Apple is reportedly investing heavily into Nvidia servers for AI development

AppleInsider

Perhaps defying Apple's very public claim that it is using Apple Silicon servers to run Apple Intelligence, an analyst insists the company is now spending $1 billion to buy Nvidia systems as well.

Siri icon in a datacenter

In April 2024, there were the rumors that Apple would use its own Apple Silicon processors to run its AI servers. Then in June 2024, the expectation was that whole data centers would be run on Apple Silicon chips.

By September 2024, it was certain. Apple's Craig Federighi said its Apple Intelligence servers were on Apple Silicon, and that this was crucial to making the company's AI services private.

Yet now according to Loop Capital analyst Ananda Baruah -- as spotted by Investor's Business Daily -- Apple is in the process of spending around $1 billion to order in new Nvidia servers specifically for generative AI.

"AAPL is officially in the large server cluster Gen AI game and SMCI [Super Micro Computer] & Dell are the key server partners," he wrote in a note to investors. "While we are still gathering fuller context, this appears to have the potential to be a Gen AI LLM (large language model) cluster."

Baruah claims that Apple is buying 250 Nvidia NVL72 servers at a cost of between $3.7 million and $4 million each.

According to Nvidia, its NVL72 server contains 36 Grace CPUs and 72 Blackwell GPUs. The company also says that, as of March 18, 2025, that this server is not yet available.

Doubtlessly Apple could pre-order the servers now, and it's not surprising that the company sees a need to expand its servers. Given the quantities, this may be for development purposes, and not public-facing, but there's no way to tell right now -- and that's assuming the report is correct.

If it's for more than development, this just doesn't quite fit with Federighi's claim that he thinks using Apple Silicon servers "sets a new standard for processing in the cloud in the industry."

"Building Apple Silicon servers in the data center when we didn't have any before [and] building a custom OS to run in the data center was huge," he said. "[Creating] the trust model where your device will refuse to issue a request to a server unless the signature of all the software the server is running has been published to a transparency log was certainly one of the most unique elements of the solution -- and totally critical to the trust model."



Read on AppleInsider

fred257

Apple Intelligence is a scam

koron

So the last shall be first, and the first last: for many be called, but few chosen.

abolish

This isn’t defying anything. The nvidia chips will be used for training models, which doesn’t involve any user data. The custom silicon is used for handling user queries and applying those models.

muaddib

Hopefully this development of Apple AI servers will make way its back into Mac OS so it can run LLM and Diffusion models locally better. 

discountopinion

As previous poster has said - Nvidia is likely used for training and fine tuning whereas Apple Silicon is used for inference. Scaling laws are now also relevant for inference with the advent of reasoning models. Apple still has a capex light AI strategy which is a boon for its bottom line.

mikethemartian

Doesn’t Nvidia ration out their server chips because of high demand? Has their relationship with Apple improved?

avon b7

No doubt CUDA is vital here as I haven't heard anything about a complete Apple AI training stack for use with the heavy lifting.

Nvidia has CUDA. Huawei has CANN. 

Has Apple released an equivalent solution? 

blastdoor

avon b7 said:

No doubt CUDA is vital here as I haven't heard anything about a complete Apple AI training stack for use with the heavy lifting.

Nvidia has CUDA. Huawei has CANN. 

Has Apple released an equivalent solution? 

Apple has Metal Performance Shaders and MLX. 

I'm not qualified to say whether they are 'equivalent' to CUDA. But I believe they are focused on doing the same general job. 

blastdoor

I really hope that Apple primarily uses their own silicon for both training and inference because I think that could be a nice competitive advantage for them in the long run. 

But I could see using Nvidia in a limited capacity in the short run as Apple works to catch up. 

kju3

blastdoor said:

avon b7 said:

No doubt CUDA is vital here as I haven't heard anything about a complete Apple AI training stack for use with the heavy lifting.

Nvidia has CUDA. Huawei has CANN. 

Has Apple released an equivalent solution? 

Apple has Metal Performance Shaders and MLX. 

I'm not qualified to say whether they are 'equivalent' to CUDA. But I believe they are focused on doing the same general job. 

