Nvidia Expects To Sell 'At Least' $1 Trillion In AI Chips By 2028 (techcrunch.com) 43
An anonymous reader quotes a report from TechCrunch: Nvidia CEO Jensen Huang threw out a lot of numbers -- mostly of the technical variety -- during his keynote Monday to kick off the company's annual GTC Conference in San Jose, California. But there was one financial figure that investors surely took notice of: his projection that there will be $1 trillion worth of orders for Nvidia's Blackwell and Vera Rubin chips, a monetary reflection of a booming AI business.
About an hour into his keynote, Huang noted that last year Nvidia saw about $500 billion in demand for its Blackwell and upcoming Rubin chips through 2026. "Now, I don't know if you guys feel the same way, but $500 billion is an enormous amount of revenue," he said. "Well, I'm here to tell you that right now where I stand -- a few short months after GTC DC, one year after last GTC -- right here where I stand, I see through 2027, at least $1 trillion."
About an hour into his keynote, Huang noted that last year Nvidia saw about $500 billion in demand for its Blackwell and upcoming Rubin chips through 2026. "Now, I don't know if you guys feel the same way, but $500 billion is an enormous amount of revenue," he said. "Well, I'm here to tell you that right now where I stand -- a few short months after GTC DC, one year after last GTC -- right here where I stand, I see through 2027, at least $1 trillion."
Nobody can afford them (Score:1)
To who? Consumers can't afford them, the prices of tech is going up. Businesses can't afford them, they can't justify additional costs without proof of return.
Other businesses will buy them? And do what? Nobody is doing anything significantly profitable and novel with AI.
Re: (Score:3, Insightful)
"Nobody goes there anymore. It's too crowded." -Yogi Berra
Re: (Score:2)
Re: Nobody can afford them (Score:2)
Like every other bubble not deflated on time, eventually the government will foot the bill, not a big secret.
Re:Nobody can afford them (Score:4, Funny)
To the hyperscalers. Who then rent them to the AI companies that are losing hundreds of billions of dollars as fast as they can, and surely won't by unable to pay their bills.
Re: (Score:2)
Money upfront. Let the "investors" pay the bills.
Re:Nobody can afford them (Score:5, Insightful)
The AI companies will pay with the money they don't have to put in the datacenters that haven't been built.
Re: (Score:2)
The AI companies will pay with the money they don't have to put in the datacenters that haven't been built.
Yes, the AI customers don't have the money for all the data centers. Even the money-rich hyperscalars don't have the money and have to borrow. However, these hyperscalars are self-funding for the most part and borrowing is only for around 10-20% of their spending. They are being stretched, but a case could be made that they can "handle" the financial strain. Microsoft, Alphabet, Amazon, and Meta have a combined annual profit of over $350 billion, and their actual operating cash flow is over $500 billion
Power plants? (Score:2)
Where are the power plants for these?
Re: (Score:2)
If the big AI companies can't afford them, it's fine, nVidia will give them money they can use to buy the GPUs.
Re: (Score:2)
If the big AI companies can't afford them, it's fine, nVidia will give them money they can use to buy the GPUs.
Aren't we already seeing this - nVidia provide investment capital to companies ($100B to OpenAI for instance) they can buy nVidia AIPUs?
It's like McDonalds requiring franchisees to build McDonalds restaurants on land rented to them by McDonalds - one way or (and) another they'll get their ROI.
Re: (Score:2)
Re: (Score:2)
Why GPUs? (Score:3)
Serious question, why haven't they architected something better than GPUs for running inference? Surely something specifically designed for the task that could do it faster using less power? Something like Groq ASIC (that's just one I've heard of). Why aren't these the future and eclipsing the stop-gap that is GPUs because they already existed and were the best fit at the time?
Re:Why GPUs? (Score:5, Informative)
The datacenter "GPUs" at this point have been specifically designed for the task.
The B300 is mostly dedicated to FP4. The only use case for 4-bit floating point is AI. If you want VDI or non-AI use, you want something other than a B300.
Re: (Score:3)
Eh, I wouldn't say it's "mostly dedicated" to FP4. It still works fine with everything up to FP32. Only FP64 gets thrown under the bus, via emulation.
Re: (Score:3)
Sure, it has INT32 and FP32, but the lion's share is devoted to being a FP4 beast.
