Previously, we had attempted to discern how Ai was a different beast than traditional software and services; primarily, from business point of view. The lower margins and higher development costs may not make it suitable for every case. The biggest business concern is that the costs associated with developing an Ai infrastructure are hidden, inexplicable and variable – fine points that will not exude confidence with neither the decision makers nor the auditors.
On-Premise to SaaS
In the beginning of time, the software was developed, loaded onto media, shipped to the buyer, who bore the cost of running it – as in the desktops and servers were owned, maintained and operated by the end-user.
With SaaS (Software as a service) – the cost is borne by the software company, who pay humungous Azure (or AWS) bills every month, and the costs are directly proportional to the complexity of the software.
In a nutshell – the cost has shifted from the end user to the developer. And those costs are high.
Artificial Intelligence is challenging and demanding
Ai requires thousands of dollars in compute resources and unlike non-ai software, that is not a one-time cost, since data that the ‘Ai model’ feeds on keep changing. This is called Data Drift – another topic for another day.
Besides Data Drift when applications operate on rich media – images /audio/video – the storage costs associated are very high – something that we have noticed during our traffic analytics exercise.
Ai applications are not easily scalable i.e. cost efficiently – so you have to transfer the trained models from one cloud to another. Piling up on egress and ingress costs.
Remember we said Ai resides in the long tail
For your Ai model to spit out meaningful results and give that competitive edge – you will require exponentially more processing, more data and more intelligent human resources, and a successful model that has the above ingredients is not every organization can afford.
Some might say, distributed computing is a fascinating solution – but it addresses speed and not costs.
Also the transistor count of NVIDIA GPUs has only grown about 4 times over the last decade – which means that even the hardware technology is not keeping pace with the growing complexity – that essentially means higher costs.
“Cloud operations in Ai and Data Analytics, are more complex and costlier than traditional software and to scale them even more so.”
Next, in the Business of Data Analytics and Artificial Intelligence, we will take closer look at the humans.