Google Open-Sourced Its Crown Jewels. Why?

Google has released TensorFlow, the system it uses internally for machine learning, under an open source licence. Free to use, free to modify, free to build a competing product with.

Every executive I’ve mentioned this to has asked the same question, and it’s the right question: why would they do that?

The Cynical Answers, and Why They’re Incomplete

“It’s not the real one.” The most common suspicion. Google is keeping the good version and released a hobbled one. There’s a grain of truth here in that the initial release doesn’t support distributed training across many machines, which is exactly what Google needs internally and exactly what a competitor would want. But the core is genuinely the thing. This isn’t a toy.

“It’s a recruiting exercise.” Partly true and not trivial. If every graduate student learns your framework, hiring gets easier and onboarding gets shorter. Given what the talent market looks like right now, that alone might justify it.

“They want your data.” It runs on your hardware. It doesn’t phone home. This one doesn’t hold up.

All of these are real and none of them is the main thing.

The Actual Logic

Here’s the strategic argument, and once you see it you’ll see it everywhere.

Google does not make money from machine learning frameworks. Nobody does. The framework is infrastructure, and infrastructure is a cost centre that happens to be necessary.

What Google makes money from is advertising, cloud services, and products that are better because they use machine learning. Every one of those improves when the surrounding ecosystem gets bigger and more capable.

So the calculation is:

If the field grows, Google wins, because it has more data, more compute, and more places to apply the results than almost anyone.
A common framework makes the field grow faster. Fewer people reinventing plumbing, more people building on shared foundations, faster circulation of techniques.
The framework being theirs makes it grow in a shape that suits them, running well on the hardware they use and eventually on the cloud they rent.

“You give away the thing you can’t sell in order to grow the market for the thing you can. The framework is the razor. Their business is everything else in the bathroom.” — Sameer Gupta

There’s a name for this pattern in economics: commoditise your complement. If your product is more valuable when some adjacent thing is cheap and abundant, then making that adjacent thing cheap and abundant is a rational investment even when it costs you money directly.

What It Means for Everyone Else

Set the strategy aside. What actually changes for a company that wants to use this stuff?

The framework question is settled, or will be shortly. There have been several serious options for a few years: Theano, Torch, Caffe, each with its own community and idioms. Google’s entry with Google’s documentation and Google’s marketing will consolidate a lot of that. Consolidation is good for you. It means more tutorials, more people who know it, more code you can copy.

The barrier moves from tooling to data and judgement. When the infrastructure is free and standard, nobody has an advantage from it. Everyone has the same tools. What differentiates you is the data you can bring and whether you know which problems are worth attacking. I’ve been making this argument for a while and each of these releases makes it more true.

Build-versus-buy shifts toward build, at the margin. Not for everything. But the class of problem where you’d previously have bought a vendor solution because building was too hard just got a bit larger on the build side.

The Caution I’d Add

I want to be careful not to oversell this, because “free framework” is going to be read by some people as “machine learning is now easy.”

It isn’t. The framework is the least difficult part of the problem.

You still need labelled data, and you probably still don’t have enough.
You still need to know what you’re doing. A free tool for building models is also a free tool for building models that look fine and are wrong. Nothing in the release helps you tell the difference.
You still need to get it into production, which is engineering work that no framework does for you.
The learning curve is real. This is not a library you call. It’s a system you express computations in, and it takes weeks before you’re productive.

Free removes a cost that was never the main cost.

Final Thoughts

I think the release is genuinely good news, and I think the strategic reasoning behind it is more instructive than the software.

The lesson I’d take into a boardroom is this: Google looked at its machine learning infrastructure, an enormous investment built over years, and concluded it was not a competitive advantage worth protecting. The advantage was elsewhere, in the data and the distribution and the ability to apply results at scale.

That’s a useful question to ask about your own capabilities. Which of the things you carefully protect are actually advantages, and which are just expensive infrastructure you’d be better off sharing the cost of?

Most companies have the answer backwards.