While it’s still popular to say “software is eating the world,” we may want to pay hardware more mind given all this Apple news.

If you’re investing in, valuing, or growing an LLM or an LLM-centric product right now, which means if you’re anyone in technology investment, you need to know what major changes are afoot, and we’ll advise you on them here.

A potential alliance between Apple and the makers of efficient, locally run LLMs could change the way you need to think about investing in AI firms and adopting their technology, as entirely new frameworks and players mean entirely new fields of diligence and expertise.

Things we’re thinking about that savvy investors and ops leaders should be thinking about, too:

This is neither financial advice nor legal advice.

On a podcast recorded a couple weeks ago, we tackled a topic that we didn’t yet know would soon become of even clearer importance: Apple’s perceived position in the AI race.

The news at the time was that they were in early-stage talks with PrismML about a partnership, in a bid to continue down the road they’ve been on for a while as the hardware that would run the software that would power the AI revolution.

This is a notion that’s been gaining steam as the main counterpoint to “Apple is falling behind by failing to deliver a big LLM product.”

It began when reports about a new wave of highly-efficient Chinese models ripped through the hyperscaler hypesphere, tantalizing us with a possible future where LLMs are so judicious in their use of compute that, for almost everything we use them for, that we can run them locally on our own computers.

The ultimate dream, of course, is that users could even run sophisticated models on their phones.

When Openclaw was suddenly and briefly ubiquitous, AI maxxers rushed out for Mac Minis, a second glimpse at an LLM powered world where Apple could have a central role to play.

If you’d asked the incumbent powers if they were worried about Apple on most given days, I imagine they would have told you that they wouldn’t even make a list of the top five competitive threats.

Seeming to some like a nail in Cupertino’s coffin, OpenAI even acquired io Products, the company founded by Apple’s legendary former design chief, Jony Ive, and announced a forthcoming AI hardware device.

A few hours after we’d recorded our podcast, we learned that OpenAI might actually be extremely bothered by visions of a future in when Apple eats into their AI revenue pie and not so sure Jony Ive had walked out with all of the company’s secret design sauce; news broke that Apple had filed a trade secret lawsuit against OpenAI and several individual defendants.

How we got here and where it means we’re heading is fairly apparent when you consider just how central the iPhone is to humans’ modern digital existence.

Apple’s role in the “AI” era and what it has to do with any of this is a fair question-they make computers, smartphones, and related accessories, plus the operating systems and software that power these (iOs, MacOS, etc).

They’re also, of course, a huge players in music and movies and TV, although they don’t break this revenue stream out by each distinct segment, lumping them under Services.

Where did the expectation come from that that Apple need to be a horse in the LLM race, exactly?

The softest notion, although not one worth dismissing entirely, is simply that they are an Important Technology Company, and as such, they need to be involved in every important tech paradigm shift.

I say we shouldn’t dismiss this casually because it’s impossible to deny that the stock market viewing a company as AI aligned, or not, has consequences that even non-tech companies are heeding. Jersey Mikes put “AI” in their S-1 for their IPO, and CEO Damien Adamolekun says he wants to make Red Lobster “the most AI forward restaurant that exists,” and so whoever we think should be spared the need to claim to be riding the AI wave, I don’t think we can say it’s Apple.

More meaningfully, Apple makes money selling us devices which we use to consume media, and use the internet, which means they make money by being the portal through which our attention flows and the hardware we do work on. Via their services division, they also sometimes make the things we’re paying attention to or working at.

An easy way to understand their position is by pondering what might happen to Apple Music in a world where people prefer to generate their own music on Suno, and Suno inks a partnership with OpenAI’s smart speaker and a hypothetical line of wearable devices. This would cut Apple entirely out of at least three different huge pies when you factor in the possible disruption of AirPods that happens in this hypothetical.

That’s a bit of a reductio ad absurdum, or at least a pretty severe worst case scenario for Apple, but I think it hammers home the tangible fears for their incumbency here.

In fact, before we fully assess this, we should consider the total power Apple’s position has afforded them historically and still does today, so we can understand the flip side of the above scenario: how Apple might continue to dominate in the future using its broad, highly defensible positions in multiple consumer sectors.

Figures vary based on how you want to track it, but by current active device OS share, iPhone has about 58.3% of the US, and about 31% of global device shipments.

If the global number sounds low to you, it’s helpful to consider this still puts them at the top of the charts.

For this reason, most of the internet giants of the modern era have been pretty dependent on the iPhone for their own growth-a reliance reflected in their own SEC filings, which consistently list iOS availability and Apple platform changes as material business risks.

Let’s run through FAANG to see how Apple impacted its peers in the pre-”Magnificent 7” era.

