AI is going to be a highly-competitive, extremely capital-intensive commodity market that ends up in a race to the bottom competing on cost and efficiency of delivering models that have all reached the same asymptotic performance in the sense of intelligence, reasoning, etc.
The simple evidence for this is that everyone who has invested the same resources in AI has produced roughly the same result. OpenAI, Anthropic, Google, Meta, Deepseek, etc. There's no evidence of a technological moat or a competitive advantage in any of these companies.
The conclusion? AI is a world-changing technology, just like the railroads were, and it is going to soon explode in a huge bubble - just like the railroads did. That doesn't mean AI is going to go away, or that it won't change the world - railroads are still here and they did change the world - but from a venture investment perspective, get ready for a massive downturn.
There is a pretty big moat for Google: extreme amounts of video data on their existing services and absolutely no dependence on Nvidia and it's 90% margin.
I have yet to be convinced the broader population has an appetite for AI produced cinematography or videos. Independence from Nvidia is no more of a liability than dependence on electricity rates; it's not as if it's in Nvidia's interest to see one of its large customers fail. And pretty much any of the other Mag7 companies are capable of developing in-house TPUs + are already independently profitable, so Google isn't alone here.
If you think they are going to catch up with Google's software and hardware ecosystem on their first chip, you may be underestimating how hard this is. Google is on TPU v7. meta has already tried with MTIA v1 and v2. those haven't been deployed at scale for inference.
I don't think many of them will want to, though. I think as long as Nvidia/AMD/other hardware providers offer inference hardware at prices decent enough to not justify building a chip in-house, most companies won't. Some of them will probably experiment, although that will look more like a small team of researchers + a moderate budget rather than a burn-the-ships we're going to use only our own hardware approach.
Well, anthropic just purchased a million TPUs from Google because even with a healthy margin from Google, it's far more cost effective because of Nvidia's insane markup. That speaks for itself. Nvidia will not drop their margin because it will tank their stock price. it's half of the reason for all this circular financing - lowering their effective margin without lowering it on paper.
It's in Nvidia's interest to charge the absolute maximum they can without their customers failing. Every dollar of Nvidia's margin is your own lost margin. Utilities don't do that. Nvidia is objectively a way bigger liability than electricity rates.
And yes, all their competitors are making custom chips. Google is on TPU v7. absolutely nobody is going to get this right on the first try among their competitors - Google didn't.
Bigger problem for late starts now is that it will be hard to match the performance and cost of Google/Nvidia. It's an investment that had to have started years ago to be competitive now.
> AI is going to be a highly-competitive, extremely capital-intensive commodity market
It already is. In terms of competition, I don't think we've seen any groundbreaking new research or architecture since the introduction of inference time compute ("thinking") in late 2024/early 2025 circa GPT-o4.
The majority of the cost/innovation now is training this 1-2 year old technology on increasingly large amounts of content, and developing more hardware capable of running these larger models at more scale. I think it's fair to say the majority of capital is now being dumped into hardware, whether that's HBM and research related to that, or increasingly powerful GPUs and TPUs.
But these components are applicable to a lot of other places other than AI, and I think we'll probably stumble across some manufacturing techniques or physics discoveries that will have a positive impact on other industries.
> that ends up in a race to the bottom competing on cost and efficiency of delivering
One could say that the introduction of the personal computer became a "race to the bottom." But it was only the start of the dot-com bubble era, a bubble that brought about a lot of beneficial market expansion.
> models that have all reached the same asymptotic performance in the sense of intelligence, reasoning, etc.
I definitely agree with the asymptotic performance. But I think the more exciting fact is that we can probably expect LLMs to get a LOT cheaper in the next few years as the current investments in hardware begin to pay off, and I think it's safe to assume that in 5-10 years, most entry-level laptops will be able to manage a local 30B sized model while still being capable of multitasking. As it gets cheaper, more applications for it become more practical.
---
Regarding OpenAI, I think it definitely stands in a somewhat precarious spot, since basically the majority of its valuation is justified by nothing less than expectations of future profit. Unlike Google, which was profitable before the introduction of Gemini, AI startups need to establish profitability still. I think although initial expectations were for B2C models for these AI companies, most of the ones that survive will do so by pivoting to a B2B structure. I think it's fair to say that most businesses are more inclined to spend money chasing AI than individuals, and that'll lead to an increase in AI consulting type firms.
> in 5-10 years, most entry-level laptops will be able to manage a local 30B sized model
I suspect most of the excitement and value will be on edge devices. Models sized 1.7B to 30B have improved incredibly in capability in just the last few months and are unrecognizably better than a year ago. With improved science, new efficiency hacks, and new ideas, I can’t even imagine what a 30B model with effective tooling available could do in a personal device in two years time.
I, personally, use chatGPT for search more than I do Google these days. It, more often than not, gives me more exact results based on what I'm looking for and it produces links I can visit to get more information. I think this is where their competitive advantage lies if they can figure out how to monetize that.
What exactly is "second" place? No-one really knows what first place looks like. Everyone is certain that it will cost an arm, a leg and most of your organs.
For me, I think that, the possible winners will be close to fully funded up front and the losers will be trying to turn debt into profit and fail.
The rest of us self hoster types are hoping for a massive glut of GPUs and RAM to be dumped in a global fire sale. We are patient and have all those free offerings to play with for now to keep us going and even the subs are so far somewhat reasonable but we will flee in droves as soon as you try to ratchet up the price.
It's a bit unfortunate but we are waiting for a lot of large meme companies to die. Soz!
OpenAI is (was?) extremely good at making things that go viral. The successful ones for sure boost subscriber count meaningfully
Studio Ghibli, Sora app. Go viral, juice numbers then turn the knobs down on copyrighted material. Atlas I believe was a less successful than they would've hoped for.
And because of too frequent version bumps that are sometimes released as an answer to Google's launch, rather than a meaningful improvement - I believe they're also having harder time going viral that way
Overall OpenAI throws stuff at the wall and see what sticks. Most of it doesn't and gets (semi) abandoned. But some of it does and it makes for better consumer product than Gemini
It seems to have worked well so far, though I'm sceptical it will be enough for long
Because almost everyone involved in AI race grew up in "winner takes it all" environments, typical for software, and they try really hard to make it reality. This means your model should do everything to just take 90% of market share, or at least 90% of specific niche.
