- anthropic IPO repricing
- pareto frontier gap decreasing by the month between open and closed models
- SpaceX disappoinment
- enterprise buyers as main consumers of closed tokens are being displaced by OS competition for pennies
- job loss and entry level hiring collapse
- massive discounts of current data center contract debt
- consumer open weights model efficiency improving constantly
- the Chinese arms race
- a possible rapid downward spiral that may force governments to step in and subsidize
Am I missing anything?
- margins shrinking at Anthropic
- Google getting thwacked by missed earnings
- OpenAI missing funding target
- price war
- Microsoft revising capex projections down
- Plateauing of model capability improvement
What's weird to me is that you don't see a lot of people making deals that seem way too good to be true but yet the total numbers are absolutely gigantic. Throw on the fact that the stranded assets will have a shelf life, it really seems terrifying. Yet everything is being priced as pretty darn risky. I think maybe there can be a big slowdown without a dramatic pop. A sort of grind down. But there's not a lot of room for error to prevent a pop so I guess it will happen? I don't know its very weird.
Also of course this is all predicated on models not making another dramatic leap. The signals out of the cutting edge labs don't really suggest they have a model that would be as dramatic of a leap as Mythos was when it came out in the hopper. Their code quality seems about as mediocre as what the publicly available models can do. The OpenAI math results are very impressive but that's ripe picking for AI. Also they keep talking about a coordinated training slowdown which is textbook antitrust so they seem to want the government to force their hands off the "spend money on R&D" button.
So I don't really know. I guess I would say purely on vibes that this doesn't feel like a bubble but the size of the build out is just so massive that there really isn't a big window for a success case.
In that context I don’t think I’d describe 10% as “high risk.”
So maybe it's not high risk enough for you, but it is definitely high risk enough for the market.
But the best approach is just comparing rates. Go to https://archive.epa.gov/epawaste/nonhaz/municipal/web/pdf/bo... , and you will see a huge list of companies whose bonds are classified by armies of analysts as "less risky" than these 10% Anthropic bonds. Is Anthropic less risky than AT&T, or BP, or Firestone Tire & Rubber Co? The market says "no".
Recently CoreWeave's bond yields spiked to 16% which represent an 18.8% probability of default per year.
currently US long term bonds have very low yield and are being issued less as usual buyers move to gold/stock -- trust in US gov. has eroded especially in fiscal matters. The budget keeps growing, interest rates are still too low to combat inflation and debt is ever-expanding, without drastic measures taken to stabilize the situation: thus, high bond yield - we pay you premium to have at least someone buy it due to trustworthiness not being the sole attractor anymore
This is what keeps me up at night. That fiber that got overbuilt in 1999 is still there spitting photons. That Vera Rubin gathering dust on a shelf will be a paperweight in 5 years.
Why is this? Competing on price, or compute somehow getting more expensive?
Anthropic has no value proposition - just being 3-6months ahead in frontier models vs. open-weight doesn't justify trilion dollar valuation.
My company buys Claude because it’s a trusted name with a user-friendly interface and integrations with a lot of business apps.
This “open weight models will take over based on price” is analogous saying that Linux and LibreOfficr will displace Windows and Microsoft Office in the workplace because they’re drastically cheaper.
Another analogy is that people talk a lot about the resurgence of audio CDs, but if you look at sales numbers it’s about 4% of peak audio CD sales from the year 2000. Sometimes in society a trend is more that people keep saying that a trend is happening rather than a trend is really happening.
You can't just literally drop LibreOffice into an organization and have it automatically and seamlessly work.
LLMs are, definitionally, fungible. Their input is tokens. Their output is tokens. They have minor variations in chat templates and can be trained to perform better in a particular harness (like training GPT in Codex).
Do you or your friends at work use redis? Postgres? Are they hosted at "Redis Inc" and "Postgres Inc"? No, they're commodities hosted at a huge number of places.