I haven't personally tried to touch a B300 FP64 wise, because I was told even the 'emulation' was a no-go as the emulation still banked on *some* real FP64 units to work, and B300 has zero. So B200 with FP64 emulation leveraging some of the real FP64 was what I was told was the best to hope for in FP64 from nVidia's line.
Re: (Score:3)
True, this [glennklockwood.com] does make it seem pretty terrible at 32 bits as well. I hadn't realized that "TF32" was really a 19-bit format, and that's the biggest data type that actually runs well.
Re: (Score:2)
The beastliness of the B300 in FP4 is for compute.
It is, like all large model inference hardware, strictly limited by its memory bandwidth for inference.
That's why the B200 and the B300 get nearly identical t/s inference (obviously not on ttft- since the B300 does indeed have double the compute performance for FP4) - because they've got the same HBM3e @ 8TB/s, and you just can't cache data that is read sequentially.
Re: (Score:2)
In fact, "GPUs" are evidently whatever nVidia wants them to be.
nVidia has been obnoxiously pushing the narrative that they invented the very first GPU ever in 1999. Despite not being the first or even the first to use the acronym, but their marketing picked GPU to describe the Geforce 256 and obnoxiously the marketing groups selection of an acronym that Sony used earlier is the biggest thing retained in their "history" from that era.
Re: (Score:2)
I presume Google's TPU would be classed as dedicated. There are others, like the mind bogglingly large Cerebras WSE series, but most don't get a lot of interest.
Re: (Score:2)
Serious question, why haven't they architected something better than GPUs for running inference? Surely something specifically designed for the task that could do it faster using less power? Something like Groq ASIC (that's just one I've heard of). Why aren't these the future and eclipsing the stop-gap that is GPUs because they already existed and were the best fit at the time?
The answer is that everyone is already doing exactly what you said. Groq is now essentially Nvidia, so even Nvidia is expanding their product portfolio. They offer GPUs (Blackwells and Vera Rubins), inference systems (Groq), CPUs (Vera), and networking (including InfiniBand and ethernet, where Nvidia now has higher data center networking revenue than Cisco). Currently about 70% of Nvidia data center revenue is from GPUs, and that percentage will drop when the Groq systems ramp up.
Re: (Score:2)
You are using models gigabytes to terabytes in size. It's an uncacheable workload.
This means the actual bottleneck to inference, is the speed of the RAM.
Your question would better be posed as: Why haven't the architected better RAM than X.... and the answer is, they are.
I only care about an RTX6070 (Score:4, Insightful)
Re: (Score:2)
No, you are now supposed to just rent remote GPU time from nVidia instead.
Re: (Score:2)
Why can't the loyal customers do something for their loyal supplier? You paid money, and got a graphics card.
Re: (Score:2)
Or maybe it just seems that way to me because I'm an impatient human.
Oh BALLS! I just checked. 2nd half
Have you seen Evita? (Score:2)
Re: (Score:2)
But if Jensen is reading this, hey, I'm open to it if you are!
Re: (Score:2)
But that's my opinion, I won't tell you not to enjoy something you like. Have at it. Or don't. I don't get to tell you what to do.
Re: (Score:2)
Now would be the time for Intel to step into the market. They already have the Arc line.
Re: (Score:2)
The real question is (Score:2)
Re: (Score:2)
nah they are going to sell 3 for the trillion dollars
1st is to fuckerburg for the new AR glasshole anal plug
2nd is to bezos to create minime ai bots for his own self pleasure
3rd is for elmo so he can say he is a leet gamer
profit
Frenzy (Score:2)
Re: (Score:2)
Owned hardware means potential for locally run AI, which can be open source and therefore no subscription costs for use. I'm not just talking PCs creating stuff, or a calendar wall display summarising my schedule, but robots performing household tasks.
Very likely (Score:2)
" last year Nvidia saw about $500 billion in demand for its Blackwell and upcoming Rubin chips through 2026"
Very good revenue, and then this next generation already available that is "3.5x faster than the Blackwell architecture on model-training tasks and 5x faster on inference tasks". If you are building AI data centers this translates to very serious money. The Rubin chip architecture could easily be worth twice the price of Blackwell.
Dr. Evil (Score:2)
Did he do the pinky-to-corner-of-mouth thing when he said it?
"One TRILLION dollars" /pinky