The problem is, they can't find the moat, despite searching very hard, whatever you bake into your AI, your competitors will be able to replicate in few months. This is why OpenAI is striking deal with Disney, because copyright provides such moat.
Alice changed things such that code monkeys algorithms were not patentable (except in some narrow cases where true runtime novelty can be established.) Since the transformers paper, the potential of self authoring content was obvious to those who can afford to think about things rather than hustle all day.
Apple wants to sell AI in an aluminum box while VCs need to prop up data center agrarianism; they need people to believe their server farms are essential.
Not an Apple fanboy but in this case, am rooting for their "your hardware, your model" aspirations.
Altman, Thiel, the VC model of make the serfs tend their server fields, allow them control of foundation models, is a gross feeling. It comes with the most religious like sense of fealty to political hierarchy and social structure that only exists as hallucination in the dying generations. The 50+ year old crowd cannot generationally churn fast enough.
Totally agree, people love to talk about how hopelessly behind Apple is in terms of AI progress when they’re in a better position to compete directly against Nvidia on hardware than anyone else.
> your competitors will be able to replicate in few months.
Will they really be able to replicate the quality while spending significantly less in compute investment? If not then the moat is still how much capital you can acquire for burning on training?
Because as with the internet 99% of the usage won’t be for education, work, personal development, what have you. It will be for effing kitten videos and memes.
Openrouter stats already mention 52% usage is roleplay.
As for photo/video very large number of people use it for friends and family (turn photo into creative/funny video, change photo, etc.).
Also I would think photoshop-like features are coming more and more in chatgpt and alike. For example, “take my poorly-lit photo and make it look professional and suitable for linkedin profile”
If Gemini can create or edit an image, chatgpt needs to be able to do this too. Who wants to copy&paste prompts between ai agents?
Also if you want to have more semantics, you add image, video and audio to your model. It gets smarter because of it.
OpenAI is also relevant bigger than antropic and is known as a generic 'helper'. Antropic probably saw the benefits of being more focused on developer which allows it to succeed longer in the game for the amount of money they have.
>Also if you want to have more semantics, you add image, video and audio to your model. It gets smarter because of it.
I think you are confusing generation with analysis. As far I am aware your model does not need to be good at generating images to be able to decode an image.
It is, to first approximation, the same thing. The generative part of genAI is just running the analysis model in reverse.
Now there are all sorts of tricks to get the output of this to be good, and maybe they shouldn't be spending time and resources on this. But the core capability is shared.
> Who wants to copy&paste prompts between ai agents?
An AI!
The specialist vs generalist debate is still open. And for complex problems, sure, having a model that runs on a small galaxy may be worth it. But for most tasks, a fleet of tailor-made smaller models being called on by an agent seems like a solidly-precedented (albeit not singularity-triggering) bet.
> But for most tasks, a fleet of tailor-made smaller models being called on by an agent seems like a solidly-precedented (albeit not singularity-triggering) bet.
not an expert by any means, but wouldn't smaller but highly refined models also output more reproducible results?
I think you're partially right, but I don't think being an AI leader is the main motivation -- that's a side effect.
I think it's important to OpenAI to support as many use-cases as possible. Right now, the experience that most people have with ChatGPT is through small revenue individual accounts. Individual subscriptions with individual needs, but modest budgets.
The bigger money is in enterprise and corporate accounts. To land these accounts, OpenAI will need to provide coverage across as many use-cases as they can so that they can operate as a one-stop AI provider. If a company needs to use OpenAI for chat, Anthropic for coding, and Google for video, what's the point? If Google's chat and coding is "good enough" and you need to have video generation, then that company is going to go with Google for everything. For the end-game I think OpenAI is playing for, they will need to be competitive in all modalities of AI.
Because for all the incessant whining about "slop," multimodal AI i/o is incredibly useful. Being able to take a photo of a home repair issue, have it diagnosed, and return a diagram showing you what to do with it is great, and it's the same algos that power the slop. "Sorry, you'll have to go to Gemini for that use case, people got mad about memes on the internet" is not really a good way for them to be a mass consumer company.
It's like half the poster on here live in some parallel universe. I am making real money using generated image/video advertising content for both B2C and B2B goods. I am using Whisper and LLMs to review customer service call logs at scale and identity development opportunities for staff. I am using GPT/Gemini to help write SQL queries and little python scripts to do data analysis on my customer base. My business's productivity is way up since GenAI become accessible.
But how much more profitable are they? We see revenue but not profits / spending. Anthropic seems to be growing faster than OpenAI did but that could be the benefit of post-GPT hype.
This article doesn’t add anything to what we know already. It’s still an open question what happens with the labs this coming year, but I personally think Anthropic’s focus on coding represents the clearest path to subscriber-based success (typical SaaS) whereas OpenAI has a clear opportunity with advertising. Both of these paths could be very lucrative. Meanwhile I expect Google will continue to struggle with making products that people actually want to use, irrespective of the quality of its models.
Bart was a flop.
Google search is losing market share to other LLM providers.
Gemini adoption is low, people around me prefer OpenAI because it is good enough and known.
But on the contrary, Nano Banana is very good, so I don't know.
And in the end, I'm pretty confident Google will be the AI race winner, because they got the engineers, they tech background and the money. Unless Google Adsense die, they can continue the race forever.
If Google is producing very good models and they aren’t gaining much traction, that seems like a pretty bad sign for them, right? If they were failing with bad models, the solution would be easy: math and engineer harder, make better models (I mean, this is obviously very hard but it is a clear path). Failing with good models is… confusing, it indicates there’s some unknown problem.
OK, but Gmail, Google Maps, Google Docs, and Google Search etc are ubiquitous. `Google' has even become a verb. Google might take a shotgun approach, but it certainly does create widely used products.
What Google AI products do people not want to use? Gemini is catching up to chatpt from a MAU perspective, ai overviews in search are super popular and staggeringly more used than any other ai-based product out there, a Google ai mode has decent usage, and Google Lens has surprisingly high usage. These products together dwarf everyone else out there by like 10x.
