The Not-So Tidy Economics of an AI Startup
Technologies like AI, ML and Deep Learning probably need no introduction. Thanks to Andrew Ng’s super famous course on Machine Learning almost everyone can and have tried their hands on AI.
All of this in today’s time looks pretty simple. For almost every problem there is a GitHub repository that can get you started with the solution. And in case you happen to be someone who knows how AI algorithms and systems work behind the scenes, you really have an edge in coming up with solutions better than most of the other ‘copy-pasted’ run-off-the-mill programs.
How much can AI solve and to what degree of precision and accuracy is a question that deserves a separate in-depth discussion. However, there is something which is rarely discussed in all the online courses on Deep Learning or in all those meet-ups that happen all over the world every day.
How are the economics of an AI startup, especially in a place like India where not many are bent towards hardcore AI research but are more inclined towards taking a pre-baked solution, tweaking it and scaling it up to the next level.
An interesting story happened with someone recently. This group of startup founders was very keen on riding the on-going AI wave. Their goal was simple: make money somehow through all this AI hype.
So, I asked, what’s the plan.
The plan was also pretty simple and straightforward. They didn’t want to get into the more common product development scenario which startups generally are into. Working on a product for couple of months and try selling it and eventually (barring those 3% successful ones) not many stay afloat for long. The unit economics of a product based startup are again a separate discussion space but enough to say that these founders had seen how pockets can stay empty for long trying to sell a tech product. So, they wanted to get into AI Services, just like typical IT Services where client wants something and you make it for a price.
Plain and Simple.
But they were stuck. AI Services have a little difference than the traditional IT Services. Here you have to show you can or have developed a software solution that can or is doing something that is remarkable in some way. This is where things start getting little weird. The startup founders felt it will take close to a decent 7–8 months even to come up with a workable model with some respectable accuracy in case they want to develop a model which can be shown. And this is just capability show-off, not actual sales.
But that’s the route of a more common product based startup. There will be no revenue for close to 7 months and you don’t know how much will customer agree to pay even after you have developed a fairly strong core solution.
The other problem that popped up was that clients in general aren’t ready to completely procure the solution, rather they generally like to go for something like a revenue share model which brings in the additional onus of developing a high accuracy AI model without realizing money up-front. That’s completely different than say helping a client in adopting cloud or micro-services based architecture throughout their organization. Here the client pays based on pre-decided milestones.
Why is this happening? Because no one knows how accurate or useful that AI solution really will be. AI Services has started looking more like an oxymoron, something that can be rendered and sold like a service is something which is pretty much standardized and has a lot of community support and evidence that it works. That’s not the case with AI and ML based systems, at least not now.
Now let’s take a little deeper look at how the economics work in a so called AI Services startup.
It’s pretty clear until your solution and model are top-notch and doing something very accurately no client will offer up-front money. You have to invest time, money and energy in developing something that works.
Good thing about this is you work hard once and then if you are able to develop something that really solves a problem in a needed way, you just have to sell it with minimal incremental effort. It can be your cash cow, but, that’s a big ‘if’.
Now let us say you are able to develop something fairly workable. But how much money can it give you. There is no one answer here, startups have seen million of dollars coming to them during acquisitions and there are startups whose AI models fetch them merely INR 50,000/Year from a B2B client. That’s a pretty huge difference. On the other hand, AI developers have a pretty standardized cost and this is a big cost.
That’s one of the big reason why IT Services are way different than AI Services firms. Traditional IT Services operate in a model where something gets developed at a certain cost and the client pays you something which generally helps you realize the profit. You know up-front how much a certain piece of software will be sold at. Plus you do not really have to build something end-to-end beforehand. When clients say yes you start developing and in many cases even charge them on an hourly basis. This works because everyone kind of knows how much unit effort produces how much unit output.
AI systems are traditionally research driven. Even if three kickass developers work for two whole months you don’t know if they will be able to build something that will work. Plus there are other add-on issues which are alien to the traditional IT Services world.
Most important ones are Data and Training, both of which consumes a lot of power, money and time. Procuring data is something client do not care for, for them the AI provider has to bring a solution that is working. AI startups which are anyways low on cash cannot easily go and start collecting data in expensive mode. Even if you have the right data in the right format, training your AI programs on GPUs is a development cost which you simply cannot push to the client right away (until the client agrees to buy your full model at a huge premium, which is rare).
Interestingly, one of the startup did develop a niche AI solution for a client and the client agreed to use their model only in a mode where the data will not leave client’s ecosystems. That was troubling. The startup didn’t really have much choice but to either securely lock the containers or pack the softwares in a .exe or .dmg or something like that. The solution I am talking about here was of extreme value so the founders knew anyone will do anything to crack open the software and get access to the source code. Either the startup had to take the risk of giving away the solution (although securing it but it can be broken or just convince client in sending the data to their servers in a classic API-as-a-Service mode).
So, what’s the takeaway. AI startups are hard to bootstrap, they need huge up-front funding to hire engineers, run GPUs, maintain the infra and what not. And all this cannot ensure how much will a one year runway result into. The core to success perhaps lies in developing something that solves a very critical problem at scale, and it really solves it in a way that the solution can work well in real world out there. The AI industry is still at its infancy and there is hope that as the AI ecosystem matures solutions to these and many other problems will emerge.
Until then, it’s an exciting journey to be part of.
Later,
Team TechStreet
