OpenAI's cash burn will be one of the big bubble questions of 2026

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.

It is not a railroad and the railroads did not explode in a bubble (OK a few early engines did explode but that is engineering). I think LLM driven investments in massive DCs is ill advised.

  > There's no evidence of a technological moat or a competitive advantage in any of these companies.

I don't use my Google Pro subscription much. Gemini 3.0 Pro spends 1/10th of the time thinking compared to GPT 5.2 Thinking and outputs a worse answer or ignores my prompt. Similar story with Deepseek.

The public benchmarks tell a different story which is where I believe the sentiment online comes from, but I am going to trust my experience, because my experience can't be benchmaxxed.

The cost of entry is far beyond extraordinary. You're acting like anybody can gain entry, when the exact opposite is the case. The door is closing right now. Just try to compete with OpenAI, let's see you calculate the price of attempting it. Scale it to 300, 500, 800 million users.

Why aren't there a dozen more Anthropics, given the valuation in question (and potential IPO)? Because it'll cost you tens of billions of dollars just to try to keep up. Nobody will give you that money. You can't get the GPUs, you can't get the engineers, you can't get the dollars, you can't build the datacenters. Hell, you can't even get the RAM these days, nor can you afford it.

Google & Co are capturing the market and will monetize it with advertising. They will generate trillions of dollars in revenue over the coming 10-15 years by doing so.

The barrier to entry is the same one that exists in search: it'll cost you well over one hundred billion dollars to try to be in the game at the level that Gemini will be at circa 2026-2027, for just five years.

Please, inform me of where you plan to get that one hundred billion dollars just to try to keep up. Even Anthropic is going to struggle to stay in the competition when the music (funding bubble) stops.

There are maybe a dozen or so companies in existence that can realistically try to compete with the likes of Gemini or GPT.

I don't know why people always imply that "the bubble will burst" means that "literally all Ai will die out and nothing will remain that is of use". The Dotcom bubble didn't kill the internet. But it was a bubble and it burst nonetheless, with ramifications that spanned decades.

All it really means when you believe a bubble will pop is "this asset is over-valued and it will soon, rapidly deflate in value to something more sustainable" . And that's a good thing long term, despite the rampant destruction such a crash will cause for the next few years.

Eventually the curves cross. Eventually the computer you can get for, say, $2000, becomes able to run the best models in existence.

The only way this doesn’t happen is if models do not asymptote or if computers stop getting cheaper per unit compute and storage.

This wouldn’t mean everyone would actually do this. Only sophisticated or privacy conscious people would. But what it would mean is that AI is cheap and commodity and there is no moat in just making or running models or in owning the best infrastructure for them.

Google was built on the shoulders of a lot of infrastructure tech developed by former search engine giants. Unfortunately the equity markets decided to devalue those giants instead of applaud them for their contributions to society.

Try “@gmail” in Gemini

Google’s surface area to apply AI is larger than any other company’s. And they have arguably the best multimodal model and indisputably the best flash model?

As a loan officer in Japan who remembers the 1989 bubble, I see the same pattern. In the traditional "Shinise" world I work with, Cash is Oxygen. You hoard it to survive the inevitable crash. For OpenAI, Cash is Rocket Fuel. They are burning it all to reach "escape velocity" (AGI) before gravity kicks in.

In 1989, we also bet that land prices would outrun gravity forever. But usually, Physics (and Debt) wins in the end. When the railway bubble bursts, only those with "Oxygen" will survive.

I don't expect AGI or Super intelligence to take that long but I do think it'll happen in private labs now. There's an AI business model (pay per token) that folks can use also.

- take your data

- make a model

- sell it back to you

Eventually all of the available data will have been squeezed for all it's worth the only way to differentiate oneself as an AI company will be to propel your users to new heights so that there's new stuff to learn. That growth will be slower, but I think it'll bear more meaningful fruit.

I'm not sure if today's investors are patient enough to see us through to that phase in any kind of a controlled manner, so I expect a bumpy ride in the interim.

If we consider "AI" to be the current LLM and ImageGen bubble, I'm not sure we can say that.

We were all wowed that we could write a brief prompt and get 5,000 lines of React code or an anatomically questionable deepfake of Legally Distinct Chris Hemsworth dancing in a tutu. But once we got past the initial wow, we had to look at the finished product and it's usually not that great. AI as a research tool will spit back complete garbage with a straight face. AI images/video require a lot of manual cleanup to hold up to anything but the most transient scrutiny. AI text has such distinct tones that it's become a joke. AI code isn't better than good human-developed code and is prone to its own unique fault patterns.

It can deliver a lot of mediocrity in a hurry, but how much of that do we really need? I'd hope some of the post-bubble reckoning comes in the form of "if we don't have AI to do it (vendor failures or pricing-to-actual-cost makes it unaffordable), did we really need it in the first place?" I don't need 25 chatbots summarizing things I already read or pleading to "help with my writing" when I know what I want to say.

They only lasted a couple of decades as the main transportation method. I'd say the internal combustion engine was a lot more transformative.

I think we will end up with market similar to cloud computing. Few big players with great margins creating cartel.

This comparison keeps popping up, and I think it's misleading. The pace of technology uptake is completely different from that of railroads: the user base of ChatGPT alone went from 0 to 200 million in nine months, and it's now- after just three years- around 900 million users on a weekly basis. Even if you think that railroads and AI are equally impactful (I don't, I think AI will be far more impactful) the rapidity with which investments can turn into revenue and profit makes the situation entirely different from an investor's point of view.

