Report: Apple Asks Google to Run Siri on Its Servers

[Big_D]

First thought was this doesn't sound good for privacy.

It depends on how it is done, if the Gemini servers are set upto the privacy standards required by Apple, there should be no problems. If Siri is just running on a normal Gemini instance, where the data is passed into the learning dataset, that is a huge problem and a bad look for Apple.

 

[svish]

This approach does not look good. Hopefully there will be some form of privacy. Instead of this, a custom version of Gemini could run on Apple' servers.

 

[Piggie]

As apple has always gathered data on it's own users for it's own use/reasons. (1st party monitoring)
Do we presume that Google is collecting the data Apple wants about users and than passing this back to Apple?

 

[DrRadon]

Because some people (myself included) prefer iOS to Android.

Wait, you did not sign up for an apple branded forum to constantly complain about apple? Weirdo.

 

[Spike1999]

Oh boy this is going to be fun...

 

[Puiz]

"Its." Next time you write such a headline, maybe specify whose you mean? 🤔😎

 

[manhattanboy]

Does Anthropic own data centers, or are they leasing space in the same way that Apple would be doing?

Apple owning Anthropic doesn't solve the issue of Apple not being able to run its own models in its own hardware.

But if said hardware is sitting in warehouses, why? They are not specced for the job?

I now think Apple will take back its own crown in two years, not one.

If they have unused server hardware why not release it to consumers? I would buy something that is more powerful than their current m3 ultra. The proceeds can then be used by Apple to buy nvidia hardware or fund development of in house server chips. Letting hardware rot unused and unopened reeks of mismanagement.

 

[Trioptre]

It looks like a lot of indecision. Almost more than my ex wife, which I thought was impossible

Your ex wife sounds like a real piece of work. Congrats on leaving her - hopefully you’ll get a better wife next time.

Anything else you’d like to add about your ex wife?

Sounds like you need to move on and think before you post petty digs at her on MacRumors.

Please keep your thoughts your ex wife to yourself.

 

[Trioptre]

I asked Gemini, "does gemini get the "Heavy World Knowledge" requests via PCC and is it protected or made anonymous at a level that makes it impossible for google to ID a person? keep in mind, there are indirect ways to figure it out such as location data and knowing where an anonymous source sleeps" 😄

Answer, typically verbose...

This is the "billion-dollar question" of the 2026 AI era. You are right to be skeptical—anonymity is often an illusion if you have enough data points like "where a phone sleeps."

However, the architecture of the Apple-Gemini partnership is designed specifically to solve the "Sleeping Location" problem. Here is how it works technically:

1. The "Privacy Buffer" (The Identity Eraser)

When you ask a "Heavy World Knowledge" question (e.g., "What's the best hiking trail near me right now?"), the request doesn't go straight to Google. It goes through an Apple-controlled Privacy Buffer.

• IP Masking: Apple strips your IP address and replaces it with a generic one from their own relay. Google sees a request coming from "Apple California," not your house.
• Tokenization: Your Apple ID is removed. To Google, the request is just a random string of numbers.
• Fuzzing Location: If location is needed, Apple doesn't send your GPS coordinates. It sends a "coarse" location (e.g., "North Seattle") that is just enough for the AI to answer the question but not enough to pinpoint a specific home.

2. Is it truly impossible to ID you?

The goal is to make it computationally and legally impossible. While a "rogue" engineer could theoretically try to piece things together, several layers of defense prevent the "sleeping location" tracking you mentioned:

• Stateless Processing: In the Enterprise-grade Gemini instances Apple uses, the session is "stateless." The moment the answer is sent, the server's memory is wiped. There is no "history" for Google to look back on and say, "This same anonymous user was at Point A last night and Point B this morning."
• No Cross-App Linking: Unlike using the Gemini app directly, where Google can see your Maps history, YouTube searches, and Gmail, the Apple version is isolated. Google has no way to link an anonymous Siri query to your actual Google account.
• Independent Audits: Apple has made the source code for the Private Cloud Compute (PCC) nodes available to security researchers. They can verify that the data really is being deleted and that no "backdoor" identifiers are being leaked to Google.

