Isn't that risk balanced by a healthy reward of controlling their verticals and possible secret sauce?

And their chips give "1600 sparse INT8 TOPS" vs the Orin's "more than 1,000 INT8 TOPS" -- so comparable enough? And going forward they can tailor it to exactly what they want?

Orin is Nvidia's last generation. Current gen is Thor at 1k TOPS. Rivian's announcement specifies TOPS at the module level. The actual chip is more like 800 and probably doubled. Throw two Thors on a similar board and you're looking at 2000 sparse int8 TOPS.

I've been involved with similar efforts on both sides before. Making your own hardware is not a clear cut win even if you hit timelines and performance. I wish them luck (not least because I might need a new job someday), but this is an incredibly difficult space.

Mostly it costs hundreds of millions to develop a chip; it relies on volume to recover the cost.

NVIDIA also tailor their chips to customers. It's a more scalable platform than their marketing hints at... Not to mention that they also iterate fairly quickly.

So far anyway, being on a specialised architecture is a disadvantage; it's much easier to use the advances that come from research and competitors. Unless you really think that you are ahead of the completion, and can sell some fairly inflexible solution for a while.