Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Apple's Siri Chatbot May Run on Google Servers

Nyx1 said:
why build your own server in Houston if you don't use it
Click to expand...

It is likely a matter of where does the 'overflow' go. The downside of using. Google ( or Anthronpic Claude or OpenAI ) is that they are in a 'arms race' to be the largest , multimode modal models possible. Everything and the kitchen sink thrown in. If Apple is going to occasional dispatch a subset of user requests to something like that then the Apple servers probably won't have the 'chops' to handle that.

Apple's servers are not all that heavy weight compute.


SSEHQHMXD5KDJJ3FA5LSFSQYQI.jpg

www.reuters.com

Apple begins shipping AI servers from Houston factory

Apple has started shipping artificial intelligence servers built in a factory in Houston, it said on Thursday, part of the company's plans to invest $600 billion in the U.S. in the next few years.
www.reuters.com www.reuters.com


Also.

108216265-1761252279956-AppleFactory_Houston6691.jpg

https://www.cnbc.com/2025/10/23/apple-american-made-ai-servers-texas.html


Not much of a huge networking connection there (doesn't look like doing many-to-many direct connections in the rack. More so, it looks like data streams in from Apple device and and goes right back out. ) Also not much of Air throughput (if that second picture is the 'front' then no obvious liquid chilling connections. )

There is a pretty good chance Apple built something that has better performance/watt then the competition. But 'chatbot' AI has taken on more of a power consumption pissing contest where the objective is to burn as much power as possible. I don't think Apple has that or even wants to do that at all. A power consumption pissing contest is very likely going to be outsourced.


It is probably also a huge mistake to offer a super long list of very fancy features of Apple Intelligence entirely for free. I don't think Apple can wait forever before introducing a paid tier. If only the paid stuff is dispatched to Google servers then. PCC could be enough for all of the folks getting 'free' queries.



P.S. from the looks there is a pretty good chance that Apple's PCC servers are just repackaged and refactored Mac Studios ( stripped of ports and video output stuff that and end user would use and flatten out heat sink that covers a bigger 2D footprint than Apple would tolerate on a desktop. The 'slab' is big enough for two units ( i.e. 2+ Ultras and some auxiliary stuff. ). Revise the Mac Studio motherboard to make it easier to fit in the slab and stripping the parts they don't need (e.g. Wi-FI , USB ports , etc. )

If liquid cooling , then probably max 4 Ultra in a Apple's 'cluster' fashion. But that only scales in cluster compute to 4 nodes. Google TPU nodes are capped that narrow.
 
Last edited:
  • Haha
Reactions: BigDO
Even if Craig himself said to the camera: “Hey, look, we’ve actually rented those servers to Google, so that makes them technically our servers… under our privacy standards”, the impression most people are going to get is that the new Siri is ran by Google… I’m not even sure I’d trust it, honestly. I’d prefer them to focus on “On-device” AI even if they do it with third party models.
 
  • Like
Reactions: HazeAndHahmahneez
This is definitely a mistake if followed through with. I would rather they just tell us they aren't ready and just allow existing AI makers to replace Siri than give me some half baked compromise that doesn't do what the others can but provides "Kinda Private" protection.

Apple is losing its unique features slowly but surely. Eventually its going to be a battle of whos hardware is better and cheaper. If it comes to that I believe Apple will lose. Apples greatest advantage has always been its operating system and eco system. Lose those by "Being Just like everyone else" and I might as well go buy a Samsung and use the extra money on a Gemini subscription.
 
  • Like
  • Angry
Reactions: HazeAndHahmahneez, turbineseaplane and Brother Cavil
Lol, at this rate by the time it happens it’ll be ‘Apple gives all user data to Google in the clear and lets them use it to target advertising, but with a very solemn pinky promise that it’ll all be ok and everyone will be nice about it, so nothing to worry about…’

I’m so glad I don’t care about AI or Siri.
 
I feel like people do not really understand how hosting of another company data works for the service provider. Seeing that I work at one of the "hyperscaler" companies that literally everyone heard of, and we host data for bajillion companies, I can confidently tell you that it is absolutely possible to have data in one's datacenter where the company hosting it does not have access at all, even if they tried.

Your iCloud backups today could already be hosted in Google datacenter.

This is not a problem that is unsolved; Apple could run literally everything they have in a Google datacenter with a dedicated model, private encryption keys etc. and it would make zero difference to anyone here.

The question is "would Google have access to customer data on application layer" and I am going to guess the answer to that will be "No" just like today your iCloud backups are not something Google can read.
 
  • Like
Reactions: montuori, HazeAndHahmahneez, neuropsychguy and 7 others
Be interesting to see the structure of what is done here. The security (or lack there of) will all depend on how the relationship is structured, and what services Apple is using.

I think many are overreacting regarding privacy not understanding how managed services or dedicated cloud computing works. Meta, AWS, and many others do direct, singular farms for dedicated clients without any crosstalk into other systems, which includes specific isolation and other protections for their data.

Apple already uses non apple equipment and 3rd party services to support many of their online offerings. They also use colocation datacenters across the country from places like Databank and others, often using additional 3rd party network and system infrastructure to support their services.

The days of apple doing everything in-house, on their own servers built by just them has long since been dead.
 
