How AI platforms are stacking up

Many traditional vendors are announcing platform around AI, but what does AI platform entail and how does this impact the IT stack?

Over the past year, there has been a steady stream of AI platform announcements. One recent example is Intel’s AI platform launch, supported by a $1 billion investment in an AI ecosystem, including a scalable family of processors for AI workloads, vision technologies, and field programmable gate arrays for deep learning. Intel’s offering joins a list of variations on the AI platform theme being announced by other industry heavyweights such as Microsoft, Amazon, NVIDIA, and IBM.

According to an Intel spokesperson, until recently, the timing has not been right for AI platform development. But three converging factors have now made this possible. First, compute capability has progressed to the point where it has crossed the threshold needed to support the intense demands of machine intelligence. Second, smart and connected devices have unleashed a “data deluge” which is required to train many AI algorithms and is ready to be mined for fresh insights. Third, a surge of innovation has pushed industry over the tipping point from research to mainstream use.

Steve Irvine, CEO and founder,

This growing commitment to AI platform development by major players is a telling sign that the technology stack as it is known today will be transformed, said Steve Irvine, CEO and founder of, who left an executive role at Facebook in the US to establish his own AI platform development company in Toronto.

“This is the first time we are seeing companies like Amazon, Microsoft and Google making really big foundational commitments to build in the AI stack. [Google’s] TensorFlow is a good example. It’s only a couple of years old, but was built based on the belief that machine learning will be at the core of every piece of software built in the future,” he explained.

A key driver behind all this activity is the massive shift that is taking place in how programming is going to be done, Irvine added. “We’re moving from a deterministic world where you basically write scripts to say when something happens and do this, to a probabilistic one where algorithms can be trained to predict what will happen in future, based on historical data. Deterministic computing is not nearly as intensive as probabilistic, where you have to take data, train on it and make guesses. It’s much more complex, and because of that, requires a lot more power at different levels.”

When faced with explaining AI platforms – which he often is – Irvine said his go-to approach is to present on a white board where he draws a stack moving from the bottom to the top.

  • The bottom layer, the hardware platform, is where players such as Intel and NVIDIA are offering GPU (graphics processing unit) technologies that deliver the processing power needed for machine learning and AI. (Quantum chips will be the next evolution.) “When talking platforms at this layer, it’s the hardware and the chips capable of handling the big data and complicated math of AI,” Irvine said. “Simply put, you can’t do machine learning without having more computing power. Without it, running processes will take forever. In other words, if you don’t have enough horsepower at the bottom, you won’t have enough power to run the top layers.”
  • The next layer of the stack is new generation cloud platforms, including AWS, Azure and Google Cloud, among others. These will deliver the numerous core services that AI developers will build on. “They are now being powered more by GPU hardware clusters versus traditional CPU servers in the data centre,” Irvine explained. “This provides the hardware power that will allow developers and startups to build businesses without having to amass the hardware. They are also adding ‘packages’ that have not been available previously. For example, Google Cloud is not only more powerful, it has the ability to run frameworks like TensorFlow. These new generation platforms are being built to run neural networks and advanced machine learning. That part of the stack didn’t exist before.”
  • Above cloud services sit the frameworks that are specific to the AI/machine learning world. Examples include TensorFlow, Hadoop Caffe, Theano, Torch, Deeplearning4j, Paddle, MxNet, Keras, and CNTK. These frameworks allow you to abstract a lot of sophisticated math involved in machine learning and to build machine learning from the ground up without requiring a PhD, Irvine said. “To write core algorithms or build neural nets without any sort of library or framework would require an insane amount of work and very sophisticated expertise. You need to be able to build frameworks and libraries for repeatable tasks like running algorithms so developers aren’t starting from scratch.”
  • At the top of the AI stack is the service layer. In some cases, the whole platform can be adapted to support specific AI-related service layers such as’s machine learning-based predictive modeling for businesses serving consumers. Or it can deliver more broad capabilities such as IBM’s Watson. To build a platform that would help doctors understand if an illness is cancer or not, for example, the technology user might choose to deploy IBM Watson or to support his/her own particular use case by providing a backbone for third party applications that end users can consume through APIs.


The scope and complexity of the AI stack, begs the question; at what point does an enterprise begin to make the transition to AI?

There are several ways to address this challenge. An organization with an in-house data science team, for example, would be well advised to have the team familiarize itself with machine learning frameworks as a starting point.

Another approach is to identify a specific business issue and connect with a company that has an AI application that can solve that problem. “It might be that the app provider is also the company that builds the platform beneath it,” Irvine said. “With this approach, you can get results quickly and see how AI can be utilized to deliver business results before making a multi-year commitment to reworking your whole technology stack.”

Perhaps the most important consideration is that none of this will matter if companies can’t access their data, Irvine said. “If it is not in some sort of data lake or consumable format, AI adoption is going to be challenging. It’s no longer about structured data in a transactional or CRM database. The engine behind machine learning is the data sets that algorithms train on. You first need to learn what good data means in this AI-first world before building out the rest of the stack.”




  1. Thanks for this detailed guide, it is very interesting. I was involved with Artificial Intelligence and Machine Learning development in the past, and I am always curious how it works. Choosing the right AI platform is a complex task by itself, but this is only one layer, and you did a great job explaining the whole process. I am looking forward to build another business solution, and complete understanding of the process will help me to communicate better with my partners and get the best results.


Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.