AIAAS profitability

Climate change is making many businesses less profitable than before and using Artificial Intelligence could be a good solution to improve their profitability.

After describing How to create your first large-scale Artificial Intelligence As A Service (AIAAS), let’s see how to make them profitable, which are the profitability criteria and how to deal with hardware limits...

AI is one of the best solutions to fight against climate change. Photo by Matt Palmer on Unsplash

What is profitability?

Profitability is: “The fact that something produces or is likely to produce a profit.”

According to the Cambridge dictionary.

In other words, something that can produce a financial advantage compared to the current market competitors.

This advantage also applies to innovation projects like Artificial Intelligence, and it is often difficult to assess it, because:

• The hardware cost is higher than classic Software As A Service (SAAS).
• Satisfactory AI models are difficult to deploy in production.
• The resulting benefit vs investment is often not clear.

Eventually, it is a matter of difference between the perceived and the invested value as shown in the picture below:

The aim is to increase the perceived value and reduce the invested one.

How could we evaluate the criteria to maximize profits in AIAAS?

AI can explore new fields like creativity or fun, which were until recently unexplored. Photo by Dragos Gontariu on Unsplash

Profitability criteria

The main criteria to evaluate AIAAS profitability are the following:

AIAAS profitability criteria. Image by the author.

Some of them like performance and cost are directly linked to the AI model, while some others like usability and usefulness increase the service interest.

The sales criteria are the most crucial ones, and they should compensate the customer's investment with clear benefits as much as possible.

Remember that time = money, and video ads could be a good solution to compensate GPU costs, and even get benefits from them. In addition, you can limit the free service with some constraints to motivate people to pay a subscription plan. Those subscription plans could have additional functions to ease the decision to buy them as much as possible.

Be careful to have a good GPU auto-scaling, and its cost. GPU autoscaling is crucial to answer demand peaks and ensure good performance.

Accessibility and usability are key factors in AIAAS profitability. Most AIAAS creators are technical experts that forget to make the functions as easy as possible. For instance, some text inputs can be replaced by a list of options including the most frequent or useful requests. Some AI criteria like randomness or the number of output results can also exist as parameters for the customer.

Security is mandatory, and more important than classic SAAS because the results of calculation power are more expensive and could be stolen by hackers or bots. If a slowdown or shutdown occurs due to an attack, this could end the whole service. That’s why the AIAAS should check quite frequently if the user is a human and protect its system as much as possible.

Inspiring, interesting, and fun criteria are subjective, and that’s why they can be evaluated through tests done by humans or inquiries.

In all cases, all those criteria should be regarded in the long term as a continuous improvement. Model sizes, interesting functions, and hardware costs should be continuously improved. AIAAS is more subject to improvement than SAAS due to its complexity and innovation capabilities. Many starting AIAAS in production might have some errors or imperfections at the beginning but they are corrected later.

If your service is the only service to offer a new and exclusive innovation, this is a market advantage indeed, but it should be continuously improved over time to remain competitive and attract new customers.

This is a marathon, not a sprint.

Finally, many AI models must be big enough to have good results and several options are possible to deal with that.

Most advanced models require a lot of training to reach satisfactory results, just like humans. Photo by Filip Mroz on Unsplash

Profitability thanks to model quality

The more the model can answer interesting needs (accurate weather, smart chatbots, beautifully generated pictures, etc.) the more profitable it is, but the more memory and computing power are needed.

That’s why cheap and accessible AI services usually have worst results than expensive ones.

AI model quality highly depends on:

• Data scientist skills. An expert in Data Science will always have solutions to improve models in many unexpected ways. For instance, a model can be improved just by improving the data in the input by increasing its entropy or by removing useless information. A model can also be improved if transformed into a lighter model architecture such as ONNX or TensorFlow Light. She or he would know the model limits and be able to play with different hyperparameters to improve results.
• Data size and quality. The more quality data you have, the better chances you have to get good results. This includes good data variability that can be useful to differentiate cases.
• Training. The model training could be a very tedious task because it is improved through several long tries, failures, and tests. Many parameters are crucial like the learning rate or the dropout. And many technical solutions are possible like AdamW or training strategies (batch size improvement, objective function, etc.).
• Hardware. Nowadays, cloud services have a wide range of hardware options. You can have from low-cost CPUs with classic hard drives and 4GB of RAM for small models to very fast TPU with SSD and 20GB+ of RAM. Low-cost options are generally easily scalable automatically and are highly available. High-cost options like TPUs are incredibly fast and could process massive requests in a very short time.

Time management is a key factor in AIAAS. Photo by Laura Ockel on Unsplash

How to deal with big models?

Big models having more than 2GB could require better GPUs to offer a good level of service.

Depending on the calculation needs, instead of having expensive GPUs always available, you could set up an on-demand service that would lower the GPU use.

For instance, using a TPU 24 hours a day for one month would cost several thousand dollars. However, if you book a TPU for one hour per day that would process all the requests during that hour, it would cost a few hundred dollars.

In this way, it is possible to deliver high-quality service at a low cost, even if it is a daily service and not a real-time one.

On the other hand, there are many options for small to medium models.

Do not underestimate the marketing

Like any web service, an AIAAS needs to be promoted and the marketing is often a mandatory inversion but it is not necessarily expensive.

In fact, for a few hundred dollars, you can promote your AIAAS massively through ads campaigns on main social networks with hundreds of thousands of impressions per month (Facebook, Google, LinkedIn, etc.). The chosen social network highly depends on your aim profile. For instance, Google if you aim for the general population, LinkedIn if you aim at companies, and Facebook if you aim at people with a social ambition.

Created articles or videos can also increase visits, but new contents are necessary often.

Conclusion

AIAAS profitability is a key and complex field that will improve with time.

It is highly linked to skills, algorithms, and hardware as well as a good sales strategy including marketing to increase profit and lower investment.

And you, what are the solutions you are applying to increase your AIAAS profitability? Feel free to share them in the comments.