Apple Silicon: First Among Equals Is No Longer “Good Enough.” Tales of Silicon, The Myth Of Local AI, And A Deluded Commentariat.
Apple Silicon is excellent. That was never the problem. The problem is its “exceptionalism:” the belief that good integration excuses fifteen years of Siri drift, two years of AI delay, and a market now mistaking local inference as a reality. Apple is at the front of the pack, not streaks ahead.
I had a read of the Apple blogosphere today, and steamed a bit after browsing all the armchair AAPL experts frothing at the mouth attempting to comfort themselves with “nobody but Apple can do this,” so let’s dispel these myths once and for all.
Because, sometimes a picture is worth a thousand misplaced blog comments.
Not because pictures are cleverer than words, but because they make certain “evasions” by armchair experts harder to maintain. The “drop” image at the end of this article is not intended as a technical benchmark chart. It is a visual map of a delusion: the belief that Apple is somehow uniquely positioned to own local AI because Apple Silicon is excellent.
Apple Silicon is excellent. That is not the argument. It is no longer the exception though. Apple is strong, but the rest of the world has not been preserved in formaldehyde since the M1 launch six years ago.
Want to skip the sermon and skip straight to the infographic which illustrates this in a jaw dropping and humbling glance?
Go to the end of the article.
If you want to educate yourself a bit about Apple’s performance relative to its competitors when it comes to LLMs, AI and Silicon, read on for a deep dive.
If you’d like to read the rest of this article, free of charge but for providing your email, just press the button and select the free tier option. You’ll receive articles direct to your inbox. Don’t worry, you can turn off this feature later, if you don’t want notification.
Note: If you are already a member, but de-subscribed from the newsletter, you will need to re-subscribe to it to access this article. Just go Account settings and re-enable the Newsletter!
The argument is that excellence has been inflated into exemption, and Apple are somehow, uniquely positioned.
For years, since 2022, the Apple faithful treated AI as hype, LLMs as parlour tricks, and AI-assisted writing as evidence of moral or intellectual softness. If you don’t like LLM output, stop reading now, because I unashamedly use it. They even ridiculed me for using it - openly - to help research and produce skeletons of long-form articles. And now judging by the length of some forum comments, they’re all using it themselves now.
Then in 2023 the market turned. Apple stumbled.
Siri remained Siri. Apple Intelligence became a promise in 2024 wrapped in a delay, wrapped in a keynote, wrapped in a valuation premium. Suddenly, the same people who once sneered at AI are producing thousand-word essays that read suspiciously like LLM-assisted catechisms explaining why Apple was right all along (and of course, they along with it).
There is a certain theatre to it. “X is Y, and “Alpha is Beta.” And “That matters, because:”
Intelligence is replaced by inference. Due diligence is replaced by hyper-vigilant pattern hallucination. Technical understanding is replaced by word count. The result is not analysis so much as devotional sludge: confident, padded, beautifully over-long, and wrong in all the important places. And GPT output is recognisable at a glance - without needing be so neurodivergent as to be a good patter spotter.
Just like the paragraph above.
The latest version of this is the myth of Apple exceptionalism in local AI.
The claim goes something like this:
- data-centre AI is doomed,
- only on-device AI matters, and
- only Apple has the hardware, software, silicon and operating system to make it work.
Therefore “Apple is not behind. Apple is merely early to the correct future while everyone else burns cash in GPU cathedrals.”
Neat. Comforting. And nonsense.
Apple does have a serious advantage in integrated consumer hardware.
- The M-series chips remain beautifully engineered.
- Unified memory is useful.
- The Neural Engine matters.
- Apple’s control of silicon, OS, battery, privacy model and developer frameworks gives it a cleaner path to on-device features than most competitors.
- In 2025, the M4 Neural Engine was rated by Apple at up to 38 trillion operations per second, and Apple was perfectly entitled to boast about it when it arrived.
But the wider market did not stop moving. It narrowed, and the Walled Garden suddenly become a lot less exceptional, just nicely pruned.
- Qualcomm’s Snapdragon X Elite arrived with a 45 TOPS Hexagon NPU,
- Intel’s Lunar Lake/Core Ultra 200V platform brought an NPU rated up to 48 TOPS, and
- Microsoft’s Copilot+ PC baseline itself requires a 40+ TOPS NPU for many local AI features.
