NVIDIA Reportedly in Talks to Lease Data Center Space for its own Cloud Service

[AleksandarK]

The recent development of AI models that are more capable than ever has led to a massive demand for hardware infrastructure that powers them. As the dominant player in the industry with its GPU and CPU-GPU solutions, NVIDIA has reportedly discussed leasing data center space to power its own cloud service for these AI applications. Called NVIDIA Cloud DGX, it will reportedly put the company right up against its clients, which are cloud service providers (CSPs) as well. Companies like Microsoft Azure, Amazon AWS, Google Cloud, and Oracle actively acquire NVIDIA GPUs to power their GPU-accelerated cloud instances. According to the report, this has been developing for a few years.

Additionally, it is worth noting that NVIDIA already owns parts for its potential data center infrastructure. This includes NVIDIA DGX and HGX units, which can just be interconnected in a data center, with cloud provisioning so developers can access NVIDIA's instances. A great benefit that would attract the end-user is that NVIDIA could potentially lower the price point of its offerings, as they are acquiring GPUs for much less compared to the CSPs that receive them with a profit margin that NVIDIA imposes. This can attract potential customers, leaving hyperscalers like Amazon, Microsoft, and Google without a moat in the cloud game. Of course, until this project is official, we should take this information with a grain of salt.



View at TechPowerUp Main Site | Source

 

[HisDivineOrder]

Good. Let these companies get Nvidia'ed the way gamers have. Hopefully, this encourages them to invest in and buy other options that eventually drive down profitability for Nvidia once they're a competitor.

 

[DY69SX]

Sounds like more better cloud gaming aka GF Now with DLSS10 is coming!

 

[Verpal]

Good. Let these companies get Nvidia'ed the way gamers have. Hopefully, this encourages them to invest in and buy other options that eventually drive down profitability for Nvidia once they're a competitor.

That would be too much fair play, alternatively, certain government entity will intervene and either fine NVIDIA or forced technology transfer, worst case scenario NVIDIA can even be split into multiple company.

 

[Assimilator]

This can attract potential customers, leaving hyperscalers like Amazon, Microsoft, and Google without a moat in the cloud game.

Far more likely, NVIDIA subcontracts its GPU cloud to AMZN/MSFT/GOOG. That way the latter still get to offer GPU compute services to customers, and NVIDIA gets to worry about managing the GPU hardware. As noted, NVIDIA can always procure its own GPUs at lower cost than anyone else, which would likely translate into cost savings for the big cloud providers, making such an arrangement highly attractive (of course, it's not going to translate into cost saving for the end-users of those services).

The concern there of course, is NVIDIA becoming a monopoly on cloud GPU computing, which honestly... is kinda the case already? So not much of a change, really, excepot it would be formalised.

 

[wNotyarD]

Taking the EVGA debacle to a whole new level, I see.

 

[Denver]

"As the dominant player in the industry with its GPU and CPU-GPU solutions"

It's obvious that it has the largest share of the GPU market, but is Nvidia using ARM cores a superior solution to monstrosities like the Mi300?

I really doubt it. But anyway, the main point is that Nvidia will compete with its own partners who buy its products in high volume. Hmm Yeah, I think Huang is arrogant enough to shoot himself in the foot like that eventually.

 

[thesmokingman]

Far more likely, NVIDIA subcontracts its GPU cloud to AMZN/MSFT/GOOG. That way the latter still get to offer GPU compute services to customers, and NVIDIA gets to worry about managing the GPU hardware. As noted, NVIDIA can always procure its own GPUs at lower cost than anyone else, which would likely translate into cost savings for the big cloud providers, making such an arrangement highly attractive (of course, it's not going to translate into cost saving for the end-users of those services).

The concern there of course, is NVIDIA becoming a monopoly on cloud GPU computing, which honestly... is kinda the case already? So not much of a change, really, excepot it would be formalised.

Most cannot afford to build a 10K gpu cluster so they have to suffer the software as a service route. I hope Tesla can get Dojo scaled...

 

[Assimilator]

But anyway, the main point is that Nvidia will compete with its own partners who buy its products in high volume. Hmm Yeah, I think Huang is arrogant enough to shoot himself in the foot like that eventually.

Partners aren't going to stop buying NVIDIA GPUs regardless of what NVIDIA does, for the simple reason that NVIDIA is the de facto only GPU compute name in town. AMD is trying but they're a decade or more behind at this point in terms of ecosystem (CUDA).

 

[Denver]

Partners aren't going to stop buying NVIDIA GPUs regardless of what NVIDIA does, for the simple reason that NVIDIA is the de facto only GPU compute name in town. AMD is trying but they're a decade or more behind at this point in terms of ecosystem (CUDA).

Huh ? In fact, all of them, the big partners, are already moving to no longer depend exclusively on Nvidia, whether through partnerships or creating their own solutions. In terms of hardware, in most aspects, the mi300 is already a level above the best Nvidia has today. Plus, Nvidia also does not have infinite production capacity contracted in advance.

 

[Assimilator]

In fact, all of them, the big partners, are already moving to no longer depend exclusively on Nvidia, whether through partnerships or creating their own solutions.

"creating their own solutions" LOL how, by building their own GPUs?

In terms of hardware, in most aspects, the mi300 is already a level above the best Nvidia has today.

Nobody cares about the hardware if it has no ecosystem, and AMD's hardware effectively doesn't. This is the same reason why Windows beat all other operating systems, despite not being objectively superior, and why it always will. Inertia is an incredibly powerful forced to overcome.

Plus, Nvidia also does not have infinite production capacity contracted in advance.

Neither does AMD, so what's your point?

 

[Denver]

"creating their own solutions" LOL how, by building their own GPUs?


Nobody cares about the hardware if it has no ecosystem, and AMD's hardware effectively doesn't. This is the same reason why Windows beat all other operating systems, despite not being objectively superior, and why it always will. Inertia is an incredibly powerful forced to overcome.


Neither does AMD, so what's your point?

Tesla Packs 50 Billion Transistors Onto D1 Dojo Chip Designed to Conquer Artificial Intelligence Training

D1 deliver 362 TeraFLOPs of power

www.tomshardware.com

Asic solutions such as those created by Tesla, Meta, Google etc... Nvidia's ecosystem is not necessary in most cases, current demand leaves room for competition.

If Demand exceeds production capacity, the market tends to absorb all the products it finds on the market. Microsoft's words, not mine.

 

[Assimilator]

Tesla Packs 50 Billion Transistors Onto D1 Dojo Chip Designed to Conquer Artificial Intelligence Training

D1 deliver 362 TeraFLOPs of power

www.tomshardware.com

Asic solutions such as those created by Tesla, Meta, Google etc... Nvidia's ecosystem is not necessary in most cases, current demand leaves room for competition.

D1 has been in development for half a decade and only now just starting to enter full production, now Tesla has to build an ecosystem around it. Hardware is always the easy part, software less so, which is why custom ASICs are only a solution if you intend to commit to them exclusively - a massive investment with a massive risk of failure. Plus it's not like Tesla doesn't have the same constraints with fabbing as NVIDIA does, since both utilise TSMC.

And Tesla are still using 10,000 NVIDIA GPUs anyway.