3. The "Google Cloud" Nuance

You mentioned Google Cloud earlier. The crucial distinction is who owns the "keys" to the data. Even when Apple uses Google Cloud servers for massive scale, they use Confidential Computing (specifically Google’s "Titanium Intelligence Enclave"). This encrypts the data while it is being processed.

Think of it like this: Google is the landlord providing the room (the server), but Apple brings their own safe (the encryption). Even Google, the landlord, can't see what's happening inside that safe.

Ask for a less verbose answer next time.

 

[clg82]

Siri has never been the reason anyone does or does not purchase an iPhone.

Thank goodness for that, cause they would never sell any.

 

[clg82]

Not sure how many more ways Apple can get caught with its pants down when it comes to Siri. It’s kind of morbidly fitting that Steve Jobs passed away the day after Apple announced the first Siri-enabled iPhone in 2011.

Caught with it's pants down with Siri, how about AI as a whole!

 

[SamSwede]

It's probably the cheapest option.

Google's Tensor Processing Units (TPUs) provide a significant cost advantage over Nvidia GPUs, with estimates suggesting they cost between a half and a tenth as much to operate. TPU v6e offers up to 4x better performance per dollar for AI workloads, allowing Google to potentially reduce AI query costs to only twice that of traditional search.

https://www.economist.com/business/2025/10/30/google-v-microsoft-the-battle-of-ai-business-models

 

[mjschabow]

I also use the Gemini app currently and I would prefer it baked in.

Now we have both our opinions out there.

My thoughts too. I’m excited for Gemini to be baked into the iOS experience.

 

[Macalway]

Your ex wife sounds like a real piece of work. Congrats on leaving her - hopefully you’ll get a better wife next time.

Anything else you’d like to add about your ex wife?

Sounds like you need to move on and think before you post petty digs at her on MacRumors.

Please keep your thoughts your ex wife to yourself.

Ah yes, a sign of the times! Exactly like something my ex would say. Uncanny! (kidding, lighten up)

 

[jole]

Only because Apple arbitrarily restricts some things such and imessage and FaceTime while also choosing not to implement things like full RCS 3.0 which brings encryption and in-line replies for texts

People only stick with ios due to ecosystem lock in. Majority of things like notifications for instance are just better on Android

I am deeply in Apple ecosystem — as are many people around me. Even with that background, I believe FaceTime, SharePlay, ... are dead in the water. Anything communication related _must_ be interoperable to be competitive. Apple somehow keeps pretending that they do not understand this and are losing every time on products that could have taken over the world.

 

[nt5672]

Why would it matter to you whose servers hold your encrypted data? That is the point of data encryption. This not some old box in a corner running Windows and with its data open to the world. Modern cloud servers run everything with encryption to precent access to the data.

Do you hold the encryption keys? No, then someone has them or has access to them that is not you and can decrypt your data.

If processing has to be done on the server, then the server has to have the ability to decrypt my data.

If the server can decrypt my data for Siri processing, then the server can decrypt my data for malicious use.

I know this goes against Apple advertising, but this is reality, not Apple pie-in-the-sky cult messaging.

Simple computer 101.

 

[transpo1]

It's kind of like this: Apple might be buying servers from Dell, or HP. Does this mean that you have to trust Dell or HP with your data? The analogy is not perfect but it communicates my point. Apple has not had a massive data center presence for many years and used various large companies to do stuff for them. And yet it was OK and their privacy commitments hold...

Now, if they change their commitments to privacy... that would be a different story (but I have not seen this).

With Dell and HP, however, we and our data are not the product or the focus of their business model; so I would trust those companies much more than Google. And also, we assume that any privacy solution Apple has for their custom Siri/Gemini is foolproof in terms of data privacy; as a longtime Apple customer, I need assurances. Apple needs to make privacy a focus when they launch the new Siri.

In other words— “Trust but verify.”

 

[Silverado]

Does anyone know English anymore? The following two sentences are ambiguous. It’s not clear which company “its” is referring to!!

Report: Apple Asks Google to Run Siri on Its Servers

Apple has asked Google to investigate setting up servers in its data centers to run a future version of Siri powered by Gemini

It could mean either Apple or Google. I can’t believe the quality of editing used here!

