AWS and Google at least own their own hardware (Trainium and TPUs, respectively). It's a moat in the sense that designing, building, and deploying your own chips at scale is quite a feat and not easily replicated. The vertical integration will allow them to continue to be profitable once the models get good enough and competitors' prices race to the bottom. Google has Gemini; AWS may not deploy its own models (yet?), but that's not necessarily a losing position, as long as the market is able to run models sourced elsewhere on Trainium and the price is right.

▲

foobiekr an hour ago | parent | next [-]

Their "own" as in built by Marvell and Broadcom. Especially Trainium but also TPU4.

▲

amlib 2 hours ago | parent | prev | next [-]

Isn't specialized hardware also a big risk? GPUs are more amenable to any big changes that may happen in the next 5, 10 years of AI research. Maybe we won't even be talking about LLMs anymore. Maybe matrix multiplication won't even be the main primitive.

▲

moduspol 2 hours ago | parent [-]

If matrix multiplication isn't the main primitive, I think we have a lot of pain coming our way.

	▲	amlib an hour ago \| parent [-]
		Maybe that is far fetched, but I could see them specializing for some super high dimension multiplication and meanwhile 5 years later turns out "all you need" are 3x3 matrices and suddenly 90% of your specialized hardware is now dark silicon :)

▲

rsoto2 an hour ago | parent | prev [-]

That's exactly what giant train corporations thought. "We own all the railways, we've squashed the competition"

and they STILL went out of business because they over-estimated the demand for their shitty rails they built to the middle of nowhere. Same with "AI."

https://www.youtube.com/watch?v=2J2Fb1bBufA