Would that be the most expensive single thing ever built? The ISS cost around $150B and is commonly said to be the most expensive single item, but that does include running costs.
The loans will just take longer to repay. There is a market for Anthropic & OpenAI, it just likely doesn't have the 200B profit each year required for the maths to make sense.
If shit hits the fan, the companies collapse, then Nvidia gets their money from the investors anyways.
it'll be a rack based GPU without video ports. The only thing getting sick GPU's will be landfill when they all burn out and the data center rationalization happens.
More like rack-sized GPU that takes in half-megawatt of power as 800 volts DC, and requires 10 gallons per second of liquid cooling or it'll catch fire.
If Nvidia sells hardware for $100B with 75% cross margin, and provides $50 billion in backstop for that same hardware, it would be still be nicely profitable deal ($25B) if the backstop capacity would be a total write-off recovering $0. Reselling that capacity in some large discount below already low backstop price would increase the profits.
It's all those pension funds, sovereign wealth funds and Softbank getting into that $500 billion deal that will be hurt.
Pretty much, I have said it for a while now, Softbank and Oracle are the ones I would be worried about. Both of them have put their companies wealth behind this, if it goes down so will they.
Others have played it fairly smart in terms of insulating potential issues.
SoftBank has always managed to squeeze by after every mistake selling some early huge wins Alibaba, Nvidia, arm. Wonder if they still have any of those left in the back pocket.
Businesses aim to make the most profit possible with their resources. If they can make a 25% margin that is good, but if they can turn around sell thr same thing for a 50% margin, that is much better.
Basically what i am saying is maybe there is a better buyer than openai.
Uh no, NVDIA helping startups get financing so they can buy NVDIA chips is inherently damaging because eventually the debtors will not help with the financing and startups will not be able to buy chips.
A Toyota dealership arranges a loan for you. Through a bank for a used vehicle, sometimes through Toyota itself for new cars.
A house builder will routinely take on part of the loan providing burden to get some of the interest.
Even someone selling you their thirty year old house will often provide seller financing.
You may have ideological opinions against this, which is fine. There are billions of people, for example that are fundamentally opposed to the idea of interest. But like it or not, this is how it has worked for the last ~500ish years.
No, I get this; it is not a problem. It becomes one when they cannot pay back this financing, in the event that they cannot build a sustainable business, which they cannot, because the capital cycle leads to overinvestment, meaning the financiers cannot meet their returns.
This is probably a lot more related to the fact they want to make GPUs an asset class. Nvidia is banking on the fact there will be an entire market that will guarantee whatever anyone needs.
What would happen to Nvidia, Anthropic, OpenAI, if tomorrow someone released an open weights model on HuggingFace that matched performance and accuracy of Opus 5 running locally on an RTX 5070? That won’t happen tomorrow, but it will likely happen someday… what’s the plan beyond “don’t be the one holding the bags?”
There’s no reason to assume frontier-level intelligence eventually collapses all the way onto a midrange consumer GPU. In fact, there are quite a few reasons not to assume that (information-theoretic constraints, etc).
I'm pointing out that there's no known information theoretic constraint about the impossibility of frontier AI models being improved to fit/run on a small GPU.
Please do not make up plausible sounding science facts.
I will not claim a 5070, but there is already evidence in nature that you can get very good general intelligence with an order of magnitude less wattage.
There are constraints of course- training takes way longer.
But, it could happen for a coding-focused model, or an accounting-focused model, etc. most tasks only need a subset of the total model to be done effectively.
Workloads will inflate just as they have been. Remember when llm assisted development used to be good only for a function, then a whole file, then a handful of files, then a code base, then a full stack, etc etc etc.
People will claim to have “enough” even though they already have the equivalent of last years capabilities locally.
probably not all that much... the market would dip, just like every time a new open weights model gets announced. but hundreds of millions of people aren't going to immediately self-hosting their own models.
the biggest winner in that scenario would be ai providers, who suddenly have a capable model that they can serve much more efficiently. and the incumbents have a whole lot of compute. wouldn't anthropic and openAI just start offering that open weights model at prices that nobody else could compete with?
They could but then their valuation is no longer justifiable, which breaks a lot of things downstream (loans being the biggie). They'd rather lose money than start making money in a non defensible way.
Those companies will be quick to copy the tech, inference cost would plummet and there is a greater chance that these companies could make it to solvency. At least in the short term. Long term it might not be so great as consume hardware catches up.
Nothing would really change IMO? 99% of users don't have anything like a RTX5070 (mobile especially).
Even if it did, it still doesn't make much economic sense running a model locally vs on a datacentre.
