Apple looks better and better as every day passes. Other players going deeply into debt to build out massive infrastructure, VCs pumping up the model ecosystem that is now a total commodity.
Apple’s just been sitting there with solid cashflow waiting for all this to implode and then have on-device models with their own chips.
Apple will own the end device while the rest of the world is fighting over a pure commodity. It will burn hard and Apple will laugh all the way to the bank.
Isn’t Apple potentially under a lot of pressure due to the DRAM prices as well as having to compete with AI for fab capacity? Before all the price hikes their products were already priced at the top of the market so I would assume that memory getting more expensive will impact their sales or margin quite a bit as they can’t subsidize hardware like AI companies do (they can of course eat into their own margins a bit which I think they already do). So not sure if what happens right now is great for them, if the memory market collapses it will of course be good.
> Isn’t Apple potentially under a lot of pressure due to the DRAM prices as well as having to compete with AI for fab capacity?
I'd qualify this with for now. It's not entirely clear what happens in the event the AI industry implodes, but I do think there will be a glut of memory and hardware up for grabs. It very well could crash the consumer hardware market as a result. The AI obsession and investment has burrowed so deeply across the US economy that I can't think of a single industry that won't be impacted, including agriculture.
Yes it does impact all hardware makers and they raised prices as a result. But again remember the financial reality here… DRAM is in short supply because of the demand for models propped up by VC subsidies and infrastructure providers going deeply into debt to buy all this stuff just as folks are realizing we’re on track to over build capacity. So yes Apple gets impacted a bit in the short term but in the long term they’ll gleefully watch it all burn.
No Apple was not sitting there and waiting as if its some grand plan. They tried their VR device which they don't even talk about now. They were late to AI and yes if the bubble bursts they will be at a better position but it's not some grand master strategy. Even inside apple there's been a lot of talk about how badly the company handled AI.
As for laughing all the way to the bank, I am sure memory manufacturers and Nvidia was definitely join them regardless of what happens.
Also comparing Apple (literally one of the largest companies in the world) to a company burning VC money makes no sense.
As for the rest, they’re the only major tech company to not have a serious case of AI FOMO and dumping all their cash into that FOMO. Thats not an accident.
I think it is pretty clear at this point that Apple has gone all-in on being the ideal edge silicon for AI. This leverages their core competencies, requires only modest investment, and will likely pay out no matter how the AI market eventually shakes out. They are in one of the only parts of the obvious future AI market where there isn't really a fight for greenfield turf with other big companies.
Staying in their lane is arguably the optimal business decision for Apple and they lose nothing by it.
> He also described a shift toward running AI locally rather than in the cloud – a move motivated by privacy, security, and the rising cost of inference as agents consume more tokens. However, Brooks envisions a hybrid future in which agents decide what runs on-device and what gets sent to the cloud.
And coupled with their efforts in auditably-private cloud computing, that's a strong pitch.
I generally agree. Xserve didn't work out for Apple and they dropped it pretty quickly--think I still have an A/UX coffee cup. Wasn't really in their DNA even if there was an argument before Linux took off. For better or worse, they're basically a client company and barely a software (OK and basically a software services) one--although they spend a fair bit on (roughly) user experience.
They've been investing in NPUs for a long long time. Since the first iPad Pro at least (2018). Apple Silicon is a natural fit for tensor/matrix operations due to unified memory, so if anything them going "all in" on that has been a nice side effect.
I don't know if Ed Zitron is right about allhis analysis, but it's nice to have an alternative, well-argued narrative to the gushing torrent of AI company propaganda.
Read a little more of what he's written before you praise him, so you don't end up looking as dumb as I did.
Ed may be right about some of his claims, but he is 100% a crank, and he makes so many incoherent claims I'd say that if he's right, it's in the nature of a broken clock.
Let me guess, you disagree with his views on the utility of Ai (which he's pretty bad about) and you use that as cover and/or to negate his real journalism that is based in facts (Ai being an unprecedented & unrecoupable bubble littered with unprofitable companies thats likely to crash our economy unless these Ai conglomerates cozy up lawmakers to situate themselves for a bail out)
Okay, but you have to back your insults up with something substantive or else you're just another Kool-aid drinker. Show us on the doll where Zitron touched you.
I find it weird/funny you are getting so many hate responses towards Ed, still no counter-arguments at all. When asked why he is wrong, they vanish. From the outside it looks like a coordinated campaign to try to discredit him.
Happy to read why is he wrong, and change my mind, as long as the arguments provide the same level of analysis he provides.
At this point I have yet to hear a solid argument against his arguments that use half as much evidence as he does.
That isn’t to say that he’s right, that’s just to say that while he gets creative in his phrasing numbers don’t lie and I haven’t seen anyone else present competing numbers that make sense.
If anyone has them to the degree with which he provides them then please, by all means, I’m interested.
Ya, there are already three pure ad hominem comments here with no examples. Zitron isn't even anti-AI per se. He clearly talks out his ass at times (like AI doesn't work for developers... which he's not saying so much anymore, I don't think) but his main talking points are around numbers and profitability.
As a C++ developer, AI does not work well for software development in my experience. Maybe LLM developer tools work better if you write React or HTML/CSS, but they have been inadequate for my use case since their arrival.
Casey Muratori, who doesn't use AI, made this point on a podcast. Specifically he said that he doesn't blame web developers for wanting to use AI because it's such a mess of an ecosystem and impossible to keep everything you need to know to be successful in your head.
I'm a web developer to myself who has done zero game or systems programming so I can't speak to that side of it, but you reminded me of it!
I find it bad for generating code but alright for review, I put it as a tool to run along static analysis to catch if I missed something on a code change or merge request (being specifically about C++)
I don’t have numbers, but I would like to point out that he's missing a crucial number as well. The fact that maxing out a subscription's limit makes a user generate an amount of tokens that costs much more than the subscription itslef isn't enough to back his claim. What we need to know, is the average use of subscriptions. Think about Gyms. My gut feeling is that, if we run the numbers, we see that by maxing out a gym subscription we can get more value out of it than what we paid for. And still, gyms are profitable because the average customer is using less than what they're paying for. Now, I still think AI companies will have to rise their prices (they are already doing that), but the situation could be less dramatic than what people pointing at those figures based on subscription maxing are claiming.
