News Nvidia's business model 'may fall apart' if corporations can't make AI pay, says SK Group boss

[Admin]

SK Group Chairperson Chey Tae-won says that Nvidia may lose its place among the top corporations globally if AI technologies fail to guarantee profits in a couple of years. However, Nvidia's competitors pose a bigger threat if they develop chips that could perform just as well but at a lesser cost.

Nvidia's business model 'may fall apart' if corporations can't make AI pay, says SK Group boss : Read more

 

[ET3D]

NVIDIA's business model is likely to fall apart in any case. AI inference, and even training, can be done more efficiently than with a general purpose accelerator, and is easier to create. Although NVIDIA has a head start, it might not be able to hold it against the competition.

 

[bit_user]

NVIDIA's business model is likely to fall apart in any case. AI inference, and even training, can be done more efficiently than with a general purpose accelerator, and is easier to create.

Certainly not more efficiently. You get greater efficiency by going in the direction of specialization, not generalization.

Although NVIDIA has a head start, it might not be able to hold it against the competition.

I'm waiting for them to drop features like fp64, which you really don't need for the vast majority of AI use cases. They could also cut way back on fp32. If they streamlined their current architecture to be better tuned specifically for AI, I think they shouldn't have much trouble remaining competitive. The only remaining question is whether they could maintain such substantial margins.

 

[bit_user]

AI wasn't even around when Jensen Huang thought of Nvidia.

Are you serious, bro? Not only does the field of AI go back to the 1940's (most prominently highlighted in the writings of Alan Turing), but even the earliest work on artificial neural networks dates back about that far!

https://en.wikipedia.org/wiki/History_of_artificial_intelligence
​

By 1982, Hopfield nets were drawing new interest to digital artificial neural networks. That was some 10 years before Nvidia's founding.

Yeah, people didn't start talking about harnessing graphics hardware for other computational problems until the late 1990's (a practice that came to be known as GPGPU), and therefore Nvidia's founders wouldn't even have considered it, but that's a far narrower claim than AI being some recent invention.

Wow, that was sure a whopper!

 

[Pierce2623]

Certainly not more efficiently. You get greater efficiency by going in the direction of specialization, not generalization.


I'm waiting for them to drop features like fp64, which you really don't need for the vast majority of AI use cases. They could also cut way back on fp32. If they streamlined their current architecture to be better tuned specifically for AI, I think they shouldn't have much trouble remaining competitive. The only remaining question is whether they could maintain such substantial margins.

You do realize he literally said the small matrix math accelerators they’re adding to everything these days are MORE efficient than a GPU not less, right?

 

[Pierce2623]

While it’s not nearly as profitable as commercial accelerators, Nvidia can always fall back on owning 80% of the discrete GPU market. When Nvidia will really be in trouble in when x86 systems go mostly in the APU direction. I mean who would buy a $500 GPU and a $300 CPU over a $500-$600 APU that’s competitive on both fronts?

 

[bit_user]

You do realize he literally said the small matrix math accelerators they’re adding to everything these days are MORE efficient than a GPU not less, right?

Who said? I don't see that phrase in either the article or the comments.

Also, I disagree with that assertion. I've looked at AMX benchmarks, if that's what you've got in mind, but go ahead and show us what you've got.

 

[ThomasKinsley]

I'm hearing mixed signals on this topic. On the one hand, consumer AI is supposedly a hot new market with phones supporting AI garnering increased sales. On the other hand, these sales increases still pale in comparison to pre-covid phone sales. I would characterize this as wildly successful for businesses with only mild interest among consumers.

 

[watzupken]

NVIDIA's business model is likely to fall apart in any case. AI inference, and even training, can be done more efficiently than with a general purpose accelerator, and is easier to create. Although NVIDIA has a head start, it might not be able to hold it against the competition.

