Report: Apple Plans to Make On-Device AI a Key WWDC Focus

[Ishimura]

They’re not. Subscriptions barely cover the cost of the first few tokens. The vast majority of AI users use the free plans.

Every AI company runs LLMs at a huge loss. OpenAI will end up being bought out by Microsoft. Grok is having to rent out its data centres because not enough people are using them and burning so much capital then its bosses had to merge with SpaceX to siphon funds. Anthropic recently turned a profit it’s true but even that looks like they cooked the books.

Nobody is currently making any money from AI.

Ed Zitron is a clown who is widely regarded as a joke within the AI and media industries.

Did you actually read the piece, by the way? Here’s a direct quote from it:

‘While I wouldn’t say this was cooking the books..’

Yet, strangely, you’ve written that Anthropic’s apparent single quarter of profitability is them having ‘cooked the books’.

Re data centres you said that nobody is paying for them and my response was to point out that that is a laughable statement when these companies have many paying customers.

Anthropic’s revenues are exploding, rocketing up even beyond what they had expected. Why? Because they have many paying customers who are using their product a lot. Of course everybody knows that the companies haven’t yet turned a profit but that doesn’t mean it makes any sense to say that nobody is paying for the data centres.

Every paying Anthropic customer (almost entirely enterprise) is paying for them. Every OpenAI customer is paying for them. These are two of the fastest growing companies in history.

 

[Ishimura]

The average life of the GPUs in these data centres is also 3 years

No it isn’t. GPUs olds than that are in heavy use, maxed out in fact.

Everybody is desperate for GPUs, including old ones. Various examples but the A100 came out in 2020 and is still in heavy use.

 

[microChasm]

Private yes, but if you’re mature enough to understand scale then you’ll understand datacenters are more efficient and cheaper than millions of devices trying to do the same thing. They can invest in renewable energy, or buy energy at fixed rates for many months and they can bulk purchase hardware components at discount every three years.

Who cares about scale other than AI companies. It seems like your comment comes from that perspective.

 

[microChasm]

My concern is RAM and Battery.

What AI can do 'on device' will be limited unless there is a huge swath of memory available. How useful this will be without access to cloud services is questionable to me. I suspect this feature will be using the cloud more than Apple is acknowledging at the moment.

How often are you actually AImaxxing on an iPhone or iPad? Not very much. Now agentic AI is a different story. That is where AI companies need to get their act together. That is the consumer sell for them that they don’t seem to care about

Apple likely sees an opportunity to utilize Apple Silicon on devices and computers to share Private Compute for agentic AI use while keeping data private secure while keeping personas local and controlled by the human and not a company.

 

[Ctrlos]

Ed Zitron is a clown who is widely regarded as a joke within the AI and media industries.

Did you actually read the piece, by the way? Here’s a direct quote from it:

‘While I wouldn’t say this was cooking the books..’

Yet, strangely, you’ve written that Anthropic’s apparent single quarter of profitability is them having ‘cooked the books’.

Re data centres you said that nobody is paying for them and my response was to point out that that is a laughable statement when these companies have many paying customers.

Anthropic’s revenues are exploding, rocketing up even beyond what they had expected. Why? Because they have many paying customers who are using their product a lot. Of course everybody knows that the companies haven’t yet turned a profit but that doesn’t mean it makes any sense to say that nobody is paying for the data centres.

Every paying Anthropic customer (almost entirely enterprise) is paying for them. Every OpenAI customer is paying for them. These are two of the fastest growing companies in history.

They’re only exploding because of the circular financing model. Using VC funding to buy loads of GPUs so that Nvidia buy shares so that they can buy more GPUs is a dangerous house of cards pump and dump. They need investment because they’re burning money hand over fist. Their subscription income is but a drop in the ocean for staff wages. They’d have to charge customers £2000 a month if they wanted to remotely break even.

These companies are building out unwanted capacity. The majority of people use it for silly things like generating clip art. It says a lot when if these products disappeared from the planet tomorrow the world would largely keep spinning. You could not have said the same about the iPhone 4.

 

[Ctrlos]

No it isn’t. GPUs olds than that are in heavy use, maxed out in fact.