You do really need to see the useage numbers for this. Movie pass famously thought they could price movie tickets like gym memberships - I assume they looked at how often the average person watched movies in theaters at the start - but then found out having a subscription to see movies for free / cheap any time you want changed that rate substantially and they ran out of vc money.
Having a subscription might make people run agents overnight, or casually generate images for fun, or start more coding projects. The reason gym memberships work is that people like the idea of going to the gym a lot but don't really enjoy actually doing it very much and fall off over time, leaving the gym for the rare person who's really into fitness to use it cheaply.
is AI like that where most users will get bored? or is it like movies where they'll try to get their money's worth?
I gotta say that if their best bet is “people forget to use it” I don’t think that that’s going to drive the adoption they need.
Especially if part two of your theory comes to fruition and they have to raise prices. I might of forgotten my Netflix subscription when it was 10 dollars a month but certainly not if it was 200 you know?
You're biasing for quantity of argument vs. quality of argument. That's Zitron's MO: writing 10k word rambles that cover 20 topics, ostensibly with tons of links to sourcing, but with most of the sources not actually saying what Zitron implies they do.
Nobody has the patience to read a comprehensive point by point rebuttal to any of it, it's just too tedious.
You say "numbers don't lie". But they do, when the numbers being presented have been adversarially chosen. You just find numbers (no matter how low quality) that fit the chosen narrative, you throw out the numbers (no matter how high quality) that rebut it. If you're trying to predict the future, you can't afford that of bias. If you're an anti-AI grifter, you can't afford to not have that bias, your entire livelihood depends on only showing things that support the narrative.
So I’ve heard this argument before but I’ve not even seen someone demonstrate HOW the numbers are “adversarial”.
This is the crux of my argument: Zitron provides numbers and he provides a lot of them sourced from reliable and trustworthy sources. He provides counterarguments to numbers-sparse arguments by using even more actual, real numbers with a logical theory formed from them.
This should make it rather easy to dispute his claims, no? So then where are the equally rational disputes?
"The biggest thing we’ve learned from the large language model generation is how many people are excited to replace human beings, and how many people just don’t understand labour of any kind"
Really? I think we've heard C-suite blabbermouths and borderline nontechnical tech CEOs being amplified by social media making these assertions, but boy did the weakly efficient market deliver a relatively swift correction to that mindset, no?
Q: "What would it take to change your point of view?"
A: "[AI] would have to solve all hallucinations forever, which they are completely incapable [of]."
Can you produce a human that is infallible? I'll wait.
My hot take is AI is increasingly less unprofitable as the cost of serving tokens drops and Nvidia's ongoing offers to guarantee profitability is a sign that it isn't stopping anytime soon.
You either believe in the underlying science and technology or you don't. But in a world where AI is a fad like cabbage patch kids and beanie babies, what's next?
I thought Ed Zitrons argument was just corroborated by Nikkei last Friday with the new report that 1.6 Trillion in SPVs accounts for off-books debt, on top of the $1.4 Trillion of on-books debt that's in these AI companies?
The main argument of Ed Zitron is just that. Huge huge amounts of debt that's due "soonish". Every SPV is different but we are fast approaching the point where this all starts to become real.
Now I dunno why Ed Zitron feels the need to write 10,000++++ word articles that only says "more leverage than people expect", but that's really the crux of things.
“Per my own analysis, NVIDIA’s predicted $1 trillion in Blackwell and Vera Rubin GPU sales (by the end of 2027) represents around 40GW of data center capacity, which will, assuming a PUE of 1.35, result in around 30GW of usable capacity. At a cost of around $12 million a megawatt, that works out to around *$435 billion in global annual compute revenue to make these data centers necessary.*“
> He's a college drop out with no experience in programming or technology.
I mean Zuckerberg is a drop out.
zitron has been around since before flash was dead, more over hes worked in PR so like anyone with heavy involvement with c-suite, sales and journalism, you get a good nose for bullshit.
That said, do I agree with his assertions? no not entirely. But. I suspect its a bit of an audience capture thing. He makes wild statements, then goes on a flight of fancy to back them up. I appreciate that, but it also lets me decide that he's going down a path that I feel is not backed up by his data.
This but instead it's people insisting that AI will replace my job (been three months out for at least 3 years now) because they know it involves words.
To be fair he has a background in technology PR and Marketing (which he’s been very critical of) which I would count as more worthwhile when assessing the actual value of the product.
I mean you can say he’s not a “science guy” but he’s undeniably “smart”.
A resolute anti-AI perspective is a breath of fresh air compared to the middle-of-the-road "AI made me 10x more productive BUT"-style comments.
My professional opinion is that LLM technology doesn't work very well for software development and I'm not interested in hemming and hawing over its supposed benefits any longer.
Talking with a friend of mine who's been in development for some time and I had been talking a little bit about the skill required to wield AI effectively and that people were expecting it to be something that required less skill than it does.
He said his version of this observation is "your agents are only as good as you are." I think he's right, and the key is to practice and build the skill.
As I mentioned in this thread the other day, it's not worth debating AI skeptics anymore in 2026, with step changes like Fable and Sol. Maybe it was the case in 2025 and before but the field has improved rapidly that it doesn't make sense to listen to people with AI experiences prior to that who haven't touched it since. Does AI make mistakes, sure, but increasingly fewer.
I don’t think people like Linus Torvalds or Stephen Wolfram feel like language models beat them, just they’ve managed to find specific cases where flawed brainless workhorses serve their needs.
I'm generally pretty accepting of trust me bro style arguments where we some claims presented without evidence. Especially for AI, sure some model just rm -rf ed your home directy bro sounds bad. It works fine for online discourse.
This is less agreeable. You kind of just made statements without even having claims (even without requirement of evidence or details) to back them up.
