In many investment theses - like Nvidia's bet that demand for compute will keep growing - the first order assumption is usually correct. Yes, demand for more compute, chips, infrastructure is huge and each year some additional data centers will be built. Where such investment bets usually fail is in the second-order assumptions: Ie. the expectation of the growth of demand. This is where there's a high chance that the current expectations are likely exaggerated. So: demand is likely to persist for the foreseeable future but not increase every year. And that can upend the whole investment story. That can be enough to make these bonds a huge burden for Nvidia in the end. Not because people stopped buying more compute but because they stopped buying more every year.
What makes this insanely hard to predict is that the compute needed for the same quality output has roughly gone down 90% every 18 months for ~5 years.
1) We don't know how long that trend will continue, but you do know where to look for when it may end (if smaller sized models continue to compress the knowledge effectively of larger models).
2) We don't know when the appetite for higher cost models might go down and by how much if smaller models get "good enough" and price becomes far more important.
It is entirely possible that 5 years from now, there's >100x LLM inference going on - but demand for AI chips (including memory) is only 2x or less.
It is also entirely possible that at some size - LLMs pick up some emergent capability that doesn't scale well to smaller sizes - and that there's an incredible boost to demand to get that capability.
I think efficiency is unlikely to result in lower demand for compute, instead more useful compute per watt increases the value of that compute; and we are not going to run out of economically useful things to do with it anytime soon on the demand side.
The harder thing to forecast for me is if we hit a wall on increasing efficiency, either on the model weights side or silicon side, with current approaches. If we have to switch to something like burning the model weights into silicon to continue to make gains, then the current math on general purpose accelerators might be upside down.
What's really interesting is that if you scale it to higher densities (eg 3nm and stacked die) with ComputeInMemory for fp8 you can reasonably start to fit 30B-70B models. With MoE and multiple stacked die, just like HBM, you could fit an open weight near frontier 1T model (like GLM5.2) at similarly much lower power <10kW and high token rates >2ktps. For running a bunch of agents where fill rate and speed/latency are important it may not matter that you're 6-18mo behind on weights. The process for the chip design could be largely automated, and new silicon pumped out as new weights are available (with a 3-6mo delay).
> efficiency is unlikely to result in lower demand for compute, instead more useful compute per watt increases the value of that compute
I buy that. Jevon's Paradox, sure.
> and we are not going to run out of economically useful things to do with it anytime soon on the demand side
This I don't buy. Not fully, at least. Whether or not there's demand for LLMs in some particular field is one thing, whether or not there is a sustainable business model to be built out of that demand is another thing entirely.
There is a staggering amount of money pouring into startups looking for novel use cases for LLM-based agents. As usual, 99% of them will fail, but those other 1% are going to have to look harder and harder to find a novel use case that can actually be served profitably.
First of all, there's only so many places where a chatbot is going to sell. But, that also seems to be the only interface anyone can come up with that allows a user to steer an agent through a long-running task reliably. I'd love to be proven wrong here.
Also, if current trends plateau and large datacenters are still needed for complex tasks, that would stimy growth of LLM usage across entire industries.
But, if present trends continue, then local inference will become feasible for most tasks. That would lower the barrier to entry across tons of heavily-regulated and/or cost-sensitive industries. But, widespread local inference will almost certainly come with a painful market correction centered around hyperscalers, which would itself dry up the pool for ventures into new markets.
Replace "chat-bot" with "Human Being" because the models I've been using are significantly better than 90% of the human-chat-bots that I must talk to on the phone while scheduling and coordinating my internet installation for example.
Now for every human replacement, that is 1 unit less of communication and bureaucratic burden (HR, middle management etc) that the org requires.
> the models I've been using are significantly better than 90% of the human-chat-bots that I must talk to on the phone while scheduling and coordinating my internet installation for example.
I guess I don't know what to say except that my experience is the polar opposite of yours.
I moved to a new state at the beginning of the year. Needed a new doctor, needed to schedule apartment tours, needed to talk to my employer about insurance and relocation stuff, etc etc. Lots of chatbots, a handful of humans. Humans consistently did what I needed them to do, the chatbots just didn't. I could list examples but I'd be typing all night.