Everybody is desperate for GPUs, including old ones. Various examples but the A100 came out in 2020 and is still in heavy use.

I mean they recycle them further down the food chain from training to inference but newer models demand higher compute. It’s like Doc Brown adding more horses to the DeLorean in Back to the Future 3. Doing so doesn’t wear out the horses but it’s never going to be as good as the cutting edge steam train they need.

 

[Ctrlos]

How often are you actually AImaxxing on an iPhone or iPad? Not very much. Now agentic AI is a different story. That is where AI companies need to get their act together. That is the consumer sell for them that they don’t seem to care about

Apple likely sees an opportunity to utilize Apple Silicon on devices and computers to share Private Compute for agentic AI use while keeping data private secure while keeping personas local and controlled by the human and not a company.

I think these companies are still betting the farm on a product nobody asked for. Humans are by their very nature control freaks when it comes to their own lives. They like to have as much say as possible over the things they do. Automating parts of your private life to an LLM which has inaccuracy baked into its very nature is a fools errand because the moment it makes a single mistake on something simple like setting an appointment people will never use it it again and those stories spread like wildfire.

An LLM doesn’t work on absolutes like facts, but rather tries to visualise and predict what it ‘thinks’ is the right answer. Doing so with something simple like a sentence gives them the illusion of intelligence but it’s all still a prediction. This is why they require near constant human checking for hallucinations.

In automation if you cannot guarantee that it’s going to be perfect 99.9% of the time (there is no such thing as an error free machine anywhere on the planet) then you’ve failed. The LLM is just too inaccurate a product to be left to its own devices.

 

[name99]

They’re only exploding because of the circular financing model. Using VC funding to buy loads of GPUs so that Nvidia buy shares so that they can buy more GPUs is a dangerous house of cards pump and dump. They need investment because they’re burning money hand over fist. Their subscription income is but a drop in the ocean for staff wages. They’d have to charge customers £2000 a month if they wanted to remotely break even.

These companies are building out unwanted capacity. The majority of people use it for silly things like generating clip art. It says a lot when if these products disappeared from the planet tomorrow the world would largely keep spinning. You could not have said the same about the iPhone 4.

The entire economy is "circular financing" as is the biological ecosystem... You say circular financing, I say innovative financing.
It's only a problem for *America* if what is being produced along the way is not useful. That seems unlikely.

If you are an investors, the concerns you give may be important. But as simply a "consumer" of the economy, nah. We've had similar rollouts (the obvious recent example is fiber in the late 90s, but there are earlier examples going all the way back to electricity and oil, and then the mother of them all, railroads, which as a fraction of US economic spend in the late 1800s dwarf anything going on today).
Even the most disastrous of these (consumer airlines, in the sense that the companies are always in crisis and never actually make money for the investors) still generates value for the rest of us, just not for the investors.

And no, "The majority of people [will not] use it for silly things like generating clip art". Every day new people are figuring out something truly useful that they can do with LLMs, and telling their friends. It takes time for people to change, even more so for businesses. Even something as obvious as google took time; there were plenty of people in 2005 thinking that the work you looked something up was to go to the library.

Most things in life are a matching problem. I want a job, you want an employee. I want a house, you are selling one. I want an education, you offer courses. I want a partner, someone else wants a partner. I want a couch, you want to sell me a couch. If you can sit between the two sides of a matching problem there is both tremendous value to be delivered (how much are the best house, the best job, the best partner for YOU worth to you?) and tremendous profit to be made. The somewhat limited, somewhat sub-optimal matching provided by google classic was a money machine the likes of which the world had never seen before. LLMs (once we get past the initial teething stages) will be Google ramped up to 11 as far as being a matching machine. And that's just one element of their value.

 

[name99]

I think these companies are still betting the farm on a product nobody asked for. Humans are by their very nature control freaks when it comes to their own lives. They like to have as much say as possible over the things they do. Automating parts of your private life to an LLM which has inaccuracy baked into its very nature is a fools errand because the moment it makes a single mistake on something simple like setting an appointment people will never use it it again and those stories spread like wildfire.

