They're not building on someone else's platform as much as leasing completely commoditized compute from whomever gave them the best deal. I wouldn't be surprised if this compute will actually be happening on xAI hardware [1].

[1] https://finance.yahoo.com/sectors/technology/articles/google...

Apple's AI infrastructure is a hub device like AppleTV, Mac Mini, iMac, or MacBook connected to iPhone and Watch. They will be the first to benefit from cheaper to run, eventually embed in chip models, and are positioned to be a key player in offloading demand from data centers.

They’re the bridge to the frontier models until iPhone 21 ships with then modern frontier model baked into Apple Silicon.

> until iPhone 21 ships with then modern frontier model baked into Apple Silicon.

Citation: It Came To Me In A Dream

Uh huh everyone denied the value of the cloud ...grey beards that are all dead now would never allow their cherished benefactor small business owners private data be stored in a homogeneous data center... heard it all every decade when the state management semantics change with the generations of tech bro butt in seats

I am sure this time the Ruby and Terraform script kiddies will hold back progress just like all those BSD admins staved off Linux like they promised to do on Slashdot 30 years ago

Now that this site has no social moat to boast from the other side of, how out of touch yall are is ridiculously obvious

Glad I went with applied physics, electronic and computer engineering degrees. Will take hands on skill setting up experimental research of propagation of information through various physical materials at various power levels over skill in copy paste of ultimately meaningless syntax.

Instead of feeling insulted, you should just refute me. Why are ASICs a desirable alternative to GPU/NPU inference here?

The "baked-in" weight technology you are describing uses enormous amounts of memory. There is not enough SRAM or flash ROM in the world to put a "modern frontier model" on every iPhone, even if you quantized it to 1 bit. It is scarce and expensive, and we are in a global shortage of memory and flash storage.

If we lived in a world where Cerebras, Taalas and Groq could dethrone Nvidia, they would have done it years ago. Turns out, cutting architectural corners has consequences, and they're extremely offputting for the majority of inference providers.

"If we lived in a world where LLMs could work we'd have done it decades ago instead of in the last few years."

I wasn't offended but secondhand embarrassed for you. You didn't think up LLMs before they happened but you're so sure about the future nonetheless.

I mean frontier like ability with less data.

You're doing the same old IT thing where ya act like your grasp of today is where progress terminates. It's cheesy.

But they could rent from any of the 200+ neo clouds right?

I equate it to partnering with AT&T on the first iPhone. Very inconsistent with Apple’s preference of owning the entire stack, but if your bet is this a short term stop gap until the market is more “unlocked” - less of an issue, and better to get the direct market learnings along the way.

> I equate it to partnering with AT&T on the first iPhone.

In terms of strategy, they picked the underdog carrier (Cingular) specifically to avoid having to bundle bloatware and carrier customization at the cost of exclusivity. Likewise, when Apple brought the iPhone to Japan, they did the same thing with the underdog carrier here, SoftBank.

The strategic importance of this was as a bargaining chip with other carriers and fundamentally gave them a lot of control they would not have had if they had gone with Verizon or similar.

Context for Europeans: the US mobile market was basically total vendor lock-in hell up until the mid-2010s. There were only two carriers (AT&T and T-Mobile[1]) you could bring-your-own-phone with, the other two (Verizon and Sprint[2]) only allowed the phones they sold on their network. Phone service and plans were bundled together.

In fact, one of the critical arguments Apple made as to why they weren't a monopoly was "Sure, we demand 30% - but the phone company monopoly we made obsolete used to charge 70%!" Mobile app stores on other platforms were absolutely hell.

One extra wrinkle of this is that Apple didn't actually partner with AT&T, they partnered with Cingular, the smallest national[0] carrier, who only agreed to let Apple ship an unmolested device because they got 5 year exclusivity. Then Cingular got bought out by AT&T. AFAIK AT&T didn't want the iPhone, they wanted the AT&T (insert terrible name here) Powered by AppleTM, with like 20 AT&T services preinstalled and unremovable. 'Cause that's literally what happened with the first Samsung Galaxy phone, every carrier gave it a different name and their own bloatware.

And yes, this hatred was mutual: the 5 year exclusivity meant iPhones were shackled to a terrible network in the US, and there was a significant amount of effort put into baseband jailbreaking just so people could violate Apple's exclusivity agreement and use their iPhone with their choice of carrier - which, to be clear, meant "use it on T-Mobile".

Also, while I'm giving a lot of credit to Apple, the thing that really broke the vendor lock-in was:

1. Google deliberately bidding up spectrum to trigger an FCC open-access pledge on LTE bands, which forced Verizon to allow third-party networks

2. Qualcomm giving up on UMB, forcing Verizon to adopt LTE, and by extension, SIM cards[3]

3. Clearwire / Sprint's WiMAX service totally failing, also forcing Sprint to adopt LTE and SIMs

[0] The US has a few "regional" carriers like US Cellular that only operate towers in some markets and license spectrum outside of that area.