While I am not a fan of Nvidia, I feel if AI continues to boom, Nvidia will continue to thrive. You can tell from retail consumer's behavior where AMD hardware may offer a better bang for buck, but people tend to prefer to pay more for a lesser performant product just because of RT, DLSS, and mainly the perception that Nvidia products are better. So competition will chip some sales away, but I feel most people and companies will stick with Nvidia.

Having said that, the comment from the SK group boss is common sense. At this point, the AI master race is going on, but the real question is how they will be monetizing AI services, and how receptive will consumers be to a paying model? Again, OpenAI/ ChatGPT became widely popular because its free. Switching to a paying model (and likely one that is not cheap), is going to mean that only a minority of people will be willing to use the paid for service. But will that be enough to recoup the sunk and running cost of training and maintaining? I think it is not a matter of whether the AI bubble will implode, but rather when.

 

[bit_user]

Switching to a paying model (and likely one that is not cheap), is going to mean that only a minority of people will be willing to use the paid for service.

A lot of companies are already paying for it. They have an incentive not to use free services, because they want to protect their IP from leaking out, as well as preventing their IP from becoming tainted by external IP being fed to their employees via AI. So, this requires models that are trained on carefully vetted material, as well as a guarantee not to use queries or other metadata in training of models shared with other customers.

But will that be enough to recoup the sunk and running cost of training and maintaining? I think it is not a matter of whether the AI bubble will implode, but rather when.

Corporate users pay for lots of services. Office suite, email, cloud storage, github, bookkeeping software, customer relationship tools, etc. The real question is how much value they can derive from it, because that will ultimately determine how much they're willing to pay.

 

[thestryker]

I think it's mostly just a situation for as the market exists nvidia is significantly overvalued as part of the current AI bubble. The question is certainly whether monetization is figured out or the bubble bursts. Nvidia making moves with the Grace line ought to also keep them going even if the AI rush collapses as they should be able to be used for HPC.

Of course there's also the dominance in consumer discrete graphics as it stands today, but we'll see what Strix Halo ends up looking like.

Nvidia has had to deal with two crypto booms so I'm sure management is doing everything they can to minimize risk here. Just a wild guess but I'd assume their biggest additional cost right now is just in sheer volume of hardware manufacturing and that they're not doing a ton of additional hiring.

 

[hotaru251]

he is right...the only reason ai sells is casue its "fomo" and everyone is tossing it on stuff.

ai has no real profitability though to average joes. While the companies spending $ to advance it on the hope it will pay out in future is beneficial to the industry (for good and bad) it has no future as a profit genre.

Even in gaming and the so called "ai npc's" that react and generate stuff as you communicate is niche and 99% of players wouldnt care much less pay for it.

ai/llm's training and "theft" (its theft as its not paying for content & many are using it for profit reasons) is still a toss up until it gets into court (and if llm's lose that would effectively kill any grain of profit for the companies pushing it so hard and start losing them more $)

 

[wr3zzz]

Are you serious, bro? Not only does the field of AI go back to the 1940's (most prominently highlighted in the writings of Alan Turing), but even the earliest work on artificial neural networks dates back about that far!

​

https://en.wikipedia.org/wiki/History_of_artificial_intelligence​

​

By 1982, Hopfield nets were drawing new interest to digital artificial neural networks. That was some 10 years before Nvidia's founding.

Yeah, people didn't start talking about harnessing graphics hardware for other computational problems until the late 1990's (a practice that came to be known as GPGPU), and therefore Nvidia's founders wouldn't even have considered it, but that's a far narrower claim than AI being some recent invention.

Wow, that was sure a whopper!

"AI" has gone through a few boom-bust cycles. I was in grad school during the nadir of last AI bust in the early 90s. IBM had closed its NYC mid-town AI public showcase and gave the space to Sony. Researchers and marketers avoided the word AI like a plague for fear of being laughed at. Instead, the discipline was renamed to its many branches of specialties like fuzzy logic, stochastic scheduling, neural network etc. Like in previous cyclical downturns engineering continued to progress just under different names. This current cycle started rebounding when IBM Deep Blue beat Kasprov in chess in 1997 and the word AI started getting acceptable but mega corps like Facebook and Tencent still referred to their evil machinations as algorithms or machine learning. When IBM Watson won Jeopardy in 2011 that's when rebranding started snowballing. Now everything is AI again.