An LLM doesn’t work on absolutes like facts, but rather tries to visualise and predict what it ‘thinks’ is the right answer. Doing so with something simple like a sentence gives them the illusion of intelligence but it’s all still a prediction. This is why they require near constant human checking for hallucinations.

In automation if you cannot guarantee that it’s going to be perfect 99.9% of the time (there is no such thing as an error free machine anywhere on the planet) then you’ve failed. The LLM is just too inaccurate a product to be left to its own devices.

"Humans are by their very nature control freaks when it comes to their own lives."

Some are. And some let their wife or husband make all the decisions. Some delegate to a life coach, some to a guru.
The world is more varied than you might imagine.

What you read in the papers is not perfect 99.9% of the time. And yet people use it to plan holidays, buy products, even decide on careers, let alone buy stocks or gamble.
I'm sure there are hikikomori who are in fact terrified to make any decisions under any circumstances. But they are not the average person.

 

[Happy_John]

The entire economy is "circular financing" as is the biological ecosystem... You say circular financing, I say innovative financing.
It's only a problem for *America* if what is being produced along the way is not useful. That seems unlikely.

If you are an investors, the concerns you give may be important. But as simply a "consumer" of the economy, nah. We've had similar rollouts (the obvious recent example is fiber in the late 90s, but there are earlier examples going all the way back to electricity and oil, and then the mother of them all, railroads, which as a fraction of US economic spend in the late 1800s dwarf anything going on today).
Even the most disastrous of these (consumer airlines, in the sense that the companies are always in crisis and never actually make money for the investors) still generates value for the rest of us, just not for the investors.

And no, "The majority of people [will not] use it for silly things like generating clip art". Every day new people are figuring out something truly useful that they can do with LLMs, and telling their friends. It takes time for people to change, even more so for businesses. Even something as obvious as google took time; there were plenty of people in 2005 thinking that the work you looked something up was to go to the library.

Most things in life are a matching problem. I want a job, you want an employee. I want a house, you are selling one. I want an education, you offer courses. I want a partner, someone else wants a partner. I want a couch, you want to sell me a couch. If you can sit between the two sides of a matching problem there is both tremendous value to be delivered (how much are the best house, the best job, the best partner for YOU worth to you?) and tremendous profit to be made. The somewhat limited, somewhat sub-optimal matching provided by google classic was a money machine the likes of which the world had never seen before. LLMs (once we get past the initial teething stages) will be Google ramped up to 11 as far as being a matching machine. And that's just one element of their value.

Is “innovative financing” the new “creative accounting”? Both sound suspiciously like euphemisms for fraud.

Because the financial state of the AI industry in 2026 is decidedly unhealthy.

Looking at where AI through a purely financial lens, it’s obvious that many of the “players” are not going to survive, and across the board there’s going to be a painful finial “correction”.

AI does not not cover running costs with revenue. It covers running costs through endless investment cycles. That’s not sustainable.

But charging end-users a fee that does cover costs is not viable either - people are not going to pay $1,000 per user a month, but that is what it would take to make the current industry model ro become sustainable.

Right now, investors have realised that the only way they can recoup their investment is to play along, hype the companies, and push for an IPO, immediately after which they’ll offload their stock and run the hell away as fast as they can.

The “big models” are just too expensive to run as an “always on” service to end users. Possibly the one way “big model” companies can survive is to use ‘big models” exclusively to build and train smaller models for local and local network use, and then sell licences to use those small models.

 

[Ishimura]

They’re only exploding because of the circular financing model. Using VC funding to buy loads of GPUs so that Nvidia buy shares so that they can buy more GPUs is a dangerous house of cards pump and dump. They need investment because they’re burning money hand over fist. Their subscription income is but a drop in the ocean for staff wages. They’d have to charge customers £2000 a month if they wanted to remotely break even.

These companies are building out unwanted capacity. The majority of people use it for silly things like generating clip art. It says a lot when if these products disappeared from the planet tomorrow the world would largely keep spinning. You could not have said the same about the iPhone 4.

What are you talking about? I’m talking about revenues and not investment, which is entirely separate. Their revenues are exploding because more and more companies are paying for their product and using their product more.

