Ed Zitron: Apple Will 'Watch Everything Burn' When the AI Bubble Bursts

Memory prices person doubled, Macs and iPads person gone up, and iPhones are expected to follow. Ed Zitron – who writes the Where's Your Ed At newsletter, hosts the Better Offline podcast, and has been described by Politico arsenic the AI boom's astir "acerbic gadfly" – has spent years arguing the buildout driving those costs volition ne'er wage for itself.

We asked him what happens to Apple if he's right.


You've been calling AI a bubble since earlier it was fashionable. For MacRumors readers who mostly cognize it arsenic ChatGPT oregon Apple Intelligence connected their iPhone, what, successful plain terms, is really breached astir the economics of the LLM industry?

At their precise core, Large Language Models' costs tally contrary to fundamentally each exemplary of selling software.

Consumers and enterprises alike person been trained to wage a monthly interest for a service, and portion these services mightiness person limits oregon strictures, fundamentally cipher buying bundle expects to person a metered service, fto unsocial 1 that's some metered and with hard to measurement costs.

LLMs pain tokens astatine a per-million complaint careless of whether oregon not you get the effect you want, oregon whether it does what you inquire it to do. If you inquire a coding cause to bash immoderate benignant of bundle task and it goes disconnected and spins its wheels successful a loop, you're paying for the tokens regardless.

AI companies knew that consumers would ne'er wage the existent outgo of their AI services, truthful they have, for the astir part, sold them monthly subscriptions with vague complaint limits that let them to pain mode much successful tokens than the outgo of their subscription. SemiAnalysis recovered that you tin pain hundreds of dollars connected a $20-a-month subscription and thousands of dollars connected a $200-a-month subscription, and portion AI boosters volition assertion that these companies person "70% gross margins connected tokens," determination is small impervious that this is the case, and my ain reporting shows that OpenAI mislaid $20.9 cardinal connected $13.07 cardinal successful gross successful 2025.



This means the precise basal economics are broken. If Anthropic and OpenAI believed customers would really wage the existent outgo of AI tokens, they wouldn't person to springiness distant 20 to 40 times the magnitude of tokens to subscribers.

Meanwhile, backmost successful March of this year, some moved their endeavor customers implicit to token-based billing. Within a fewer weeks, it came retired that Uber had spent its full yearly token fund successful the abstraction of a quarter, and its COO said that it was getting "harder to justify" the outgo of AI due to the fact that it was hard to way the outgo of AI to immoderate existent utile features shipping. Sam Altman would yet accidental it was a "huge issue" but declined to accidental however it mightiness beryllium fixed.

This is simply a occupation crossed fundamentally each azygous AI-powered startup, which has to wage the per-million token rate. Perplexity, Cursor, GitHub Copilot (which moved to token-based billing successful June) – each azygous AI startup is unprofitable due to the fact that their users don't privation to wage the existent outgo of AI.

Another contented is that AI services are conscionable not that utile oregon differentiated. While radical get immoderate benignant of payment retired of AI-generated code, these tools really extremity up making them slower, and are filling codebases afloat of slop. Otherwise, an LLM is an LLM is an LLM – it tin generate, it tin summarize, it tin search, and that's astir it, which means that each AI work is efficaciously the same. That's wherefore 89% of each AI revenues are Anthropic and OpenAI, and wherefore each AI startup talks successful presumption of "annualized revenue" (monthx12) – due to the fact that actual revenues are precise depressing. Even then, astir are hardly astatine $100 cardinal annualized.

Then determination are the information centers. An AI information halfway is very, precise costly to build, takes 18 to 36 months, and costs billions of dollars, which means efficaciously anyone gathering 1 volition beryllium raising indebtedness and lone get paid erstwhile a lawsuit moves in... but determination aren't truly immoderate customers for AI information centers extracurricular of Anthropic and OpenAI, some of whom are truthful unprofitable that they've had to rise hundreds of billions of dollars even erstwhile Microsoft, Google and Amazon built each their infrastructure.

The lone crushed everybody isn't freaking retired astir this is due to the fact that AI-related stocks person done well, adjacent though nary of the hyperscalers really stock their AI revenues.

You've argued that AI's request communicative is fundamentally a mirage – that astir of the information halfway capableness is being absorbed by OpenAI and Anthropic themselves, which is masking the lack of existent endeavor demand. If that's right, who bash you deliberation volition really carnivore the outgo erstwhile the full happening unravels?

Honestly, it's going to beryllium a batch of backstage recognition funds, due to the fact that they're the ones backing the information centers, and they're funded by pension funds similar the SF teachers money oregon CalPERS, which makes maine really, truly disquieted astir the systemic contagion.

People volition reason that this means there's going to beryllium a bailout, but this isn't truly a bailoutable thing. These information centers are funded by task financing, which means that the wealth is fundamentally gone and the lone mode to "make them whole" would beryllium to either bargain retired the indebtedness oregon provender them revenues. While you could theoretically bail retired these peculiar intent vehicles (SPVs), doing truthful would beryllium to the tune of hundreds of billions of dollars and beryllium governmental cancer.

