The obvious read from Chris Lattner's appearance on The TWIML AI Podcast with Sam Charrington is that chip portability is solved. I do not buy it yet.
Lattner, now EVP of advanced AI software and platforms at Qualcomm after its Modular acquisition, made the claim fairly. He said Modular rebuilt the stack from silicon up with Mojo as the language and Max as the serving layer, according to the outlet. Batching, KV cache handling and speculative decoding stay common, Lattner said, while only the inner matrix math gets tuned per chip, which is how six architectures from six vendors were supported ahead of ModCon.
Portability without penalty has never existed.
What Lattner did not show is the hard part. There were no performance comparisons, no cost of the specialized kernels, and no named user running the same model in production on Nvidia and Qualcomm. Max remains gated too, since Lattner said Mojo is now Apache licensed but an open alliance around Max is coming later this year. That means the portability layer itself is still not fully open.
For the obvious reading to hold, the common code would have to be fast enough and the per-chip kernels would have to stay current with Nvidia CUDA and Google TPUs. The fair concession is that Qualcomm has incentive to try, since Qualcomm said in its closing announcement that Modular will keep its open, multi-hardware approach. That pledge matters because a startup could only afford to chase cloud, while Qualcomm needs software for data center and phones, laptops, watches and IoT.
If you are choosing silicon this quarter, do not bet your roadmap on write once.