> ai overviews in search are super popular and staggeringly more used than any other ai-based product out there
This really is the critical bit. A year ago, the spin was "ChatGPT AI results are better than search, why would you use Google?", now it's "Search result AI is just as good as ChatGPT, why bother?".
When they were disruptive, it was enough to be different to believe that they'd win. Now they need to actually be better. And... they kinda aren't, really? I mean, lots of people like them! But for Regular Janes at the keyboard, who cares? Just type your search and see what it says.
1. Google books, which they legally scanned. No dubious training sets for them. They also regularly scrape the entire internet. And they have YouTube. Easy access to the best training data, all legally.
2. Direct access to the biggest search index. When you ask ChatGPT to search for something it is basically just doing what we do but a bit faster. Google can be much smarter, and because it has direct access it's also faster. Search is a huge use case of these services.
3. They have existing services like Android, Gmail, Google Maps, Photos, Assistant/Home etc. that they can integrate into their AI.
The difference in model capability seems to be marginal at best, or even in Google's favour.
OpenAI has "it's not Google" going for it, and also AI brand recognition (everyone knows what ChatGPT is). Tbh I doubt that will be enough.
Google's most significant advantage in this space is its organizational experience in providing services at this scale, as well as its mature infrastructure to support them. When the bubble pops, it's not lights-out or permanently degraded performance.
There is no doubt that OpenAI is taking a lot of risks by betting that AI adoption will translate into revenues in the very short term. And that could really happen imo (with a low probability sure, but worth the risk for VCs? Probably).
It's mathematically impossible what OpenAI is promising. They know it. The goal is to be too big to fail and get bailed out by US taxpayers who have been groomed into viewing AI as a cold war style arms race that America cannot lose.
Bailing out OAI would be entirely unnecessary (crowded field) and political suicide (how many hundreds of billions that could have gone to health care instead?)
If it happens in the next 3 years, tho, and Altman promises enough pork to the man, it could happen.
It is the term "mathematically impossible" that caught my attention. Since it is about the future promise of OpenAI, one could debate the likelihood or "statistically improbable", but "mathematically impossible" implies some calculation, proof and certainty. Hence my curiosity.
I've seen some calculation I think from an HSBC analyst that it would take a monthly subscription of $200/mo. from some large portion of the US population for some insane number of years to break even.
on the one hand, i understand you are making a stylized comment, on the other hand, as soon as i started writing something reasonable, i realized this is an "upvote lame catastrophizing takes about" (checking my notes) "some company" thread, which means reasonable stuff will get downvoted... for example, where is there actual scarcity in their product inputs? for example, will they really be paying retail prices to infrastructure providers forever? is that a valid forecast? many reasonable ways to look at this. even if i take your cynical stuff at 100% face value, the thing about bailouts is that they're more complicated than what you are saying, but your instinct is to say they're not complicated, "grooming" this and "cold war" that, because your goal is to concern troll, not advance this site's goal of curiosity...
Correction: OpenAI investors do take that risk. Some of the investors (e.g. Microsoft, Nvidia) dampen that risk by making such investment conditioned on boosting the investor's own revenue, a stock buyback of sorts.
Apparently we all have enough money to put it into OpenAI.
Some players have to play, like google, some players want to play like USA vs. China.
Besides that, chatting with an LLM is very very convincing. Normal non technical people can see what 'this thing' can already do and as long as the progress is continuing as fast as it currently is, its still a very easy to sell future.
I don't think you have the faintest clue of what you're talking about right now. Google authored the transformer architecture, the basis of every GPT model OpenAI has shipped. They aren't obligated to play any more than OpenAI is, they do it because they get results. The same cannot be said of OpenAI.
The fact is nobody has any idea what OpenAI's cash burn is. Measuring how much they're raising is not an adequate proxy.
For all we know, they could be accumulating capital to weather an AI winter.
It's also worth noting that OpenAI has not trained a new model since gpt4o (all subsequent models are routing systems and prompt chains built on top of 4), so the idea of OpenAI being stuck in some kind of runaway training expense is not real.
Didn't they create Sora and other models and literally burned so much money with their AI video app which they wanted to make a social media but what ended up happening was that they burned billions of dollars.
I think you are messing up things here, and I think your comment is based on the article from semi analysis. [1]
It said:
OpenAI’s leading researchers have not completed a successful full-scale pre-training run that was broadly deployed for a new frontier model since GPT-4o in May 2024, highlighting the significant technical hurdle that Google’s TPU fleet has managed to overcome.
However, pre-training run is the initial, from-scratch training of the base model. You say they only added routing and prompts, but that's not what the original article says. They most likely still have done a lot of fine tuning, RLHF, alignment and tool calling improvements. All that stuff is training too. And it is totally fine, just look at the great results they got with Codex-high.
If you got actually got what you said from a different source, please link it. I would like to read it. If you just messed things up, that's fine too.
The GPT-5 series is a new model, based on the o1/o3 series. It's very much inaccurate to say that it's a routing system and prompt chain built on top of 4o. 4o was not a reasoning model and reasoning prompts are very weak compared to actual RLVR training.
No one knows whether the base model has changed, but 4o was not a base model, and neither is 5.x. Although I would be kind of surprised if the base model hadn't also changed, FWIW: they've significantly advanced their synthetic data generation pipeline (as made obvious via their gpt-oss-120b release, which allegedly was entirely generated from their synthetic data pipelines), which is a little silly if they're not using it to augment pretraining/midtraining for the models they actually make money from. But either way, 5.x isn't just a prompt chain and routing on top of 4o.
Prior to 5.2 you couldn’t expect to get good answers to questions prior to March 2024. It was arguing with me that Bruno Mars did not have two hit songs in the last year. It’s clear that in 2025 OpenAI used the old 4.0 base model and tried to supercharge it using RLVR. That had very mixed results.
>It's also worth noting that OpenAI has not trained a new model since gpt4o (all subsequent models are routing systems and prompt chains built on top of 4)
At the very least they made GPT 4.5, which was pretty clearly trained from scratch. It was possibly what they wanted GPT-5 to be but they made a wrong scaling prediction, people simply weren't ready to pay that much money.
> The fact is nobody has any idea what OpenAI's cash burn is.
