News Nvidia's next-gen Blackwell AI GPUs to cost up to $70,000 — fully-equipped servers range up to $3,000,000: Report

[Admin]

HSBC analysts estimate that Nvidia will significantly increase pricing of next-generation Blackwell processors for AI and HPC.

Nvidia's next-gen Blackwell AI GPUs to cost up to $70,000 — fully-equipped servers range up to $3,000,000: Report : Read more

 

[Giroro]

Has anybody, other than Nvidia, figured out how to make money with one of these things?

 

[Lucky_SLS]

^ They will do that when they figure out how to be competitive with Nvidia's offerings...

 

[spongiemaster]

^ They will do that when they figure out how to be competitive with Nvidia's offerings...

I don't think he meant the competition, I think he meant the customers. How are companies using these systems for AI tasks to generate revenue.

 

[Lucky_SLS]

I don't think he meant the competition, I think he meant the customers. How are companies using these systems for AI tasks to generate revenue.

When they learn how to be competitive with OpenAI I guess XD

Those GB200 chips are definitely not marketed towards budget startup AI companies.

 

[palladin9479]

Has anybody, other than Nvidia, figured out how to make money with one of these things?

Yes and no, it depends what market you are in. While everyone is busy drooling over a better chatbot, the real money is in data analytics and forecasting. I've said it before but all these "AI" algorithms are doing is predicting "yes and". The "training" is just analyzing ridiculous amounts of data and building a model that correlates every data point with every other data point to form a giant probability mode. Afterward you can feed it a string of data elements and it'll predict what the next elements should be. You say [F,B,E,H] and it responds [A,Q,B,D] as the most likely following elements. The entire thing is just a pattern prediction system, in the case of natural language processing (NPL), which is what ChatGPT does, the patterns are in our English language lexicon.

Now imagine feeding it the daily/hourly percentile price change of the S&P 500 for every day going back thirty or more years as "Training" data. Then putting in this weeks daily/hourly price and letting it "guess" next weeks price values. See how accurate it was, adjust the model, feed more data, then do another estimation run. Keep doing this until it starts to accurately predict the patterns. Congrats you can now almost print money.

 

[umeng2002_2]

Has anybody, other than Nvidia, figured out how to make money with one of these things?

Take grant money from governments for research and sell half-working AI technology to other companies.

 

[neojack]

it's just the same as the internet bubble back in the 90's
remember AOL ? no me neither

major players will have crashed on the ground when the dust will settle.

It's a gold rush, and NVidia sells the shovels....

 

[ekio]

LOL what a plain sight scam…

 

[slightnitpick]

Now imagine feeding it the daily/hourly percentile price change of the S&P 500 for every day going back thirty or more years as "Training" data. Then putting in this weeks daily/hourly price and letting it "guess" next weeks price values. See how accurate it was, adjust the model, feed more data, then do another estimation run. Keep doing this until it starts to accurately predict the patterns. Congrats you can now almost print money.

Really bad idea. You'd also want to feed it the real-time news feed that the traders were trading on back in the day. As well as any other particulars that impact buying and selling on the market that you can think of and that you have means of getting real-time or close to real-time data of today. Random events such as obituaries of company owners, cyclical events such as benefits open enrollment, et cetera.

You'll still miss some black swan events, but at least you won't be trading algorithmically based solely on prior trades. That sort of trading is already locked up, and isn't going to be realistically improved with a modern AI engine on modern hardware.

 

[Dntknwitall]

To me this is all just a cash grab. It is marketed as the future technology that has value just like all past technology has, and it gets companies like Nvidia filthy Rich then they just keep pumping out new ideas that sell to keep generating flow in a technology that they master and everyone wants. The world is just a game of Monopoly, everyone is trying to get rich but only a few really get there. This is what the meaning of life has become and honestly the future looks really boring and I think that is why major powers are trying to reset the balance, but some (the evil ones) just want to balance it towards themselves. Great future the next generations have to look forward to.

 

[jp7189]

Has anybody, other than Nvidia, figured out how to make money with one of these things?

Maybe not directly, and I can't say the efficiency improvements are worth it either, but I can give some neat examples.

Signal propagation of cell towers in complex (city) environments to suggest optimization of locations and antenna properties.

Error correction of very long distance signals - predicting what damaged bits should have been. Many signals work with "good enough" and don't need perfect. Machine learning is great for that.

Optimizing parameters for experiments where direct observation is hard. E.g. in artificial diamond growth chambers with high heat and pressure.

 

[jp7189]

I do wonder about the total addressable market for these things. I would expect the average business to rent time rather than buying outright... which I guess makes this profitable for cloud providers to buy these, but how many do they really need?

 

[palladin9479]

Really bad idea. You'd also want to feed it the real-time news feed that the traders were trading on back in the day. As well as any other particulars that impact buying and selling on the market that you can think of and that you have means of getting real-time or close to real-time data of today. Random events such as obituaries of company owners, cyclical events such as benefits open enrollment, et cetera.

You'll still miss some black swan events, but at least you won't be trading algorithmically based solely on prior trades. That sort of trading is already locked up, and isn't going to be realistically improved with a modern AI engine on modern hardware.

That wouldn't work, "AI" isn't some magical quantum aether cloud enabled blockchain powered magic fairy dust. What we call "AI" is nothing but algorithmic datapoint analysis. If you read my previous post I explained how all it's doing is predicting "what next" based on "what happened before".

 

[slightnitpick]

That wouldn't work, "AI" isn't some magical quantum aether cloud enabled blockchain powered magic fairy dust. What we call "AI" is nothing but algorithmic datapoint analysis. If you read my previous post I explained how all it's doing is predicting "what next" based on "what happened before".

When it comes to trading the point is to either look at fundamentals and invest for the long term (e.g. Berkshire Hathaway's strategy) or to get in front of other's expected trades so as to profit off of their buying or selling.

You don't need magic fairy dust for that, you need fast computation and salient data. The algorithmic trading based on prior trades that you were talking about has already been optimized with data analysis approaches, and has serious holes as shown by Long-Term Capital Management's implosion. Doing this with the newest, expensive AI machines isn't going to net you much, and will hurt you in the long run.

Looking at the news or chatter that causes people to trade, and speculating (e.g, your improv "yes and" AI strategy) on how stocks and bonds will react to that news, would be a more viable strategy. At least until others start doing it.