Ask HN: Could AI be a dot com sized bubble?
AI hype has landed in the laps of retail investors and in general anyone passively investing as NVIDIA and Microsoft shares have risen and become part of major ETFs in the expectation that the AI-driven demand for their products will continue unabated. Other tech stocks seem to be seeing similar treatment around AI hype.
I do expect the current trajectory of generative models will eventually be incredibly important just like the internet was and is, but it seems there’s a lot of high expectations of how useful it can be in the near future with fuzzy ideas around business models like in the dot com era.
If these near future expectations don’t pan out, could companies slow down their R&D expenditures which are floating NVIDIA, Microsoft, et al and lead to a sizable stock market correction?
136 comments
[ 3.1 ms ] story [ 207 ms ] threadhttps://www.reuters.com/technology/amd-launches-new-ai-chips...
I respect AMD for trying to get people to care about Open Source alternatives to CUDA, but the rest of the industry would rather fight Nvidia for scraps than work together to dethrone them. CUDA represents almost a decade of concentrated software design, and it's going to take a lot of work, proprietary or Open Source, to put it in it's grave.
It's so hard that with trillions of dollars at stake, nobody can be bothered to compete. There are really two alternative routes that the industry can take to kill CUDA. I've mentioned it a few times in this thread but their choices are:
1) Revive OpenCL as a cross-platform GPGPU solution. Fund Khronos and get them to brush up the spec and make OpenCL into what Vulkan was for DirectX. There's a lot of work OpenCL needs before it catches up with CUDA, but a Free and Open Source, cross-platform GPGPU API would make a lot of Nvidia's software monopoly irrelevant.
2) Compete with proprietary, hardware-specific CUDA alternatives that implement only parts of CUDA's spec piecemeal. Basically entails trying to beat Nvidia at their own game, something the rest of the industry hasn't been very good at historically.
Option 1 makes the most sense; everyone wins if everyone's device speaks the same language. But the industry is too greedy to desire a true solution to CUDA's problem. OpenCL is derelict today because Apple, Microsoft and Google all feel as though they have better odds fighting CUDA alone. None of them are even close to Nvidia's performance profile or software capability, which is why I really feel that Option 1 is the only way forward.
https://finance.yahoo.com/news/single-customer-made-19-nvidi...
MSFT is hosting and investing in Open AI, but other than offering AI in Azure and O356 products, what is the business model and future for AI?
GOOG, META, and TSLA now have their own AI LLMs and chat gpts, but only GOOG has sort of a plan for AI, which is to replace search with it.
AMZN is the only competitor to Open AI and MSFT with Claude.
Again, what is the business model and plan for revenue?
Meta is sorta trying to make an AI assistant that lives on your glasses. I'm not sure they are going to make it practical before instgram dies.
You don't have to ascribe a business strategy to that any more than you ascribe a business strategy to whatever wack technology Apple patents next. These companies are using their R&D to test new technology, it's not hard to believe they do it for kicks since they already write-off hundreds of non-product developments a year.
80% datacenter sales are to less than 8 customers.
marginal buyers - coreweave and xai are funded with speculative monies to the tune of $15B
so, expect pain.
LLMs have absolutely put us in a similar place. People are getting customers/money by saying "Look at our AI-driven features!" But give it 5-10 years, or maybe less, and that will seem silly. People will be using LLMs for focused use cases and making sure they help the end users. You will be evaluated with a higher standard - of what you actually accomplish with your products. Saying your product is "AI-driven..." will be as stilly as saying your product is "RDBMS-driven..." would be today. Because it will be a foundational piece of the tech stack, not a marketing blurb.
I doubt it will be as big of a bubble as the dotcom crash, though. We may all still fall victim to hype cycles, but it is hard to explain just how low the bar was for investment during the original dotcom bubble. We are collectively wiser now, even as we stumble along the way.
This day cannot come fast enough in my opinion. Hype is tiring and ultimately detrimental in the long term (more disappointment than fulfilled wishes), in terms of opportunity cost of what could have been invested in instead.