We are probably near the peak of current AI cycle and when it goes bust just like previous cycles the word AI likely will be avoided again. The current tech-bros think they gave birth to AI because none were even born the last time "AI" was viewed as a glory discipline. This bubble will burst because we don't have the computing power price-performance to do what the tech-bros are selling to the public. We have been through this in the early 60s and again in the late 80s.

 

[slightnitpick]

Yeah, people didn't start talking about harnessing graphics hardware for other computational problems until the late 1990's (a practice that came to be known as GPGPU), and therefore Nvidia's founders wouldn't even have considered it, but that's a far narrower claim than AI being some recent invention.

From one of NVIDIA's founding employees: https://blog.dshr.org/2024/07/accelerated-computing.html

I don't disagree with what Huang said. Despite the need to focus on gaming, we did have a vague idea that in the future there would be other application areas in which custom accelerators could make an impact. And it is true that Nvidia's VCs, Sutter Hill and Sequoia, gave us the time to develop a multi-generation architecture rather than rushing out a "minimum viable product". I do quibble with the idea that this was "genius foresight".

 

[ttquantia]

There is a very strong and completely unfounded belief in the IT industry that the same technologies that make it possible to generate images and text are going to solve a large number of actual real AI problems, which have not been solved so far, even partially.
The thinking is, "if AI can generate realistic video, then AI can control a humanoid robot, replace human programmers, and fly an airplane". This thinking is wrong, flawed, and it will be catastrophic to the current "AI" as soon as a large fraction of the people responsible for the current massive over-investment in AI realize that none of this is going to happen.
We are now seeing the first stages of the current "AI" death throes with Microsoft attempting to push "AI" in their products. Some of this stuff is nice and cute, but it is in reality only a silly fad that will not survive long.

 

[JRStern]

NVidia's "business model" is just "give the customer what they want", and don't ask a bunch of questions about why they want it. Just like the old pickaxe companies must have been pleased when someone ordered 300,000 pickaxes, ROFLMAO.
Selling aspirations.
Sell the sizzle not the steak.
Congratulations you have just earned your Marketing 101 certificate!

 

[KyaraM]

While it’s not nearly as profitable as commercial accelerators, Nvidia can always fall back on owning 80% of the discrete GPU market. When Nvidia will really be in trouble in when x86 systems go mostly in the APU direction. I mean who would buy a $500 GPU and a $300 CPU over a $500-$600 APU that’s competitive on both fronts?

It will likely be a while until we are at that point, though. Memory bandwidth is a real issue here, as is size and other factors. So far, desktop APUs didn't exactly wow me, and I will be honest, while Strix Halo looks great on paper, I highly doubt it will be able to actually compete with a 1070, as suggested by computing benchmarks. And by today's standards, a 1070 isn't a great card anymore, as much as it hurts me to say that. It's my personal GOAT of the cards I had so far. Also, it's not as if Nvidia is completely unable to enter that market; they have the money, and let's not forget that they already developed a (handheld) CPU before, and it looks like that one will soon get a successor. Then they also have CUDA and other tech that is deeply entrenched in professional workloads. Yeah, no, I doubt they will be in major trouble anytime soon...

 

[thestryker]

So far, desktop APUs didn't exactly wow me, and I will be honest, while Strix Halo looks great on paper, I highly doubt it will be able to actually compete with a 1070, as suggested by computing benchmarks.

Just a note you're referring to Strix Point, not Strix Halo. If the leaks regarding Strix Halo are to be believed it will have a 256-bit bus and exclusive cache for the GPU which should put it at least at the 7600/4060 range. We're also a long way from this being standard, but if Strix Halo parts happen to be successful perhaps we'll see more of it.

I definitely agree on Strix Point though it ought to be faster than what we have now, but probably not significantly so which is what would be required to reach even the 1070.

