At the time ONNX was announced, it seemed like a good move by Microsoft to partner with a savvy, but not competitive, AI company like Facebook in order to stick it to cloud computing rivals AWS and Google Cloud. Now, with AWS supporting ONNX, the partnership (from a business perspective, at least) looks more like a joint effort to fend off the Google threat.
That would make sense: in the world of deep learning frameworks, Google’s TensorFlow is the most popular, seemingly by a wide margin. Its popularity is aided by the popularity of Keras
, another deep learning tool designed to simplify the programming of models using a few different frameworks – but mostly TensorFlow.
And if there’s a company more synonymous with AI than Google, I don’t know what it is. Beyond search, Google Now, Google Home and the company’s various AI cloud services and APIs, there is DeepMind. The AI research company that Google bought in 2014 regularly makes news (sometimes bad news
) by pushing the boundaries of what AI systems can do. Its latest effort was an advanced version of its AlphaGo system
that mastered the ancient game of Go in just 3 days, using just a handful of Google’s custom-built Tensor Processing Unit chips. Google, by the way, plans to offer up those same TPUs to its cloud customers
as a way to run faster, cheaper AI models.
So, it makes perfect sense that Microsoft and AWS might see some advantage in working together on ONNX as a way to boost their respective AI frameworks and products.
However … everything I’ve mentioned in this post so far is open source, which says something about the nature of how AI technologies are being adopted
. Google could announce its support of of ONNX next week, and then this whole competitive argument would be moot. Everybody would claim they’re in it for users, and that they want to differentiate their businesses not by low-level tooling but by products, performance and user experience.
But the big difference there is that Kubernetes had such a lead in mindshare that AWS and Microsoft didn’t have much a choice but to embrace it. There’s no way of knowing for sure (although we can read between the lines) how either company would have reacted had other platforms or their own services been able to win that popularity contest. Also, AI is not infrastructure.
In AI, there’s no way of knowing if Google sees ONNX as a big enough threat to TensorFlow, or an important-enough industry standard, that it must embrace it. Or maybe the next move is the launch of a third-party TensorFlow foundation to try and nip in the bud any shift in momentum toward those other frameworks.
As much as open source is about altruism, it’s also about business strategy when you’re operating at the scale of AWS, Microsoft and Google. Considering the importance with which these companies speak about AI, it will be interesting to watch how collaborative they’re willing to be with their crown jewels as the businesses around them pick up.