Their investors surely do (absent outrageous fraud).
> For all we know, they could be accumulating capital to weather an AI winter.
If they were, their investors would be freaking out (or complicit in the resulting fraud). This seems unlikely. In point of fact it seems like they're playing commodities market-cornering games[1] with their excess cash, which implies strongly that they know how to spend it even if they don't have anything useful to spend it on.
RAG? Even for a "fresh" model, there is no way to keep it up to date, so there has to be a mechanism by which to reference eg last night's football game.
Yes it was, op didn't read the reporting closely enough. It said something to the effect of "Didn't pretrain a new broadly released, generally available model"
OpenAI has #5 traffic levels globally. Their product-market fit is undeniable. The question is monetization.
Their cost to serve each request is roughly 3 orders of magnitude higher than conventional web sites.
While it is clear people see value in the product, we only know they see value at today’s subsidized prices. It is possible that inference prices will continue their rapid decline. Or it is possible that OAI will need to raise prices and consumers will be willing to pay more for the value.
Yes, but that is the standard methodology for startups in their boost phase. Burn vast piles of cash to acquire users, then find out at the end if a profitable business can be made of it.
it's a simple problem really. what is actually scarce?
a spot on the iOS home screen? yes.
infrastructure to serve LLM requests? no.
good LLM answers? no.
the economist can't tell the difference between scarcity and real scarcity.
it is extremely rare to buy a spot on the iOS home screen, and the price for that is only going up - think of the trend of values of tiktok, whatsapp and instagram. that's actually scarce.
that is what openai "owns." you're right, #5 app. you look at someone's home screen, and the things on it are owned by 8 companies, 7 of which are the 7 biggest public companies in the world, and the 8th is openai.
whereas infrastructure does in fact get cheaper. so does energy. they make numerous mistakes - you can't forecast retail prices Azure is "charging" openai for inference. but also, NVIDIA participates in a cartel. GPUs aren't actually scarce, you don't actually need the highest process nodes at TSMC, etc. etc. the law can break up cartels, and people can steal semiconductor process knowledge.
but nobody can just go and "create" more spots on the iOS home screen. do you see?
depends if they can monetize that spot. So either ads or subscription. It is as yet unclear whether ads/subscription can generate sufficient revenue to cover costs and return a profit. Perhaps 'enough ads' will be too much for users to bear, perhaps 'enough subscription' will be too much for users to afford.
For what I use them for, the LLM market has become a two player game, and the players are Anthropic and Google. So I find it quite interesting that OpenAI is still the default assumption of the leader.
ChatGPT dominates the consumer market (though Nano Banana is singlehandedly breathing some life into consumer Gemini).
A small anecdote: when ChatGPT went down a few months ago, a lot of young people (especially students) just waited for it to come back up. They didn't even think about using an alternative.
When ChatGPT starts injecting ads or forcing payment or doing anything else that annoys its userbase then the young people won't have a problem looking for alternatives
That is different because all of the players I mentioned have credible, near-leading products in the AI model market, whereas nobody other than Google has search results worth a damn. I wouldn't recommend anyone squander their time by checking Kagi or DDG or Bing more than once.
Burn rate often gets treated as a hard signal, but it is mostly about expectations. Once people get used to the idea of cheap intelligence, any slowdown feels like failure, even if the technology is still moving forward. That gap is usually where bubbles begin.
why does the article used words like burn and incinerate, implying that OpenAI is somehow making money disappear or something? They’re spending it; someone is profiting here, even if it’s not OpenAI. Is it all Nvidia?
Because typically one expect a return on investment with that level of spending. Not only have they run at a loss for years, their spending is expected to increase, with no path to profitability in sight.
IIRC, current estimates are that OpenAI is losing as much money a year as Uber or Amazon lost in their entire lifetime of unprofitability. Also, both Uber and Amazon spent their unprofitable years having a clear roadmap to profitability. OpenAI's roadmap to profitability is "???"
I have lived through Amazon’s rags to riches and there was never a clear plan to profitability. Vast majority of people were questioning sanity of anyone investing in Amazon.
I am not saying OpenAI is Amazon but am saying I have seen this before where masses are going “oh business is bad, losses are huge, where is path to profitability…”
I think you're saying that just running up huge losses is sufficient to create a successful company? But that you personally wouldn't want to run up huge losses? Not sure.
I suspect most of it is going to utilities for power, water and racking.
That being said, if I was Sam Altman I'd also be stocking up on yachts, mansions and gold plated toilets while the books are still private. If there's $10bn a year in outgoings no one's going to notice a million here and there.
Tragically I don't make CEO money so I also don't have one but I presume you'd want to have at least one per mansion and another one in the office. Maybe a separate one for special occasions.
“Burn rate” is a standard financial term for how much money a startup is losing. If you have $1 cash on hand and a burn rate of $2 a year, then you have six months before you either need to get profitable, raise more money, or shut down.
On the radio they mentioned that the total global chocolate market is ~100B, I googled it when I was home and it seems to be about ~135B. Apparently that is ... all chocolate, everywhere.. OpenAI's valuation is about 500B. Maybe going up to like 835B.
I'd love to see the rationale that OpenAI (not "AI" everywhere) is more valuable than chocolate globally.
Ignoring that those numbers aren't directly comparable, it did make me wonder, if I had to give up either "AI" or chocolate tomorrow, which would I pick?
Even as an enormous chocolate lover (in all three senses) who eats chocolate several times a week, I'd probably choose AI instead.
OpenAI has alternatives, but also I do spend more money on OpenAI than I do on chocolate currently.
I am just trying to help you write better. Your writing says "if I had to give up either AI or chocolate [...] I would probably choose AI". However, your language and intent seems to be that you would give up chocolate.
If you really wanted to know you could stop eating chocolate or stop using ai and see if you break. Or do both at different times and see how long you last without one or the other.
You are already paying for several national lab HPC centers. These are used for government/university research - no idea if commercial interests can rent time on them. The big ones are running weather, astronomy simulations, nuclear explosions, biological sequencing, and so on.
No chance they're going to take risks to share that hardware with anyone given what it does.