Central to Google’s monopoly trial proceedings was the amount they paid to be the default search engine on Apple’s browser Safari, which had ballooned to over $20 billion a year by 2022. It seems like Google thinks Apple’s devices are a pretty important venue!

That’s without accounting for YouTube, which serves vast amounts of mobile video on iPhones; additionally, many creators have adopted Apple's suite of creative tools, including Logic Pro and Final Cut Pro, as key parts of their workflows, though Apple's dominance in this space has ebbed and flowed significantly over time.

Meta (then “Facebook”) invented the first good mobile ad IMHO, and today derive ~92% of their ad revenue from mobile devices. Meta’s official earnings documents don’t break that out between OS or device for us but it’s fair to infer a significant portion of spend is on iOs users.

You know what was the last, and biggest, thing that ever really socked Meta’s stock in the stomach? The iOs 14.5 version update, which severely hampered mobile app user tracking.

Amazon’s attempt at a phone went very badly, and while Alexa is technically a thing lots of people use, so is Siri, to an extent such that it’s hard to say Amazon has really edged Apple out of the Objects You Yell At market. Also: Alphabet haters, look away from these numbers!

Netflix are basically the on exception to this, but that’s because they essentially derive no revenue from mobile phones, something it appears they’re trying to change.

This is valuable context-now we understand the completeness of the device prison that the next wave of world-striding internet giants hope to cast off on their way to trillion-dollar valuations.

A challenger on this front would be aiming to break into the now highly competitive (if sometimes also highly concentrated, especially in the US) global smartphone, laptop, home speaker, home assistant + network, and general wearables markets if they want to do more than simply align with a different partner and/or challenge Apple.

Now we get to another question: why would OpenAI, or Anthropic, or any other frontier LLM purveryor for that matter, not simply align with the 800 lb. gorilla of the global hardware market?

Sure, they can be a tough taxman, as Epic Games and their legal action against the App Store’s 30% cut will tell you.

That being said, Epic did massive business on the iPhone, as many other software companies have inside Apple’s ecosystem, and most without making a serious effort to break out of any part of it.

A potential cut taken from subscription revenue or any other expensive “cost of doing business” payment to apple could be an existential problem, however, for OpenAI.

OpenAI's most expensive subscription tier is itself a money-loser, and the deeper structural problem is that compute costs dwarf the subscription revenue the company does collect.

I probably do not need to tell you about OpenAI’s capex and profit figures if you’re reading this, but documents leaked a little over a month ago that were subsequently audited by multiple parties indicate they’re currently losing about $21 billion a year.

A March 2026 survey by Bank of America indicated that roughly 3% of American households pay for some kind of AI product. This would not be, by any measure, a good penetration rate for any mass consumer technology product in terms of paid user base.

Here is where those who think this doesn’t matter would say, as they did in the crypto era, “we’re so early!”

The issue here is that while it’s hard to say objectively where the LLM movement is temporally as a consumer market, probable US market penetration in terms of regular LLM chat bot usage sits closer to 56%.

So in terms of some kind of adoption, over half of the US is regularly using these tools, and has not yet decided to spend money on them.

It’s also worth noting I’m talking about public generative AI usage and OpenAI a bit interchangeably here, which is quite dated-while some people use the word “chat” as shorthand for any LLM chatbot they’re talking to, it’s not quite the Coke or Kleenex it used to be when it comes to market share.

If we say a good freemium consumer software free-to-paid conversion rate is 10%, and great is 20%, we could say 5.6%-10.2% of Americans paying for some kind of LLM is a place we could plausibly get to, on this one wobbly leg, at least.

Where things start to get messy are two relative unknowns about LLM product monetization and how they could upend that potential.

We have seen we have seen a mixed pricing picture on LLM subscription plans, including a price hike on ChatGPT Team (+20% in November 2024), the introduction of new ultra-premium tiers at $200–$250/month, and most recently a price reduction on Google AI Ultra from $250 to $200/month.

Companies are also experimenting with usage based models, albeit more on the B2B side of things.

Subscriptions persist as the main recurring revenue model for B2C software because they’re widely regarded as simple and predictable, and ergo, easy to sell to people.

While there’s no evidence the headache of confusing pricing might come to the consumer side, it evokes the specter that hangs over all LLM pricing: the growing notion that all of this has been costing a lot more money than they could ever charge.

Usage based pricing solves the problem of a product where a segment of users account for a disproportionate amount of the software’s total usage. This can be key to address indirect and theoretical costs, like technical support and the need for innovative new feature development that benefits a small user segment, or it can address a very direct cost such as energy and materials.

