an at an aggregator. The reason why so many the reason why so many companies now call open router or versali gateway instead of the openi API or the cloud API or the perplexity API individually is because these aggregator services provide a unified platform. They provide one single API.

uh if if you want if you want to integrate five five different models into your solution before this you needed five different APIs with five different rule sets with five different rate limiting conditions with five different parameters that you had to had to take care of with with something like the virt gateway or open router or in the case of integrations my favorite composio I I cannot tell you how core a part of the product is verseli gateway and open router and composio literally I would This is a this is a guesstimation but like 50% of everything I use overlay personally for uses a tool call that it directly uh that it directly takes from from composio and the reason why why these these companies win is because they operate on the assumption that more models is better more services are better having them all under one roof in one solution in one place is better it just has taken so much operational overhead and cognitive load away from me that I can just focus on the core product. Yeah, this is this is a story I think is worth telling. This is Founders Inc. I'm in Founders Inc.

and they had a dev day where where Composi happened to come into our offices just to support developers and I I did not even need the swag. I didn't even want this swag. I just wanted to go and tell the team how much time and and money they've saved me in building overlay. I'm a solo developer who has built overlay across four surfaces now.

the web, Mac OS which is desktop, Windows electron so across cross crossplatform iOS Android so web desktop mobile and a chrome extension and each one of them treats my one API as a single source of truth and this one API just calls Verselia gateway open router and composio and the first time I actually integrate composio I didn't even know this they had a sandbox feature built in it they had a workbench feature that that executed code for me which was so magical uh because I didn't even expect it to. And that was that that was that was the first wow moment I had in many many months of going deep into software and and so I just went I just went to to thank them uh for for making such a wonderful product. I will tell you this I think that there are like around 150 teams uh in founding somewhere around there. I could be wrong 120 to 150ish.

I have whichever team is doing anything remotely related to to software where they integrate services I have actively evangelized people to use these three services which is versel AI gateway composio and open router vers gateway more than open router because versa gateway does not have a platform fees but yeah I'm a huge fan of such services wow thank you for such kind words the team's going to love to hear that and so kind of reiterating what's the thing composio does that you just can't replicate by stitching together a bunch of other tools. I've thought a lot about this because I as someone building an open source solution as a solo dev um I've always had the impulse to think why cannot this be built inhouse and I've built many things in house that that I could just uh outsource to an API including my own memory system and there exist many many uh memory systems out there uh that are very very good uh that of course have comparative advantage because they just focus on memory the reason why I made memory system on my own is because it is a core part of the product which which I could manage and enterprises that adopt overlay would need to customize and extend. The reason why I thought about making my own integrator of MCP tools or making my own aggregator of MCP tools but didn't choose to do so is one to be useful you need many many tools in one place. I know Composio Composio like at least the ones I use into my platform even without an O config.

Composia has around 400 of those tools and uh with an O config they have even like I think closer to 800 which is nuts. Uh by the way if I were to be expected to make files that do that it would not only it would not not only take more time on my end it would not only probably lead to more bugs on the system. It would probably also lead to worse user experience because let's say I I don't I don't integrate Gmail well enough. That is such a core thing people use that if I messed up my user experience would tank and especially with consumer or business oriented product software products.

It only takes one bad experience to hurt retention and retention is the name of the game. I'm obviously going up against people like at OpenAI orthropic. They're already behemoths in the AI space. The only way I can I can get any any ground is by being open, extensible, customizable and having very very good retention.

And last of all, there is there are a lot of benefits that coming from being an aggregator service uh that operates at scale. You can you can negotiate better prices, better API prices uh with each of the with each of the providers. um that I as a as a one person team can't uh not only is it not uh not only is the is the is the return of interest uh not only is the rate of return not not good enough it is also actively against my best interests to focus on that more than the core product which is chat files context automations and agents also post has a fantastic free tier like they have a very generous fee tier I'm still on the free tier and it is it is more than it is more than enough uh to to to help startups get off the ground. If I if I were to be if I were to be integrating separate services into um into into my uh into into overlay natively, I would have to manage these APIs one at a time.

And for 400 integrations, that would mean 400 APIs, 400 accounts. I would have to create 400 uh 400 different o configs, which is just not it's just not possible. I I and I I really deeply reason about this from first principles. It's just not viable.

Sure, it's possible. It's just not viable for any team that that that values moving fast. Yeah. Which is why Composure saved me a bunch of time and I'm so grateful.

If a founder or enterprise dev is watching this and they're building with agents right now, what's the one thing you tell them not to get wrong? There are so many things. What are a few top three things? Top three things.

One, make it open source. Two, be an aggregator and a customizable and extensible solution. Three, do not make things natively that you can white label. and there are good solutions to white label.

Especially if you're at a startup or an enterprise, there is immense incentive to choose solutions that solve a massive part of the problem at a fraction of the cost. The reason why aggregated services also win is economies of scale. As I mentioned previously, use that. Do not be arrogant and think, oh, I could just make that.

I think I think that is that is a that is a that's a fool's errand to think you could be an aggregator of everything and also manage that. " No, you're increasing long run cognitive load. You're increasing long run operational costs for maintenance because all software that you make, you also have to maintain with also with something as as important as MCP, security is so paramount. Oh my god.

you know, every day in the news, uh you hear uh an MCP server, you know, call gone wrong and and it it compromised secrets because it wasn't integrated properly with with people like Composio actively taking care of it. I have never had to worry about about that happening. Yeah. Yeah.

So those three things would be would be very very if possible make it open source if uh if possible make it modular and extensible and three do not do not build things in house that you could white label easily. That's amazing. Last question. What are you most looking forward to for yourself and for also the industry moving forward?

Yeah, I for myself I would love I I think the product is at at a at a mature stage right now. Now it's distraction mode. I'm just reaching out to enterprises, reaching out to um uh to to uh uh to institutions uh that are in regulated industries like education, healthcare, law and finance um that would benefit from making their own solution in uh within their premises but also white labeling a solution so that the their development costs go down to basically zero. So it's just in traction mode right now.

I'm loving my time in in San Francisco in Founders Inc. They've been nothing but fantastic to me. Um, and also being around such cool people is is is bound to make you excited all the time. Right.

As for the industry, here are my bold predictions. Open source models are going to gain parity and intelligence with the best closed source models very very soon. In which case, at the risk of sounding like a a broken horn, it becomes so much more important to offer all of those models in one place. offer offer all AI services that humans and agents use in one place and three as the cost of producing software goes to zero software is going to be innately customizable and it would need to be extensible to even uh stand a chance in the market and fourth open source and aggregator solutions are going to be valued much more highly because of that fact because of of things commoditizing so that when there's equal access the the solution with the best access wins and the best access is is in solutions that aggregate services.