I think this conflates together a lot of different types of AI investment - the application layer vs the model layer vs the cloud layer vs the chip layer.

It's entirely possible that it's hard to generate an economic profit at the model layer, but that doesn't mean that there can't be great returns from the other layers (and a lot of VC money is focused on the application layer).

And among them the overwhelming majority of companies in the sectors died. Out of the 2000ish car-related companies that existed in 1925 only 3 survived to today. And none of those 3 ended up a particularly good long term investment.

I may add that investors are mostly US-centric, and so will the bubble-bursting chaos that ensues.

I think this is analysis is too surface level. We are seeing Google Gemini pull away in terms of image generation, and their access to billions of organic user images gives them a huge moat. And in terms of training data, Google also has a huge advantage there.

The moat is the training data, capital investment, and simply having a better AI that others cannot recreate.

I don't see how Google doesn't win this thing.

As stated in TFA, this simply has not been demonstrated , nor are there any artifacts of proof. It's reasonable to suspect that there is no special apparatus behind the curtain in this Oz.

From TFA: "One vc [sic] says discussion of cash burn is taboo at the firm, even though leaked figures suggest it will incinerate more than $115bn by 2030."

Why "soon"? All your arguments may be correct, but none of them imply when the pending implosion will happen.

The other, highly invested, companies (if openai and anthropic) may be in for a free fall.

You never wake to be left in the wake of "the next big thing".

Have you thought that there was a massive physical infrastructure left behind by the original railroad builders, all compatible with future vehicles? Other companies were able to buy the railroads for low prices and use.

Large Language Models change their power consumption requirements monthly, the hardware required to run them is replaced at a rapid rate too. If it were to stop tomorrow, what would you be left with? Out of date hardware, massively wasted power, and a gigantic hole in your wallet.

You could argue you have the blueprints for LLM building, known solutions, and it could all be rebuilt. The thing is, would you want to rebuild, and invest so much again for arguably little actual, tangible output? There isn't anything you can reuse, like others that came after could reuse the railroads.

Practically, what I'm finding is that whenever I ask Claude to search stuff on Reddit, it can't but Gemini can. So I think the practical advantages are where certain organizations have unfair data advantages. What I found out is that LLMs work a lot better when they have quality data.

There is tons of money to be made at the application layer, and VCs will start looking at that once the infrastructure layer collapses.

Here's a blog post I wrote about that: https://parsnip.substack.com/p/models-arent-moats

We just don't know who will win in which area yet. It doesn't mean there is no moat.

I'll be sad when $20 a month Claude goes away.

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.

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.

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 entertainment industry is by far the easiest way to tap into global discretionary income.

[1] https://arxiv.org/pdf/2509.20328

[2] https://deepmind.google/blog/genie-3-a-new-frontier-for-worl...

[3] https://arxiv.org/pdf/2509.24527

You could imagine an entirely new cultural engine where entire genres are born off of random reddit "hey have you guys every considered" comments.

However, the practical reality seems to be that you get tick toc style shorts that cost a bunch to create and have a dubious grasp on causality that have to compete with actual tick toc, a platform that has its endless content produced for free.

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.

>whereas OpenAI has a clear opportunity with advertising.

Personally, having "a clear opportunity with advertising" feels like a last ditch effort for a company that promised the moon in solving all the hard problems in the world.

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 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

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.

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...

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.

This isn't really accurate.

Firstly, GPT4.5 was a new training run, and it is unclear how many other failed training runs they did.

Secondly "all subsequent models are routing systems and prompt chains built on top of 4" is completely wrong. The models after gpt4o were all post-trained differently using reinforcement learning. That is a substantial expense.

Finally, it seems like GPT5.2 is a new training run - or at least the training cut off date is different. Even if they didn't do a full run it must have been a very large run.

https://www.theinformation.com/articles/openai-says-business...

https://epoch.ai/blog/training-compute-of-frontier-ai-models...

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.

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.

GPT5.2 Codex is the best coding model right now in benchmarks. I use it exclusively now.

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

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.

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.

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?

Hmm, what about member-owned coöperatives? Like what we have for stock exchanges.

They need a better marketing strategy.

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.

  There is a hardware cost bubble though. I'm betting OpenAI will get a lot more bang for its buck in terms of hardware by 2030. It won't be NVidia taking most of that revenue.

With more compute, you can train better models, serve them to more users, serve them faster. The more users, the more compute you can buy. It's a run away cycle. We're seeing only 3 (4 if you count Meta) frontier LLM providers left in the US market.

Nvidia's margins might come down by 2030. It won't stay in the 70s. But the overall market can expand quicker than Nvidia's profits shrink so that they can be more profitable in 2030 despite lower market share.

Is it necessary to a point you want to make?

You can just point to behavior of a given entity, such as to conclude it's untrustworthy, without the problematic area of armchair psychoanalysis.

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...

Banks get bailed out because if confidence in the banking system disappears and everyone tries to withdraw their money at once, the whole economy seizes up. And whoever is Treasury Secretary (usually an ex Wall Street person) is happy to do it.

I don't see OpenAI having the same argument about systemic risk or the same deep ties into government.

Banks needed bailout to keep lending money. Auto industry needed one to keep employing lot of people. AI doesn't employ that many.

I just don't believe bailout can happen before it is too late for it to be effective in saving the market.

Elon owns a competitor and bought the White House.

But it might mean that LLMs don't really improve much from where they are today, since there won't be the billions of dollars to throw at training for small incremental improvements that consumers mostly don't care to pay anything for.

People take old things for granted often. Explains the Coolidge effect, and why plenty of people cheat.

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.