How does one escape this?
Morgan Housel (a writer at the collab fund) has written about how cycles of boom and bust are inherent to markets.
https://collabfund.com/uploads/Collaborative%20Fund%20--%20T...
The first section "1: The Inevitability of Insanity Among Sane People" is probably the most relevant, and a more summarized version is from Housel directly: https://collabfund.com/blog/the-laws-of-investing/
And another great level-headed financial writer Howard Marks has a great "memo" about how bull markets (sometimes read: bubbles) rhyme. Specifically the section "Optimistic Rationales, Super Stocks, and the New, New Thing".
https://www.brookfieldoaktree.com/sites/default/files/2023-0...
I can't find the exact quote, but I feel that one of them basically once said that you wouldn't really want a market to not have bubbles, if something didn't have the ebbs and flows that financial markets have, there wouldn't really be any room to find niches or to innovate because innovation requires risk, and risk inherently destabilizes any existing stability (re: the first article).
I can’t stop beating this analogy to death:
A market that is never allowed to fall into recession is like a forest whose fires are always snuffed out immediately… One of these days it’s gonna go and when it does the resulting conflagration will be many, many times worse than if It had been allowed to burn a little bit here and there…
If that's reflective of the broader human condition, we only avoid hype by having a scapegoat to demonise… or possibly the causality is the opposite direction and we only avoid demonising scapegoats by hyping something up.
There's a major difference between different kinds of "crazy new stuff".
One kind is things that are just beyond the current frontier, and things people obviously kind of want, like say self-driving cars. It's another thing if we're talking about haphazardly bolting the new hotness in a distracting, pointless, or detrimental manner onto every single tech stack in the entire world, regardless of whether people actually want it, and regardless of whether it hurts end-users.
Lots of the current activity with AI is clearly more in the 2nd category than the 1st. In terms of end-users, things like job searches, home/apartment searches, loan applications, college admissions, etc are just a few of the most obvious examples where things are about to get muuuuuch worse for almost everyone (and it was already kind of awful).
Businesses/institutions generally don't care about end-users of course, but I would argue that it won't really help them either. It will just become another very large cost of doing business in the modern world, something you have to do to even be involved, similar to marketing/advertising/legal-council and the rest of the boring-but-necessary overhead. A necessary evil but one with very ambiguous ROI.
But AI hype in 2024 is not even close to dotcom hype in early 2000.
During the dotcom bubble, companies IPOed with no revenue and no profit. In 2024, OpenAI is probably at $2 - $3 billion in revenue right now. All the companies benefiting the most from AI are established companies such as Microsoft, Apple, Nvidia, TSMC. We haven't had any pure AI companies IPO with no revenue in 2024 as far as I know. Heck, I'm not even sure if there is an AI company that IPOed in 2024.
I think AI hype is in the 1996 equivalent of hype - not 1999.
We're at the baby steps of the LLM revolution. There are so many things I want an LLM to do but it hasn't been done yet. I want Slack to be integrated with an LLM so I can ask it business logic discussions that I can't find using its search engine. I want Outlook to summarize long email chains that I just got cc'ed into. I want an powerful but private LLM to ingest all my digital data so it takes into account all those things before doing things for me or answering my requests.
That's cheating; you moved from "companies" to "a specific company". Some specific companies had revenue in the dotcom bubble. There are plenty of AI companies today being valuated sky-high with currently no revenues (dunno how many have gone public, that has changed a lot since 2000), and there are plenty of "real", normal companies getting their valuations goosed by saying "AI" and not having any current revenues to show for that valuation rise. Most recently Apple put on something like 300 billion in valuation for their deal with OpenAI, but if you dig into the deal, neither of them are paying the other anything.
The people selling shovels are doing well, if not necessarily as well as their stocks indicate, but so far there really isn't that much gold being found for all the shovels being sold. Not zero, not claiming it's zero, but it's nowhere near what the valuations imply.
It was more likely that someone would throw money at your idea back then if it ended with ".com" than it is for someone to fund your idea today that ends with "using AI".