OpenAI and the other LLM companies are dealing with the latter situation in which product usage directly incurs cost, which makes usage based pricing a natural fit.

It has to be said, however, that even if the cost incurred varies wildly user to user, if a whole product is still profitable across its whole user base, a one-cost subscription model can still technically work.

The Occam’s Razor conclusion here is that if a company is aggressively fiddling with pricing, it may be a signal that the whole picture does not add up to profit.

The longstanding assumption around “how all this works in the end” for the LLM business is very simple:

Frontier LLMs are currently expensive, but they will become so essential to everyday life, and decline dramatically in cost, such that the freemium conversion model will work. An LLM subscription will be about as optional as car insurance, and be seen as a reasonable expenditure for the value it delivers. The only people who won’t pay for it will be very poor people and we may consider subsidizing (a very diminished version of) it for them because they’ll need it to even have a glimmer of hope in their lives.

What may be surprising to many people currently watching the AI space is how quickly things could get cheaper, who might be the responsible for that efficiency, and who stands to really benefit.

As with many modern global sociopolitical and macroeconomic transformations, it largely revolves around China.

Apple could have been happy with a world in which OpenAI, Anthropic, and Alphabet deliver the consumer AI experience.

Apple already receives a hefty sum from Alphabet to keep Google as the default search engine in Safari., and if the new players could learn to accept the Cupertino Cut, everyone could be happy creating mountains of shareholder value.

It seems, however, that OpenAI chose war, first with their acquisition of Jony Ive’s io Products, and if the lawsuit allegations ever prove true, with trade secret theft. What the actual legal outcome will likely be doesn’t matter much for their relationship, as Apple has clearly decided what OpenAI’s intents towards them are and drawn the battle lines.

OpenAI have signaled they think it is vital in some measure to not be contained in the hardware ecosystem of another party, and to sell your own hardware. I don’t want to discount that they may be intent on making a lot of profit on their devices, and not just using them as a tool to jailbreak out of other peoples’ machines, especially given the future of LLMs I am about to deem highly plausible.

In turn, this means that what Apple possibly wants is an entity besides OpenAI to “win” LLMs, who Apple believes is more interested in a symbiotic partnership. They could potentially have this in Alphabet, who seem happy enough ruling just the Android side of phone world to make minimal inroads into iOs’s application layer.

What could also be acceptable to Apple is a bit of a “scorched earth” or “no loss” scenario in which Apple don’t monetize LLMs heavily, but see to a world in which the ultimate leaders in the LLM space do not threaten any of Apple’s current or future profit streams.

A future like this was already taking shape in the high efficiency open parameter models coming out of China, which Apple could involve themselves with in order to accelerate their impact on American consumer LLM usage.

The now well known “killer trait” of these efficiency focused open parameter models, Chinese and otherwise, is that they’re well positioned to capitalize on a simple truth about consumer LLM usage: much of it does not require 2026’s most advanced models, not by a long way.

Putting Apple aside for a moment, the biggest challenge this kind of model faces will be twofold:

If the price of mass consumer LLMs do continue to rise relative to the general cost of living, these issues may work themselves out via consumer demand for free or very cheap LLMs. People won’t mind a little managing and metering of something when there’s visible, meaningful money savings.

As far as the carrot, I don’t think I know any company in the world better equipped to sell the American people a device with an “unconventional” technical performance story, or something made in China, than Apple.

In 1998, Apple launched the now iconic Bondi Blue iMac with a bang. Most of us were sold on it because of its wild design, but back then there actually was a performance narrative (that some say have cherrypicked a bit, sure) widely regarded as a key part of the marketing.

A mass market PC comparably priced at the time had better numbers on paper, and if you factored in the top end of the custom PC range that a hobbyist could build with $1,300 in parts, it would look even more like a win for PCs.

However, the iMac used a RISC chip architecture that allowed the 233mhz G3 chip to outperform Intel chips with significantly higher cycle speed on some key benchmarks, and came with a large full-speed L2 cache that many PCs still lacked, which alleviated a data bottleneck on memory-bandwidth-limited tasks that diminished the impact of higher speed chips on PCs without the technology.

By leaning into some select areas Apple knew the iMac did well (e.g. photo editing and other graphics work) and not getting into fights they would lose (e.g. gaming) they positioned it as a superior performance product. Plus, many users and reviewers saw the iMac as a tidily and beautifully integrated product, unlike the fiddly mess of a megatower PC.

It’s been a long time since those heady days, but this narrative is so clear that it won’t take a second coming of Steve Jobs for Apple to dust it off and reuse it well.

It goes like:

“The frontier labs make big, complicated, expensive software models they want to charge you an arm and a leg for.