For example, I managed to just about squeeze a Q2 quant of Qwen 3.7 27b on my 9070XT. I get around 60tps decode (slightly faster prefill). _but_ it uses 300W of power to do so. At UK electricity rates of 30c/kWh this works out at something like 42c/MTok. I can get far far better models on openrouter cheaper than that, plus I'm not horrendously constrained on context length.
I dunno a lot of things said about AI economics sound like an IBM executive making reassuring statements about their terminal/mainframe business before the personal computer took off.
Like even if you run it in a datacenter in this scenario, you could do it on a cheap GPU instance in Azure, you still wouldnt need OpenAI or Anthropic specific clouds.
>uses 300W of power to do so.
There are plenty of people with phat electricity pipes in their on prem server rooms that have been vacated for cloud. Companies who want the benefits of AI but dont want the risk of sending their data to foreign API endpoints.
It seems very very unlikely that an Opus 5 matching local model that runs on a 5070 will be released within the next 5 years (I don't want to say "ever").
If it does happen then NVidia will sell a lot of 5070s though!
There's a release from DoE about it: https://www.energy.gov/articles/fact-sheet-department-energy...
That's a horrible amount of gas energy generation.
https://www.datacenterdynamics.com/en/news/openai-in-talks-t... has more details. The whole campus build could be as much as $500B.
Would that be the most expensive single thing ever built? The ISS cost around $150B and is commonly said to be the most expensive single item, but that does include running costs.
If shit hits the fan, the companies collapse, then Nvidia gets their money from the investors anyways.
If Nvidia sells hardware for $100B with 75% cross margin, and provides $50 billion in backstop for that same hardware, it would be still be nicely profitable deal ($25B) if the backstop capacity would be a total write-off recovering $0. Reselling that capacity in some large discount below already low backstop price would increase the profits.
It's all those pension funds, sovereign wealth funds and Softbank getting into that $500 billion deal that will be hurt.
Others have played it fairly smart in terms of insulating potential issues.
Goose value 71 here: https://group.softbank/media/Project/sbg/sbg/pdf/ir/investor...
Basically what i am saying is maybe there is a better buyer than openai.
It is not the first time, either; the capital cycle will prevail.
https://s-1.vercel.app/posts/the-capital-cycle-theory/
You pay Apple for a MacBook, Apple uses it to develop a better MacBook.
What goes wrong is leverage. We haven't seen much hint of the 10x leverage kind of deals that brought down the house in 2008.
A Toyota dealership arranges a loan for you. Through a bank for a used vehicle, sometimes through Toyota itself for new cars.
A house builder will routinely take on part of the loan providing burden to get some of the interest.
Even someone selling you their thirty year old house will often provide seller financing.
You may have ideological opinions against this, which is fine. There are billions of people, for example that are fundamentally opposed to the idea of interest. But like it or not, this is how it has worked for the last ~500ish years.
https://s-1.vercel.app/posts/the-capital-cycle-theory/
Neither of those look likely yet.
It's almost a given that whatever is frontier intelligence today will run on a potato in a few years.
I'm pointing out that there's no known information theoretic constraint about the impossibility of frontier AI models being improved to fit/run on a small GPU.
Please do not make up plausible sounding science facts.
There are constraints of course- training takes way longer.
People will claim to have “enough” even though they already have the equivalent of last years capabilities locally.
the biggest winner in that scenario would be ai providers, who suddenly have a capable model that they can serve much more efficiently. and the incumbents have a whole lot of compute. wouldn't anthropic and openAI just start offering that open weights model at prices that nobody else could compete with?
Even if it did, it still doesn't make much economic sense running a model locally vs on a datacentre.
For example, I managed to just about squeeze a Q2 quant of Qwen 3.7 27b on my 9070XT. I get around 60tps decode (slightly faster prefill). _but_ it uses 300W of power to do so. At UK electricity rates of 30c/kWh this works out at something like 42c/MTok. I can get far far better models on openrouter cheaper than that, plus I'm not horrendously constrained on context length.
Like even if you run it in a datacenter in this scenario, you could do it on a cheap GPU instance in Azure, you still wouldnt need OpenAI or Anthropic specific clouds.
>uses 300W of power to do so.
There are plenty of people with phat electricity pipes in their on prem server rooms that have been vacated for cloud. Companies who want the benefits of AI but dont want the risk of sending their data to foreign API endpoints.
RTX 5070 prices go up ~N times. Nvidia makes more money because it's easier to make these things than it's to make a GB300.
If it does happen then NVidia will sell a lot of 5070s though!