I don't know why. Maybe it's interest rates. Maybe it is because the AI hype is so recently after the cryptocurrency hype. It was quite easy to make people dump their money into a cryptocurrency, and the crypto winners are still actively fishing for bigger fools while the AI hype is going on.
Really struggling to decide if AI is ultimately a winner take all market. Of course the very best models will require trillions in capital to train, but seems there also will be a long tail of local, smaller, and specialized models doing a majority of the workloads.
Revenue has gone up, but fewer publicly traded companies are making a profit than ever before [0].
I could never understand why so many people just talk about revenue. Revenue without profits is meaningless. There's the old logic of "get enough revenue and then figure out profits and you're highly profitable", but it's very clear that switching the "profit switch" is not so easy in practice.
Investors are still basically waiting for the fed to drop rates, which means that people have abandoned rationally thinking about businesses and are just holding until the free money starts pouring in again.
I honestly don't think the AI bubble is anything like the dotcom bubble. There's something much stranger happening here since the entire market is basically hallucinating and AI is just one manifestation of that.
0. https://finimize.com/content/beware-the-rise-of-unprofitable...
It's detailed in the article, but that graph is way misleading because it isn't weighted by revenue size and the unprofitable part is dominated by tiny companies.
That chart is deceiving. (from your link)
If you look carefully, you'll see that "very profitable" companies over the decades is unchanged.
What changed is the balance between "barely positive" and "negative".
What guarantee is there that "free money" will come back again?
Wasn't the last free money printer run something like a first time in history, and supposed to be only a temporary measure that went on for far too long leading to inflation and assets spiraling out of control creating various speculative bubbles like crypto, Gamestop fiasco, housing, and dozens to hundreds of crappy overhyped "start-ups" adn food delivery apps, that were never able to be very profitable on their own but still grew like crazy thanks to that free money and gullible investors to stay afloat, leading to an artificial over demand of SW devs which also crashed with them.
Seeing all it lead to, do we even want/need it to come back again? And "But this time will be different" doesn't scan for me as a believable answer since we all know it'll definitely be the same.
Companies are taking on massive investment larger than any IPO take, with zero revenues! OpenAI’s investment predated its revenue push.
I’ve seen people suggest this cycle is malinvestment… if you think the goal of AGI which is exceeding unlikely to emerge in this investment cycle.
Get ready to pick up a ton of cheap hardware in a year or two…
That's not a bold statement since that's the natural life cycle of hardware anyways. I think the "year or two" is a bit soon, but also tracks with the price of hardware holding value for that term too.
Why? As far as I know, we're hugely bottlenecked by hardware in both training and inference even for current models.
Do you have any source for this estimate? Everything I find online tells a different story (single digit million per year).
We live in a very messy information space, both publicly and privately, and we navigate it with messy, sometimes-inaccurate brains. I don't expect LLMs to operate at a higher level of formal logic than a human brain does.
People make mistakes: even the best people. We accept that risk as colleagues/employers/employees because there's a human connection, but that doesn't make it nonexistent. We're now trying to figure out how to deal with that when you're "employing" AI, and whether the risk tolerance is higher or lower alongside the reduced operational cost
The market is much smarter to scrutinise and punish companies attempting to IPO and underperform in the public markets like what we have seen with WeWork and the other SPAC companies that failed to go anywhere.
Once again, companies with little to no-revenue and especially no-profit are getting valued at >$1B all because they slap 'AI', 'LLM' nonsense to inflate their valuations. That is already a dotcom bubble level hype.
The reality is, the existing big tech companies (Microsoft, Apple, Meta, Google, Nvidia, etc) will take all the value of LLMs and will bankrupt the late comers except for the advanced few such as (OpenAI, Anthropic and Cerebras Systems)
It’s been applied to crypto as an analogy and now here to “AI” though I think you actually mean LLMs.
The thing about the initial web bubble was that the potential using already proven tech was just absolutely gigantic.