Some developers work better with LLMs in their workflows than others. Some problems are easier for LLMs to generate a reasonable solution for than others. Some folks prompt minimally and see what the vibes bring. Others start with a detailed architecture and implementation plan.
The individual results will depend quite a lot on the person, the problem, and the approach. It's not a guaranteed winning formula for everyone.
The vast majority of what I see out there is relentlessly negative about AI. It is kind of absurd to call this the "alternative" unless the implication is that the rest of AI skeptical writing is not "well-argued."
While I don't really care one way or another, Zitron isn't arguing AI as a technology is going to fail and go away, but that the numbers of the American companies make no sense and there is a good chance they, or at least some of them, will go away. He does argue that there are large amounts of people who just don't care enough about "AI," and this is likely true. The thing to keep in mind though is that he is not talking about software developers here. What is evident about some of the comments in this thread (and HN comments in general) is that many people on this site seems to think that whenever anyone says anything along the lines of "AI doesn't work for me" that they automatically assume they are talking about "in software development" when they aren't.
the pro-AI side on the other hand has poured billions into ads and marketing and CEOs are forcing it on people due to a combo of FOMO, personal investments in AI (CEO, board, investor), etc
I think AI discourse is revealing some new contours to the information bubbles that we all live in.
folks in the US have been pretty aware that bubbles based on political opinions -- we're all surrounded by news from "our side" yet we know the other side exists in some other bubble. but for AI, we're all either surrounded by pro- or anti- opinions and its a little shocking to learn that theres a parallel internet with the opposite stance. especially shocking when you find somebody that lives in teh same political bubble but opposite AI bubble.
Both can be true. AI is very, very polarizing. The entire Internet is also filling up with AI generated content, some slop, some okay. There is also a very pro-AI managerial class that wants to replace everyone and everything with a Claude project so they can "just ask Claude!"
One big difference I have with Ed Zitron is I look at companies suddenly getting worried that they're spending millions of dollars on tokens and think "wow those AI vendors are going to make SO MUCH MONEY".
The best price for a product is what I call the "suck air through your teeth" price. You want your customers to suck air through their teeth... and then pay the full amount anyway.
Uber set their per-developer token allowance to $1500 per developer per tool. That suggests to me that they think they can get at least that much ROI out of AI tooling.
Selling $1500/employee/month plans to companies is a great business to be in.
Well hold on. "They" also thought tokenmaxxing was a good idea, until they didn't.
Whether companies pay will come down to the bottom line once the hype around the country club dies down a bit and for that it's still TBD on the actual product cost impact.
Not sure about that. As I replied elsewhere, it's a forcing function:
> People misunderstand the point of the tokenmaxxing time period, it was to force people to use AI so as to not have them stuck in their way, as some people are, and then to evaluate how it can help the company.
that $1.5k is ostensibly not coming out of productivity gains from the AI though.
it's coming out of the salary budget from what i've seen. i've seen a company both say "AI max. it's the future! don't be late" and then go on to increase limits to compensation across the board.
Very true and I don't think you can completely discount the end consumer market either. A large number of the people not currently paying for a subscription are becoming dependent on AI enabled search results. I imagine monetization of those users is in its infancy and will rival the search result ad business eventually
I totally think the AI companies are very overvalued, I feel that's obvious. But I don't understand why apple should have such a high run in the last year. Sure, it avoided the pitfall of overinvesting in AI -- but what did it do in for positive development? Avoiding waste isn't (imo) enough to warrant increased price. What's going on, what's the bull case for apple?
> If Anthropic and OpenAI believed customers would actually pay the real cost of AI tokens, they wouldn't have to give away 20 to 40 times the amount of tokens to subscribers.
One logical gap in the SemiAnalysis 40x cost of tokens versus subscription: I don't know anybody actually maxing out their account limits.
Sure, if you're somehow always running stuff, you can max it out, but subscriptions like this allow people to max sometimes (or always), while others never come close to the max.
What is the average usage of subscribers? Only Anthropic and OpenAI know, as far as I can tell.
I'm of the opinion that the market will need to expand, horizontally. That's what's happening, with low-cost LLMs. I already know lots of "regular folks" that are totally hooked on LLMs (mostly ChatGPT, at the "regular mensch" level). They usually use the free tier, but a lot of them are doing the $20/month Pro tier.
If anyone is in my age group, they remember when ATMs were free. It was the "heroin dealer" business model. It worked out well (for the banks). The Chinese did it with manufacturing.
Getting people hooked on subsidized junk, is an age-old (and highly effective) business model.
Think of all the shops that will soon be composed of people that simply can't even get out of bed, if their LLM is not available. If the LLM dealer starts raising the price, there's no choice. I suspect many of the valuations are taking this into account.
> "Junk is the ideal product... the ultimate merchandise. No sales talk necessary. The client will crawl through a sewer and beg to buy."
> Think of all the shops that will soon be composed of people that simply can't even get out of bed, if their LLM is not available.
I'm biased because I don't use AI, but I don't understand why you wouldn't fire these LLM-addicted people and replace them with people who can code or write an email without the use of an AI assistant.
I use LLMs all the time, but I'm also a highly capable and experienced engineer that can definitely work without. I have just found that the LLM is a force multiplier, like I haven't had before.
Just to counterpoint the adjacent replier. I have claude, chatgpt, and gemini subs and rarely even use even close to 50%. I code c++ everyday. But im not full in on agentic workflows which is probably why. I mostly do fairly targeted things to my various code bases.
This is quite a short sighted analysis. I do think the valuations are quite and they would need to meet the reality, but don’t think there’s gonna be a crash or we’d ever go back to pre-AI era. It would more or less would be a correction to valuations.
The future of AI would be on-device models which are as powerful as current frontier models and also I can imagine companies have their own deployments of inference of open weighted models for most of the use cases and use the frontier models for extremely niche or higher intelligence tasks.
