Ask HN: Why don't GPU/TPU manufacturers commoditize their RAM complement

Posted by DoctorOetker 2 days ago

Counter4Comment11OpenOriginal

Wouldn't say nvidia sell more GPU's if RAM prices were lowered? What prevents them from buying up IP / RAM fab labs and commoditizing the blueprints for its manufacture? Perhaps to the public at large, or perhaps just relicensing it for very low prices to entities in friendly nations?

Comments

Comment by kristianp 2 days ago

In terms of data centre gpus, HBM is far from a commodity as only a one or two manufacturers can fabricate the stacks of RAM.

https://newsletter.semianalysis.com/p/scaling-the-memory-wal...

Comment by Sevii 2 days ago

RAM is already a commodity.

Comment by DoctorOetker 2 days ago

"commoditize your complement" is an expression, and thus means more than just the individual meanings of the words combined.

Comment by nrhrjrjrjtntbt 2 days ago

This is the reference for anyone interested: https://gwern.net/complement

I think commoditize means "make cheaper, maybe by commoditizing" in this context.

Memory is and was commoditized. It just got more expensive.

Comment by wmf 2 days ago

I don't think there are any steps they could take to sustainably reduce DRAM prices.

Comment by theandrewbailey 2 days ago

RAM isn't their business, and that market has lots of ups and downs. What happens to Nvidia's hypothetical RAM investments when the AI rush is over and RAM is cheap and can't compete in that market? They've come to the logical conclusion to let the RAM companies do their thing.

Comment by DoctorOetker 2 days ago

Sure the market will have ups and downs, commoditizing your complement is not an investment in the complement, its an investment in their business to increase sales of their primary market by decreasing costs for their customers.

They can choose how strong they commoditize their complement: perhaps buying up and expanding operations of RAM manufacture (an investment indeed, or a revenue insurance if a GPU competitor springs up), or more expensive buying up IP and licensing IP below market prices to novel RAM factories in friendly nations perhaps with exclusive priority or agreed prices for RAM at specified volumes etc., or even more expensive buying up IP and putting in public domain. The most optimal depth of commoditization would have to be calculated by such a GPU manufacturer.

Comment by raw_anon_1111 2 days ago

No commoditizing your complements is buying up your complements, it’s either working with a standards committee together with the also rans to take on the market leader or open sourcing your competitive product to sell more of what you care about.

Comment by onion2k 2 days ago

They spend a billion or twenty persuading everyone that they need more RAM, and software developers bloat their apps until it's true. The market for memory continues to thrive.

Comment by brudgers 1 day ago

Ok, I'll byte...

Nvidia's business is cored on selling hardware to price insensitive customers. Those customers are price insensitive because the value created with Nvidia's GPU's is a non-fractional multiple of their capital cost and the capital cost of the GPU's is a fraction of the capital cost of the overall infrastructure.

And this is really important: if you maintain your margins, higher prices mean more net revenue. 30% of $2500 is less than 30% of $3000 while logistic and marketing and development costs are the same.

The retail market is price sensitive but is not the core of Nvidia's business. Large price insensitive customers are what drives development direction and what goes into retail is condensed from the exhaust fumes.

As a customer for RAM fabs, Nvidia drives the direction of RAM development without capital expenditure on infrastructure and R&D. It also mitigates the investment risks from political changes and disruption caused by technological breakthrough (the inertia of obsolescent RAM fabs driving design decisions).

Or to put it more simply, the age where near-leading-edge GPU hardware was readily accessible to consumers are past. The leading edge is increasingly military-industrial scale...with all that implies. Good luck.

Comment by axegon_ 1 day ago

Everything that is currently going on is a result of people who are convinced they know what they are doing. Spoilers: they don't. The sooner the AI bubble bursts, the better.

Comment by Getlazy 2 days ago

[dead]