Like I used to have to go to an office in a building to buy a plane ticket. Then I didn’t. People used to have to mail me a 500 page catalog for me to order via mail, and there was no way to change price or availability.
To interact with most businesses it had to be synchronous, via phone call. Or very very slow or resource intensive via mail or in person and that was that.
Even immediately it was clear these things would change a lot. You’d be able to browse an online catalog of flights or books, you’d be able to email a business and ask a detailed question and copy and paste the request to whoever.
It was right there. It worked. It was clunky and adoption took a minute but nobody honestly was all that confused about what was happening.
The bubble was almost exclusively about how fast people thought change would happen, and which people would benefit.
The LLM thing feels different. It has clear use cases that work, no doubt. I use it to help draft business docs all the time. That could even be a huge market.
But there’s also an assumption that the cognitive ability of these tactics will grow without bound. We don’t actually know that.
Maybe maybe not. But that’s a difference that makes the dotcom bubble a partial analogy at best.
If that doesn’t happen then it’s not going to be an internet-sized change. Maybe something more on the scale of GPS or something.
Like GPS is still an insanely cool technology that is massively important/useful/valuable but it's still on a completely different playing field than "the internet".
I think they're the same magnitude. I might give the edge to LLMs.
I think LLMs have the potential to make every field significantly more productive whether you are an accountant or software engineer or lawyer.
The hype around LLMs is that for the first time every, a tool has the potential to replace a massive number of white collar tasks.
One caution about using this as a metric: in the last 25 years, for a variety of reasons, early IPOs have, in general (and not just amongst hyped-up tech companies) become way less desireable. A company that might have IPO'd in the 90s today might simply take a few hundred million from VCs. And there have been a number of no-revenue, no-profit VC investments at that level in the current bubble.
Granted, a VC-driven bubble is less dangerous to the general public than an IPO-driven bubble; when it collapses it's mostly VCs holding the can, rather than peoples' retirement funds.
Nvidia and TSMC are making insane profits, but that just means there's a demand for specialized compute. Like with the crypto bubble, their success is unrelated to the quality of the final AI product sold to the consumer.
OpenAI is in a similar position due to their SaaS model. It's all about pumping the hype and getting other businesses to build AI products on their platform. Not getting good results? Must be your poor prompt engineering!
The real slaughter is going to be in the AI startups, and the companies trying to pivot to AI in an attempt to stay relevant. The general public is already starting to get tired of the whole AI hype, and we haven't seen anyone provide genuinely groundbreaking products yet. All of it is somewhere on a spectrum of "kinda neat, I guess" to "dystopian hellscape".
Unless we see someone come out with a truly innovative must-have product, this hype is probably going to end sooner rather than later.
Just like those here are old enough to remember the Cisco stock price and Intel's stock price in the 1990 when the internet was hyped up.
It is unsustainable and insiders will start or already have been taking profits out to derisk and anticipate any outside market forces that will correct these prices such as; for example: A declaration of war and an invasion.
This will be no different. Even the Nvidia CEO is taking out profits. Soon employees will do the same (and are already doing so), then Pelosi will do that ahead of the event.
If one of the AI companies use their windfall investment funds wisely they'd invest in getting fusion power working, something that small scale crypto power wasters don't have the organisation for.
That is, in fact, the only thing I can figure out what Sam Altman needs $7 trillion for.
At least then when it bursts we'd have something to show. And he'd get some place in history.
Even the company I work for, which makes and sells ERP software, literally changed its main domain from .com domain to .ai
But chatbots in all the things? That will definitely collapse. I am interested to see what cream rises to the top out of all this and if we'll see an actual bubble burst like we did then.
1 software engineer can become 10x with a better ai.
In the further future 20x.
People are worried about the investment bubble. I’m worried about the employment bubble.
It’s sort of opposing metrics if ai does pan out your job is at risk. If it doesn’t, your ai investments are at risk. I think ultimately it’s your job that will be obsolete.