As an example I use Claude code heavily for every day development and Opus 4.8 was already good enough for my use cases and never used Fable. Also note that I use AI as a tool to help with my work and I do not offload everything I have to do to AI in a single prompt
> I do think the valuations are quite and they would need to meet the reality, but don’t think there’s gonna be a crash or we’d ever go back to pre-AI era. It would more or less would be a correction to valuations.
Something to consider: would your description also apply to the dot-com boom of the late 1990s? The internet was real, the ideas for internet business were real, and we were not going to the previous reality. But the valuations weren't quite right and a "correction" happened at some point.
When people talk about AI crash, that's what they mean. Not that AI is a hoax, but that the correction could be quite violent and have effects on the broader economy.
> but don’t think there’s gonna be a crash […]. It would more or less would be a correction to valuations. [...] The future of AI would be on-device models which are as powerful as current frontier models […]
That’s the crash… that’s pretty much exactly Ed Zitron’s thesis
> As an example I use Claude code heavily for every day development and Opus 4.8 was already good enough for my use cases and never used Fable.
I think the flaw in this logic is thinking about how AI is currently used only. Yes, Opus is good enough for the task you are asking it to do, but that doesn't mean that is all you will ever need.
As AI gets better and better, it will open up new use cases that require the better performance.
You don’t “need” any LLMs to do software development. Plenty of us were productively writing code for years before these models were released, and many of us continue to do so. The “AI” companies want to sell you hype. But you don’t “need” them - not if you’re a legit, professional programmer.
The narrative that LLMs are essential to the future of the trade often feels like an assault on my professional expertise.
Whatever your experience is, mine is that LLMs don't work well for software development. I wish people who use AI would be more willing take that perspective seriously.
The money being poured into AI infrastructure means there is a market for new ways of doing things that take 1/1000 of the power or are 1000x faster, or both.
And there are many such moonshot startups.
AI on GPUs is an efficient as gaming on CPUs.
All that math where perfect precision is not required means that you can’t tell do things in different ways.
Why is on-device AI the future? What is your reasoning behind this? Look at the proportion of things we compute on someone else’s computer relative to what we compute on our own device. Why would this change for LLMs?
I think the reasoning is similar to why Amazon notification emails are very generic: user data is used for various purposes that a growing number of users and 3rd parties have concerns about.
Back of a very tiny envelope: suppose there are 100 million jobs that benefit from spending $10k/year in tokens. That's a trillion in annual revenue; at a 30% margin that's $300 billion in profits, which can easily support $6 trillion in valuation.
Thats great but this is like when 3g came out. Sure it was the future, but it was half baked, impersonal, expensive, unreliable and required a culture shift to be adopted.
Onboard decent LLM performance thats _power efficient_ is at least two/three hardware generations away. (assuming linear performance)
but, the valuations, with debt trade and private credit obscuring exposure is a recipe for disaster.
His read on the Apple Vision Pro (both the circumstances of its creation and its path forward) are incorrect in quite the discouraging manner for me. AVP wasn't released early (late, if anything), and it wasn't a dud because of Tim Cook's disinterest.
I also think it's irresponsible to not broach the obvious implication of "PC components becoming prohibitively expensive" + "untold amounts of compute sitting in compute warehouses with nothing to do because the AI companies that used to own them folded". You probably won't even notice when everything in Best Buy becomes a thin client.
> AVP wasn't released early (late, if anything), and it wasn't a dud because of Tim Cook's disinterest.
Depends on what you mean by early.
WAs it released before the market was proven? yes.
Was the whole user experience up to scratch before it was released? no.
As someone who worked on the Quest pro/3 ecosystem, We were shitting our selves, because we knew that apple would only release something when the user experience was right.
Oculus just shat out features as and when performance cycles needed juicing.
Sure the UX was good, but it was less good than I expected. The headset was bulky and surprisingly off balance. It wasn't an all day wearable (i mean the quest pro was actually comfortable by comparison)
The resolution wasn't that great and the only thing you could do with it was either look like a twat recording video or have spreadsheets pasted everywhere. I was expecting an OS where we could put things on work tops, create augmentations to my room, decorate or be creative. Instead we got OSX XR.
Tim Cook was probably mostly disinterested because it didn't seem like a real business opportunity. Wherever AVP fell on the VR/AR spectrum, VR (while having other opportunities) was mostly a niche for high-end gamers and AR, whatever the possibilities in some future idealized form with fashionable glasses that give you lots of information about what/who you're looking at is a long way from reality.
I think it's the opposite. None of the players in XR wanted to be Apple'd, least of all Apple. That is, create a base product that someone else would refine in order to drink their milkshake. Apple delayed the AVP multiple times hoping to avoid getting sniped by Meta and Google et al., each time requiring a partial rebuild as the underlying tech advanced. They would have loved to have been able to get something out when Robert Scoble was initially raving in the late 2010s, but everyone was on the same page about not wanting to sink billions into development only to have to cede the market to someone responding to your strategy. So, instead, they all dragged their feet.
If XR companies had been as "reckless" with their product strategies as mobile companies in the 90s and 2000s, AVP would have had its iPod moment 6 years ago and we'd be hurtling towards the facephone. The market suffered because of the entrenchment of incumbents with all-too-clear memories.
There is a difference between "AI is overvalued" and "AI isn't valuable". We've seen entire industries deliver real tech progress while still going through harsh valuation resets.
It's the same decision-making process that tried to do a electric car for years, and canceled it after spending time and money. In that exact timeline, China made great electic cars and new brands, Tesla advanced and Spacex launched dozens or more rockets. Apple is so far from any big new tech, they and market don't even realize it.
Apple had more money and influence than combination of all of them, yet failed with car and VR. Why would anyone think they can't fail much much bigger with AI and LLMs ?
For me, the economics are the exact opposite. I'd pay $1000 for a good gaming VR headset, but not $4000 -- because gaming is just a hobby that isn't really worth that much to me. Gaming VR headsets have been around for well over a decade, I've "been there, done that", and it's just not something that's worth that much money to me. However, I have no problem paying $3500+ for something that creates a new category of productivity and entertainment experiences for me.