Apple’s got something that’s arguably even more technically impressive and innovative-the difference is that its brilliance is in how efficient and simple it is. It does everything you need well-without trying to impress you with a bunch of complicated reports with numerical scores on a whole battery of oddly named tests.

Oh, and by the way, we’ve neatly bundled it up in a cute, artsy way in a product you already love!”

Much hay will be made, and yes, legislation threatened, over the Chinese provenance of many of these efficient LLMs, as potential threats to security, and economic dominance. So long as the law doesn’t get in the way I suspect this foreign medicine will be easier for American public to swallow if it’s Apple flavored.

Just go ahead and sign tens of millions of American up for the Bondi Blue Powered By Kimi iPhone. Let’s just hope it’s not only for AT&T customers at launch all over again.

I don’t want to diminish the potential impact of the hard power of regulatory measures against open source models, from any source, or the soft power of authorities coming out against them in a way that shapes public perception significantly.

At this time, however, there’s no helpful speculation we could make about which direction all of this heads.

The current American presidential administration has pursued trade policy at a pace and scale widely described as historically unprecedented in the post-WWII era.

There is an added layer of specific policy procedures around AI models applicable even to domestic companies, with the President openly speculating about the government taking a stake in AI companies, while also subjecting their frontier AI model releases to a voluntary pre-release government review process under national security rationale, without mandatory hold or preclearance authority.

Some other newsletter out there may be able to tell you exactly what’s in store for AI labs and American public policy, and bully for them I guess, but if you’re going to heed their advice I would simply advise you that you may be getting suckered into some sort of prediction market manipulation scam.

We don’t know anything!

AI professionals and avid hobbyists well along in their LLM journey likely have experience with open parameter models and running on local machines, but a whole ecosystem and business environment that shifts away from frontier model LLMs as complete, cloud-driven products, towards LLMs as simply software running on a variety of hardware as part of a larger OS made by another party, is potentially a seismic shift.

Primary concerns historically have revolved around data sharing, privacy, and the provenance of training and context data for a given model in use.

For what it’s worth, here at Daypart AI, there are frontier models we do not interface with due to concerns around how our data that is ingested could be used.

Moving out of these environments can add a degree of control and privacy but may introduce other risks; you lose what might be called “unaccountability as a service.”

Downloading the open weight model: free, feels great, report $0 capex to investors.

Buying hardware, running, maintaining, and updating the model: expensive, lots of work, tummy hurts, ask investors a lot of concerning questions about how they define “depreciation” and “amortization” when it pertains to GPUs.

In all seriousness, this is a transition to being a “bare metal” firm and all that comes with it. “Not that there’s anything wrong with that” I say in the same tone as I would addressing someone who lists Linux as their principal hobby.

While I’m not able to discuss this topic for even a moment without making a joke, what’s no joke is the amount of IT operations planning you’ll want to see from an enterprise to ensure that locally managed open parameter LLMs are a comparable or lesser cost, and not a wildly variable and difficult to forecast expense.

Get ready to hire several of these guys:

Once you’ve set up your GWs of GPUs, you’ll want to track your L2PHs, or Legal Liabilities Per Hour, and tack those projected legal fees onto your OpEx.

I’m kidding-BUT only a little.

Liability is a wild, wild space when it comes to LLMs, and while looking at the evolving policies and licenses of sector leaders, it may be tempting to wonder aloud “what exactly wouldn’t be my fault if it’s legally wrong?” You do get some protection, though.

For extremely important matters like training data provenance, you benefit from basic indemnification, but broader context management is generally not covered under standard indemnification terms.

Once you’re off their platform, you’re largely on your own.

Privacy is an area where you do theoretically get an immediate upgrade, as you get maximum data sovereignty and can theoretically close all your own loops. The tradeoff, of course, is that coming up with your solutions is now your problem, and that includes all the wild stuff starting to manifest on the AI cybersecurity frontier.

Honestly, it sounds kind of amazing-it’s just also going to be a lot of hard, high stakes work.

Companies you’re serving, or seeking to invest in or partner with, are going to need to bring all kinds of bona fides to bear, and you’re going to need to be able to do deep, serious research and atomic claim verification in order to be sure they’ve got what it takes to handle it all.

The payoffs could be amazing, though-a world in which most portable and wearable devices provide powerful LLM driven application with real utility, all without the need for vast amounts of compute and energy infrastructure?

Don’t you dare call us “technopitmists,” we’re decidedly technorationalists or technoskeptics, but please excuse us if we briefly get a little excited about that possible future.

Well, as long as it doesn’t come with too many more 30% platform fees.

OpenAI Policies

Anthropic Policies