And yknow what, my one call with Comcast to get my internet set up was downright pleasant. The rep was knowledgeable and a good conversationalist.
Very few consumers are going to spend multiple thousands of dollars to save $10 per month. Companies absolutely will to save hundreds per month per employee, but that's not consumer hardware.
Well if the trend that the comment further up in this thread claimed continues and compute requirements keep dropping exponentially then perhaps in a few years you can have today’s frontier performance on the normal laptop you already have on your desk anyway.
If you can integrate AI accelerators into consumer cards (you can), you can have local AI for "reasonably" cheap. This is Nvidia's long term goal if you listen to what Jensen has to say.
The limitation is entirely on memory right now. Just a few years ago we could of been strapping 80-100GB to cards for under $200 (BoM).
It could be quite a while before we reach that point though. 5+ years easily.
I've been keenly interested in the ability to run local models, but the hardware is just not there. Consumer RAM speeds and capacity will have to significantly increase before local models will be able to perform as well as even the lowest end GPT-5.6 Luna model.
This is on the backdrop of RAM becoming prohibitively expensive. And without the speed and quantity of RAM, it becomes impossible to generate tokens at interactive speeds, regardless of model. There is a fundamental dependency between calculating all of the active params with the given RAM speed.
Even with a model that has been quantized all the way down to Q4, the DGX/RTX Spark chip with 128GB of RAM can only generate ~18 tokens/sec for a MoE model with only 30B active parameters. There haven't been any broadly useful models below 30B active parameters. And that is for a $5000+ piece of hardware that will be one of the best for running on-device models.
I really want to buy instead of rent my AI, but the economics are truly terrible.
It's also hard to predict how much money will be burned going down wrong avenues. The internet was the future, but it took a lot of failed companies to eventually land on a sustainable model that brought us the giants we have today.
Railways were also the future, but that didn't stop a rush to build out (often subsidized) lines that were ultimately uneconomical (either because they were corrupt or the planned settlements never arrived).
If AI is similar, then there's going to be a long slowdown on compute spend until the surplus is worked through. A good historical analogy could be the fiber optic buildouts of the late 1990s. The demand for data never really went down much, but the industry eventually commodified and took down some large companies (Nortel, especially)
I rhink there's a difference with AI because -- it brings true value because you pay for the tokens, you only pay for what you use. That is true value.
Compare to just paying for an internet connection, you have bandwidth but not sure what you can do with it that is valuable.
Let's say you use AI to produce software. There;s no limit as to how high the quality you want your software to have. And how fast you want your project to be complete. There's plenty of room for higher quality, and more performant AI. As AI becomes chepaer people will use more of it, they're not going to say "We have enough AI".
Compare to railroads. Yes you pay for the distance travelled but there's a limit to how much people wwill want to travel, how it will benefit them.
It is entirely possible that 5 years from now, there's >100x LLM inference going on - but demand for AI chips (including memory) is only 2x or less.
I doubt it. If LLM inference efficiency is 50x better than today, then there could be 1000x increase in inference volume due and we'll end up needing even more chips.
Jevons paradox should win out for a long time for AI.
When internet connections got faster than 56k modems, we didn't use the same amount of bandwidth but faster. We used more bandwidth doing things like 4k streaming. I see the same in AI inference. If AI inference is that much more efficient, it will just enable more use cases for AI.
Nvidia's great superpower is flexibility. You can easily run models of very different types on the same card; and their hardware is great for R&D.
However, at some point AI may be good enough for most people and then it makes sense to make an ASIC for the model (or group of models); and at that point you don't need Nvidia.
I suppose this scenario will happen in various moments at different levels.
Each of the hyperscalers has put like 250B each in the last year for infra. That means that they need to be writing AI profits to the tune of 20B per year just to keep up with the cost of the cash they burned.
We are not there. But they better figure it out soon. The cash flows dried up, and everyone is taking debt to support the capex. Google for the first time in its public history is cash flow negative. Amazon too.
Plus on top of that 1st and 2nd order can be correct, but then the price is too high, meaning people lose money even if correct about the future, but over pay for it.
They also have to be feeling the heat of the ASIC vendors. AMD just acquired Taalas and they work with Cerebras all the time on special projects. ASICs outgun nVidia's chips by an order of magnitude.
