Jevon's paradox would imply that there's good reason to think that demand for shovels will increase. AI doesn't seem to be one of those things where society as a whole will say, "we have enough of that; we don't need any more".
(Many individual people are already saying that, but they aren't the people buying the GPUs for this in the first place. Steam engines weren't universally popular either when they were introduced to society.)
I also dont get how this is bearish for NVDA. Before this, small to mid companies would give up on finetuning their own model because openai is just so much better and cheaper. Now deepseek SOTA model gives them much better quality baseline model to train on. Wouldn't more people want to RAG on top of deepseek? or some startups accountant would run the numbers and figures we can just inference the shit out of deepseek locally and in the long run we still come out ahead of using oenai api.
Either way that means a lot more NVDA hardware being sold. You still need CUDAs as rocm is still not there yet. In fact NVDA needs to churn out more CUDAs than ever.
Because their biggest buyer(s) just ran into a buzzsaw. NVDA's stratospheric valuation is based on a few select customer's unrestrained ability to purchase the latest products. That unrestrained spending ability was fueled by the "AI" arms race. If those companies see their ability to be profitable with "AI" as diminished then their ability to continue spending with NVDA is probably going to be diminished as well. Anything considered bad for NVDA's top spenders is going to be viewed as bad for NVDA as well.
At a guess, Nvidia stock prices are basically fiction at this point (are there lots of short AND long selling - IIRC butterfly spreads?).
A good fundamental analysis is probably very hard to get right, and the game is probably just guessing which way everyone else will guess the herd is going to jump.
IBM probably couldn't have done things differently given the antitrust scrutiny they were under. And that was likely the best outcome for the industry anyway.
The bear case is something like “investors are going to call BS on multi-billion training DC investments”. That represents most of their short-run demand.
Not sure what is supposed to happen to the inference demand but I guess that could be modeled as more of a long-run thing, as inference is going to be very coin-operated (companies need it to be net profitable now) whereas training is more of a build now profit later game.
Jevons was talking about coal as an input to commercial processes, for which there were other alternatives that competed on price (e.g. manual/animal labour). Whatever the process, it generated a return, it had utility, and it had scale.
I argue it doesn't apply to generative AI because its outputs are mostly no good, have no utility, or are good but only in limited commercial contexts.
In the first case, a machine that produces garbage faster and cheaper doesn't mean demand for the garbage will increase. And in the second case, there aren't enough buyers for high-quality computer-generated pictures of toilets to meaningfully boost demand for Nvidia's products.
I recently had a discussion with a higher ranked executive and his take on AI changed my outlook a bit. For him the value of ChatGPT:tm: wasn't so much the speed up in any particular task (like presentation generation or so). It's a replacement for consultants.
Yes, the value of those only exists mostly if your internal team is too stubborn to change its opinion. But that seems to be the norm. And the value (those) consultants add is not that high in the first place! They don't have the internal knowledge of _why_ things are fucked up _your particular way_ anyways. That part your team has to contribute anyhow. So the value add shrinks to "throw ideas over the wall and see what sticks". And LLMs are excellent at that.
Yes, that doesn't replace a highly technical consultant that does the actual implementation. Yes, that doesn't give you a good solution. But it probably gives you 5 starting points for a solution before you even finish googling which consultancy to pick (and then waiting for approval and hoping for a goodish team). And that's a story that I can map to reality (not that I like this new bit of information about reality..)
If we accept that story about LLM value, then I think NVIDIA is fine. That generated value is far greater than any amount of energy you can burn on inferring prompts and the only effect will be that the compute-for-training to compute-for-inference ratio decreases further
"Throw ideas and see what sticks" sounds very entry-level. Maybe it saves time it would take for one of your team to read first two chapters of a book on the topic.
That exec was hiring consultant and no longer is, in meaningful proportion, thanks to LLM?
The point isn't that the result from an LLM is particularly valuable. The point is that the advice that you get from your typical (management) consultant isn't particularly useful. And that's the only bar you need to clear.
