Trade secrets? Like how to invent a trillion dollar technology and then sit on it for years while others eat your lunch with it? Like how to consistenly release inferior quality models to others despite infinite compute and engineering talent and insane profitability in your legacy businesses?
Not really sure what you're talking about. Apple just licensed Gemini for Siri, Google and their TPU hardware is starting to hit primetime audiences that OpenAI can only dream of.
The Apple deal is a good sign. Before that, it was like yeah Gemini might have the lowest cost per query but that doesn't matter when its users largely aren't paying for it and never will.
Google only won the deal because they have great cards in their hand. Apple's homemade PCC nodes were never going to scale, but the TPUs would. Investors expected market displacement in all the wrong places.
The tea leaves were pretty easy to read, too; LLMs will eventually be commoditized. Once that happens, the only road to profitable AI will be paved with 1) branding and 2) cheap compute. Apple does not possess cheap inference hardware whatsoever. Their GPUs are raster-focused, inefficient and expensive for their dedicated compute role. Unless Apple crawls back to Nvidia with their tail between their legs, TPU inference was their only real option. The deal was even more obvious when you used Apple's local foundation models; they're downright eclipsed by the quality of Gemma. Once Apple Intelligence was announced, it became clear that Apple was the branding component in search of a strange bedfellow to thumb their nose at Nvidia.
Would expect LLM training/inference infra to become commoditized before LLMs do.
There are multiple top-tier LLMs on the market right now. The competition is intense and the product differentiation is minuscule in many cases. Claude Code had a half-life of 1 or 2 months before getting sherlocked by the community.
We don't really see this type of opportunism in the hardware space. Case in point, the TPU is a pretty big accomplishment that successfully competes for inference and stimulates the need for cheaper compute.
My favourite part of their strategy here is its profitability