Their investment rounds have nothing to do with that. Yes obviously they need investment as they spending a lot of money on training and serving their models, everybody knows that.

It is total nonsense to say they are building unwanted capacity. None of the major labs can build capacity quick enough to meet the demand they are seeing, particularly Anthropic. Their ability to supply their product is below the voracious demand for it. How you spin this as an unpopular and unwanted product is beyond all reason.

By the way Anthropic don’t offer any image models at all. Nor video.

The world would have absolutely stayed spinning without the iPhone 4, what a silly thing to say.

I noticed that you’re not actually addressing anything I say.

You didn’t explain if you read the Zitron piece where he directly declines to call it cooking the books. Yet you cited this piece while describing it as cooking the books lol.

 

[Ctrlos]

What are you talking about? I’m talking about revenues and not investment, which is entirely separate. Their revenues are exploding because more and more companies are paying for their product and using their product more.

Their investment rounds have nothing to do with that. Yes obviously they need investment as they spending a lot of money on training and serving their models, everybody knows that.

It is total nonsense to say they are building unwanted capacity. None of the major labs can build capacity quick enough to meet the demand they are seeing, particularly Anthropic. Their ability to supply their product is below the voracious demand for it. How you spin this as an unpopular and unwanted product is beyond all reason.

By the way Anthropic don’t offer any image models at all. Nor video.

The world would have absolutely stayed spinning without the iPhone 4, what a silly thing to say.

I noticed that you’re not actually addressing anything I say.

You didn’t explain if you read the Zitron piece where he directly declines to call it cooking the books. Yet you cited this piece while describing it as cooking the books lol.

They might be bringing in revenue but it doesn’t even come close to keeping the lights on. The money people pay into these companies isn’t being used to run data centres but likely accounts for other running costs. The majority of users are free users costing these companies a fortune.

These companies are running on VC and debt, now amounting to over $1tn in value. That’s more than the banks had in bad debt during the 2008 sub-prime crash. This is all mid-sold AAA debt.

AI remains a massive threat to the economy that doesn’t add to worker productivity. Yet it’s being used as a massive debt and stock grift to undermine the economy it’s built on.

 

[richinaus]

They might be bringing in revenue but it doesn’t even come close to keeping the lights on. The money people pay into these companies isn’t being used to run data centres but likely accounts for other running costs. The majority of users are free users costing these companies a fortune.

These companies are running on VC and debt, now amounting to over $1tn in value. That’s more than the banks had in bad debt during the 2008 sub-prime crash. This is all mid-sold AAA debt.

AI remains a massive threat to the economy that doesn’t add to worker productivity. Yet it’s being used as a massive debt and stock grift to undermine the economy it’s built on.

It does add to worker productivity if they know how to use it in an effective way - my own productivity has skyrocketed and I will be able to deliver far more on my own than I ever could just a year ago. The issue is more that most people have no clue whatsoever, which is nothing new.
In my experience, most just learn surface level tools in tech, and profess themselves as using it. AI is no different.

 

[triumvirate]

Apple's whole AI thing is so sketchy. First, they massively overpromised, and more than two years later, they still haven't delivered on their promises. Now, it's going to be a watered-down version of what they promised, and their whole Private Cloud Compute seems to have been vaporware all along. Even if they keep the branding, it's not going to happen on their own infrastructure, which is completely different from what they promised. I want less Google deals, not more. On top of that, I am sure there will be a 'Siri+' or something similar to get those sweet, sweet service revenues they're lusting for. I do think the on-device capabilities will be quite limited, yet they're still going to hog all the RAM and battery life, even if you don't want to use them. After all, they needed years just to get simple timers to run on-device with Siri.

I think this raises a key problem - RAM. Apple, for years now at this point, has shipped devices with 8gb RAM promising they were AI Ready. The Apple AI RAM models take 2-4gb RAM to run locally from what I have been reading, and this will only increase and future OS updates will also eat more RAM. It feels like Apple lied and cheapened out shipping 8gb ram units, even "Pro", for 2-3 years now promising users they would handle AI perfectly fine.