There are basically two reasons for consultants:
Either you need some once-removed thing (be that you selling/buying some part to/from a competitor, accounting, lawyering). That part sensible people will not replace by LLMs. But here it isn't that you yourself lack the expertise at all. Here it is absolutely necessary, that someone else does the actual implementation.
Or you have some general "we need to do better" feeling. And here you have again two options: 1) you know _what_ your problem is and you just need the best solution there is. This is (obviously somewhat tongue-in-cheek) essentially corporate espionage. Again you cannot replace that with an LLM (or maybe you can I don't know), but you will pay a lot of money for it. Or 2) you don't know what the problem is. Now you are competing in finding an appropriate starting point with 20-somethings fresh from university that are not wanted in your organisation and therefore won't get access to the relevant information anyhow. So yeah, I'm willing to believe that the typical success rate of those consultancy projects is low to negative.
If you are only given a 3 week crash course in $BUSINESS, you won't be able to produce much more than a generic set of "have you thought about that?". And THAT is something that I believe LLMs to be reasonably good at. And they are dirt cheap and instantly available compared to any kind of human consultant.
Now I don't think that will necessarily a net-negative for consultants in general. I do think, similar to ATMs, that consultations are mostly becoming cheaper by that.
Thing is, most code is written by entry-level/junior programmers, as the whole career path has been stacked to start grooming you for management afterwards, and anything beyond senior level is basically faux-management (all the responsibilities, none of the prestige). LLMs, dirt-cheap as they are and only getting cheaper, are very much in position to compete with the bulk of workforce in software industry.
I don't know how things are in other white-collar industries (except wrt. creative jobs like copywriting and graphics design, where generative AI is even better at the job as it is at coding), but the incentives are similar so I expect most of the actual work is done by juniors anyway, and subject to replacement by models less sophisticated than people would like to imagine they need to be.
The other thing is that if this pushes the envelope further on what AI models can do given a certain hardware budget, this might actually change minds. The pushback against generative AI today is that much of it is deployed in ways that are ultimately useless and annoying at best, and that in turn is because the capabilities of those models are vastly oversold (including internally in companies that ship products with them). But if a smarter model can actually e.g. reliably organize my email, that's a very different story.
When you get more marginal product from an input, it's expected you buy more of that input.
But at some point, if the marginal product gets high enough, the world needs not as many, because money spent on other inputs/factors pays off more.
This is a classic problem with extrapolation. Making people more efficient through the use of AI will tend to increase employment... until it doesn't and employment goes off a cliff. Getting more work done per unit of GPU will increase demand for GPUs ... until it doesn't, and GPU demand goes off the cliff.
It's always hard to tell where that cliff is, though.
Not even Microsoft Copilot 365 had a successful launch. The same happened with Apple Intelligence.
People talk like the end user demand part of the equation is really solved when invoking an Econ 101 magical interpretation of the Law of Supply and Demand or Jevons Padadox.
I sure hope not, because what they've released so far is worse than useless. The "notifications summaries" in particular are hilariously bad. It's the same problem with Google's "AI Overview"--wrong or misleading enough that you simply can't trust any of it.
But to be clear, none of these things are commercial products. They're gimmicks. Google is an ads company, they make their money selling ads. Apple is a computer company, they make their money selling computers. These "AI products" are a circus sideshow.
Yeah, lots of them. I never thought I'd be paying for a search subscription but after a few months of using ChatGPT I expect to be paying for the privilege from now on. Maybe not paying OpenAI, but someone. There isn't much of a moat there, so there are going to be many companies basically on-selling GPU time. And even if for some weird reason there is no commercially successful AI-specific product it is causing shockwaves in how work is done, most people I know who are effective have worked it into their workflow somehow.
Are they losing billions on training or inference? If their current products - ChatGPT and the API - are profitable ie the inference cost is less than they charge, they have a long term sustainable business.