The scaled down version of El Capitan is used for non-classified workloads, some of which are proprietary, like drug simulation. It is called Tuolumne. Not long ago, it was nevertheless still a top ten supercomputer.
Like OP, I also don't see why a government supercomputer does it better than hyperscalers, coreweave, neoclouds, et al, who have put in a ton of capital as even compared to government. For loads where institutional continuity is extremely important, like weather -- and maybe one day, a public LLM model or three -- maybe. But we're not there yet, and there's so much competition in LLM infrastructure that it's quite likely some of these entrants will be bag holders, not a world of juicy margins at all...rather, playing chicken with negative gross margins.
if datacenters are built by the government, then i think it's fair to assume there will be some level of democratic control of what those datacenters will be used for.
This is literally the current system... adding more democratic controls is a good thing, the alternative is that only rich control these systems and would you look at it only the rich control these systems.
That's like every government initiative. Same as healthcare? School? I mean if you don't have children why do you pay taxes... and roads if you don't drive? I mean the examples are so many... why do you bring this argument that if it doesn't benefit you directly right now today, it shouldn't be done?
There are arguments aplenty that schooling and a minimum amount of healthcare are public goods, as are roads built on public land (the government owns most roads after all).
What is the justification for considering data centers capable of running LLMs to be a public good?
There are many counter examples of things many people use but are still private. Clothing stores, restaurants and grocery stores, farms, home appliance factories, cell phone factories, laundromats and more.
How is that distinct from any of my other examples which listed factories? Very few factories in the US are publicly owned; citing data centers as places of production merely furthers the argument that they should remain private.
Last-mile services like roads, electricity, water, and telecommunications are natural monopolies. Normal market forces fail somewhat and you want some government involvement to keep it running smoothly.
I have no idea why you're being downvoted because you're right. The entire point of taxation is to spread the cost among everyone, and since everyone doesn't utilise every government service every tax payer ends up paying for stuff they don't use. That like, the whole point...
> The Innovative and Novel Computational Impact on Theory and Experiment, or INCITE, program has announced the 2026 Call for Proposals, inviting researchers to apply for access to some of the world’s most powerful high-performance computing systems.
> The proposal submission window runs from April 11 to June 16, 2025, offering an opportunity for scientific teams to secure substantial computational resources for large-scale research projects in fields such as scientific modeling, simulation, data analytics and artificial intelligence. [...]
> Individual awards typically range from 500,000 to 1,000,000 node-hours on Aurora and Frontier and 100,000 to 250,000 node-hours on Polaris, with the possibility of larger allocations for exceptional proposals. [...]
> The selection process involves a rigorous peer review, assessing both scientific merit and computational readiness. Awards will be announced in November 2025, with access to resources beginning in 2026.
Not sure OpenAI/Anthropic etc would be OK with a six month gap between application and getting access to the resources, but this does indeed demonstrate that government issued super-computing resources is a previously solved problem.
Well, people bid for USA government resources all the time. It's why the Washington DC suburbs have some of the country's most affluent neighborhoods among their ranks.
In theory it makes the process more transparent and fair, although slower. That calculus has been changing as of late, perhaps for both good and bad. See for example the Pentagon's latest support of drone startups run by twenty-year-olds.
The question of public and private distinctions in these various schemes are very interesting and imo, underexplored. Especially when you consider how these private LLMs are trained on public data.
In a completely alternate dimension, a quarter of the capital being invested in AI literally just goes towards making sure everyone has quality food and water.
As we all know, throwing money at a problem solves it completely. Remember how Live Aid saved Ethiopia from starvation and it never had any problems again?
you'll never win that argument, but I absolutely agree.
people have no idea about how big the military and defense budgets worldwide are next to any other example of a public budget.
throw as many pie charts out as you want; people just can't see the astronomical difference in budgets.
I think it's based on how the thing works; a good defense works until it doesn't -- the other systems/budgets in place have a bit more of a graceful failure. This concept produces an irrationality in people that produces windfalls of cash availability.
Without capital invested in the past we wouldn’t have almost anything of modern technology. That has done a lot more for everyone, including food affordability, than actually simply buying food for people to eat once.
Datacenters are not a natural monopoly, you can always build more. Beyond what the public sector itself might need for its own use, there's not much of a case for governments to invest in them.
That could make sense in some steady state regime where there were stable requirements and mature tech (I wouldn’t vote for it but I can see an argument).
I see no argument why the government would jump into a hype cycle and start building infra that speculative startups are interested in. Why would they take on that risk compared to private investors, and how would they decide to back that over mammoth cloning infra or whatever other startups are doing?
In a better parallel universe, we found a different innovation without using brute-force computation to train systems that unreliably and inefficiently compute things and still leaves us able to understand what we're building.
Same reason they should own access lines: everyone needs rackspace/access, it should be treated like a public service to avoid rent seeing. Having a data center in every city where all of the local lines terminate into could open the doors to a lot of interesting use cases, really help with local resiliency/decentralization efforts, and provide a great alternative to cloud providers that doesn't break the bank.
Smells like socialism. Around here we privatize the profits and only socialize the costs. Like the impending bailout of the most politically connected AI companies.
Prediction: on this thread you'll get a lot of talk about how government would slow things down. But when the AI bubble starts to look shaky, see how fast all the tech bros line up for a "public private partnership."
That's malinvestment. Too much overhead, disconnected from long term demand. The government doesn't have expertise, isn't lean and nimble. What if it all just blows over? (It won't? But who knows?)
Everything is happening exactly as it should. If the "bubble" "pops", that's just the economic laws doing what they naturally do.
The government has better things to do. Geopolitics, trade, transportation, resources, public health, consumer safety, jobs, economy, defense, regulatory activities, etc.
In 2008 though the US government ended up making more money then they spent though (at least with the tarp), because they invested a ton of money when everything collapsed, and thus was extremely cheap. Once the markets recovered, they made a hefty sum selling all the derivatives they got at the lowest point. Seems like the epitome of buy when low and sell when high tbh.
Not really. It was not about stocks. It was the collapse of insurance companies at the core of 2008 crisis.