Besides that, I actually disagree that the Vision Pro is not good for gaming. It's not good for traditional VR gaming, but using apps like Portal it's phenomenal for, e.g., playing existing PS5 games on a massive virtual screen, which in many ways I actually enjoy more than VR gaming, which is far too limited in comparison.
Maybe a dumb question, but isn't the idea that the efficiency of the models will improve with time, such that you won't be burning hundreds of dollars of tokens for most queries?
Fable is great, but Opus can handle most coding tasks for a fraction of the cost, and Sonnet is good enough for average questions or word processing tasks.
Interestingly, that METR study has updated data for 2026 that shows a speed up, although they admit that the data may not be reliable because of changed pay rate for participation, but this quote is telling:
> The primary reason is that we have observed a significant increase in developers choosing not to participate in the study because they do not wish to work without AI, which likely biases downwards our estimate of AI-assisted speedup.
So compared to just last year, they had a hard time finding participants because too many didn't want to work without AI.
Welcome to Ed Zitron. There is a reason this man doesn't heavily short the same companies he criticizes. Be wary of anyone that won't put their money where their mouth is.
> There is a reason this man doesn't heavily short the same companies he criticizes
yes, because in order to take a short position you have to predict exactly when the bubble is going to pop, which is different from predicting that at some point it will
Not quite. You could buy a long-dated put if implied vol and current rates are amenable. This protects you against the underlying going up significantly. You might also predict that it won’t go up too much, and just short the underlying with a stop loss.
Unless Apple figures out AI software execution and opens up their walled garden a bit, their hardware edge won't matter.
In 2020 I could write audio apps that worked incredibly well on a Pixel: despite LLMs! I still can get Siri to put tasks in the right todo app 70% of the time.
I updated to iOS 27 beta hoping something had changed, but nope: the biggest change is a massive black orb search bar.
That fabled line one must cross to achieve profit is ridiculously out of alignment with the current valuations on AI. Its why SPCX is falling like a meteor, currently at $111 per share.
The people who are going to end up making the most money on this are creditors and future businesses. When the AI bubble does pop there will be a massive glut of data centers and hardware available. Both Apple and Microsoft are realigning their entire businesses to brace for this. When the AI bubble pops businesses that sell hardware, like Apple and Microsoft, will face an immediate price shock because they have had to raise prices to account for more expensive hardware. That shock will be short lived and prices will fall accordingly with disruption to supply chain but otherwise minimal disruption to margins.
I look forward to the bubble popping because when retail hardware becomes cheap again all kinds of new business opportunities will open in the self-hosted service market.
I've never heard a compelling thesis for why this is a bubble that will pop. From my perspective AI adoption is just getting started, and every organization I see that starts on the journey only ramps up tokens dramatically. Even just sticking with current model capabilities we have years of business automation, personal automation, and so much more to saturate data centers. As the models improve that only makes more sense. Rather than Apple watching anything burn I think they are going to end up looking like nitwits who missed the boat.
You’ll note that Mr. Zitron’s analysis is based on unfounded assertions (API prices are the ‘real costs’ of tokens, and model providers are margin negative on subscriptions) and outdated figures (OpenAI negative profit in 2025, ignoring at least Anthropic’s recent turn to profitability.)
Just more wishful thinking from our favorite AI skeptic
IBM stood on the sidelines of the dot com boom; their share price still halved in the resulting bust.
HSBC didn't engage in the unwise practices that led to the great financial crisis; their share price still dropped by 75% in 2008.
If the AI bubble does burst chaotically, then I'd expect all tech stocks to decline to at least some extent and for even the strongest survivors to remain in the doldrums for years (in the cases listed above, the share price of both IBM and HSBC remained flat for almost a decade).
> ..because the entire AI bubble is a result of everybody running out of hypergrowth ideas, mostly because we're flat out of new interfaces.
Really? I thought the AI bubble happened because something genuinely novel had been invented. People immediately saw practical uses for it, and then it took on a life of its own - funding, superfunding, $trillions becoming part of day to day lingo.. etc.
AI is not in the same class as Vision Pro, which Ed Zitron is comparing it to. I personally found the overall tone to be a bit hyperbolic.
This bubble too will pop, and life will go on. Costs will slowly drop down, just like the dot com days, but some useful things will be come out of this.
>At their very core, Large Language Models' costs run contrary to basically every model of selling software.
>Consumers and enterprises alike have been trained to pay a monthly fee for a service, and while these services might have limits or strictures, basically nobody buying software expects to have a metered service, let alone one that's both metered and with hard to measure costs.
Has Ed Zitron not heard about the cloud? Unpredictable AWS bills?
Unpredictable AWS bills are a supply-side problem. He's talking about consumer behavior, and consumers of software do not expect a metered service.
But most consumer's aren't paying per token for access to models, so unless that changes and the labs start charging API pricing to everyone, it's kind of a moot point.
Enterprise certainly wouldn't fit. This argument seems to be targeted at the consumer market, which Apple deals in. Consumers are not paying varying Disney+ subscription prices based on how much they watch. So many points are still interesting. Though, I'm not convinced the consumer market is the primary target for AI firms.
I'm so tired of this shallow analysis claiming vendors are losing tons of money on subscriptions. How do you even judge that?
Users go on vacation, they slack off, they spend the day talking to each other. There are very few people who are really effective at burning tokens. how do you know the ratio? do you have insides? No :)
The biggest target is enterprise, and the economics for an LLM vendor look like this: price per token = R&D + inference + infra investments. When you buy a subscription, you are quite often buying a year ahead. That lets the vendor predict future infra investments against hard commitments, and sell expensive per token pricing to everyone else. And when a hard commitment sits unused because the user is busy, they sell it twice. It is loyalty in exchange for predictability, in exchange for the promise to always deliver SOTA to users.
Vendors control the harness. Tomorrow they simply roll out a router where reading the code and doing the final edits goes to a cheaper model, and their math suddenly becomes very sexy.
Isn't that hard to predict that their economic model is very easy to tune? and this is just first baby steps.