 

[bit_user]

Nvidia making moves with the Grace line ought to also keep them going even if the AI rush collapses as they should be able to be used for HPC.

Compared to their current AI revenues, that market is tiny.

 

[thestryker]

Compared to their current AI revenues, that market is tiny.

Absolutely, but I meant more in the context of them being able to move existing product. Grace gives them a more viable exit plan for moving stock if there's a collapse. In theory the biggest threat to the health of the company would be if nvidia was left holding the bag on hardware they couldn't sell.

 

[Pierce2623]

It will likely be a while until we are at that point, though. Memory bandwidth is a real issue here, as is size and other factors. So far, desktop APUs didn't exactly wow me, and I will be honest, while Strix Halo looks great on paper, I highly doubt it will be able to actually compete with a 1070, as suggested by computing benchmarks. And by today's standards, a 1070 isn't a great card anymore, as much as it hurts me to say that. It's my personal GOAT of the cards I had so far. Also, it's not as if Nvidia is completely unable to enter that market; they have the money, and let's not forget that they already developed a (handheld) CPU before, and it looks like that one will soon get a successor. Then they also have CUDA and other tech that is deeply entrenched in professional workloads. Yeah, no, I doubt they will be in major trouble anytime soon...

Strix Halo is supposed to perform along the lines of laptop 4070. Yeah bandwidth is likely to be the limiting factor but it will have nearly as much bandwidth as 6600xt so it won’t be a horrible limitation.(just over half the actual memory speed but a bus that’s twice as wide makes it up. It’s sLao expected to have an Infinity Cache which will go a long way towards insulating from bandwidth issues at a resolution like 1080p.

 

[ET3D]

"if AI can generate realistic video, then AI can control a humanoid robot, replace human programmers, and fly an airplane". This thinking is wrong, flawed, and it will be catastrophic to the current "AI" as soon as a large fraction of the people responsible for the current massive over-investment in AI realize that none of this is going to happen.

AI has been controlling robots for a few years now, and is already doing quite a bit of work for programmers. I'm not sure how much of it exists in airplanes, but they pretty much already run themselves.

AI is pretty good at replacing other forms of heavy computing and also at replacing humans. AI is as good as the best doctors at detecting some things.

While there are a lot of misconceptions, some of them are with people like you, who just don't realise how much AI is being used, or, as has been done repeatedly over the years, classify anything that already works as "not AI".

 

[bit_user]

From one of NVIDIA's founding employees: https://blog.dshr.org/2024/07/accelerated-computing.html

It's hard to know exactly what to make of that quote. For all I know, he could be referring to the way the NV1 implemented multiple facets of multimedia acceleration.

"The design goal was to deliver the highest value multimedia solution on a single chip, optimized for concurrent acceleration of video game systems standards and multimedia content. Since NV integrates audio with graphics I will delve into this territory a bit as well. The engine is 350 MIPS, supporting in hardware 32 channels of 16 bit 48 kHz wavetable sound with several effects. VL-Bus and 32 bit local memory bus were planned options, but all cards I've seen used PCI and 64 bit data width. All processing is indeed happening in one chip ..."

Source: https://vintage3d.org/nv1.php
​

At this point in time, all rasterization was low-precision integer arithmetic. There were no programmable pixel shader pipelines, let alone floating point. It would've been a huge leap of faith to imagine it would even be feasible. Furthermore, to handle scientific and financial compute jobs you really want fp64, which was yet another huge leap. So, please don't try to tell me they had any notions of one day building compute engines like the vector processors used in a few supercomputers of the era. A linear path to that, from where they started, would not have been visible at the time.

 

[bit_user]

The thinking is, "if AI can generate realistic video, then AI can control a humanoid robot, replace human programmers, and fly an airplane".

Who are you quoting?

I would point to self-driving cars. If AI can drive a car, then it doesn't seem like a reach to suggest that it would be able to power a humanoid robot. Same with drones/airplanes.