EXECUTIVE SUMMARY

This Intersect360 Research report presents the 2018 total market model and future outlook for infrastructure spending specific to the training of machine
learning models, including deep learning and other models for artificial intelligence (AI). This report also gives an assessment of the adoption of
machine learning training as a workload in shared environments, and how this affects system configurations and budgets.

Machine learning has been in a very high growth stage. Spending on infrastructure for training machine learning models has grown over 50% per year the
past two years, and it will soon surpass $10 billion. In addition, many systems not 100% dedicated to machine learning are serving training needs as
part of their total workloads, and in these cases, machine learning can be influencing spending and configuration.

For this report, it is critical to understand that while AI is a major IT trend, it does not constitute a “market” in the normal sense of the word. While
many organizations are investing resources in the training of machine learning models, in most cases these efforts overlap with IT initiatives already
in place. The training of AI models might involve systems, storage, or networks that were already in place or already budgeted. In the extreme, this
can even include supercomputers that are purchased with AI research as a primary directive, provided that it is reasonable to assume that some supercomputer
with some other primary directive would have been purchased in the same timeframe, regardless of AI.

This report identifies the primary market opportunityin which there are distinct, new budgets for infrastructure whose primary purpose is the
training of models for machine learning. Most of this spending comes from organizations in the Hyperscale market, and these training systems are a
subset or segment of the Hyperscale market. (See also, Worldwide Hyperscale Market Model: 2018 Spending and Future Outlook.) In some cases,
organizations that are part of the High Performance Computing market also have discrete spending on systems for AI, and these are counted here as well.
These systems are a subset or segment of the HPC market. (See also, Worldwide High Performance Computing 2018 Total Market Model and 2019–2023 Forecast: Products and Services andWorldwide High Performance Computing 2018 Total Market Model and 2019–2023 Forecast: Vertical Markets.)
Finally, we also address the secondary market opportunityfor HPC systems that are budgeted or configured in part, but not exclusively, for
machine learning.

In some cases, non-Hyperscale, non-HPC organizations may be investing in machine learning training. Our research suggests that as of 2018, this work—where
it exists—is done predominantly through cloud resources, and so again it can be considered part of the Hyperscale market. In addition, we monitor
how the adoption of machine learning training as a workflow into existing on-premise infrastructures is affecting budgets and configurations. While
this is not a distinct market segment, it does imply an opportunity for selling solutions with training capabilities.