That surprised me. I didn't expect K3 to be as good as it was and now with 3.1, it's clear (imo) that the gap is closing. Anthropic still has a minor edge but from the 'feel' of these open models, that gap is negligible, not a moat.
> I'm not convinced the AI bubble exists
You're arguing against the bubble popping. So you agree the bubble exists?
On what are you possibly basing this? The fact that there hasn't been a new model release in 6 weeks?
Opus 5.5 is capable of completing very long running tasks and managing many threads of work with subagents without getting confused in a way that simply wasn't possible before.
Opus 5 vs. 5.5 is a clear stepup.
Problem is, you need to do a diverse set of experiments to recognize it. Blender support got relevant better,t he amount of corrections got less etc.
But everyone has a different view of this, it'll be interesting to see how it plays out.
Before Opus 5.0, anything with 3d was not doable at all. Starting with Opus 5, it can build you cloth simulators and more.
But I do expect a combination of fast progress to continue alone due to the fact that hardware will catch up and will be faster.
And when we reached a real bottleneck, the next thing will be to put the frontier model on a chip and advancing world models and other architectures.
We are still 'just' steamrolling LLMs.
You can now have a model work on a task non-stop for a day or more. This was unthinkable 12 months ago, and unstable 6 months ago.
If you are searching for evidence of capability, just look at the amount of projects, blog posts and videos on game remixes, hardware reverse engineering, new compilers and runtimes that has started coming out in the last 6 months. Slop? Sure, but this kind of slop didn't exist at this level 6 or 12 months ago.
https://news.ycombinator.com/item?id=50422069
It's really very hard to pin down progress in this domain, companies are definitely targeting benchmarks.
One bit of evidence I would look at is the maths results this last year which have been impressive, that was a huge expensive brute force effort, but it is still good progress.
https://dealroom.co/news/169124-revolut-moves-2-000-engineer...
Two things:
1) it requires a massive money pit to make that compute worth anything
2) it requires charging a premium price
To compare to two previous bubbles; railways in the UK and USA.
running trains was really quite cheap and engines weren't expensive per passenger served. Compared to stage coaches, the number of staff needed was < 50%. the Massive fucking problem was that it needed GDPs worth of money to make the railway in the first place. Moreover as more railways were put in it significantly altered profitably of railways near them. Lines that were there for produce moving suddenly lost profitably because the produce they were moving fell out of fashion
That was true a quarter ago, when Elon still wasn't charging them. Now that he is they're making about $0.96 for ever $1 of compute.
Genuinely asking.
Just throwing numbers at eachother without back up isn't helpful to anyone reading
I also think that some businesses may find that it might not (yet) be cost effective to focus AI use on every problem.
The pushback against data centers is getting bigger too. But america at least is usually pretty industry friendly so I'm not too sure how much it'll delay planned projects. I've read the limiting factor is sometimes just finding a source of power that can consistently deliver the electricity needed for the DCs that companies want to build.
- Is it uncommon or alarming for an IPO to be repriced?
- I’m not sure I see strong evidence to show that open weight models are making significant progress to catching up (and I’d argue that the model quality isn’t even where the business value is)
- What does “SpaceX disappointment” mean exactly?
- Are enterprise buyers buying raw tokens or are they buying productivity tooling for their employees? Not many companies switched to LibreOffice because it was cheaper than Microsoft Office.
- Look up BLS unemployment data, what job loss and entry level hiring collapse are you talking about?
- Does discounts on debt point to anything? Genuinely don’t know.
- What about the Chinese arms race is concerning? Isn’t an arms race evidence of value? There wouldn’t be an arms race if this technology wasn’t worth investing in. Doesn’t China’s desires to compete in this market bolster its merit?
- What is “possible downward spiral” and how do we know that’s real? Isn’t there a “possible downward spiral” with every industry at all times? This seems like pure speculation.
It looks like the end goal of replacing labor by capital is working.