I personally pay per token ( do not have subs for work ). I did have once a $25k/mo worth of tokens, since i knew it was free so i was doing crazy experiments. Now , 2 month later, my bill was barely $1.5k since i moved into different stage with project. I do have team members who burn $500-600. pre router, pre optimization.
I switched recently to grok 4.5 and cursor router and my bill will go even further down. It rotates 4-5 different vendor models cheap and expensive too, depends on the task. Routers will flip entire LLM economy upside down.
It's fascinating that you simultaneously argue that margins will expand... while posting about personal behavior all of which points toward commoditization and intense price competition.
I think Ed's argument is that given the infra investment the revenue won't be enough to pay back the original investors. In fact cheaper models make this worse for them. Who cares about the investors? Well it turns out the infra was financed in large part by debt that was securitized and a bunch of regarded investors that bought the debt looking for higher returns will be taking a huge haircut when the bubble pops. The amount of securitized debt is roughly the same as the mortgage backed securities back on 2007, ergo the prediction is a big recession.
What's unclear to me is if this is a systemic issue that's going to cause credit to freeze up but imo the opacity of the shadow banking "system" does not help here. If you see one cockroach, etc.
Apple’s just been sitting there with solid cashflow waiting for all this to implode and then have on-device models with their own chips.
Apple will own the end device while the rest of the world is fighting over a pure commodity. It will burn hard and Apple will laugh all the way to the bank.
I'd qualify this with for now. It's not entirely clear what happens in the event the AI industry implodes, but I do think there will be a glut of memory and hardware up for grabs. It very well could crash the consumer hardware market as a result. The AI obsession and investment has burrowed so deeply across the US economy that I can't think of a single industry that won't be impacted, including agriculture.
As for laughing all the way to the bank, I am sure memory manufacturers and Nvidia was definitely join them regardless of what happens.
Also comparing Apple (literally one of the largest companies in the world) to a company burning VC money makes no sense.
As for the rest, they’re the only major tech company to not have a serious case of AI FOMO and dumping all their cash into that FOMO. Thats not an accident.
I think it is pretty clear at this point that Apple has gone all-in on being the ideal edge silicon for AI. This leverages their core competencies, requires only modest investment, and will likely pay out no matter how the AI market eventually shakes out. They are in one of the only parts of the obvious future AI market where there isn't really a fight for greenfield turf with other big companies.
Staying in their lane is arguably the optimal business decision for Apple and they lose nothing by it.
> He also described a shift toward running AI locally rather than in the cloud – a move motivated by privacy, security, and the rising cost of inference as agents consume more tokens. However, Brooks envisions a hybrid future in which agents decide what runs on-device and what gets sent to the cloud.
And coupled with their efforts in auditably-private cloud computing, that's a strong pitch.
Ed may be right about some of his claims, but he is 100% a crank, and he makes so many incoherent claims I'd say that if he's right, it's in the nature of a broken clock.
Comment read sensibly until this bit.
Happy to read why is he wrong, and change my mind, as long as the arguments provide the same level of analysis he provides.
That isn’t to say that he’s right, that’s just to say that while he gets creative in his phrasing numbers don’t lie and I haven’t seen anyone else present competing numbers that make sense.
If anyone has them to the degree with which he provides them then please, by all means, I’m interested.
I'm a web developer to myself who has done zero game or systems programming so I can't speak to that side of it, but you reminded me of it!
Having a subscription might make people run agents overnight, or casually generate images for fun, or start more coding projects. The reason gym memberships work is that people like the idea of going to the gym a lot but don't really enjoy actually doing it very much and fall off over time, leaving the gym for the rare person who's really into fitness to use it cheaply.
is AI like that where most users will get bored? or is it like movies where they'll try to get their money's worth?
Especially if part two of your theory comes to fruition and they have to raise prices. I might of forgotten my Netflix subscription when it was 10 dollars a month but certainly not if it was 200 you know?
Nobody has the patience to read a comprehensive point by point rebuttal to any of it, it's just too tedious.
You say "numbers don't lie". But they do, when the numbers being presented have been adversarially chosen. You just find numbers (no matter how low quality) that fit the chosen narrative, you throw out the numbers (no matter how high quality) that rebut it. If you're trying to predict the future, you can't afford that of bias. If you're an anti-AI grifter, you can't afford to not have that bias, your entire livelihood depends on only showing things that support the narrative.
This is the crux of my argument: Zitron provides numbers and he provides a lot of them sourced from reliable and trustworthy sources. He provides counterarguments to numbers-sparse arguments by using even more actual, real numbers with a logical theory formed from them.
This should make it rather easy to dispute his claims, no? So then where are the equally rational disputes?
Really? I think we've heard C-suite blabbermouths and borderline nontechnical tech CEOs being amplified by social media making these assertions, but boy did the weakly efficient market deliver a relatively swift correction to that mindset, no?
Q: "What would it take to change your point of view?" A: "[AI] would have to solve all hallucinations forever, which they are completely incapable [of]."
Can you produce a human that is infallible? I'll wait.
Finally: https://martinalderson.com/posts/no-it-doesnt-cost-anthropic...
My hot take is AI is increasingly less unprofitable as the cost of serving tokens drops and Nvidia's ongoing offers to guarantee profitability is a sign that it isn't stopping anytime soon.
https://newsletter.semianalysis.com/p/nvidia-gpu-debt-backst...
You either believe in the underlying science and technology or you don't. But in a world where AI is a fad like cabbage patch kids and beanie babies, what's next?
He has zero numbers and is completely making stuff up to a gullible audience.
The real numbers show high demand for all these services to the point where the companies are rate limiting the services because demand is TOO high.
The main argument of Ed Zitron is just that. Huge huge amounts of debt that's due "soonish". Every SPV is different but we are fast approaching the point where this all starts to become real.
Now I dunno why Ed Zitron feels the need to write 10,000++++ word articles that only says "more leverage than people expect", but that's really the crux of things.
You’re thinking of a different person.