AI is already affecting the junior dev/intern space. Seniors and mid developers really don't need juniors around, there is a GPT that you can ask questions to and get the boilerplate back that you would usually send to the juniors to think through and develop for you.
Otherwise when all current seniors die there would be nobody left to do programming.
It's possible this may be true in some domains now, but not in any I work in
Everything that I have seen from ai makes me very secure that I’ll have work for far longer than desired.
Also you’re biased. I talk in terms of economics and jobs in general. You talk about you feeling “secure” and you having “work”. You’re thinking of your own security. I’m thinking about the population from a dispassionate perspective.
You need to be more logical.
Whatever GPT-5 turns out to be will set the tone. If there is a clear increase in intelligence over GPT-4, I don’t think it’s a bubble. If it turns out to be a modest improvement in capabilities (like a GPT-4.5), then it’s probably a bubble.
In other words, if the exponential gains continue as AGI/ASI proponents expect, then there is every reason for these crazy valuations to be real.
That said, I think most these AI startups which are simply wrappers around an LLM are obviously going to die either way.
There are lots of potential applications for _some_ LLM models. However, if your business is dependent on openai to differentiate you from another business, you're probably fucked.
I don't think nvidia is going to be the only shop in town for much longer. The GPUs are too expensive and too power hungry.
Meta and Amazon are going headlong into AI. Most of the stuff they are embedding into product are _shit_ despite meta accidentally seeding self hosted LLMs with llama "leaking"
Google are probably dead in the water. They have the talent, but not the vision or the leadership structure to drive meaningful change in their products.
Meta is slightly better placed. but they also have no real internal product leadership (shrimp jesus, and all the utter shit changes they've made to the facebook app, and the poison they've been pissing into Instagram. ) However they have got distracted from AR by Mixed reality, and even more recently by AI (ooo I can ask generic questions about a photo, and get a bullshit answer, how great!)
I'm dying to hear what you think is more efficient for floating-point computation. ASICs fabbed on TSMC 5nm?
yup, flops are the only thing that matter. Nothing else makes a difference.
700watts a unit, plus the ancillary heat from the processor and infiniband/fancy ethernet is challenging to house at scale.
as to my original point, they are fucking expensive. Some of the ones we have are literally gold plated.
The real "only thing" that matters is CUDA. CUDA scales from Raspberry Pi boards to desktop cards to datacenters without skipping a beat. It's a killer product in an age where OpenCL is rotting and dead.
density. Datacentre space is expensive. 9.6kw per machine means putting liquid cooling to the die, if we want to have more than 3 machines per rack. We've already lost a load of space in order to house an unspecified thousands of "h100"
We need to have some space left to put in storage, otherwise we're going to wasting power for funsies. We need that storage to be close, because otherwise it gets _really_ expensive and really slow quickly.
Literally, what are your other options? Hot aisles get hot, I don't know what to tell you there.
You are misunderstanding what I originally saying: there is an opportunity for other people to bring in new types of accelerators.
You took that to mean that I was saying nvidia are shit.
GPUs TPUs and other accelerators are a pain in the arse to integrate at scale. NVIDIA have the current winner because CUDA is well understood, and the GPUs they make are fast and plentiful for the desktop.
They also make a loverly goldplated "mini computer" in the form of the DGX. Again great for any startup that's just got funding.
But, just like IBM, DEC & HP, who also made really fast, capable, large scalable machines, they were eaten by slower, cheaper pizzabox clusters. Not because they were better, or more efficient, but because they were cheap, plentyful and easy to sling in a room.
And my point isn't that I resent people insulting Nvidia; I think the world would be a better place if CUDA didn't exist. But there literally are not HPC alternatives that exist in the mainstream today. Seriously, what are you suggesting? 20,000w Beowulf clusters that accelerate everything on 10-year-old Xeons? A few thousand Macs running GPU-shader accelerated int8 inference? These solutions just aren't out there, and it's why Nvidia owns the industry right now. The sheer incompetence of every other commercial hardware vendor has cemented their dominance for the foreseeable future. The industry would rather watch each other die fighting Nvidia than work together to stand a solitary chance.