[1] Yes, they were owned by Deutsche Telekom at one point

[2] At one point owned by Softbank, now owned by T-Mobile

[3] Strictly speaking, the SIM card spec includes CDMA/3GPP2[4] support, but the largest CDMA market was the US, and any carrier that wanted to support foreign phones used GSM/3GPP track standards

[4] Yes, Qualcomm was really that petty that they named their 3G standards body "3G 2"

> Strictly speaking, the SIM card spec includes CDMA/3GPP2

I think the only market that had R-UIM cards (CDMA SIM cards) was Japan, but even with those you couldn't use non-carrier phones since they had all the configuration hard-coded into the devices. It did allow roaming abroad by putting your card into a foreign GSM phone.

edit: apparently it was also in use in China

> I equate it to partnering with AT&T on the first iPhone. Very inconsistent with Apple’s preference of owning the entire stack

They went to Verizon first but were turned down. Apple needed distribution, which meant partnering with an established carrier.

Regarding "owning the entire stack" didn't make any sense in 2007, especially for a company that was new to the cell phone business.

Three years later, the first iPhone with Apple's silicon was the iPhone 4 with the A4 processor.

Apple can pivot Siri pretty easily.

> Apple renting its AI from Alphabet (Gemini with Siri)

this is not how it works.

It’s not a long term risk at all. Long term, models are going to be running on local hardware.

The AI data center is analogous to the mainframe.

If I can run Qwen 3.8 Flash Next on my DDR4 closet server with a $500 16GB GPU at 50 tok/s (which I can) then I just about guarantee you that in 2-3 years a pretty acceptable AI model is going to be running off of iPhones and M7/M8 Macs.

The other option that reduces Apple’s long-term risk is that they could just buy an AI company.

Yeah I can kinda see why Altman is out there trying to shame the public for having compute. The longer they carry those massive debt obligations, the more the models are going to economize.

It would not be surprising that Apple just waits and snatches their compute up from insolvent companies sliding past.

Models running locally doesn’t benefit Apple unless they can control the model. Otherwise they’ll end up being dependent on Google, Meta etc the same way they are today for apps and services. And getting out of that hole won’t be as simple as building a maps app.

I think it does if their chips are the only ones that run the models well on phones and competitors need to charge customers more money or give away less functionality for free.

Or open models become competitive? They have an entire OS based on OpenStep and FreeBSD after all.

There is exactly zero chance Apple lets open, constrained models run on iPhones. We are talking about a company that doesn’t allow apps that contain bad words. People holding out hope for this are delusional.

How much would it cost to license one? Probably less than renting servers.

The on-device models that Apple distilled from Google's server based models are already available to developers to use in their apps at no cost.

We're in a time where devices are becoming more capable and developers are figuring out how to do on-device inference using less resources.

The companies that spent a fortune on AI data centers might as well have built a huge pyre of money for all the "moat" they built with it.

That’s not what I’m saying. They could take an open model, create their own variation of it, then embed that on devices. Just like they have with their OS.

In return for Google getting a discount on how much they pay Apple to be the default search engine in Safari, Apple has the right to distill Google's Gemini server models and to create their own models that fit on-device.

> Apple is cutting down Google Gemini's massive models into smaller and more secure parts through distillation, to create elements that are more suited to on-device Apple Intelligence processing.

https://appleinsider.com/articles/26/03/25/apple-is-distilli...

Not to mention the fact that Apple, unlike most of the industry at times, has never seen itself as a software company. Someone needs to build all the devices used to interact with all this AI, and there's no evidence any major "AI company" other than Google, and to a much smaller extent Meta, has any interest in doing so.

If I were Apple and wanted to make a big bet, I'd be dumping money into a top-secret robotics lab.

> but Apple renting its AI from Alphabet (Gemini with Siri) might be a long term risk, since they're building on someone else's platform.

Apple isn't renting; they distilled Gemini to create their own model for "broad world knowledge". Some of this runs on Google's infrastructure because Apple didn't have enough compute.

All of the personal context stuff is Apple's Foundation Models.

Apple is always not-going to have enough compute of their own, they have billions of users and whatever they are investing in datacenters to service them is negligible.

That said renting storage from Google for iCloud reportedly carries an ~80% profit margin, there's no reason they can't also put a 5x markup on what Google charges them for PCC and Siri since they can still prevent rivals from providing the same service in most markets.

> ... Apple renting its AI from Alphabet (Gemini with Siri) might be a long term risk...

Similar to Apple (and everybody else) using Google maps. They survived.

I think this will depend on the commoditization, or lack thereof, of the AI models. If they are more or less interchangeable, Apple will have made the right bet.

Apple doesn't pay Google.

Google gets a small discount how much they pay Apple to be the default search engine on Safari.

I assume Apple is buying time.

What has apple really lost not investing billions in AI? IMO nothing. At any point they could just buy or partner with any missing piece. Google is the safe bet for Siri but this could be any number of providers in the future at which one of them might even pay apple like google does for search in safari. Apple needs to tackle component costs to increase revenue, especially memory which they have to start thinking about their own fab long term.

Conversely, Google is “renting” from Apple to be the default search engine in Safari for ~$12B annually.

I’m sure there’s some balance of trade going on here.