Most recent free article: https://wheresyoured.at/the-subprime-data-center-crisis/
Number after number, e.g.:
I mean Zuckerberg is a drop out.
zitron has been around since before flash was dead, more over hes worked in PR so like anyone with heavy involvement with c-suite, sales and journalism, you get a good nose for bullshit.
That said, do I agree with his assertions? no not entirely. But. I suspect its a bit of an audience capture thing. He makes wild statements, then goes on a flight of fancy to back them up. I appreciate that, but it also lets me decide that he's going down a path that I feel is not backed up by his data.
Additionally, as another commenter has pointed out, his thoughts and criticisms are being echoed by people who actually have the money.
I mean you can say he’s not a “science guy” but he’s undeniably “smart”.
My professional opinion is that LLM technology doesn't work very well for software development and I'm not interested in hemming and hawing over its supposed benefits any longer.
He said his version of this observation is "your agents are only as good as you are." I think he's right, and the key is to practice and build the skill.
This comment is so obviously false (in my experience) I'm wondering what sort of niche environment you're working in
https://news.ycombinator.com/item?id=49051369
i have yet to have an LLM beat me. i am forced to try regularly so i don't look like an AI anti at a company that is very much SV pilled
This is less agreeable. You kind of just made statements without even having claims (even without requirement of evidence or details) to back them up.
This is like the illusion of contribution.
Some developers work better with LLMs in their workflows than others. Some problems are easier for LLMs to generate a reasonable solution for than others. Some folks prompt minimally and see what the vibes bring. Others start with a detailed architecture and implementation plan.
The individual results will depend quite a lot on the person, the problem, and the approach. It's not a guaranteed winning formula for everyone.
Implying there is a "gushing torrent" of pro AI narrative is bizarrely out of touch. We both know this isn't true.
the pro-AI side on the other hand has poured billions into ads and marketing and CEOs are forcing it on people due to a combo of FOMO, personal investments in AI (CEO, board, investor), etc
folks in the US have been pretty aware that bubbles based on political opinions -- we're all surrounded by news from "our side" yet we know the other side exists in some other bubble. but for AI, we're all either surrounded by pro- or anti- opinions and its a little shocking to learn that theres a parallel internet with the opposite stance. especially shocking when you find somebody that lives in teh same political bubble but opposite AI bubble.
The best price for a product is what I call the "suck air through your teeth" price. You want your customers to suck air through their teeth... and then pay the full amount anyway.
Uber set their per-developer token allowance to $1500 per developer per tool. That suggests to me that they think they can get at least that much ROI out of AI tooling.
Selling $1500/employee/month plans to companies is a great business to be in.
Whether companies pay will come down to the bottom line once the hype around the country club dies down a bit and for that it's still TBD on the actual product cost impact.
> People misunderstand the point of the tokenmaxxing time period, it was to force people to use AI so as to not have them stuck in their way, as some people are, and then to evaluate how it can help the company.
https://news.ycombinator.com/item?id=49047448#49047826
it's coming out of the salary budget from what i've seen. i've seen a company both say "AI max. it's the future! don't be late" and then go on to increase limits to compensation across the board.
One logical gap in the SemiAnalysis 40x cost of tokens versus subscription: I don't know anybody actually maxing out their account limits.
Sure, if you're somehow always running stuff, you can max it out, but subscriptions like this allow people to max sometimes (or always), while others never come close to the max.
What is the average usage of subscribers? Only Anthropic and OpenAI know, as far as I can tell.
If anyone is in my age group, they remember when ATMs were free. It was the "heroin dealer" business model. It worked out well (for the banks). The Chinese did it with manufacturing.
Getting people hooked on subsidized junk, is an age-old (and highly effective) business model.
Think of all the shops that will soon be composed of people that simply can't even get out of bed, if their LLM is not available. If the LLM dealer starts raising the price, there's no choice. I suspect many of the valuations are taking this into account.
> "Junk is the ideal product... the ultimate merchandise. No sales talk necessary. The client will crawl through a sewer and beg to buy."
-William Burroughs
I'm biased because I don't use AI, but I don't understand why you wouldn't fire these LLM-addicted people and replace them with people who can code or write an email without the use of an AI assistant.
I use LLMs all the time, but I'm also a highly capable and experienced engineer that can definitely work without. I have just found that the LLM is a force multiplier, like I haven't had before.
The future of AI would be on-device models which are as powerful as current frontier models and also I can imagine companies have their own deployments of inference of open weighted models for most of the use cases and use the frontier models for extremely niche or higher intelligence tasks.
As an example I use Claude code heavily for every day development and Opus 4.8 was already good enough for my use cases and never used Fable. Also note that I use AI as a tool to help with my work and I do not offload everything I have to do to AI in a single prompt
Something to consider: would your description also apply to the dot-com boom of the late 1990s? The internet was real, the ideas for internet business were real, and we were not going to the previous reality. But the valuations weren't quite right and a "correction" happened at some point.
When people talk about AI crash, that's what they mean. Not that AI is a hoax, but that the correction could be quite violent and have effects on the broader economy.
That’s the crash… that’s pretty much exactly Ed Zitron’s thesis
I think the flaw in this logic is thinking about how AI is currently used only. Yes, Opus is good enough for the task you are asking it to do, but that doesn't mean that is all you will ever need.
As AI gets better and better, it will open up new use cases that require the better performance.
The narrative that LLMs are essential to the future of the trade often feels like an assault on my professional expertise.
Whatever your experience is, mine is that LLMs don't work well for software development. I wish people who use AI would be more willing take that perspective seriously.
And there are many such moonshot startups.
AI on GPUs is an efficient as gaming on CPUs.
All that math where perfect precision is not required means that you can’t tell do things in different ways.
In addition to on-device, Apple is making efforts to secure computations that need to occur off-device, see: https://security.apple.com/documentation/private-cloud-compu...
Onboard decent LLM performance thats _power efficient_ is at least two/three hardware generations away. (assuming linear performance)
but, the valuations, with debt trade and private credit obscuring exposure is a recipe for disaster.
In this case, it’s really irresponsible.
