But Mobileye isn't sitting still, so if the relationship hadn't been terminated, Tesla would've been stuck with whatever Mobileye's latest and greatest offering might be.
What's interesting is that if you compare the HW2 sensor suite to Mobileye's next-gen sensor plans, they are very similar, almost the same. They were both working on something together before the relationship went sour.
Musk hinted recently that 1 PX2 is not enough, and that 2 may be necessary for L4. There's certainly no consensus as to how much compute is needed, the more the better, but I'm skeptical that the hardware costs will be low enough near term to justify putting it in privately owned autonomous cars. I think Tesla has made costly strategic errors in the Autonomy department, and that Musk in particular has repeatedly underestimated the challenge of developing a fully autonomous vehicle. I don't think he's making that mistake anymore.
You couldn't train a good autonomous driving agent even if you used a massive GPU cluster.
I'm not saying better hardware doesn't help, it definitely helps incrementally but it's not going to be enough to solve this problem.
However I'd wager to say that betting on a powerful Nvidia GPU/chipset (piggybacking on general ML/CUDA ecosystem progress) to be able process rapidly advancing autonomous driving algorithms is a more future proof bet for Tesla than relying on already outdated hardware/algorithms from Mobileye.
"It's been a bear to get done, but prob 10 days or so, depending on full speed autosteer test results" https://twitter.com/elonmusk/status/841041891747545089