 

[deconstruct60]

I did a Google search of data related controversies associated with Google. I recall some of the news on this list, but... Google has faced numerous controversies regarding data privacy, including allegations of collecting user data without consent, even when privacy settings are enabled. Notable incidents include the tracking of users in "Incognito Mode" and the continued collection of location data despite users opting out of such features, raising significant concerns about user trust and transparency. This is copy and paste of what the search told me.

Incognito mode is inside of Chrome. How much do users pay to use Chrome? Nothing. Google search how much do people pay to use it? Nothing.

Google Cloud services ( except for some educational, 'get to know the system' instances ) are not free. Google is now making Billlions directly charging some customers money upfront to use services. There is zero need to sniff the data to make money. Or do pop up ads in the middle of those paying customer's workload.

"One other thing is that google has such a massive infrastructure all around the world..."
- Exactly. Google is so massive. Who is watching the watchers? If a company wants plausible deniability, being a massive infrastructure with 1000's of people watching or conveniently not watching... means Google can go near whatever they elect to go near.

Folks who don't pay , typically don't watch. Folk who do pay probably will watch. Especially, if they are regulated to protect privacy ( e.g., HIPPA compilance, government data disclosure laws, etc. ). Those are 'features' that are going to be minimal requirements in order to get their cloud hosting business.

That whole 1,000's of folks 'watching' open source is a false sense of security. There are exploited open source systems/libraries also. It isn't large numbers that matters most. It is diligence (which usually costs money to be a 24/7/365 active process. )

 

[deconstruct60]

Modern cloud servers run everything with encryption to precent access to the data.

Do you hold the encryption keys? No, then someone has them or has access to them that is not you and can decrypt your data.

This is flawed. Modern cloud servers have the functional equivalent of a secure enclave. Do you physically have the security key to your iPhone/Mac/Apple device? No. It is stored inside of the secure enclave. Do you have control over how the secure enclave uses the key? Yes. That is what actually matters.

"... Using Cloud KMS with the Cloud HSM backend, you can own keys that are generated and stored in Google-owned and operated hardware security modules (HSMs). If you require a hardware-protected key, you can use the Cloud HSM backend to help ensure that your keys are only used in FIPS 140-2 Level 3–validated HSMs. ..."

Cloud Key Management Service encryption | Security | Google Cloud Documentation

Whitepaper that describes how Cloud KMS architecture provides Google key management of HSM, CMEK, BYOK, and Autokeys for data-at-rest encryption.

docs.cloud.google.com

" ... Confidential VMs with H100 GPUs help ensure data remains protected throughout the entire processing pipeline, from the moment it enters the GPU to the moment the results are generated. ..."

https://cloud.google.com/security/products/confidential-computing

[ works with Google TPU VMs also. ]


If Apple's keys stored in the cloud server are controlled by Apple then Google does not have access to the data. there is a process of moving/creating/managing keys that needs to mastered, but the physical operator of the computer doesn't need to know what the keys are to do what they are being paid for.

In Apple's general system Apple doesn't particularly have the keys to users data.

There are cheaper cloud instances/services that do not have security and the host operators can 'peek and poke' at virtual machines from the host supervisor, but Apple doesn't 'have to' use those with a proper budget.

Could there be security holes in a AMD/Intel processor security cobbled with a Google hardware key storage unit and some physical hack by some operator that Apple Silicon doesn't have. Probably. However, that isn't particularly Google trying to look at the data.

If processing has to be done on the server, then the server has to have the ability to decrypt my data.

If the server can decrypt my data for Siri processing, then the server can decrypt my data for malicious use.

Not necesarrily. Apple PCC servers has keys to unlock data and it really isn't an issue because of the key managment process.

Private Cloud Compute: A new frontier for AI privacy in the cloud - Apple Security Research

Secure and private AI processing in the cloud poses a formidable new challenge. To support advanced features of Apple Intelligence with larger foundation models, we created Private Cloud Compute (PCC), a groundbreaking cloud intelligence system designed specifically for private AI processing...

security.apple.com

Google doesn't have exactly the same names for the mechanisms, but something functionally similar could be composed with a customized VM that Apple composes to run on Google's hardware.

Simple computer 101.

Simple computer 101 is actually looking at what you have to work with first.

 

[deconstruct60]

Yeah, but that doesn't say anything about the Nuvia technology.