The same can happen now on the side of private credit that gradually offloads its junk to insurance companies (again):
As a result, private credit is on the rise as an investment option to compensate for this slowdown in traditional LBO (Figure 2, panel 2), and PE companies are actively growing the private credit side of their business by influencing the companies they control to help finance these operations. Life insurers are among these companies. For instance, KKR’s acquisition of 60 percent of Global Atlantic (a US life insurer) in 2020 cost KKR approximately $3billion.
The simple evidence for this is that everyone who has invested the same resources in AI has produced roughly the same result. OpenAI, Anthropic, Google, Meta, Deepseek, etc. There's no evidence of a technological moat or a competitive advantage in any of these companies.
The conclusion? AI is a world-changing technology, just like the railroads were, and it is going to soon explode in a huge bubble - just like the railroads did. That doesn't mean AI is going to go away, or that it won't change the world - railroads are still here and they did change the world - but from a venture investment perspective, get ready for a massive downturn.
It already is. In terms of competition, I don't think we've seen any groundbreaking new research or architecture since the introduction of inference time compute ("thinking") in late 2024/early 2025 circa GPT-o4.
The majority of the cost/innovation now is training this 1-2 year old technology on increasingly large amounts of content, and developing more hardware capable of running these larger models at more scale. I think it's fair to say the majority of capital is now being dumped into hardware, whether that's HBM and research related to that, or increasingly powerful GPUs and TPUs.
But these components are applicable to a lot of other places other than AI, and I think we'll probably stumble across some manufacturing techniques or physics discoveries that will have a positive impact on other industries.
> that ends up in a race to the bottom competing on cost and efficiency of delivering
One could say that the introduction of the personal computer became a "race to the bottom." But it was only the start of the dot-com bubble era, a bubble that brought about a lot of beneficial market expansion.
> models that have all reached the same asymptotic performance in the sense of intelligence, reasoning, etc.
I definitely agree with the asymptotic performance. But I think the more exciting fact is that we can probably expect LLMs to get a LOT cheaper in the next few years as the current investments in hardware begin to pay off, and I think it's safe to assume that in 5-10 years, most entry-level laptops will be able to manage a local 30B sized model while still being capable of multitasking. As it gets cheaper, more applications for it become more practical.
---
Regarding OpenAI, I think it definitely stands in a somewhat precarious spot, since basically the majority of its valuation is justified by nothing less than expectations of future profit. Unlike Google, which was profitable before the introduction of Gemini, AI startups need to establish profitability still. I think although initial expectations were for B2C models for these AI companies, most of the ones that survive will do so by pivoting to a B2B structure. I think it's fair to say that most businesses are more inclined to spend money chasing AI than individuals, and that'll lead to an increase in AI consulting type firms.
I suspect most of the excitement and value will be on edge devices. Models sized 1.7B to 30B have improved incredibly in capability in just the last few months and are unrecognizably better than a year ago. With improved science, new efficiency hacks, and new ideas, I can’t even imagine what a 30B model with effective tooling available could do in a personal device in two years time.
For me, I think that, the possible winners will be close to fully funded up front and the losers will be trying to turn debt into profit and fail.
The rest of us self hoster types are hoping for a massive glut of GPUs and RAM to be dumped in a global fire sale. We are patient and have all those free offerings to play with for now to keep us going and even the subs are so far somewhat reasonable but we will flee in droves as soon as you try to ratchet up the price.
It's a bit unfortunate but we are waiting for a lot of large meme companies to die. Soz!
Studio Ghibli, Sora app. Go viral, juice numbers then turn the knobs down on copyrighted material. Atlas I believe was a less successful than they would've hoped for.
And because of too frequent version bumps that are sometimes released as an answer to Google's launch, rather than a meaningful improvement - I believe they're also having harder time going viral that way
Overall OpenAI throws stuff at the wall and see what sticks. Most of it doesn't and gets (semi) abandoned. But some of it does and it makes for better consumer product than Gemini
It seems to have worked well so far, though I'm sceptical it will be enough for long
The problem is, they can't find the moat, despite searching very hard, whatever you bake into your AI, your competitors will be able to replicate in few months. This is why OpenAI is striking deal with Disney, because copyright provides such moat.
Been saying this since the 2016 Alice case. Apple jumped into content production in 2017. They saw the long term value of copyright interests.
https://arstechnica.com/information-technology/2017/08/apple...
Alice changed things such that code monkeys algorithms were not patentable (except in some narrow cases where true runtime novelty can be established.) Since the transformers paper, the potential of self authoring content was obvious to those who can afford to think about things rather than hustle all day.
Apple wants to sell AI in an aluminum box while VCs need to prop up data center agrarianism; they need people to believe their server farms are essential.
Not an Apple fanboy but in this case, am rooting for their "your hardware, your model" aspirations.
Altman, Thiel, the VC model of make the serfs tend their server fields, allow them control of foundation models, is a gross feeling. It comes with the most religious like sense of fealty to political hierarchy and social structure that only exists as hallucination in the dying generations. The 50+ year old crowd cannot generationally churn fast enough.
Will they really be able to replicate the quality while spending significantly less in compute investment? If not then the moat is still how much capital you can acquire for burning on training?
(The obvious well-paying market would be erotic / furry / porn, but it's too toxic to publicly touch, at least in the US.)
As for photo/video very large number of people use it for friends and family (turn photo into creative/funny video, change photo, etc.).
Also I would think photoshop-like features are coming more and more in chatgpt and alike. For example, “take my poorly-lit photo and make it look professional and suitable for linkedin profile”
If Gemini can create or edit an image, chatgpt needs to be able to do this too. Who wants to copy&paste prompts between ai agents?
Also if you want to have more semantics, you add image, video and audio to your model. It gets smarter because of it.
OpenAI is also relevant bigger than antropic and is known as a generic 'helper'. Antropic probably saw the benefits of being more focused on developer which allows it to succeed longer in the game for the amount of money they have.
I think you are confusing generation with analysis. As far I am aware your model does not need to be good at generating images to be able to decode an image.
Now there are all sorts of tricks to get the output of this to be good, and maybe they shouldn't be spending time and resources on this. But the core capability is shared.
I think that hasn't been the case since DeepDream?
An AI!