  • Like
  • Love
  • Sad
Reactions: HazeAndHahmahneez, DefNotAnLLM, surfbutler and 2 others
MacRumors said:
In yesterday's report detailing Apple's plans to turn Siri into a chatbot in iOS 27, Gurman said that the company is in discussions with Google about hosting the forthcoming Siri chatbot on Google-owned servers powered by Tensor Processing Units (TPUs), a class of custom chips designed specifically for large-scale artificial intelligence workloads. The arrangement would mark a major departure from Apple's emphasis on processing user requests either directly on-device or through its own tightly controlled Private Cloud Compute infrastructure.​
Click to expand...

Another potential stumbling block is that Apple's. "AI server" specific chip that they were doing with Broadcomm stumbled somehow. Slide backwards in deployment timeline or Apple decided it costs too much to be in the arms race ( need resources for other chips for products they are actually going to sell directly to end users. Unless Apple AI makes lots of money... that is just a huge cost sink have to bundle into the price of products (generate subsidy program). Might need that money to pay for more expensive RAM/NAND. )

On Google's server doesn't mean Apple has to give up control of the encryption encoding/decoding. Managing the systems/cluster will be different.

Another issue is that Apple may have scaled PCC to a different workload. If running Siri-AI for 40 million users probably isn't the same deployment than 400 million users. If getting Apple server chips from TSMC ... can't really go and ask for more chips at this point. They have over a year backlog. Anyone who didn't reserve capacity a year or two ago doesn't have anymore. And not going to scarifice Mac or iPhone wafers for Server silicon. The Mac/iPhone customers actually pay money. This AI stuff isn't paying anything directly. No way they are at the 'front of queue' for additional wafers Apple has access to.
 
ArtOfWarfare said:
Does anyone complaining about this actually routinely use any of the cloud providers?
Click to expand...

Google's primary business model relies upon collecting and aggregating user data to be leveraged by advertisers.

Why would anyone assume that this is somehow would be different?

It doesn't matter if Google has the keys to the encrypted data, it's the optics of Google being an evil company and lack of any respect for users or their data.
 
  • Like
  • Disagree
Reactions: fatTribble, HazeAndHahmahneez, jwdsail and 1 other person
I guess it's time to research Ollama. To get past the gimmicky stuff, I'm looking forward to contextual AI. I was holding out to see where Apple goes with this, but with this announcement, it's not looking good. I think I'd rather invest in my own system where there is no privacy concern.
 
  • Love
  • Like
Reactions: HazeAndHahmahneez and Ursadorable
I haven’t seen this spoken about much but I’ve been curious about how Apple might want to handle Siri on non Apple Intelligence devices. Siri didn’t run on PCC prior and I’m wondering if this might bring a new Siri to all devices regardless of hardware.

Apple Intelligence devices could lean more on local processing and PCC as needed but at least then Siri becomes more unified. Apple Intelligence already runs a model locally to determine how to route requests. Older devices could always hit the cloud but then what happens with requests that can already be handled offline
 
MacUserFella said:
Why would the company that cares about privacy run their chatbot on servers owned by the company that couldn’t care less for privacy
Click to expand...

Asinine reply.

Apple has been storing your iCloud data on Google servers (and others) for ages now. Google has ZERO ability to tell what the content it’s storing is or who it belongs to. Your privacy is 100% intact.

Why would this be any different?
 
  • Like
  • Sad
Reactions: fatTribble, DefNotAnLLM and revs
mikebenton said:
I guess it's time to research Ollama. To get past the gimmicky stuff, I'm looking forward to contextual AI. I was holding out to see where Apple goes with this, but with this announcement, it's not looking good. I think I'd rather invest in my own system where there is no privacy concern.
Click to expand...

I use LM Studio on the Mac (I've also used Ollama before switching to LM Studio). It works great for local LLMs. It uses Ollama as it's core, but provides a friendly user interface, and has access to many MLX optimized models from Hugging Face for easy download right from LMS.
 
  • Like
  • Haha
Reactions: HazeAndHahmahneez and mikebenton
This is the company that sells its ability to keep your data private, handing over your most valuable date to a company that survives by selling your data and everything about you to anyone who will pay. And don't tell me about firewalls, there is no such thing. This has to be some horrible twisted joke. Please tell me it is. "This is the way the world ends, not with a bang but a whimper" (with acknowledgments to T. S. Eliot and his poem The Hollow Men).
 
Last edited:
MallardDuck said:
Now that apple admits that their chips are inferior for AI/ML/GPU workloads, maybe we'll get eGPU support back in MacOS?

And maybe I'll sprout wings.

On the privacy side, I'm sure Apple is *keenly* aware of the perception that this will cause, and will engineer things so that it's safe.
Click to expand...

For textual LLMs (gemma, mistral, llama, etc.), Apple Silicon is very fast. My M3 Pro performs just as fast as my RTX 4070 Ti Super, while consuming less than 10% of the power of the 4070. Sure it's not data center spec, but for personal use, it's frankly amazing considering the substantially less power consumption. And don't discount what unified memory brings to the table in terms of model size you can run. A Mac Studio with 96GB of RAM can comfortably run 80GB models, something your 32GB RTS 5090 nVidia card can only dream of.

Image generation on the other hand, is quite a bit slower. Pytorch isn't very well optimized for Apple Silicon, nowhere near what the performance of Cuda is.
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back