If the bubble pops (meaning these massive compute costs never turn into actual profits and the VC money dries up) what does the tech landscape look like?
A lot of us use Copilot, Claude, or ChatGPT daily for coding and docs. If the subsidized cheap access vanishes because these companies can't eat the losses anymore, do the tools just disappear? Because if a tool like Claude Code (or any other LLM) suddenly cost $1,000 a month to reflect what it actually costs to run, would people keep paying for it out of pocket? Would their companies?
I’m especially curious to hear from anyone who lived through 2000 or 2008. Does a postbubble world mean we just abandon the tech entirely or is it a move toward expensive solutions?
Additionally, new tech could make all the CAPEX in data centers or businesses useless in the fairly near future, but this can be hard to predict (black swan). I think of quantum computers or advancements like DeepSeek's LLM reinforcement learning style, but it will likely be something that isn't in anyone's mind yet. There is also a vast opportunity cost to all this investment.
Those with wealth will be fine (why do you think they're building bunkers), everyone else will pay the price.
I guess similarly in an AI bust some companies that rely on new investor money would go bust but things would go on.
The LLM companies are profitable on the current gen models. Inference is profitable, rather than subsidized.
They are raising the biggest chunk of capital to buy data center compute that will come online ~2 years from now and be an order of magnitude larger.
The bear case for the labs is that they're Cisco, not Pets.com.
And that was before he gave his updated depreciation figures here: https://futurism.com/future-society/ai-data-centers-finances
Now, we're still talking about "some analyst" and the most undifferentiated pure play for the underlying economics of inference itself as a whole, but I think that the latter, at least, should remain relevant because if the underlying inference doesn't work on current subscriptions, then nothing built on top of it without significant additional charge will.
https://en.wikipedia.org/wiki/Gartner_hype_cycle
At the moment we’re at the peak of inflated expectations - we might stay there for a quite a while.
And to confuse things more, different people/groups go through the cycle at different times/pace.
LLM research will go back to (government funded) research labs with government funded supercomputers. All AI investment will need to be written off.
Running the LLMs the research generated will of course be possible, e.g. via AWS bedrock or alternatives. Initially there will be no "flat rate" subscriptions like currently (similar to early Internet), those will come once the prices are low enough further out. Running the LLMs is a low-margin business not justifying high multiples.
If the AI tool companies increase their rates 2x, 5x, 10x, is it worth it? They aren't going to lower prices.
Consumer AI tool usage isn't going to get a lot of adopters that will pay, people outside of a work environment will see it as a fun toy, much like social media and will be fine with being served ads and letting their loss of privacy be the cost.
I work in consulting and I would have had to scope projects with myself as the lead and at least a couple of juniors to do the grunt work while I do some of the work myself and do the tech lead type work. Now I can do it all myself.
Give it a try.
To get started: https://simonwillison.net/2024/Nov/12/qwen25-coder/
I wouldn't find it hard to personally justify $200/month or $300/month for the single best LLM tool available to me. Right now I have $100/month spread out over a few different tools and it's a bargain.
I'm curious why do you keep many subscriptions around? Asking because I do too though don't have a good reason to keep anything other than Anthropic/Claude.
Personally:
Perplexity has replaced Google searches for me – basically looking things up
ChatGPT/Claude is what I use for generating code (I'm not a programmer but I like scripting tasks at work)
Claude I think is better at writing than ChatGPT
MidJourney is $10 and I like it best for image generation which I do for work and like to keep it around.
As you correctly state, the cost of AI as a Service (AIaaS) will increase for end users, but this isn't necessarily a bad thing. It will allow the "real" users to continue having access to it and sieve out the ones who are just playing around. Prices for RAM, GPUs, SSDs will normalize a lot and more people will move towards local models.
Similarly to what happened with the dot-com bubble (I saw it happening), it doesn't mean that everything will disappear, but that it will change/adapt. All of us AI realists are currently being treated like technophobes when we say things like that ;-)
When the AI/LLM bubble pops, LLMs will still exist and be used. They just won't be hyped and pushed everywhere.
:)
I also was looking for a job in 2008 - again in Atlanta, I was 34 and again had no trouble finding your regular old dev job.
The AI bubble busting means absolutely nothing to me as far as career prospects. While every project I do now is related to AI working in cloud consulting + app dev, AI is just another tool in my tool belt.
“All of the spending” means nothing to Amazon, Google, Meta, and Microsoft.
They all are spending out of free cash flow and the hardware they are buying have a high failure rate and will be worthless in 3 years anyway. That just means they won’t replenish the hardware.
