The AI revolution is here, and as the ecosystem around AI matures, more startups are focusing on AI to solve different problems. We have collectively started to refer to these startups as "AI startups".
However, AI has become more of an umbrella term rather than a specific industry, similar to the fact that almost every product has a software element or deals with software during its production.
As startups and users are exploring the possibilities that AI unlocks, naturally, investors are also shifting their focus to invest in promising AI startups.
AI startups can be placed into 4 different categories. Sometimes a startup can be placed under more than one category, but that's more an exception than the rule, and it often happens in later stages of a startup rather than the earlier ones.
Every AI startup should be clear about the nature of its product and to which category it belongs, as each category of startups faces different challenges, both in capturing the market and raising capital; two things that often go hand in hand.
Category 1: AI infrastructure
The current AI models are very advanced and sophisticated compared to even the ones made a couple of years ago. This advancement is partially due to the ever-evolving infrastructure, pipeline, and workflows.
The first category of startups focuses on developing tools and infrastructure needed for data processing, training, execution, and distribution of the AI models, without directly offering an AI model as its product. These products can be hardware, software, or a combination of both.
These may not be the sort of companies that come to mind when we talk about AI startups, but they play a crucial role in the ecosystem and control a good portion of the market.
If you are a startup in this category, your primary customers are often not the end-users of the AI models, but the companies creating or offering them. This does not necessarily mean you need to be exclusively a B2B operation, but the main customers will be businesses relying on your tool while dealing with the lifecycle of their AI products.
The software startups in this category, very commonly, take the approach of open-source core with a paid layer. The open-source core will attract users, validate the product, gain a place in the market, and then lead users towards the paid layer.
The downside of this category is that you have the largest incumbents to compete against. If you are a startup designing AI-optimised semiconductors, you suddenly have to compete with the likes of Nvidia. The good part of this existing incumbent ecosystem is that if you manage to create a unique product (and survive), you also have the highest chance of getting acquired.
While raising capital, you need to approach a different set of investors than other categories.
If you are working on hardware products, the VCs with the highest chance of investment in your startup are those focusing on hardware infrastructure and compute.
If you are a software startup, the VCs with the highest chance of investment in your products are those focusing on developer tools, data, analytics, and cloud.
While raising capital, the founders in this category often make the mistake of focusing on the incremental improvements their products make.
If you are, for example, developing a new database that performs a certain type of query 5% faster than an established player, that's not going to lead to funding.
Even though a product that has a slight technical advantage is always appreciated and important, you need to shift your focus toward the problems your product solves, its scale, and the willingness of the market to pay for your product's advantage.
The biggest challenge you will face during fundraising is demonstrating that your product solves such an important problem that companies and teams are willing to migrate away from their existing stack into yours.
Category 2: AI models
At the core of the AI industry, there are AI models that have been trained on a dataset to perform a certain function.
There are many unsolved problems in our modern world that require a custom AI model. An AI model that is trained on a unique dataset and takes a new approach in connecting the underlying to the real-world solutions.
If you are a startup in this category, you then have a series of AI models you have trained and offer as the core of your product(s).
Startups in this category can have a moat that no other category can possess, and that's proprietary data. It doesn't matter what your competition does, as long as they don't have enough data, they won't be able to reproduce your models. That's why a company like Google was the forerunner of AI advancement for the past decade, as they had access to one of the largest data collections in the world.
However, these startups face three major challenges:
They need to have access to a proprietary dataset, which is not only difficult to obtain in the first place but is also getting more difficult by the day to use for AI training, as more jurisdictions are passing laws to regulate the data pipelines used for model training. There is a large corpus of AI models in the market, some available for free and open-source, some paid. If your custom AI model is at the core of your product, your model should have a clear distinction or advantage over the ones already available. Even though it may be insulting to some founders, they often have to answer whether what they make cannot be replicated with a few prompts in ChatGPT or Claude. LLMs have become interchangeable with AI, and people, both customers and novice investors, will compare any model with LLMs, for better or worse.
One very common mistake that startups in this category make is that they confuse the usefulness of their models with their ability to be sold as a product. Your custom AI model is rarely the product; the value it delivers to the consumers is the product.
Many companies have developed their custom AI model throughout the years, and those models have been very effective and useful as part of a larger product, but haven't been able to sell their models as a product. LLMs were the exception to that rule as the NLP processing allowed companies like OpenAI to revolve their product mostly around their models.
Category 3: Post-AI
This category of startups neither trains custom models nor develops the infrastructure for one, but rather focuses on the way AI is integrated into the existing systems that are yet to be redesigned for AI, or couldn't/shouldn't be redesigned for AI.
Throughout the past few decades, a whole series of industries have emerged for the way software should be developed and consumed.