Metal Performance Shaders? No. MLX? Yes. The issue with CUDA is that it has been around since 2007, no one else had an incentive to put out a competitor for most of that time so the best software toolkits have been built on top of what was the only game in town, plus it is what the most experienced programmers, engineers etc use. So what Apple needs to create competitive AI applications doesn't work with MLX and not enough people know the stuff that does work with MLX. Now if this was 3 years ago before Microsoft triggered this genAI boom and had Apple made this a huge priority it wouldn't have been a problem: Apple could have created whatever they need for MLX/Apple Silicon and trained enough developers to do the work. As neither was the case and Apple finds itself needing to make as much progress as quickly as possible, they need to go with a solution that allows them to just plug it in, hire proven programmers and engineers who have done this job in the past and get going. 

Everyone is just going nuts over this because it is Nvidia, who is on the rather long list of companies that Apple fans are supposed to despise (along with Microsoft, Google, Intel, Samsung, Qualcomm, Masimo, Amazon and I am certain that I am leaving out a lot more) despite Apple's own, er, history of doing stuff. Such as Steve Jobs accusing Nvidia of IP theft and Apple getting upset at Nvidia's refusal to make custom MacBook GPUs under terms that likely would have bankrupted Nvidia. But honestly, it is only 250 servers for less than $1 billion. Lots of companies far smaller than Apple are paying far more to buy way more.

They are just going to be used to cover gaps that Apple can't immediately fill with their own tech: short term stuff. Other companies have already spent far more time and money being among the first to do what Apple needs to get done now. Apple will be able to trace their steps at a fraction of the time and cost while avoiding their mistakes. Once they are finished using the servers and CUDA to play catch-up they'll be done with them and will probably donate them to some university or nonprofit for a tax writeoff, and the engineers that they hire to work on this will make top dollar for a relatively brief gig and will leave with the Apple experience on their resumes that will allow them to work wherever Apple's noncompete clause allows. And yes, this means next time they will actually go with Nvidia when they want to instead of when they have to, which is the way that it should be anyway. As Apple is working with companies that they have literally sued (or threatened to) like Microsoft, Samsung, Google and Amazon then there was never any reason to try to freeze Nvidia out in the first place. That MacBook GPU thing? Well Apple wound up using AMD GPUs that weren't nearly as good, which forced a ton of people who needed the best graphics to buy Windows machines with Nvidia cards instead. So Apple really showed them, didn't they?

edited March 25

danox

mikethemartian said:

Doesn’t Nvidia ration out their server chips because of high demand? Has their relationship with Apple improved?

No, it hasn’t improved what are the odds of that it’s like saying Apple is going to use Intel again, which isn’t gonna happen either, as a previous poster said more than likely, it’s for AI training, but that doesn’t do you any good if you ain’t gonna get delivery before the end of the year, the M5 MacBook Pros will be out sooner before Apple gets any significant number of Nvidia GPUs’s, which again highlights, the fact that Apple is gonna have to roll up it’s sleeves and build them themselves in the long run, (Apple you have to build trucks Servers/Mac Pros too). The flighty nice looking iPhone/laptop/Apple watch/iPad hardware can’t be the only thing you build capturing some of the AI mind share in AI hardware Is more important than selling/having 2% of future iPhones sold being folding iPhones.

Hurts even more because Apple is probably just two generations away from the top spot on the Blender benchmark table. The next uplift (generation) of the ultra M series chip will probably put Apple in the top three on that list using less than 138 watts to do it, somewhere around 11,700.

https://opendata.blender.org/benchmarks/query/?compute_type=OPTIX&compute_type=CUDA&compute_type=HIP&compute_type=METAL&compute_type=ONEAPI&group_by=device_name&blender_version=4.3.0

Apple is also probably two generations away from being able to train using Mac ultras, Nvidia is running into wattage trouble. They’re reaching the end of the line, the era of unlimited wattage increases is coming to an end.

https://creativestrategies.com/mac-studio-m3-ultra-ai-workstation-review/

edited March 26

bestkeptsecret

@Kju3, I found your post informative, but why the snark towards "Apple fans"?

