Nvidia strikes AI alliance with S.Korea, pledges 260,000 GPUs worth $9.8 billion by 2030
Posted by self-fix@reddit | hardware | View on Reddit | 38 comments
From-UoM@reddit
The libraries, frameworks and software stack that Nvidia offers on top of Cuda and their GPUs are the real moat.
Stuff like Omniverse, Cosmos, Culitho, TensorRT, CuDNN, Dynamo, Nemotron, Groot, Isaac, Drive, Earth-2, Arial and a lot more.
The best or worst part (depending on your view)? You get locked in the Nvidia ecosystem with these.
cryptofreak190@reddit
OmniVerse is one of the worst things to work with. I work as a Software Engineer. My colleagues use it. They complain about it all the time. There's a huge lack of proper documentation for it. What can easily be achieved via Blender, Unity or Unreal Engine is a hide undertaking with OmniVerse. It's just a bad case of trying to re-invent the wheel.
self-fix@reddit (OP)
Both NVIDIA and the S.Korean gov't confirms that this brings S.Korea to the third spot in the number of GPUs after the US and China
Visible-Advice-5109@reddit
Kind of surprising Korea giving all that business to TSMC.
snowfordessert@reddit
Taiwan is pretty weak in software.. TSMC itself doesn't need GPUs
From-UoM@reddit
Funny thing is TSMC uses Nvidia's CuLitho software
Vb_33@reddit
Can't unsee because culito (no h) in Spanish means lil butt.
Visible-Advice-5109@reddit
These GPUs are made in Taiwan. Just found it interesting because Korea usually tries to heavily favor Samsung.
Ok-Parfait-9856@reddit
They gotta get the gpus from somewhere and there’s not many options. Samsungs fab is pretty shit regarding cutting edge nodes and yields. Plus Korea is pumping out ram chips that end up in gpus and servers. So Korea benefits too. Ram is super expensive now because it’s being bought up by ai data centers.
tooltalk01@reddit
You seem to forget where all HBMs come from.
BarKnight@reddit
There is no AI bubble, corporations and governments will be throwing money at this for a very long time
upbeatchief@reddit
There is no dotcom bubble, corporations and governments will be throwing money at this for a very long time.
Yeah there's money in the space. Like it was in early dotcom. But does it justify the building of multiple powerplants and trillions in investment?
PriscFalzirolli@reddit
Considering the cognitive labor market is worth 50-100 trillion USD a year? If you think you can eventually automate even a fraction of these tasks, it's entirely justified.
AI investment is not about convincing randos to eventually pay $10 a month for Gemini or to make silly videos with Sora.
jeffy303@reddit
I do think it will happen eventually, anyone who thinks LLMs will be the one to do it and in the next 5 years is a complete idiot.
Everyone is circlejerking about AGI, I am wondering when it will stop being a complete idiot. Even as something as simple as "recommend me movies like Matrix" prompt, even the best models end up recommending Matrix if you ask them few more times. It's a statistically plausible text, not some real intelligence. AI bros who think they'll use this to replace all office work are beyond out there.
LargeSinkholesInNYC@reddit
Automation will be a very big thing and it will require a lot of GPUs.
JigglymoobsMWO@reddit
At the height of the dot com bubble, 90% of the optical fibers in the US was dark, unused.
Today, 100% of GPUs are used 24/7
ComplexEntertainer13@reddit
Because they are throwing free AI based services out there that in no shape or form are recouping the cost.
If they had handed out free internet connections back in 1998, then all that fiber would have been sitting at capacity as well.
jpeirce@reddit
From-UoM@reddit
There is a very big difference.
The infrastructure for internet wasn't used that much when it rolled out.
Its with the iPhone that things truly kicked off.
AI data centres though are at capacity from the start.
upbeatchief@reddit
I disagree. Aren't h100 and a100 renting sites have them for like 0.50 and 1.2 dollars respectively? That's below break even. And most AI services are free. Yeah if you offer me a Corolla for free i might use it. But if i have to pay. Then i will use the bus.
AI is A bubble. like all speculative investments frenzies, The size of the bubble and the returns determine if the bubble is bad or not. So we have to see how many paying customers can AI have. And do they justify the investment.
JigglymoobsMWO@reddit
Those are already amortized.
upbeatchief@reddit
I cant imagine they are. They were offering them from the start for 2.5 dollars an hour. Even with 100% uptime they wouldn't break even.
Add in the fact that crypto and corono delayed deliveries. And not everyone could have made an order. You could see them receiving orders a year or two later. So they are offering late cards for the current low market price.
They likely are burning cash to establish themselves.and liquidating so excess gpus.