Apple’s Neural Engine is not a sacred object. It is just one very good implementation of a now-standard direction of travel.
So pple Silicon may be excellent, but LLM usefulness is now a model, memory, orchestration, interface and workflow problem, not merely a TOPS problem.
At the end of this article, you won’t want to look at the image on a mobile device.
But you owe it to yourself to pull it up on a large screen monitor and look at the state of AI and GPU a year ago, last year in 2025. Let alone 2026, with other chip fabs accelerating their gains on Apple’s lead.
Apple Silicon rules, but at this point, the difference is at the margins, and down to power efficiency.
Just because, for example, an F1 car from Apple might top out at 200mph and Qualcomm’s at 180mph, doesn’t mean Apple won’t win on speed. It means that against all high performance platforms, it offer a negligible difference until you reach the top 5% of performance, and all are powerful enough to host on-device LLMs and hand off to the cloud where greater horsepower is needed. The cloud is not doomed; it’s the future, just the same way Dropbox showed us the way, when Steve Jobs declared that not buying Dropbox was one of the worst decisions he had ever made. I think dot.mac proved him right. Luckily he didn’t have the hubris to pretend he was ahead, he made sure Apple moved tangentially to mark out new territory with new paradigms to make up for an ever-lengthening list of projects which were being outflanked by other in-the-cloud efforts, from Maps to cloud to rage to music streaming.
And it moved again with the advent of Apple’s M4 in 2024/2025.
Qualcomm’s Snapdragon X2 Elite family was advertised with up to 80 TOPS of AI processing in its Hexagon NPU. Its mobile Snapdragon 8 Elite Gen 5 brings a faster and more efficient Hexagon NPU to flagship phones. Google’s Pixel phones even come with a security enclave built-in very similar to Apple’s.
The point is not that Qualcomm has “beaten Apple” in some simplistic pub argument.
The point is that Android phones and Windows laptops now have their own neural engines, their own NPUs, their own local AI pipelines, their own developer routes, and their own battery-conscious inference story.
In fact they have for years, it’s more a case of nobody having an LLM which was small enough to be capable of decent on-device use - whether Apple or anyone else - and not that Apple has a head start here. It doesn’t. In fact it’s lagged so badly, it’s had to rent the models from Google, and hook Apple Intelligence up to Google’s server farms (in what are meant to be isolated PCC servers stated not to have any ability to leak into the Googlesphere. We now how much Google respects privacy though, don’t we…)
Let me be clear:
N0 platform has a decent LLM which can work solely on-device except in carefully manicured “answering machine” demos. Which is why Apple, last years, had called its “new on device” version of Siri, the “Answering Machine” project. It was a glorified Clippy with some expert knowledge. It was never an LLM, which still eludes Apple and everyone else. And no on-device LLM will ever be capable enough of satisfying increasing user demands serviceable only by in-the-cloud high level inference and responses from the best LLMs, as we approach true AGI territory.
Because if one existed, it would already be running on every device except Apple’s already. Speed is no longer the barrier. Compressing the LLM and having enough onboard memory is.
That alone should be enough to puncture the fantasy.
Apple may package the experience better and it may integrate more cleanly. With better power efficiency It may even protect privacy more elegantly (although it seems they’re going to invoke this by giving Siri amnesia, if reports turn out to be true).
Apple may offer the least ugly consumer version of local AI when it eventually ships the whole thing properly but it is not alone in having the hardware or the OS and certainly not alone in having on-device inference. Why? Nobody does.
The AAPL preachers (as opposed to enthusiastic investors who actually underdstand the technology they’re investing in) keep confusing “best integrated” with “only option.”
That distinction really does matter (and that’s not just a GPTism) because the whole industry is now building across three compute layers at once:
1) At the phone level, Apple competes with Snapdragon-class mobile silicon that already includes dedicated AI acceleration and security enclaves..
2) At the thin-and-light laptop level, Apple competes with Intel Core Ultra and Qualcomm Snapdragon X platforms built explicitly around local AI experiences.
3) At the desktop and workstation level, the power-efficiency argument changes shape entirely. Once a machine is plugged into the wall and thermals are allowed to breathe, discrete GPUs and local inference rigs become a different category of animal. Apple’s integration still matters, but raw scalable GPU power can marginalise the elegance of that advantage.