kauairooster

Please excuse my lack of knowledge in this area but I’m not exactly sure what the issue is here. Apple uses hardware from other suppliers all the time; their screens, cameras, Wi-Fi/Bluetooth Chips all come from companies outside of Apple. If Apple is going to use Nvidia’s Servers, I can only imagine Apple requesting Nvidia to meet certain standards like any other company would to suit Apple’s needs. Servers are a little bit different ballgame than GPUs in computers and it seems like there would be more at stake from a security standpoint.

elijahg

kju3 said:

blastdoor said:

avon b7 said:

No doubt CUDA is vital here as I haven't heard anything about a complete Apple AI training stack for use with the heavy lifting.

Nvidia has CUDA. Huawei has CANN. 

Has Apple released an equivalent solution? 

Apple has Metal Performance Shaders and MLX. 

I'm not qualified to say whether they are 'equivalent' to CUDA. But I believe they are focused on doing the same general job. 

Metal Performance Shaders? No. MLX? Yes. The issue with CUDA is that it has been around since 2007, no one else had an incentive to put out a competitor for most of that time so the best software toolkits have been built on top of what was the only game in town, plus it is what the most experienced programmers, engineers etc use. So what Apple needs to create competitive AI applications doesn't work with MLX and not enough people know the stuff that does work with MLX. Now if this was 3 years ago before Microsoft triggered this genAI boom and had Apple made this a huge priority it wouldn't have been a problem: Apple could have created whatever they need for MLX/Apple Silicon and trained enough developers to do the work. As neither was the case and Apple finds itself needing to make as much progress as quickly as possible, they need to go with a solution that allows them to just plug it in, hire proven programmers and engineers who have done this job in the past and get going. 

Everyone is just going nuts over this because it is Nvidia, who is on the rather long list of companies that Apple fans are supposed to despise (along with Microsoft, Google, Intel, Samsung, Qualcomm, Masimo, Amazon and I am certain that I am leaving out a lot more) despite Apple's own, er, history of doing stuff. Such as Steve Jobs accusing Nvidia of IP theft and Apple getting upset at Nvidia's refusal to make custom MacBook GPUs under terms that likely would have bankrupted Nvidia. But honestly, it is only 250 servers for less than $1 billion. Lots of companies far smaller than Apple are paying far more to buy way more.

They are just going to be used to cover gaps that Apple can't immediately fill with their own tech: short term stuff. Other companies have already spent far more time and money being among the first to do what Apple needs to get done now. Apple will be able to trace their steps at a fraction of the time and cost while avoiding their mistakes. Once they are finished using the servers and CUDA to play catch-up they'll be done with them and will probably donate them to some university or nonprofit for a tax writeoff, and the engineers that they hire to work on this will make top dollar for a relatively brief gig and will leave with the Apple experience on their resumes that will allow them to work wherever Apple's noncompete clause allows. And yes, this means next time they will actually go with Nvidia when they want to instead of when they have to, which is the way that it should be anyway. As Apple is working with companies that they have literally sued (or threatened to) like Microsoft, Samsung, Google and Amazon then there was never any reason to try to freeze Nvidia out in the first place. That MacBook GPU thing? Well Apple wound up using AMD GPUs that weren't nearly as good, which forced a ton of people who needed the best graphics to buy Windows machines with Nvidia cards instead. So Apple really showed them, didn't they?

That whole Nvidia spat was a total joke. We ended up with crappy, hot, slow, power hungry AMD GPUs in Macs from about 2010 onwards - and Apple was even so childish that they would no longer sign new releases of Nvidia drivers Nvidia was writing for Mac Pros. That did absolutely nothing to harm Nvidia but certainly did piss off Mac owners. Nvidia also supported their cards on Macs for a long long time after Apple stopped updating the drivers it wrote for the AMD cards - drivers that weren't great to begin with.

Plus, Apple had to go to AMD with their metaphorical tail between their legs because a few years before the Nvidia spat, AMD (ATI at the time) accidentally unveiled an unreleased Mac, and pissed off Apple - making Apple switch to Nvidia in the first place.

edited March 27