Many avoid predictive analytics programs because the capital investment needed just to find out if you can benefit from it creates a risk-adjusted ROI that is a tad too shaky for cautious, capital-constrained organizations. Purchasing the machine learning infrastructure is part of that ROI calculation and if it doesn’t succeed, can the software be re-purposed or returned? To reduce the risk of getting started, Microsoft has chosen to use a zero-down pricing program for their new Azure Machine Learning (ML) service that recently became generally available.
Azure Machine Learning Pricing Factors
This is the pricing for the pay-as-you-go format as presented by Microsoft: http://azure.microsoft.com/en-us/pricing/details/machine-learning/
I took this information and used it to predict the annual charges for using Azure ML.
I am ignoring the costs for developing and training a model, as this post is only concerned with the usage charges for a model in production.
There are four inputs into the annual calculation. These two are from Microsoft:
1. The cost per 1000 predictions requested.
2. The compute costs of each prediction.
These are from your own organization:
3. The time it takes to calculate each prediction.
4.The number of predictions in a year.
From April 1, 2015, Microsoft charges $2.00 per compute hour and $0.50 for 1000 predictions. The compute time is how much time it takes to make that prediction. You are also charged for storage and egress. The charts do not speak to the storage and egress charges.
Any point on the chart is calculated so:
|Price per compute hour||$2.00|
|Number of millisecond in an hour||3,600,000|
|Compute charge for 1 millisecond||$0.00000056|
|Charge to request 1000 predictions||$0.50|
|Total charge for 1000 predictions each taking 1 millisecond||$0.5006|
|Total charge for 1000 predictions each taking 50 millisecond||$0.5278|
|Total charge for 1000 predictions each taking 500 millisecond||$0.7778|
|Total charge for 3600 predictions each taking 1 second||$3.8000|
If you have 1,000 predictions per hour, 40 hours a week, every week of the year, with 500 milliseconds per each prediction, then your annual bill will be a little over $1,600 a year.
Theoretical Azure ML Pricing Scenarios
As for all these scenarios, no server purchase required, no power bills, software licenses or maintenance to pay; no load balancing techniques to perfect and no disaster recovery plans to create and test.
Use this chart to calculate your annual costs for your standard office-hours operation of 40 hour week, 52 weeks of the year.
- Determine how many predictions an hour you would require and locate it on the horizontal axis.
- Use the legend to find the line that matches your expected compute time per prediction.
- Find the intercept and trace it to the vertical axis on the left to see the total annual charge.
If your operations runs 24 hours a day, every day of the year, then use this chart. The technique is the same as for the 40 hour a week.
For 24X365 operation it is less than $25,000 annually to run 3600 predictions every hour, with 500ms per prediction. Some would consider that a bargain to avert an unplanned shutdown, product recall or prosecution, and that is at list rates. After your Vendor Management have worked their magic and Enterprise Agreements are signed it could be less.
While other machine learning toolsets need additional features to create integration points to operationalize models, and the extra source control, infrastructure and load-balancing needed to operate those integrations, Azure ML is quite different. With Azure ML once you have learned something to your advantage, a few clicks can publish it as web services, ready to integrate into your operational systems, knowing you can confidently scale up without any further ML capital investment.
Selling your ML Program
It can be a challenge to guarantee how your will create the same value as the one in the compelling machine learning case study you are presenting as part of your ML program funding request. That uncertainty of success makes cautious, capital-constrained organizations shy away from ML programs, but with Azure ML , those concerns are eliminated and can make it far easier to obtain approval to proceed.