I besides fundamentally judge that Oracle gets killed by OpenAI. Its revenues person been stagnating for 20 years, and the lone mode it's kept its caput supra h2o is $85bn+ successful acquisitions, and adjacent then, that's conscionable kept things flat. Its bets connected AI information centers – $340bn+ with hundreds of billions successful indebtedness – necessitate OpenAI to go the largest, most-profitable institution successful the satellite by 2030, oregon Oracle runs retired of money. Good luck connected that 1 Larry.

The existent contagion from the illness volition beryllium widespread. The TWSE is heavy reliant connected Taiwanese ODMs similar Quanta and Hon Hai (Foxconn) who person seen their revenues boosted by selling AI servers. While I don't deliberation it kills these companies (especially arsenic Hon Hai makes truthful overmuch from Apple), it volition chopped their banal price, which volition successful crook chopped the livelihoods of those investing. The aforesaid goes for Korean investors connected the KOSPI, and yet American investors successful the assorted hyperscalers and semiconductor companies that are moving wholly due to the fact that of the AI bubble.

It each truly sucks.


Hyperscalers gathering retired information centers person hoovered up truthful overmuch of the world's representation proviso that DRAM prices person astir doubled this year. Tim Cook has called Apple's caller terms increases "unavoidable" – Macs and iPads are already up, and iPhone models are expected to travel soon. It seems the clearest motion yet that the bubble's costs are being passed onto mean consumers. If the buildout is arsenic speculative arsenic you say, are radical fundamentally paying much for hardware to subsidise information centers that whitethorn ne'er crook a profit?

I don't deliberation immoderate of them crook a profit, no.

Hyperscalers person spent implicit $1 trillion successful capex since 2022. If we presume that adjacent fractional of that is AI information centers, they request to marque implicit $1.5 trillion successful marque spanking caller profit, not revenue, but existent profit, to warrant immoderate of these expenditures.

And yes, we are each paying much for worldly for efficaciously nary crushed different than that everybody successful large tech has gone insane and wants to physique arsenic galore information centers arsenic possible. I wrote astir this successful the Hater's Guide To The Memory Crisis.

The hyperscalers are collectively spending northbound of $650 cardinal connected AI infrastructure this year, whereas Apple is spending astir $14 billion. It's arsenic if Cupertino is treating AI arsenic a commodity by paying Google astir a cardinal a twelvemonth for Gemini to tally Siri and doing what it tin on-device. The institution was antecedently said to beryllium dangerously down connected AI. That nary longer seems to beryllium the case, but astatine the aforesaid time, is it truly much savvy to beryllium renting a dependency connected the precise companies you accidental can't survive?

So, I deliberation Apple Intelligence was some the worst and champion happening to hap to Apple arsenic acold arsenic the AI bubble goes.

It was a mass-radicalization of its users against AI – an effort to cram a barely-functional bid of add-ons that cipher asked for successful the most-obtrusive mode possible, with summaries that were astir instantly turned into memes and a caller Siri that was, somehow, adjacent worse than the aged Siri.

And I deliberation that told Apple to pump the brakes. It's hardly spent thing connected capex. It's hardly done thing with AI. Despite header aft header claiming it's "falling behind," cipher tin truly explicate what it is it's falling down connected oregon wherefore it matters. People hatred Apple Intelligence, and I deliberation Apple knows that, and truthful they're going to jingle the keys for the markets by putting "AI" connected worldly without ever truly putting their backmost into it.

If the bubble deflates the mode you expect (write-downs, excessively overmuch compute supply, perchance OpenAI cratering), tin you locomotion america done what that would really look similar for Apple? Does thing interruption for iPhone users, oregon does Apple mostly ticker it hap from the sidelines? Could it adjacent payment Apple?

I deliberation things would look overmuch the aforesaid for Apple. I deliberation they volition beryllium connected the sidelines and ticker everything burn. I could spot them doing immoderate prime acquisitions arsenic things statesman to collapse, but I could besides spot them bash nothing.

I deliberation Apple is successful a precise weird spot astatine the moment. The Vision Pro was a dud, but it was besides the astir absorbing and future-forward happening I've seen anyone enactment retired successful a while. If they were smart, they'd tread h2o and descend arsenic overmuch wealth into making that arsenic tiny arsenic humanely imaginable – nary substance however agelong it takes – due to the fact that the full AI bubble is simply a effect of everybody moving retired of hypergrowth ideas, mostly due to the fact that we're level retired of caller interfaces.

I privation to beryllium wide that what they privation to bash with the Vision Pro requires it to fundamentally beryllium weightless and invisible and ne'er request adjustments. When it works, it's genuinely awesome. But that's a load-bearing when. One flimsy question means the full happening goes retired of focus. I can't adjacent usage excavation anymore due to the fact that it needs an update that requires you to deterioration the happening the full time. So overmuch promise, released excessively early, shoved retired the doorway by a CEO connected his mode out.

This article, "Ed Zitron: Apple Will 'Watch Everything Burn' When the AI Bubble Bursts" archetypal appeared connected MacRumors.com

Discuss this article successful our forums