Models-as-a-Service
Pt. 1 - What the Heck Are They
Software-as-a-Service, or SaaS, is one of the most recognizable business models of the twenty first century. It has been around, in some shape or form for more than two decades. Its creation has led to some of the most powerful companies in the market today. It’s hard to pin down who exactly invented this business model, as it was predicated on cloud storage and the internet, among other things, technology that bloomed thousands of great companies all in a very short period of time. However, one of the earliest adopters of the term SaaS was Salesforce, who specifically positioned their centrally managed software solution as being anti-software.
While Salesforce might be one of the most impressive examples of a SaaS business, it certainly is not alone. There are SaaS CRM tools, ERPs, Accounting tools, productivity suites, HRMs, ATSs, CMSs, LMSs, and approximately one thousand additional alphabet-soup applications, many of which have examples of public companies currently worth billions of dollars and are masters at spitting off cash flow.
Why has SaaS been so dominant? Aside from the insane growth it has experienced due to the rise of the internet and cloud storage, it is arguably the greatest business model ever created. SaaS is easy to deploy - all you need is the internet and it can be set up anywhere, anytime. On top of that, it is easy to make - as the saying goes, “build one, sell many”. It can be hard to design and develop, but once the product is baked, it doesn’t need a lot of raw materials to reproduce in many instances, creating incredible margins. (A good enterprise SaaS business should be able to generate 80-95% gross margins.) Of course, this doesn’t guarantee that the operating expenses are going to be inexpensive, but that is more a reflection of management and market conditions, and has less to do with the business model itself.
There are some very clever folks reading this who might be thinking “but aren’t all these SaaS businesses big money losers?” To which my response would be “no, while many of them are cash flow negative to drive insane growth at absurd scale, there are countless examples of SaaS companies displaying the ability to turn on profitability in extremely short time frames”. The meme-thought that all software companies are money losers is the business equivalent of grandmothers who clutch their pearls and complain that “the world seems to be getting more dangerous” despite abundant evidence to the contrary.
SaaS is now ubiquitous. To the point where many folks, even deep in the industry use the term SaaS and Software interchangeably, unfortunately. If I hear a pitch and someone mentions that their product is “on-prem”, I wouldn’t be doing my duty as a diligent investor if I didn’t spend a fair amount of time inquiring why.
There are now a lot of other -aaS type businesses (even though SaaS wasn’t the first) that are now trailing in the footsteps of SaaS. There is Banking (BaaS), Payments (PaaS), Security (also SaaS), Infrastructure (IaaS), Insurance (IaaS), and many, many more. I cannot describe how many times I have heard a pitch where a company is selling itself as offering a product and calling itself X-as-a-Service, seemingly pulling the concept out of thin-air to try and generate a SaaS halo-effect.
However, one of the more popular XaaS’s in recent weeks has been the arrival of Models-as-a-Service, or MaaS. To be very clear, I don’t think anyone is really calling it that at this point. I have heard it being referred to as AI as a Service and ML as a Service, and a host of other things. But the point is that the rise of artificial intelligence and machine learning has led several startups and even bigcos to start offering AI/ML tools as a service.
So what does that mean? For the sake of this article, I am using MaaS to reference companies who have well-defined machine learning models and provide them as resources (usually by API) to other companies to use within their various service offerings.
ChatGPT is one of the more prevalent examples, as it has an API that it sells to developers to build NLP-enhanced products. Another company can use ChatGPT’s models to develop chat- and text-based tools, like a customer-success bot, a sales navigation chat function, or much more.
There are tons of other machine-learning models out there that companies can license to make their own products better. There is Amazon Polly, which turns text into life-life speech. Or Amazon Rekognition which white labels Amazon’s computer vision tools that can identify, label, and react to pictures. Or Google Cloud’s speech-to-text tool, which transcribes spoken word into written word. There are even off-the shelf tools designed to provide data scientists with the ability to quickly build models on their own without having to worry about a lot of the backend work that would necessitate a huge team of developers.
So there are companies that are focused on building models, like text-to-speech, computer vision, etc. And then there are other companies that are buying these models off the shelf and adding them to their repertoire to enhance their own businesses. Some of these folks are likely just using it to create a feature around an already robust product, while others are likely just creating a simple wrapper to resell the model using nicer packaging.
The question becomes how much extra work does a company who is buying these models off the shelf need to do to integrate them into their actual products and processes. While these models are getting more and more robust, that doesn’t necessarily mean that they are a one-size-fits-all situation. Many times software companies will need to do a fair amount personalization to make sure that the model works effectively with their stated value proposition. You might use Chat-GPT to build a chat system for your SaaS tool, but maybe the topic it needs to address are extremely niche, like answering questions about a specific subset of parts for a maintenance repair team at an industrial site. This might require adding an additional layer of training to the model to get it to work effectively. While that might seem simple, it certainly can take a lot of work and frequently requires fairly specialized talent.
At what point does it no longer seem feasible to pull these models off the shelf, but rather, to build them on one’s own? Many would say that if all you do is use off the shelf models, you won’t have a competitive moat. I think this is still being borne out, but I feel confident that many companies will need to develop their own models based on there very specific data sets. Where will companies draw the line between the efficiencies of using off-the-shelf models and the effectiveness of building it on their own?
On the other hand, there are going to be model-building companies that are tackling big enough use cases that they are going to embed in every applicable software the way some big payments players embed in every SaaS+ tool. Models-as-a-Service will allow software companies to focus on delivering value to customers and spend less time building stuff they aren’t equipped to be best-in-breed of. Much like how Cloud services allowed software companies to focus on building product and spend less time worrying about data storage and infrastructure, MaaS is going to allow software companies to focus on design, delivery, and business concerns, instead of trying to recreate the machine-learning wheel.