So AI is a risky bet but Apple should definitely abandon its business model to chase it? Okay.
Your comment is too short for me to cleary understand your point. If by “business model” you are referring to the hardware part of my comment, AMD on Monday just followed my suggestion by releasing 2 Strix Halo CPUs for AI that are more affordable then the CPUs previously on the market.
That could be, but I would wonder if that’s true because the value of AI is widely variable between users. People who have figured it out are benefiting, and those who haven’t or aren’t using it effectively aren’t. Since AI solutions tend to be highly specific—it helps me but not others—that makes sense.
I had that thought while writing my comment, but avoided going down that particular path. On the one hand, there clearly is a lot of investment that may never be recouped in LLMs and generative AI, and companies like Microsoft and Google try to mitigate the costs by raising subscription prices to offer features that few people are asking for.
On the other hand, there just as clearly are use cases where AIs are delivering value, like training programs and customer support, or analytics that use machine learning algorithms that have become lumped in as “AI” for marketing purposes.
The real questions are which applications are delivering positive economic value today, and which applications may do so in the future (and when).
1 Like
I keep coming back to the gap between flat subscriptions and metered API list prices. On the API side the spread is huge right now: Claude Haiku 5.5 is about $0.10/$0.50 per million tokens (in/out), Sonnet 5.5 sits at $2/$10, and Opus 5.5 at $4/$20 — so a “Pro-like” month of heavy agent use can look cheap on a seat plan and expensive once you meter the same workload.
Subscriptions make sense when usage is bursty and you want a ceiling; raw API makes sense when you can route cheap models for bulk work and only pay frontier rates for the hard steps. The sustainability question for me is less “will labs keep losing money on seats” and more whether seat plans keep bundling enough headroom once agent loops chew tokens all day.