The Myth Of the Mac Mini and AI
To dispel any further myths, the Mac Mini has been a very popular choice for people wanting to run LLM’s locally. However this is chiefly an phenomenon to experiment with a very dangerous and privacy poking agentic OS called Openclaw, for which the Mac Mini is ideally suited because it’s cheap, can run isolated from sensitive data, and generally not run production level or commercial systems. It’s a hobbyist obsession, not a long term commercial advantage.
Think of it as the Cabbage Patch Doll rage, if you can remember that far back. Or if you like, Demand spike ≠ proof of monopoly,
Apple was only overwhelmed with demand because of the Mini’s relatively small production volumes - which was immediately soaked up when when the open source Openclaw arrived, followed by Perplexity launching a Mac Mini + Perplexity computer for an astonishing deal.
Hint: Openclaw runs on any platform, not just Mac minis. In other words, don’t confuse demand for exceptionalism: the mini is a great machine, but was a convenient package to buy at short notice by hobbyists and corporations alike desperately wanting a cheap small system to experiment with, and not because of its undue power. Demand is unlikely to remain at these levels for long, now the base price of the Mini has jumped 30% and cheaper PCs undercut it.
No mainstream device has a frontier-class LLM experience running solely on-device in the way the Apple faithful imply.
Small local models can work well for narrow tasks, coding assistance, summarisation, retrieval and agent support, but the serious frontier experience remains hybrid: local where possible, cloud where necessary, orchestration everywhere.
So the mythology of the recent Mac mini narrative ignores the fact that PC platforms not based on Apple’s M-series are perfectly capable of performing the same feat, on small PCs now cheaper than the base Mac Mini’s old $599 price, and which is now 30% higher at $799.
That is why the picture matters.
Apple’s edge is integration, not exemption.
It is a first-among-equals story, not an only-adult-in-the-room story.
Armchair AAPL investors hate that distinction because it removes the narcotic.
- Apple can be excellent and still late.
- Apple Silicon can be superb and still not enough.
- Siri can run on brilliant hardware and still be dim.
- A Mac mini can be a wonderful local AI box and still not prove that every other platform is technically barren farmland.
The more uncomfortable point is that the model layer has moved faster than the Apple story can metabolise.
Since Apple first promised the Apple Intelligence/Siri future at WWDC 2024, the frontier has not politely waited in an orderly queue.
- OpenAI has moved into GPT-5.5 - which unless you pay a bargain price of $20 a month for, you will never discover the incredible difference between it, and the free tier - from which your cognitive bias against LLMs probably stems.
- Anthropic has pushed restricted frontier work with Claude Mythos Preview,
- Google has deepened Gemini across its stack (and - oh the irony - performed a reverse takeover of Siri in the process), and
- Perplexity has continued to point towards the interface layer as the real battleground.
This is the bit the Apple commentariat still struggles to grasp:
The fight is not just model versus model, or chip versus chip. It is interface versus interface. It is who owns intent. It is who turns user context into action. And Apple’s platform is meaningfully more efficient than others, to make it the exception its investors and fan base seem to mistake it to be.
And That is where Apple’s problem becomes historical rather than merely technical.
- Siri launched in 2011.
Apple had the first mainstream consumer assistant. It had the phone, the OS, the iPad, the services surface, the App Store, the identity layer, the privacy story and the customer relationship. It had all the pieces. What it did not have, apparently, was the institutional will to turn Siri into the Knowledge Navigator future before others built it above Apple’s own platform and actually integrate all of those components into one stack. In fact if anything the company has been so two-left-feet that even with all the parts existing in isolation, hubris and hesitation seem to have paralysed the company into inaction. My worries have been, ever since 2023, that Apple can’t ship anymore and the last three years have, to date, borne that out.
- By 2024, Apple was no longer leading the assistant era. It was promising to re-enter it.
- By 2025, the promise had curdled into delay.
- By 2026, the stated solution to involves Gemini doing most of the thinking Siri should have learned to do years ago.
That may be commercially sensible with the benefit of a history of failure. It may even work. But it is not vindication. It is a white flag wrapped in Cupertino reality distortion amplified by Apple blogs rocking readers back and forth that “Apple will get it right in the end.”Just like they did with “Maps,” I suppose?
