AI Infrastructure Requirements: Not All AI Is Created Equal

This year, AI has taken center stage, and following the ChatGPT/Open AI hype, it has become a suitable topic for small talk at birthday parties. As industry professionals, we familiarized ourselves with the topic quite a while ago already. As have science fiction adepts, I can imagine.

Yet we all realize that the days of talking and fantasizing about the future perspective of the technology have (long) gone since the inception of the phrase ‘artificial intelligence’ in 1956; the time to apply AI within your organization is now. Even though the legal boundaries have not fully been set and social as well as societal implications are still in the open, there is a surge in the number of readily available AI tools and technologies to develop your own models.

At Leaseweb, we’ve been keeping a close eye on AI, as we do with all major technologies that impact our business and that of our clients. In fact, we have been working with many customers who deliver AI services to their end users, offering them the foundation of their AI solutions in the form of compute, network, and storage capacity. For example, to analyze extensive collections of data to provide insights for their customers.

Different AI Models, Different Requirements

It’s notable to state in this perspective that not all AI is created equal. Numerous algorithms can be used to build an artificial intelligence model, for instance, linear regression, logistic regression, and deep neural networks. The latter is a favorable method for performing Natural Language Processing as used in Chat GPT, but also in MidJourney, which is a very interesting text-to-image hosted bot.

Seeing as AI models differ, including the way they are used, the requirements for the underlying infrastructure – pipeline – will vary as well, typically including high computing capacity, storage capacity, and secure, low latency networks. To better understand the infrastructural needs of an AI model I’d like to break down its development and set out the characteristics of each phase, as well as the implications on the hardware and networks. The fundament of the underlying infrastructure for the AI will stand throughout its lifecycle, yet the different elements of it may need to be scaled up or down according to the specific stage that the model is in. Since Large Language Model (LLM) based intelligent agents (Chat GPT, MidJourney, Bard) show huge potential for application within enterprises, I’m using them as an illustration here.

Data Fitness

The process of building an AI can roughly be divided into the following stages. The first step is to collect data – in the case of an LLM, a lot of data – and to get the data in shape. To prepare the data for use in training a model, it needs to be pre-processed. A data quality assessment is the starting point for this procedure, followed by cleaning, transforming, and reducing your data. Tools to perform these actions can be Apache Spark, RapidMiner, Alteryx, Python, or others. More information on this topic can be found in our blog ‘An Introduction to AI’.

Storage capacity requirements in this phase are high, yet we’re not looking for ultra-fast SSDs. SATA-based storage servers would suffice, even though more speed is always nice. For the data crunching, we must put in some computing weight, so speed is subordinate here. Since network capacity is of importance in the data pre-processing phase, it’s advisable to use high bandwidth servers for the transportation of all the (unstructured) records. This becomes ever more relevant when the data resides in another location (on-premise) than the tooling (in the public cloud).

Application of Methods

The second phase is where we start shifting gears. To create an AI model, we want to apply several methods as described above to the (now fit) dataset. The one to choose depends on criteria such as the nature of the problem, data size and structure, and the desired accuracy of the output. Needless to say, mapping these factors in advance is crucial for a successful project, and may help to avoid overfitting i.e. running the risk of qualitatively poor output. Training the model currently happens mostly through supervised learning, unsupervised learning, and semi-supervised learning and can be done with tools like Python and TensorFlow.

This phase consists of initial training, where approximately 80 percent of the dataset is used to train the model, and validation testing, where the remaining 20 percent of the dataset (which is not used initially) is used to review the model on eventual shortcomings. To train the model, computationally, you may want to go all in. Whilst this may be a bit excessive for your purposes and may not fit your budget, it’s clear that this is the part that counts. The good news: there’s no need to maintain this top-notch infrastructure indefinitely, the model goes into production and you may be able to tone down the requirements for computing and storage in relation to the purpose of your AI. Regular retraining may be mandatory though, to maintain the quality of the model and to adapt it following the human input it receives.

To illustrate this, I’d like to refer to Isaac Asimov, who is renowned for his Foundation trilogy. The super-intelligent computer Prime Radiance records all events in the world and sets them off against the Seldon plan to save the galaxy. Every two centuries Prime Radiance must answer the question (paraphrased): “Are we still on track?” The ongoing recording demands mostly storage and networking capacity, and the bicentennial occurrence predominantly computing power. I’ve chosen this non-typical use case to stress that given the desired outcomes, the underlying pipeline for the AI will look different. Inevitably, new data will flow into the AI model to further tweak this, impacting the infrastructure once again.

Tailor Your Infrastructure

Conclusively, we can say that an AI has different characteristics throughout its lifecycle as do many applications. It’s also safe to say that not all AI is created equal. At Leaseweb, we can provide and tailor our infrastructural services to your needs across the globe, also offering direct connections to the public cloud where the Artificial Intelligence services are running. Feel free to reach out to us for a deeper understanding of how we can help you make headway in the world of AI.

Leave a Reply

Your email address will not be published. Required fields are marked *