I think their play has been building specialized hardware that executes TensorFlow better than anyone else. "You could use a GPU to do this, but check out our custom ASIC that does it 400x faster for 1/5th the cost..."
The 400x was clearly hyperbolic – and sure. But you probably don't have one-click integration with an existing battle-tested IaaS platform.
Vertical integration is powerful, and by open-sourcing Tensorflow Google is achieving useful synergies in sales and recruiting. At their vast scale, even small ROIs (as a percentage) can be massive.
And the hardware advantage is easily negated. For example, our startup is building something like this.