The specialist vs generalist debate is still open. And for complex problems, sure, having a model that runs on a small galaxy may be worth it. But for most tasks, a fleet of tailor-made smaller models being called on by an agent seems like a solidly-precedented (albeit not singularity-triggering) bet.
intuitively it sounds akin to the unix model...
I think it's important to OpenAI to support as many use-cases as possible. Right now, the experience that most people have with ChatGPT is through small revenue individual accounts. Individual subscriptions with individual needs, but modest budgets.
The bigger money is in enterprise and corporate accounts. To land these accounts, OpenAI will need to provide coverage across as many use-cases as they can so that they can operate as a one-stop AI provider. If a company needs to use OpenAI for chat, Anthropic for coding, and Google for video, what's the point? If Google's chat and coding is "good enough" and you need to have video generation, then that company is going to go with Google for everything. For the end-game I think OpenAI is playing for, they will need to be competitive in all modalities of AI.
But on the contrary, Nano Banana is very good, so I don't know. And in the end, I'm pretty confident Google will be the AI race winner, because they got the engineers, they tech background and the money. Unless Google Adsense die, they can continue the race forever.
Gemini is built into Android and Google search. People may not be going to gemini.google.com, but that does not mean adoption is low.
But you cannot use it with an API key.
If you're on a workspace account, you can't have normal individual plan.
You have to have the team plan with $100/month or nothing.
Google's product management tier is beyond me.
This really is the critical bit. A year ago, the spin was "ChatGPT AI results are better than search, why would you use Google?", now it's "Search result AI is just as good as ChatGPT, why bother?".
When they were disruptive, it was enough to be different to believe that they'd win. Now they need to actually be better. And... they kinda aren't, really? I mean, lots of people like them! But for Regular Janes at the keyboard, who cares? Just type your search and see what it says.
1. Google books, which they legally scanned. No dubious training sets for them. They also regularly scrape the entire internet. And they have YouTube. Easy access to the best training data, all legally.
2. Direct access to the biggest search index. When you ask ChatGPT to search for something it is basically just doing what we do but a bit faster. Google can be much smarter, and because it has direct access it's also faster. Search is a huge use case of these services.
3. They have existing services like Android, Gmail, Google Maps, Photos, Assistant/Home etc. that they can integrate into their AI.
The difference in model capability seems to be marginal at best, or even in Google's favour.
OpenAI has "it's not Google" going for it, and also AI brand recognition (everyone knows what ChatGPT is). Tbh I doubt that will be enough.
In my view Google is uniquely well positioned because, contrary to the others, it controls most of the raw materials for Ai.
If it happens in the next 3 years, tho, and Altman promises enough pork to the man, it could happen.
Citation is needed
It’s going to crash, guaranteed
Some players have to play, like google, some players want to play like USA vs. China.
Besides that, chatting with an LLM is very very convincing. Normal non technical people can see what 'this thing' can already do and as long as the progress is continuing as fast as it currently is, its still a very easy to sell future.
I don't think you have the faintest clue of what you're talking about right now. Google authored the transformer architecture, the basis of every GPT model OpenAI has shipped. They aren't obligated to play any more than OpenAI is, they do it because they get results. The same cannot be said of OpenAI.
For all we know, they could be accumulating capital to weather an AI winter.
It's also worth noting that OpenAI has not trained a new model since gpt4o (all subsequent models are routing systems and prompt chains built on top of 4), so the idea of OpenAI being stuck in some kind of runaway training expense is not real.
It said: OpenAI’s leading researchers have not completed a successful full-scale pre-training run that was broadly deployed for a new frontier model since GPT-4o in May 2024, highlighting the significant technical hurdle that Google’s TPU fleet has managed to overcome.
However, pre-training run is the initial, from-scratch training of the base model. You say they only added routing and prompts, but that's not what the original article says. They most likely still have done a lot of fine tuning, RLHF, alignment and tool calling improvements. All that stuff is training too. And it is totally fine, just look at the great results they got with Codex-high.
If you got actually got what you said from a different source, please link it. I would like to read it. If you just messed things up, that's fine too.
[1] https://newsletter.semianalysis.com/p/tpuv7-google-takes-a-s...
No one knows whether the base model has changed, but 4o was not a base model, and neither is 5.x. Although I would be kind of surprised if the base model hadn't also changed, FWIW: they've significantly advanced their synthetic data generation pipeline (as made obvious via their gpt-oss-120b release, which allegedly was entirely generated from their synthetic data pipelines), which is a little silly if they're not using it to augment pretraining/midtraining for the models they actually make money from. But either way, 5.x isn't just a prompt chain and routing on top of 4o.
I’m sure all these AI labs have extensive data gathering, cleanup, and validation processes for new data they train the model on.
Or at least I hope they don’t just download the current state of the web on the day they need to start training the new model and cross their fingers.
At the very least they made GPT 4.5, which was pretty clearly trained from scratch. It was possibly what they wanted GPT-5 to be but they made a wrong scaling prediction, people simply weren't ready to pay that much money.
Their investors surely do (absent outrageous fraud).
> For all we know, they could be accumulating capital to weather an AI winter.
If they were, their investors would be freaking out (or complicit in the resulting fraud). This seems unlikely. In point of fact it seems like they're playing commodities market-cornering games[1] with their excess cash, which implies strongly that they know how to spend it even if they don't have anything useful to spend it on.
[1] Again c.f. fraud
I know sama says they aren’t trying to train new models, but he’s also a known liar and would definitely try to spin systemic failure.
Doubtful. This would be the very antithesis of the Silicon Valley way.
Their cost to serve each request is roughly 3 orders of magnitude higher than conventional web sites.
While it is clear people see value in the product, we only know they see value at today’s subsidized prices. It is possible that inference prices will continue their rapid decline. Or it is possible that OAI will need to raise prices and consumers will be willing to pay more for the value.
a spot on the iOS home screen? yes.
infrastructure to serve LLM requests? no.
good LLM answers? no.
the economist can't tell the difference between scarcity and real scarcity.
it is extremely rare to buy a spot on the iOS home screen, and the price for that is only going up - think of the trend of values of tiktok, whatsapp and instagram. that's actually scarce.
that is what openai "owns." you're right, #5 app. you look at someone's home screen, and the things on it are owned by 8 companies, 7 of which are the 7 biggest public companies in the world, and the 8th is openai.
whereas infrastructure does in fact get cheaper. so does energy. they make numerous mistakes - you can't forecast retail prices Azure is "charging" openai for inference. but also, NVIDIA participates in a cartel. GPUs aren't actually scarce, you don't actually need the highest process nodes at TSMC, etc. etc. the law can break up cartels, and people can steal semiconductor process knowledge.
but nobody can just go and "create" more spots on the iOS home screen. do you see?