The more critical a piece of software becomes, the more rigid are security, validation, verification, certification, and the audit process. The more sensitive the data a piece of software handles, the more attentive we become toward the way these data are handled, stored, and processed.
The systems we have for traditional software are not perfect, but it has worked reasonably well. There are software failures in the world with varying degrees of consequences, which make it to the headlines. And that is the point. Major software failures with a catastrophic blast radius are rare enough to make it to the news, in the same way that car crashes are not unusual, but the crash of an airplane will be talked about for years.
With the introduction of AI and the astronomical pace of its adoption, the old systems no longer work. Sometimes, even the developers of an AI model can't tell you how exactly their model works, let alone an auditor who is reviewing the model as a black box.
These types of gaps are the exact problems that these startups are trying to solve. They are not selling an AI tool but the processes necessary to integrate an AI model and make sure they don't crash and burn everything we have built over the years.
The biggest problem that these types of startups face is the incumbents, existing systems, and long sales cycles, which are very much based on trust and networking. A trusted giant with a slow pace and poor product or service has a higher chance of success compared to a newcomer with a better offering.
During fundraising, these types of startups have the highest chance of raising capital from VCs who are focused on regulated industries, social impact, and security. Even though this category of startups fills a real gap in the market and their product is quite sticky, they don't have the same fast-paced growth as other categories. So, the investors have to be familiar with the slow but steady pace that these startups have to take.
Category 4: AI-enhanced products
This is the most common type of AI startup out there, where the startup does not own or train the AI models in the product but uses the functionalities of these AI models to solve unique problems.
These types of startups became popular with the introduction of the ChatGPT API in late 2022 and early 2023, which commoditized access to high-quality LLMs through an API. Initially, many of the startups were thin wrappers, providing simple interfaces while calling the ChatGPT APIs in the backend without any meaningful additions.
In the early days, there was a wave of skepticism in the community about this type of startup, as they did not own the underlying model, had no control over the pricing, and the underlying model could be used by any competitor.
Over the past few years, however, the presence and usage of AI models have been accepted as the norm, and as long as the startup can provide a tangible value to the users, the ownership of the underlying model has become less of an issue.
One of the biggest problems that these startups face is the sheer number of competitors, both from new startups and established players. The startups that do not have a defendable product may get swallowed by an added feature by Anthropic or an AI integration of an established player.
Despite having a sizable investor directory, here on FundingBanker, and having analyzed thousands of investors in the past couple of years, I cannot confidently say what sort of investors would be interested in your startup, and that's kind of the point.
Almost any investor can be interested in your product, given that your product solves a tangible problem, addresses a large market, and people are willing to pay for it.
One of the grave mistakes that I see early-stage AI founders make during fundraising is that they rely solely on the existence of AI in their product as the main selling point. I commonly hear founders pitch their startups as "AI for fashion" or "AI-based sales system," which says nothing about the product besides the existence of AI.
AI is a tool. AI has become an important and fundamental tool in a short span of time, but it is a tool nevertheless. You need to go beyond the surface-level buzzword stuffing and describe your product and vision on a deeper level. The good news is that there are an infinite number of problems in the world that need solving, a large chunk of them genuinely need an ML model, and both investors and consumers are aware of this.
What this category means for you
One of the biggest killers of a startup is the lack of direction and clarity.
We have all been there. When you're deep inside the development mode, ideas and directions get tangled together, and it's worth it to take a step back and review your strategies.
When it comes to AI startups, you have a much clearer structure, something that other startups may not possess. You need to be very clear and honest about what type of AI startup you are.
Are you training your custom models? Are you providing infrastructure tools for the AI ecosystem? Are you controlling the behavior of AI? Or are you taking advantage of the existing AI models to solve niche and tangible problems that were thought to be impossible?
Once you are clear about the nature of your product, you need to adjust your fundraising strategy and the investors you target. We have a blog post on how to shortlist the investors that may come in handy.
The majority of investors will invest in an AI startup. This is not a mere assumption, and our data proves it. While selecting an investor, you will encounter two types of investors:
- AI-first investors whose main focus is that the product is either AI-based or involved in the AI ecosystem. Not only are they very vocal about it, but their portfolio company also reflects that approach. In such cases, you need to review their portfolio, making sure they support startups in your category.
- Investors who would invest in AI companies operating in the verticals they focus on. For example, an investor might focus on Healthcare and will not invest in a Fintech AI startup, regardless of how great the product is. So, for the majority of the investors, the underlying industry is the main criterion, while your job is to prove why your AI-based solution can fix a problem others cannot.
And always remember, try not to expand your scope into multiple categories, as each category requires its own specialization.