The answer depends on the company, but I think many companies are still figuring out whether and how they can exercise this new capacity. We don't know where it will land yet at a societal level. It's a parameter shift, and it will take some time. Pray that you live in a place with sensible labor laws and safety nets.
It performs the same function as regulation, replacing "the scale to navigate and manipulate the regulatory environment successfully" with "the scale to navigate and manipulate the political environment".
It's essentially grassroots regulation instead of federal regulation.
(I don't think it will stop the bust, but moderate it, allowing more companies to survive.)
Oh? This is the first I've heard that.
> pareto frontier gap decreasing by the month between open and closed models
We need a temporary moratorium on the words "pareto frontier". Every new model is the pareto frontier by some combination of metrics. Like "Bayesian Priors", people love saying / tweeting it because it sounds Smart.
It's like watching car commercials where every single vehicle has won something from JD Power and Associates. Which, (a), who the hell are they? and (b) you can only pass out so many awards until the gluttony of awards proves the award is worthless.
This isn't targeted to you! It's a pet peeve.
There is no such thing as a moratorium on words. It's another phrase that people like to repeat (a meme) until the excitement or prestige of using it has waned.
That said, i agree with you that word-salad investobabble is a sign that the bubble has put many people in a raving fever. This current bubble manifests like the cryptocurrency mania.
Also the sadder slice is all the X people token maxxing with no hope of distribution. Cool things are cool, but getting buyers is getting harder.
This is a problem for Anthropic and OpenAI, sure, but why does it matter for e.g. Nvidia (or any of its suppliers) who will use their GPUs for training?
We need to separate the model companies from the shovel-sellers in these "AI bubble" conversations.
The reason that is important is it will fuel the competitiveness of on premises and local open weight models.
Though maybe it's less useful in the "bubble" conversation since over four more years you could fit in various booms and busts before the supply side significantly starts changing.
The big issue will be access (Re: BYD)
I'm betting many countries will allow companies to buy and deploy them in their countries
- the general lack of merit with regard to most of the words in the article
- this post being front page
so I think that's one big miss from this blog, or I don't understand datacenters as well as I think I do (which is always a possibility! hopefully someone will inform me if I am wrong)
This made sense in the past when they were not prohibitively expensive, but other than them breaking down why replace them on the 5 year schedule now? Yes, the newer stuff is better but the jumps from this point forward seem more incremental than revolutionary hardware wise. And even then, a company can simply swap out racks as parts fail.
Also, power draw is a huge cost factor, and newer models are more power efficient. If you want to stay competitive, you will have to upgrade.
Same for density. Data centers take time to build and face ever increasing opposition. By increasing density, you can improve competitiveness as well.
---
Not all tokens have equal value, and frontier labs (or at least hyperscale companies) will always be able to generate tokens at lower costs per task than enterprise companies will be able to do.
As always, reality is likely a muddled mix: there is real utility and there is wasted investment that will never pay itself back.
So, if you have massive unemployment because AI really is cheaper than labor, you'll get a massive reduction in revenue for virtually every industry, collapsing all of these valuations. Even for B2B companies, Google isn't going to be making much money selling ads to unemployed people who can't afford to buy anything.
After AI Tokens: I don't need to hire someone anymore and if, i need to know why. Because if i have an contractor for 1-2 years and explain a lot to them, i can just explain it to AI.
Also i prefered before already to not hire someone if i had the feeling they don't have potential, now i have even less incentive to do so.
The net result is unemployed or effectively unemployed people, who cut back on extra spending (games, patreon, subscriptions, etc.).
That puts pressure on those people/companies which then repeats as they have reduced revenue. So the cycle repeats.
As long as robots are not there yet, there might be enough of buffer in the whole system for non computer jobs to catch but i'm not seeing this either.
It will increase pressure on every industry.
And it might also lead to an increase in speed of adoption because everyone getting fired, will have the tools to fill any niche.
Simple enough, no?