I also think it's irresponsible to not broach the obvious implication of "PC components becoming prohibitively expensive" + "untold amounts of compute sitting in compute warehouses with nothing to do because the AI companies that used to own them folded". You probably won't even notice when everything in Best Buy becomes a thin client.
Depends on what you mean by early.
WAs it released before the market was proven? yes.
Was the whole user experience up to scratch before it was released? no.
As someone who worked on the Quest pro/3 ecosystem, We were shitting our selves, because we knew that apple would only release something when the user experience was right.
Oculus just shat out features as and when performance cycles needed juicing.
Sure the UX was good, but it was less good than I expected. The headset was bulky and surprisingly off balance. It wasn't an all day wearable (i mean the quest pro was actually comfortable by comparison)
The resolution wasn't that great and the only thing you could do with it was either look like a twat recording video or have spreadsheets pasted everywhere. I was expecting an OS where we could put things on work tops, create augmentations to my room, decorate or be creative. Instead we got OSX XR.
If XR companies had been as "reckless" with their product strategies as mobile companies in the 90s and 2000s, AVP would have had its iPod moment 6 years ago and we'd be hurtling towards the facephone. The market suffered because of the entrenchment of incumbents with all-too-clear memories.
China is doing all the R&D for free and the whole thing turned out to be unprofitable anyway.
I would love to replace my monitor with Apple Vision Pro for programming and productivity. I would gladly pay $1000 for that.
But at $4000 it really needs to put me in a Microsoft Flight Simulator cockpit.
Apple had more money and influence than combination of all of them, yet failed with car and VR. Why would anyone think they can't fail much much bigger with AI and LLMs ?
Besides that, I actually disagree that the Vision Pro is not good for gaming. It's not good for traditional VR gaming, but using apps like Portal it's phenomenal for, e.g., playing existing PS5 games on a massive virtual screen, which in many ways I actually enjoy more than VR gaming, which is far too limited in comparison.
Fable is great, but Opus can handle most coding tasks for a fraction of the cost, and Sonnet is good enough for average questions or word processing tasks.
Curious that he references a METR study from July 2025, before the leap in model and harness performance towards the end of 2025/early 2026.
> The primary reason is that we have observed a significant increase in developers choosing not to participate in the study because they do not wish to work without AI, which likely biases downwards our estimate of AI-assisted speedup.
So compared to just last year, they had a hard time finding participants because too many didn't want to work without AI.
yes, because in order to take a short position you have to predict exactly when the bubble is going to pop, which is different from predicting that at some point it will
What leveraged? yeahnah thats not going to be profitable.
Not true.
In 2020 I could write audio apps that worked incredibly well on a Pixel: despite LLMs! I still can get Siri to put tasks in the right todo app 70% of the time.
I updated to iOS 27 beta hoping something had changed, but nope: the biggest change is a massive black orb search bar.
The people who are going to end up making the most money on this are creditors and future businesses. When the AI bubble does pop there will be a massive glut of data centers and hardware available. Both Apple and Microsoft are realigning their entire businesses to brace for this. When the AI bubble pops businesses that sell hardware, like Apple and Microsoft, will face an immediate price shock because they have had to raise prices to account for more expensive hardware. That shock will be short lived and prices will fall accordingly with disruption to supply chain but otherwise minimal disruption to margins.
I look forward to the bubble popping because when retail hardware becomes cheap again all kinds of new business opportunities will open in the self-hosted service market.
Just more wishful thinking from our favorite AI skeptic
HSBC didn't engage in the unwise practices that led to the great financial crisis; their share price still dropped by 75% in 2008.
If the AI bubble does burst chaotically, then I'd expect all tech stocks to decline to at least some extent and for even the strongest survivors to remain in the doldrums for years (in the cases listed above, the share price of both IBM and HSBC remained flat for almost a decade).
This is an interview with Ed Zitron.
Really? I thought the AI bubble happened because something genuinely novel had been invented. People immediately saw practical uses for it, and then it took on a life of its own - funding, superfunding, $trillions becoming part of day to day lingo.. etc.
AI is not in the same class as Vision Pro, which Ed Zitron is comparing it to. I personally found the overall tone to be a bit hyperbolic.
Hell no.
They'll pick up companies in trouble at bargain-basement prices.
>Consumers and enterprises alike have been trained to pay a monthly fee for a service, and while these services might have limits or strictures, basically nobody buying software expects to have a metered service, let alone one that's both metered and with hard to measure costs.
Has Ed Zitron not heard about the cloud? Unpredictable AWS bills?
But most consumer's aren't paying per token for access to models, so unless that changes and the labs start charging API pricing to everyone, it's kind of a moot point.
Users go on vacation, they slack off, they spend the day talking to each other. There are very few people who are really effective at burning tokens. how do you know the ratio? do you have insides? No :)
The biggest target is enterprise, and the economics for an LLM vendor look like this: price per token = R&D + inference + infra investments. When you buy a subscription, you are quite often buying a year ahead. That lets the vendor predict future infra investments against hard commitments, and sell expensive per token pricing to everyone else. And when a hard commitment sits unused because the user is busy, they sell it twice. It is loyalty in exchange for predictability, in exchange for the promise to always deliver SOTA to users.
Vendors control the harness. Tomorrow they simply roll out a router where reading the code and doing the final edits goes to a cheaper model, and their math suddenly becomes very sexy.
Isn't that hard to predict that their economic model is very easy to tune? and this is just first baby steps.
I personally pay per token ( do not have subs for work ). I did have once a $25k/mo worth of tokens, since i knew it was free so i was doing crazy experiments. Now , 2 month later, my bill was barely $1.5k since i moved into different stage with project. I do have team members who burn $500-600. pre router, pre optimization.
I switched recently to grok 4.5 and cursor router and my bill will go even further down. It rotates 4-5 different vendor models cheap and expensive too, depends on the task. Routers will flip entire LLM economy upside down.
What's unclear to me is if this is a systemic issue that's going to cause credit to freeze up but imo the opacity of the shadow banking "system" does not help here. If you see one cockroach, etc.