You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
With AI becoming a critical factor in productivity and progress, what companies platforms will adapt the best in this fast changing ecosystem? Microsoft, Google, Meta, Amazon, Service Now and others are all in the race. Any thoughts on who the winners will be, or will it go to a newcomer in this area? Will the future AI be a universal fit or will it fractionate to several specialized platforms? Will custom chips become a major factor? Any thoughts, comments and additions sought and welcomed.
Right now most of the money flow is going to the infrastructure suppliers including chip suppliers, system builders, cloud service providers and LLM / Generative AI model builders. I have heard that 9/10s of all the big iron Gen AI chip deliveries are going to just 8 companies. Hardware time for last gen LLM model development/training cost something like $100M (within 2x) with current generation in training costing closer to $1B.
But the real money will come from the monetization of the killer apps based on the underlying technology. And I would say it’s anyone‘s guess how all that will play out, and even where the Gen AI inference is going to live is a big question - ARM, Apple, Qualcomm, Microsoft, Intel and AMD are all prepping for a lot of it to live at the edge on client devices, while some of those folks are also targeting on-prem and cloud data centers. Will the killer app be conversational search with advertising, conversational support, AI-driven human super-productivity (AI paralegals and radiologists), or as Jensen and Elon Musk speculate, trainable robots that do much more of the human labor that humans do today ?
Hardware (General purpose) —. Those making AI accelerators that anyone can and is using — think Nvidia, Intel, Others
Hardware (special purpose/new) — those creating niche or new to market companies. This may include Tenstorrent (which looks like they have some really good offerings), and others.
Software Layer
…
IMO - the Commodity layer will see benefits no matter what happens; there’s tons of support silicon required regardless of what direction AI goes, and this will continue to fund the race to more advanced nodes. And while the latest node always represents the best margin, it’s still a far lower margin than what Nvidia will see with Blackwell for example. OTOH - the volume on Commodity is a bit higher.
Hardware General purpose companies are seeing benefits and will continue to do so, but they will be subject to the boom/bust cycles that are inevitable in a buildup like this. There’s no way this will be a straight line path to success, but overall they should end up well off. Very high margins too at times..
Hardware (special purpose/new) - these are the traditional “small business” types that have a LOT of headroom for growth, but also a higher risk than the general purpose companies.
Software is all over the place, I won’t say more than that
..
I think the hardware standards are going to come from a few of the general purpose companies - either Nvidia’s going to have a permanent foot hold or some alliance between Intel, AMD, and others will drive an open standard for that side.
The software side is way more complex; I think the software still has to conform to hardware for now, but later the hardware will follow the software as we always see in mature ecosystems (i.e. 2000s vs 1980s desktops, 2020s vs 2000s mobile phones)..
TL;DR - Nvidia and/or other large companies pumping out AI / NPUs will define the standards for the next decade. The decade after will belong to whichever software standards have the most traction at that time.
Commodity+ - Chip manufacturers, Fabs; Think TSMC, Intel, Samsung
Hardware (General purpose) —. Those making AI accelerators that anyone can and is using — think Nvidia, Intel, Others
Hardware (special purpose/new) — those creating niche or new to market companies. This may include Tenstorrent (which looks like they have some really good offerings), and others.
How would you classify parts designed & owned by Data Center companies like Amazon, Meta, Microsoft & Google ?
They don't publish how much they spend on these parts (on purpose) -- and they don't care about external benchmarkets. But there are of multi-billions of dollars spent on these parts for each of these companies. (Barron's, JPM have done some limited reports on this)