For the most part, I don’t deal with VC funded unprofitable companies - see how the bubble busting didn’t affect me.
If Claude Code becomes more expensive - which I doubt, compute gets cheaper over time and if the bubble does burst, that means compute gets cheaper because of excessive capacity.
On the other hand, I don’t pay for Claude Code myself, if the company thinks it does add value, the company I work for will up my monthly allowance from $1000 a month. If I were paying myself, I would just buy a computer that could run a local model.
In 2000, internet stocks collapsed. The internet didn’t. What disappeared were business models built on fantasy unit economics. The same distinction matters here.
If AI funding compresses, here’s what likely happens:
• Speculative AI wrappers die fast • Subsidized API pricing rises toward real compute cost • Consolidation around a few model and infrastructure providers • Enterprises shift from experimentation to strict ROI enforcement
Right now a lot of AI usage is artificially cheap. Growth is being prioritized over margin. If that flips, casual usage drops. That doesn’t mean the tools disappear. It means only usage that creates measurable value survives.
If a tool suddenly costs $1,000 per month, most hobbyists won’t pay. But that’s irrelevant. The real question is whether it replaces or amplifies enough labor to justify the price.
If it saves a team $8,000 to $20,000 in monthly productivity or headcount cost, it survives. If it’s just a nice-to-have, it dies.
The bigger risk isn’t foundation models disappearing. It’s the application layer collapsing. A lot of AI startups exist purely because inference is subsidized. If API pricing normalizes, many evaporate overnight.
A post-bubble AI world probably looks less magical and less overfunded, but more disciplined. Fewer demos. More boring enterprise contracts. More focus on margins.
The internet after 2000 didn’t disappear. It matured and consolidated into utilities. AI likely follows the same path.
The real question isn’t whether AI vanishes. It’s who was relying on subsidies instead of economics.
when collected enough points, eventually punishes the authors and every comment they write will be labeled as "AI Slop"
The real question isn’t whether AI helped write this. It’s whether the reasoning makes sense and matches what happens when capital and infrastructure collide.
Yeah, words that rarely age well. But AI isn’t like the web. It is more like the transistor. Or the internet (i.e. the protocol). Or like the first printing press.
Or even cells going multi-cellular, emergence of nerve cells, ganglions, brains.
It is a step change in the life of information. Non-substrate locked information, approaching an ability to improve itself isn’t a phase. it isn’t an adoption S-curve.
It is too fundamental, and has too many paths leading forward, by organizations and individuals, to ever be as simple as a bubble.
There may well be turbulence. But any computing power overhang created by over investment in one segment will get quickly absorbed by another segment. So there may be train wrecks, but the industry as a whole is going to thrive.
The problems I see are the unlimited downstream disruption of AI on everything else. Everything else. But for AI itself, the future is bright.
API access to AI models is not "subsidized", AI companies make a profit on inference. They are only losing money because they spend a lot on training the next generation of models.
It would make sense to claim the bubble pops if AI companies' valuations were pure speculation, but revenues are increasing exponentially for OpenAI/Anthropic. They are selling a product that people are buying at a price which is profitable on margin.
I think once the dust settles n next 2-5 years, few clear winners will remain who will figure out a way to become cash positive.
> if a tool like Claude Code (or any other LLM) suddenly cost $1,000 a month to reflect what it actually costs to run, would people keep paying for it out of pocket? Would their companies?
Probably. If you're not gaining at least $1000 a month in productivity now, then you're doing it wrong. I suspect however they may become an enterprise only offering, with limited availability to "normies"
> I’m especially curious to hear from anyone who lived through 2000 or 2008. Does a postbubble world mean we just abandon the tech entirely or is it a move toward expensive solutions?
In the late 90s you could get 100k a year for being able to spell HTML, and then the bubble pop pushed all the grifters back to whatever they were doing before. Those with real skill stuck around, even though it did suppress salaries for a while.
That also meas that the worth of today's computers declines rapidly and all the money invested in compute power with it.
Compute is basically what the AI VC money is spend on. That money will be gone in a few years due to the hardware being worthless.
On the other side running a model (locally) will become cheaper and cheaper to the point that Ai stuff becomes a everyday commodity running on cheapo devices everywhere.
Then there are optimizations too. Which lowers the cost.
So it's not going away and it's not going to be expensive for the consumer in the long run.
My 5 year old rtx 2027 runs models those output would have been state of the art a couple of years ago. In a few years something running on the level of today's top models might run under your desk if that progress goes on at this pace.