The same pattern appears in the broader AI-browser and agentic-interface shift. Perplexity is not merely “search with citations” but a live example of AI as orchestration layer: model-agnostic, interface-led, able to collapse search, research, summarisation, memory and workflow into something closer to a cognitive operating system.
My own Perplexity/SenseOS argument was built around precisely this point: Apple does not need another soloist as much as it needs orchestration, and Perplexity looked uncomfortably like the thing Apple was too proud or slow to ship.
I wrote about this extensively a few days ago:
Tommo_UK
A look forward and a look back. Where will we be after WWDC 2026?
So no, the image at the end of this article is not saying Apple Silicon is poor. It is saying the mythology around it is lazy, and charts the how and the why.
- Apple’s hardware advantage is real.
- Its software execution gap is also real.
- Its privacy stance is valuable.
- Its delay problem is corrosive.
- Its integration is impressive.
- Its AI delivery has been inadequate to the point of non-existent.
All of these things can be true at once, which is why the more tribal Apple commentary has become so utterly useless and resembling more of a salon talking shop than anything intellectually engaging or inquisitively investigative.
The question is not whether Apple can run AI locally.
Of course it could if a Mac Studio had an NVDIA GPU and stack in it..
The question is whether Apple can ship an AI experience that matters for the every-person before the rest of the industry turns the OS into scenery with agentic AI sitting on top of it like freshly poured concrete.
That is why “Real Artists Ship” still bites.
Under Jobs, Apple’s greatness was not that it waited until the universe gave it permission to be perfect. It shipped, iterated and forced the market and the world to respond. The record is not spotless, but the cultural bias was clear: the product had to leave the building, working, and the task became a rapid release of the follow up. The “don’t buy a first generation Apple product” used to be legendary.
The more recent pattern under Cook has too often been delay, polish, risk management, and then a keynote trying to reframe lateness as wisdom, and then delay everything year after year until either it has to be cancelled, wrapped up in spatial anomaly or just launched five years late and without a use case.
The faithful keep saying Apple is playing the long game.
But sometimes the long game is just the short game missed repeatedly with better typography, and recycled on repeat by blogs with nothing new to say in the absence of anything actually shipping.
And sometimes a picture is worth a thousand comments because it shows the thing the comments were written to avoid:
- Apple started early,
- Stalled for years,
- Watched the LLM world take off,
… and now needs to prove that its beautiful hardware stack can support a real intelligence layer, not merely a more apologetic Siri.
Apple’s advantage is superb integration. That gives it a 15-20% performance head start.
It does not mean Apple is out on its own any longer. It means Apple is a higher-performing high performer, not the sole author of a brand-new paradigm. That is a “does it better” advantage, not a “nobody else can do it” exemption.
The myth is this sense of exceptionalism and exemption from judgement. Mythology, however carefully packaged, is not a product. It’s hopium.
”The Drop” illustration below. Courtesy of ChatGPT 5.5 as the core running on my own custom agentic AI, “Ask FYI” ((currently in, you guessed it, beta testing) and illustrating the state of play as of this time, last year, drawing on my own archive. To be updated, post-WWDC 2026
Apple does not just need to ship Siri, Apple Intelligence, another iteration of Liquid Glass, or a new M6 processor.
It needs a new OS-level integration layer that makes “on device” actually mean “across devices,” and “across apps.”
Which is what I proposed in SenseOS, first published before WWDC 2025: cross-device intelligence, ambient context, and an Apple interface that acts less like a feature drawer and more like a nervous system.
Tommo_UK
Sense OS - First published as my concept of across-device intelligence, in May 2025
Apple, have you been listening?
Or are you still going to try and awe us with Genmoji and gaslight us with “more coming soon” instead of a great ”one more thing?”
See you in a couple of weeks when the WWDC dust has settled, and the arguments are over about whatever Apple ships, does not ship, promises it will ship, or half-ships while playing footsie under the table with its fabled installed user and investor base — just enough to stop them going elsewhere out of CGI boredom.
Tommo_UK, London, Wednesday, 27th May 2026
© 2026 Tommo_UK / tommo.fyi
CONTACTING ME
💡 Reach out to me using the Confidential Drop Box form below.
CONTACT ME DIRECTLY: discreetly (and anonymously if you prefer)
