What Kind of AI Infrastructure is Best for my Business?
What Kind of AI Infrastructure is Best for my Business?

In this week's Exponential Chats, some of the team members responsible for Amalgam's development will have a chat about the various infrastructure alternatives available when it comes to training and deployment of Artificial Intelligence models. Between Cloud, Colocation, and On-Premise which one would you say is the best infrastructure for your AI needs? Come join us and participate by asking questions or giving your opinion in the live chat.

- Adriano Marques is the founder and CEO of Exponential Ventures.
- Nathan Martins is a Machine Learning/DevOps Engineer at Exponential Ventures, where he works on projects to democratize AI, as well as other cutting edge innovations.
- Rhuan Silva works as FullStack, DevOps Engineer and now he's majoring in Big Data and Artificial Intelligence.
- Fernando Camargo is a Data Scientist with in-depth machine learning expertise and wide-ranging experience in software engineering.

Creating Artificial Intelligence often requires a lot of computing resources. This is especially true when dealing with most Machine Learning techniques. When starting out, most Machine Learning engineers quickly learn how inadequate laptops and desktops are when it comes to serving as a platform for the development of AI models. In fact, the requirements are so extreme that even your typical server-grade hardware is inept for the job. That is why Engineers have been resorting to the big guns, such as GPUs, ultra-high-speed networks, and highly parallelized distributed storage solutions. All of this comes with a price tag, and I’ve got to tell you… it is not cheap.

There are three types of infrastructure to pick from: the Cloud, colocation, and on-premise.

The Cloud has become the preferred type of infrastructure for essentially any kind of business in need of hardware. The three major cloud providers are Amazon, Google, and Microsoft. Part of the appeal is the fact that you’re not required to acquire any physical hardware or supporting infrastructure, and you pay only for resources that you use. This is great for most businesses because it allows for instantaneous access to hardware and infrastructure without having to figure out how to acquire and set these up. This kind of infrastructure has facilitated an uncountable number of startups and apps that benefit us on a daily basis. It is great for almost everything, but when it comes to Artificial Intelligence, it fails dramatically in two important aspects: performance and price. Because all hardware and infrastructure are shared between different customers, you’re often unable to achieve the desired performance with the hardware you’re renting and when you do, you’re probably impacting the performance of someone else. And finally, there is the price problem. While startups are able to rent servers for as cheap as $0.002 per hour, when committing to full usage for 3 years, when it comes to hardware capable of supporting any kind of AI workload, you’ll be paying $9.057 per hour on a single server, again, when committing to full usage for 3 years. That comes up to a bill of $241,278 at the end of the 3 years with a single server and not counting the other charges related to data transfers. For that reason, Cloud infrastructure has been prohibitively expensive to most companies. Even big ones! Take it from NASA: they forgot to account for the cost of data transfers and ended up with the prospect of having to pay an extra $30 million dollars to their yearly budget if they wanted to do anything with the machines they reserved with AWS.

Colocation solves some, but not all of these problems. It gets most of the costly infrastructure issues out of the way, such as cooling and power redundancy, but leaves you having to acquire and manage the hardware on your own. Plus, access to the facility is not easy due to the extraneous access control measures, and finally, you end up sharing some resources with hundreds of other companies in the same datacenter, which can impact your business in different ways. It is also very expensive: the monthly dues for a single rack are on average $2,500 discounting other fees that add up based on the kind of access and add-ons you’re looking to benefit from.

Finally, there is On-Premise. It is admittedly, the hardest one to manage and with the highest upfront cost. But when it comes to AI it is the CHEAPEST in the long run, the one that affords the most freedom, and the most performant. Since you won’t be sharing resources with anyone else, you’re free to experiment and reach the capacity of your hardware without fear of racking up fees or enduring throttling. Moreover, you’re in complete control of your hardware and how you want to support it.

Despite the cost and performance discrepancies, I believe that different businesses will benefit more from different solutions.  Which is why we’re having this discussion today.

Exponential Chats is a live event conducted by our parent company, Exponential Ventures. In this event, our team members and guests have an in-depth conversation about Exponential Technologies, Entrepreneurship, and some of the world's most challenging outstanding problems.

This live episode streamed on Thrusday, October 1st 2020 at 1:00 PM CST. Here is a link to video on Youtube: https://www.youtube.com/watch?v=h0Ap0VP0rgE