With other providers giving away similar products for free (Google AI Studio, DeepSeek et al) right now, I'm not sure that counts as commercial success when it is not sustainable.
The same is happening in enterprise tier products, Copilot 365 is still an extra SKU to count while Google Gemini Advanced has been integrated into the Workspace offering (i.e. they actually force you for an upsell of ~20% per user license for something we didn't ask, but I digress). At least that's a better alternative that paying +20 USD per license.
Prices need to and will go down, and business models will have to change and they are already doing so. But I'm not sure if OpenAI is really ready for that.
I gave it an easy one, “How many of the actors from the original Star Trek are still alive”. It gave me accurate information as of its training cut off date. But ChatGPT automatically did a web search to validate its answer. I had to choose the search option for it to look up later info.
With ChatGPT even when it doesn’t do a web search automatically, I can either tell it to “validate this” or put in my prompt “validate all answers by doing a web lookup”.
Then I gave it a simple compounding interest problem with monthly payments and wanted a month by month breakdown. DeepSeek used its multi step reasoning like o1 and was slower. ChatGPT 4o just created a Python script and used its internal Python interpreter.
Then DeepSeek started timing out.
This is the presentation of “what are some of the best places to eat in Chicago?”
I meant 10 million paying subscribers not 10 million dollars. I put the dollar sign there by mistake. That’s $2.4 billion in revenue and growing not counting API customers.
The question is whether ChatGPT (the product) and running thr API is profitable or at at least whether the trend is that cost are coming down.
>Really? Has anyone made a useful, commercially successful product with it yet?
Aren't millions or even tens of millions students using ChatGPT for example? To me that sounds like a commercial success (and looks comparable with the usage of Google Search - a money printing machine for more almost 30 years now - in the first years)
And enterprise-wise - heard recently a VP complaining about entering expenses. As we don't have secretaries anymore in the civilized world, that means "Agents AI" is going to have a blast.
(i'm long on NVDA and wondering is it enough blood on the streets to buy more :)
It isn't because it's not making them any money. Having users doesn't mean you have a business. If you sell two dollars for one dollar having more users is not a blessing financially. Of course you could slap ads on it, like Google, but unlike Google openAI has no moat and there's already ten competitors. Competition eliminates profit and AI is being commoditized faster than pretty much anything else.
>and there's already ten potential competitors. Competition eliminates profit and AI is being commoditized faster than pretty much anything else.
we're discussing NVDA. Where are its competitors? ChatGPT having 10 competitors only makes things better for NVDA.
>Competition eliminates profit
Competition weeds out bad/ineffective performers which is great. History of our industry is littered with competition taking out bad performers, and our industry is only better for that. Fast commoditization of AI is just great and fits the best patterns like say PC-revolution (and like it i think the AI-revolution wouldn't be just one app/user-case, it will be a tectonic shift instead).
NVidia's not selling LLM subscriptions, they're selling shovels in the goldrush. I don't think 3 trillion is a reasonable valuation either, but NVidia's applications extend way beyond consumer and they've effectively become the chokepoint for any application of AI
> Aren't millions or even tens of millions students using ChatGPT for example? To me that sounds like a commercial success
I read somewhere that OpenAI brought in $3.7 billion in 2024, and made a loss of $6 billion. So... no I don't think that is an example. They want to make a commercially successful product, but ChatGPT doesn't seem to be there yet.
> And enterprise-wise - heard recently a VP complaining about entering expenses. As we don't have secretaries anymore in the civilized world, that means "Agents AI" is going to have a blast.
We don't have secretaries because the word became unfashionable. They are called PAs or executive assistants or something like that now. They're still there, but if anything the need for them has probably been reduced with (non-AI) computers (calendars, contacts, emails, electronic documents, etc.) so I'm not sure that there is some enormous unmet demand for them.
(Many individual people are already saying that, but they aren't the people buying the GPUs for this in the first place. Steam engines weren't universally popular either when they were introduced to society.)