It isn't about the technology. It is about what the financial opportunity is to pay for developing the technology.

It speaks to what Qualcomm wanted. Qualcomm wanted to turn their work into an alternative to Apple silicon in consumer products,

So both Apple and Qualcomm did not see the huge payday in server, but it really does exist. All the other folks who could have also bid on Nuvia , Microsoft , Nviida, Amazon ( folks looking to get into the Arm server business for in-house use) they all missed it too.

and is why ARM filled a lawsuit against them.

ARM sued because they gave Nuvia a 'sweetheart' deal on an Architecture license because they though that Nuvia was going to eventually pay full price once more expensive server chips actually started to get made. Plus Nuvia didn't have lot of uncore substance so likely were going to need other IP from the Arm ecosystem for I/O connectivity. ( not likely bringing a ECC memory subsystem ideas over from Apple. Very large core count bus ... nope . large PCI-e provisioning ? nope. NPU nope, GPU nope. etc. ) A very small company with a fixed sized budget isn't going to be able to do everything for everybody in the first couple of generations.

Nuvia was always structured to be bought. (now that the Nuvia founders bank accounts rose by more than 7 digits in the $1B Nuvia buyout suddenly working on phone/PC processors isn't worthwhile work. *cough* ) I think Arm just expected it to be someone from the server space since they gave Nuvia a 'server space' license. Intel fumbling in the server space meant that window wasn't going to stay open forever. (AMD would grab share , but at some point the inflection point would close).

Qualcomm represented a more rapid loss of income for Arm (a faster drop in buying design IP). Bad timing when trying to sell off Arm to the highest bidder. Archictecture license are more money up front but loss less longer term.

 

[Tagbert]

Do you hold the encryption keys? No, then someone has them or has access to them that is not you and can decrypt your data.

If processing has to be done on the server, then the server has to have the ability to decrypt my data.

If the server can decrypt my data for Siri processing, then the server can decrypt my data for malicious use.

I know this goes against Apple advertising, but this is reality, not Apple pie-in-the-sky cult messaging.

Simple computer 101.

The encryption is done using your password.

 

[deconstruct60]

The encryption is done using your password.

Typically no for these kinds of systems. Your password only grants you access to the key that is stored in the hardware key store unit. It doesn't encrypted anything (extremely likely too short in length to be effective). It is more so for authentication that you are the owner.

 

[Tagbert]

As apple has always gathered data on it's own users for it's own use/reasons. (1st party monitoring)
Do we presume that Google is collecting the data Apple wants about users and than passing this back to Apple?

It would not be possible for Google to collect user data or to pass it to Apple if they use Google’s Private AI Compute. This is essentially Google’s version of Apple Private Cloud Compute system. Both are special servers that receive AI queries, process the request in RAM and return the response, all with encrypted data and without the ability to log any data or to do any training.

Google says new cloud-based “Private AI Compute” is just as secure as local processing

New system allows devices to connect directly to secure space in Google's AI servers.

arstechnica.com

 

[johnsawyercjs]

It's a little hard for me to understand why Apple didn't realize (or just admit?) earlier that their PCC servers using M2 Ultra processors wouldn't be able to handle the workload for their customers once Siri 2.0/Apple Intelligence was fully up and running.

The scaling requirements for the amount of off-device quality LLM inference Apple wants to provide weren’t totally anticipated even by researchers in the field in late 2022 when ChatGPT was released to the public, and maybe even by mid to late 2023 either, but by mid-2024 enough real-world user data had been collected though use of ChatGPT, Bard, etc. that it was becoming fairly apparent that the level of scaling needed to provide good responses for lots of users was far larger than anticipated. Among other things, it was found that training with a lot more data was needed to create larger, better models, but this led to increased requirements to provide quality inference compute after the training was done, and so some better idea of the hardware needed should have been more apparent. But Apple began committing to using its M2 Ultra-based servers in mid-2024 anyway. Sure, Apple wanted its own hardware to do the job, but surely Apple should have known how to anticipate these things better.

Was Apple really thinking clearly when it assumed, even as late as mid-2024, that its iPhones and Macs would be powerful enough to do most of the work on-device, with cloud overflow being modest enough that data centers running its M2 Ultra servers could do the rest of the job?