But Apple also promised AI would release in 2024. I am not trusting Apple like I used to.

 

[Luis Ortega]

making on device as a key focus is something, actually delivering it is another.

We will see how it goes.

Siri 2.0

 

[rosuna]

I think this raises a key problem - RAM. Apple, for years now at this point, has shipped devices with 8gb RAM promising they were AI Ready. The Apple AI RAM models take 2-4gb RAM to run locally from what I have been reading, and this will only increase and future OS updates will also eat more RAM. It feels like Apple lied and cheapened out shipping 8gb ram units, even "Pro", for 2-3 years now promising users they would handle AI perfectly fine.

But Apple also promised AI would release in 2024. I am not trusting Apple like I used to.

Apple has always been very stingy with its hardware specifications, RAM and internal storage capacity (SSDs, or HDs in the past). They claim to compensate for it with software optimization, and I suppose they will use that strategy, but it actually makes the devices work with an arm tied behind the back, and when something like AI arrives there may be insurmountable problems. We'll see. LLMs can be installed right now on our Macs, iPads and iPhones and they are not bad. Apple's LLMs will be very optimized (trained) for their assigned tasks so... lets see.

 

[falainber]

Still waiting on Windows to be able to copy text from an image. Something iOS/macOS has done for years. That’s like the lowest of low hanging fruit in the ML space.

Windows has not one but two built-in ways to extract text from the image:

• Snipping tool
• PowerToys Text Extractor

 

[falainber]

Apple first talked about running LLMs on local devices at WWDC in 2024. It looks like they are going to keep talking. They want to milk their privacy mantra to death. In the meantime we have witnesse incredible progress in LLMs in the last two years: just not from Apple. All Apple did in these two years was relasing cheaper devices with more colors.

 

[name99]

Is “innovative financing” the new “creative accounting”? Both sound suspiciously like euphemisms for fraud.

Because the financial state of the AI industry in 2026 is decidedly unhealthy.

Looking at where AI through a purely financial lens, it’s obvious that many of the “players” are not going to survive, and across the board there’s going to be a painful finial “correction”.

AI does not not cover running costs with revenue. It covers running costs through endless investment cycles. That’s not sustainable.

But charging end-users a fee that does cover costs is not viable either - people are not going to pay $1,000 per user a month, but that is what it would take to make the current industry model ro become sustainable.

Right now, investors have realised that the only way they can recoup their investment is to play along, hype the companies, and push for an IPO, immediately after which they’ll offload their stock and run the hell away as fast as they can.

The “big models” are just too expensive to run as an “always on” service to end users. Possibly the one way “big model” companies can survive is to use ‘big models” exclusively to build and train smaller models for local and local network use, and then sell licences to use those small models.

Here's where I come from (one of many data points)

Look at
<a href="

https://twitter.com/x/status/2061328955404362156

">

https://twitter.com/x/status/2061328955404362156

</a>

Just like anyone (even very new converts) to the internet immediately had no patience for a business that was not on the internet and did not provide decent search, I suspect we will see the same behavior around kiosks (and apps, and general "computing" that does not provide a decent language interface.
It took about five years for this transition in the US economy. But it DID happen. And almost anyone who refused to be part of it is no longer part of the economy. Followed by versions of the same thing in other spaces (like Uber).

This is also why I see Apple's AFM APIs as so important, even though they're not yet widely used. They're the equivalent of adding a search box to every app, by adding the ability to use natural language in any app.
One thing I hope for in iOS27 is wide use of these APIs in all the Apple apps, so that Apple shows other developers how to use this stuff from a UI point of view.

 

[rosuna]

Apple first talked about running LLMs on local devices at WWDC in 2024. It looks like they are going to keep talking.

There are LLM models running right now on iOS and MacOS.

 

[Happy_John]

There are LLM models running right now on iOS and MacOS.

There are, and this is not that recent.

More importantly, there has been non-LLM AI on Apple device for a lot longer.

 

[falainber]

There are LLM models running right now on iOS and MacOS.

There are local LLM models running on all sorts of devices. Apple has nothing special here. Those models are of very limited use compared to the cloud based models though (for obvious reasons).