> "It's just money; it's made up. Pieces of paper with pictures on it so we don't have to kill each other just to get something to eat. It's not wrong. And it's certainly no different today than it's ever been. 1637, 1797, 1819, 37, 57, 84, 1901, 07, 29, 1937, 1974, 1987—Jesus, didn't that fucker fuck me up good—92, 97, 2000 and whatever we want to call this. It's all just the same thing over and over; we can't help ourselves. And you and I can't control it or stop it, or even slow it, or even ever-so-slightly alter it. We just react. And we make a lot of money if we get it right. And we get left by the side of the road if we get it wrong. And there have always been and there always will be the same percentage of winners and losers, happy fuckers and sad suckers, fat cats and starving dogs in this world. Yeah, there may be more of us today than there's ever been, but the percentages—they stay exactly the same."
https://youtu.be/IAqAl292ozs?t=217
Look at the performance of th qqq certificate, which invests in the largest 100 non-tech companies in the Nasdaq:
https://www.google.com/search?q=qqq
And click on "max". It goes from $52.97 in May 1999 to $485.51 today.
That is an annualized ROI of 9.21%. A better performance than many other market indices had over the same time.
QQQ is a good indicator for how tech performed since the so-called "dot com bubble". And it performed pretty well. That's why I tend to not call it a bubble.
- The current "AI" stuff is truly useful. So it isn't just hot air, like "web3" was, for example. - Does that justify Nvidia's crazy market cap? I don't think so.
Obviously, you still need to verify the output, but even with that in mind, it's a great productivity boost, at least for me.
I'm sure people in other industries can also find use for it.
https://www.theverge.com/24173858/ai-cohere-aidan-gomez-mone...
The host is clearly trying to push the CEO of Cohere on how all this is going to make money (or just be economic). The CEO is confident, but not in a very specific way. There is a great moment where he is like "we did some proof of concepts with 5 users, and they were pretty good, but when you tell CFOs about the running costs for a full user base, its not viable."
What fascinates me about AI right now is that it seems to have very different economics from traditional software/internet/SaaS businesses. Those business scale super-efficiently. They have some initial startup costs (but still relatively low, especially with cloud providers) and low running costs.
With AI, the initial capital costs to build the model are quite high. And, the running costs to handle queries are also quite high. These companies need to find use cases that generate value significantly in excess of those costs. If those use cases are out there, they must either involve really significant productivity improvements, or the costs have to come down a lot, or both.
All that said, I remember going to a talk by Adobe's founders, in which they pointed out that when they introduced Postscript, the first Apple printer that ran it was only viable because of a last minute drop in memory prices, and when they started building Photoshop, you could fit six (6!) digital images on a powerful computer.
So, I see why the investment is happening, but its a high risk investment right now hoping to identify both high-value use cases and significant cost savings simultaneously.
It's the startups building on top of LLMs that will have much lower cost.
Something which I'm kind of convinced about though is that I think after this we're going to see an end of an era in tech in that I don't foresee another "next big thing". Feels like we've kind of played all of them out, they've all more-or-less matured. Like, maybe there will be another hype cycle in AR/VR, but unless Apple really outdoes itself with the next Vision Pro launch I don't foresee it being all that buzzy. Hell, even if Apple launches a headset that's both affordable and ridiculously good, I still don't foresee AR/VR being all that buzzy.
And if AI really does live up to the hype, well, there still won't be a "next big thing", just for an entirely different reason.
I'm not personally upset about it. There's still a wealth of good ideas out there to chase that don't have the scaling potential to attract VC interest but nonetheless could potentially make a decent chunk of people rich and employ a lot of people. Perhaps society will give more of its collective attention to the problems that aren't so easily solved with a mobile app.
End users don't care whether you solved a problem with clever algorithms, a cluster of hardware, or actual black magic. Just like they don't care if their software is written in Python or JavaScript. Why does anyone think customers care that their product is "AI powered"? Customers don't care, but for some reason, investors do care. Wacky.