A small anecdote: when ChatGPT went down a few months ago, a lot of young people (especially students) just waited for it to come back up. They didn't even think about using an alternative.
This "moat" that OpenAI has is really weak
I am not saying OpenAI is Amazon but am saying I have seen this before where masses are going “oh business is bad, losses are huge, where is path to profitability…”
That being said, if I was Sam Altman I'd also be stocking up on yachts, mansions and gold plated toilets while the books are still private. If there's $10bn a year in outgoings no one's going to notice a million here and there.
That's what the words mean in this context.
I'd love to see the rationale that OpenAI (not "AI" everywhere) is more valuable than chocolate globally.
... so crash early 2026?
Even as an enormous chocolate lover (in all three senses) who eats chocolate several times a week, I'd probably choose AI instead.
OpenAI has alternatives, but also I do spend more money on OpenAI than I do on chocolate currently.
People take old things for granted often. Explains the Coolidge effect, and why plenty of people cheat.
Why would you want my money to be used to build datacenter that won’t benefit me ? I might use a LLM once a month, many people never use it.
Let the one who use it pay for it.
these things constitute public goods that benefit the individual regardless of participation.
No chance they're going to take risks to share that hardware with anyone given what it does.
The scaled down version of El Capitan is used for non-classified workloads, some of which are proprietary, like drug simulation. It is called Tuolumne. Not long ago, it was nevertheless still a top ten supercomputer.
Like OP, I also don't see why a government supercomputer does it better than hyperscalers, coreweave, neoclouds, et al, who have put in a ton of capital as even compared to government. For loads where institutional continuity is extremely important, like weather -- and maybe one day, a public LLM model or three -- maybe. But we're not there yet, and there's so much competition in LLM infrastructure that it's quite likely some of these entrants will be bag holders, not a world of juicy margins at all...rather, playing chicken with negative gross margins.
Uncanny really.
What is the justification for considering data centers capable of running LLMs to be a public good?
There are many counter examples of things many people use but are still private. Clothing stores, restaurants and grocery stores, farms, home appliance factories, cell phone factories, laundromats and more.
Why not an LLM datacenter if it also offers information? You could say it's the public library of the future maybe.
Data centers clearly can exist without being owned by the public.
This is not at all true of generative AI.
OpenAI ask for 1m GPUs for a month, Anthropic ask for 2m, the government data center only has 500,000, and a new startup wants 750,000 as well.
Do you hand them out to the most convincing pitch? Hopefully not to the biggest donor to your campaign.
Now the most successful AI lab is the one that's best at pitching the government for additional resources.
UPDATE: See comment below for the answer to this question: https://news.ycombinator.com/item?id=46438390#46439067
It would still likely devolve into most-money-wins, but it is not an insurmountable political obstacle to arrange some sort of sharing.
Edit: I meant to say over subscribed, not over provisioned. There are far more jobs in the queue than can be handled at once
https://www.ornl.gov/news/doe-incite-program-seeks-2026-prop...
> The Innovative and Novel Computational Impact on Theory and Experiment, or INCITE, program has announced the 2026 Call for Proposals, inviting researchers to apply for access to some of the world’s most powerful high-performance computing systems.
> The proposal submission window runs from April 11 to June 16, 2025, offering an opportunity for scientific teams to secure substantial computational resources for large-scale research projects in fields such as scientific modeling, simulation, data analytics and artificial intelligence. [...]
> Individual awards typically range from 500,000 to 1,000,000 node-hours on Aurora and Frontier and 100,000 to 250,000 node-hours on Polaris, with the possibility of larger allocations for exceptional proposals. [...]
> The selection process involves a rigorous peer review, assessing both scientific merit and computational readiness. Awards will be announced in November 2025, with access to resources beginning in 2026.
Not sure OpenAI/Anthropic etc would be OK with a six month gap between application and getting access to the resources, but this does indeed demonstrate that government issued super-computing resources is a previously solved problem.
In theory it makes the process more transparent and fair, although slower. That calculus has been changing as of late, perhaps for both good and bad. See for example the Pentagon's latest support of drone startups run by twenty-year-olds.
The question of public and private distinctions in these various schemes are very interesting and imo, underexplored. Especially when you consider how these private LLMs are trained on public data.
people have no idea about how big the military and defense budgets worldwide are next to any other example of a public budget.
throw as many pie charts out as you want; people just can't see the astronomical difference in budgets.
I think it's based on how the thing works; a good defense works until it doesn't -- the other systems/budgets in place have a bit more of a graceful failure. This concept produces an irrationality in people that produces windfalls of cash availability.
Hmm, what about member-owned coöperatives? Like what we have for stock exchanges.
I see no argument why the government would jump into a hype cycle and start building infra that speculative startups are interested in. Why would they take on that risk compared to private investors, and how would they decide to back that over mammoth cloning infra or whatever other startups are doing?
Everything is happening exactly as it should. If the "bubble" "pops", that's just the economic laws doing what they naturally do.
The government has better things to do. Geopolitics, trade, transportation, resources, public health, consumer safety, jobs, economy, defense, regulatory activities, etc.
2026: US AI companies pump stocks -> market correction -> taxpayer bailout
Mark my words. OpenAI will be bailed out by US taxpayers.
The same can happen now on the side of private credit that gradually offloads its junk to insurance companies (again):
As a result, private credit is on the rise as an investment option to compensate for this slowdown in traditional LBO (Figure 2, panel 2), and PE companies are actively growing the private credit side of their business by influencing the companies they control to help finance these operations. Life insurers are among these companies. For instance, KKR’s acquisition of 60 percent of Global Atlantic (a US life insurer) in 2020 cost KKR approximately $3billion.
https://www.imf.org/en/Publications/global-financial-stabili...