Lucie_Goosey-@reddit
Not having capacity being fully utilized after investment (dot com) versus utilization exceeding capacity but unprofitable are two different scenarios.
What this could become is simply a mandatory $50 monthly subscription by all the major players and problem solved, or companies subsidizing through alternative revenue streams.
China though could cause some serious hurt though by leaving access to their models free and open source, until they bankrupt US companies.
From-UoM@reddit
And how old is the A100? Its 5 years old and still going for a price.
If you look back even a few years ago, older GPUs were given for free to use for limited time within a few years.
For example you can get Nvidia Tesla T4 (Turing) and sometimes P100 for free on the Google Colab for a while now. Can get older TPUs even.
But everything after Volta is still priced. Volta just happens to be the the same GPU that trained GPT3
Chipaton@reddit
They're at capacity because they haven't figured out how to monetize it yet. That's the entire bubble. They're pissing money currently to get people hooked on AI, and the monetization comes next. You can see the same phenomenon with Uber and Gamepass for example.
They're also at capacity simply because LLMs take a massive amount of resources to operate, they're essentially guaranteed to be at capacity without significant technological improvements. Unlike the internet, the infrastructure didn't exist beforehand (in a meaningful capacity) and is being built as the tech develops.
I really do not see much separating the dotcom bubble and the current AI bubble.
pwreit2027@reddit
Bill Gates gave a good interview on the question if their is an AI bubble and used the dotcom bubble as an analogy. he said if you mean it's a bubble in the sense you'll look back and think AI as a technology is a scam and will do nothing in the future then you are very much mistaken, AI will be the biggest revolution in human history. But will it be a bubble in the sense that you'll get most companies that are a scam and will lose your money then yes. Most start ups in this space will go under. Don't forget even though we had the dotcom bubble, the internet itself was not a bubble, it redefined how society functions and accelerated human civilisation. AI is just like that.
upbeatchief@reddit
Yeah it definitely feels like that. Most "invention" is just a chatgpt wrapper. I personally bet that 99% of AI companies will be gone or bought in the next 4 years. Most having lost money. But investors still think there's a next Amazon or google in this matket so they keep threwing money at it.
pwreit2027@reddit
yes most likely it will be a very big portion that will go under but those that make it will be worth trillions of dollars. I get what you are saying about people thinking their will be new big players in the space, that can overtake the multi trillion dollar companies like Amazon and Google, but why can't their be? you still get multi billion dollar companies that make products that even google can't take market share from, isn't Open AI kind of relatively new? but you won't get many new ones. AI is kind of general. you make it better and it works across lots of areas, but you can still make stupid money if you get in on a company that will become billions of dollars worth. The thing is the amount of value (and hence money) that will be generated is almost limitless . Nvidia isn't going to 100x from here anytime soon, but get in on an early company and you can easily 100x your investment. but yeah , easily 95% will go under
jeffy303@reddit
>one of the largest buyers
> 10 bil revenue
>not even guaranteed
> by 2030
> 5 trillion evaluation for single one of these bubles
Make these numbers make sense lmao. One has to be in deep denial to not see how insanely overvalued all of these companies are. Price of GPU cloud compute is plummeting rapidly already and that's before all the hundreds of GW of AI datacenters that everyone is promising to build
Plus-Candidate-2940@reddit
Keep dreaming. I hope they all go bust tomorrow.
Rencrack@reddit
True
self-fix@reddit (OP)
Exactly. When the dotcom bubble popped, Google stayed. And because Google stayed we now have ChatGPT, Android, and YouTube.
The AI financial bubble will pop, eventually. But that doesn't mean AI is going away. People think their jobs are safe just cause they believe AI is a bubble.
Visible-Advice-5109@reddit
Google bought YouTube, lol.
But anyways, nobody is saying AI will dissappear. People judt kinda expect it to be like Intel which was double what it is now 25 years ago at the height of the bubble. People will keep buying Nvidia products.. just not at this rate.
amineahd@reddit
lmao when does it stop being a bubble? companies are govs are "throwing money" and its a bubble? who else needs to throw money for you to be satisfied? aliens?
lessadessa@reddit
i want nothing to do with this personally
From-UoM@reddit
TSMC and Samsung chips are made with Nvidia's Culitho.
Most of the dram comes from Samsung and SK Hynix. Both South Korean.
Almost all Oled screens come from LG and Samsung.
You are already personally involved in this if you have a phone and pc.
chapstickbomber@reddit
Lmao 40k per card that costs like 5k, literally just moat piracy, what a joke
Spright91@reddit
"Alliance"? You mean theyre selling them GPUs.