Hacker Newsnew | past | comments | ask | show | jobs | submit | pbmonster's commentslogin

But you would need maglev if you wanted to compete on energy per ton of cargo per mile. Ships are several times more efficient than classic trains.

NZ is also several times farther.

> enough ammonia in the air will suffocate you.

It will suffocate you on account of all that blood pulp in your lungs. And eyes, and nasal mucosa...

Having a high pressure ammonia leak indoors is a death sentence for anybody in the same room. Same building, if it's large enough...


Technically, you could also just pull refrigerant lines from the heat pump to the bath tub - you probably already run those to the indoor AC units you have in different rooms anyway - or connect the bath tub to the hot water lines you run to the radiators. Both imply a grey water heat exchanger built into the tub.

Depends on setup, of course. If there's a central AC blowing hot/cold air through ducts, you're stuck using a grey water loop. But it is the least attractive setup...


Total consumer-price inflation has been about 27.7% since 2021.

Inflation is multiplicative, so: 1.277 * 1.26 = 1.61 giving us the 61% subscriptions went up.


Though it's a little circular, because this is the kinda of thing that goes into the inflation calculation.


I don't get that mentality. Millions of units, shares many parts with other models, OK margin. This is no reason to kill a product.

For comparison, that's roughly like Ford killing the Bronco Sport, or VW killing the ID.7.


> I don't get that mentality. Millions of units, shares many parts with other models, OK margin. This is no reason to kill a product.

At 2% of sales, the opportunity cost is the reason to kill the product. Keep in mind this still needs design, manufacturing, marketing, and distribution resources.

I say all of this as a fan of the Mini series.


> At 2% of sales, the opportunity cost is the reason to kill the product. Keep in mind this still needs design, manufacturing, marketing, and distribution resources.

All this is true for many other products, specifically cars. Ford makes the Bronco Sport anyway, and they make money on it.

I get it, Apple has much better margins than Ford. And maybe a iPhone 18 Mini would reduce total margin for them. But if they move a couple of million units, it absolutely would make money.


> Ford makes the Bronco Sport anyway, and they make money on it.

The current Bronco Sport generation was introduced in 2021 with minor changes in the ensuing model years. Ford USA sold 134k Bronco Sport in 2025, of 783k SUVs (17.2%) and 2.1M total USA vehicles (6.4%). the Ford E-series truck with 3.2% Ford Truck sales or 2.0% total vehicle sales may be a better comparison. It also is a design introducted in 2021.

phones are more likely mostly new designs in each model year, but we can wait for the teardowns to see.

https://www.best-selling-cars.com/usa/2025-full-year-usa-for...


this is exactly what Steve Jobs did after coming back from Next. Simplifying is very useful, especially in a big corp that want to be differentiated


Except having a 16, 17, 17e, 18 Pro, Air, and Duo, …doesn’t feel particularly simple.


well, VW has cancelled the ID.7 for the US market, and by 2030 will have killing 50% of their models across all brands (according to topgear).


Of course, they never even brought it to the US. It has no chance there, there's barely a market for lowslung station wagons, and close to no market for EV wagons. Always was a niche in the US, while it was always the primary family vehicle in Europe.

And yes, VW is fighting for survival. The things they are killing are the low volume models specific to the Chinese market (because absolutely nobody in China is buying any VWs) and a lot of the variants and overlaps they've accumulated over the years. They'll continue making less than 100k iD.7s, and will be happy to do so (it's modestly profitable).


But this is different, right?

The equivalent would be taking a (fully offline) LLM and asking it about the ending of one specific Goosebumps book, and it revealing the twist. And although that specific book was (probably) only once in the training data, a high parameter LLM can usually "remember" the twist.


The only way it would be able to tell you the ending is if it was somehow given more importance in pretraining, loaded into context, or represented in multiple sets of training samples. I have a blog that I make very LLM friendly and usually load posts up into context when I’m working on something relevant. I’ve also opted to improve models for everyone. Despite this, the model can’t recognize my site or any of my posts when I ask it to recall without internet usage (I also turn memory off btw).


That's not correct for a SOTA model with trillions of parameters. Those have immense amounts of knowledge trained into their parameters. Try it. It "remembers" the ending of random books, Goosebumps and otherwise.

That's the entire point of why knowledge cutoff is so important (if you use them offline).

I find it entirely believable that the Navier-Stokes conversation was auto-flagged as high value training data, and burned-in the models knowlegdge base.


> Is that not self evident by the insane revenue from frontier labs?

No...? Of course not?

Because revenue is only one side of the equation. Did you ever look at total cumulative OPEX and CAPEX, and how long it will take them to even just break even at current growth?


Anthropic is growing 10x revenue every year.

They're likely over $80b ARR by now. They'll be at $800b ARR next year at the same rate. Let's say their growth gets cut down to 3x instead of 10x - that's still $240b ARR by this time next year.

When you are growing so fast, you don't need to make a net profit. You just need to make sure your unit economics are good - which it seems like they are given reports that their gross margins are at 60-70%.


> They're likely over $80b ARR by now. They'll be at $800b ARR next year at the same rate.

And they will be 800 trilion ARR in a couple of years, following that same rate! 8 quadrillion by 2029!

> When you are growing so fast, you don't need to make a net profit. You just need to make sure your unit economics are good - which it seems like they are given reports that their gross margins are at 60-70%.

If their margins were anywhere near this good, they wouldn't need to raise so much money so often.

If you create a machine that turns 1 dollar into 3 dollars, you don't dillute your ownership of the machine, you use your fabulous profits to expand your machine's capabilities.


  If their margins were anywhere near this good, they wouldn't need to raise so much money so often.
Why not? They are reinvesting into growth. There isn't a clear winner yet and Anthropic wants to make sure it is one of them. Taking a profit now while letting OpenAI take your marketshare and train better models is not very smart.


Or they bleed money like crazy, and their margins are pretty awful. Which is the correct answer.

Your $200 subscription is a major net loss for them. The vast majority that pays for that would cancel in a heartbeat the moment they had to pay API prices. Which may or may not be profitable, I am not entirely sure. But for the sake of argument, let's assume that it is.


Why are we using consumer prices when the vast majority of their revenue is from enterprise api usage?


Wihout insight on how much enterprise is paying, it is impossible to draw any conclusions. Unless you have any access to their contracts and are willing to share evidence? I find that highly unlikely.

People here throw around crazy numbers - the dude above was claiming they have some insane good margins, numberd that he took out of his ass.

The only evidence I have is that they are incredibly unprofitable, and they keep raising insane amounts of capital like crazy.

There was a leak sometime ago that they were EBITDA positive during a quarter where they didn't pay for part of their compute. And EBITDA is a cute metric to use when depreciation is actually very important to them, as a model from a year or so ago is nearly worthless.


serving models is very profitable (70%+) but the issue is you need to invest in training the next iteration. but so far all of anthropics models have been profitable fully loaded

the vast majority of the labs revenue is from enterprise api usage (theres public sources from the information and ramp). but the risk there is customer concentration, where most of the revenue comes from other tech companies and a chunk of it is from foreign labs distilling

so i am drawing a conclusion that the labs' business model is good, maybe not as great as boosters think it is. if they make real progress on the biosciences like drug discovery that could turn it into an amazing business


> serving models is very profitable (70%+)

All your argument hangs on this.

I see no evidence of this being true.


https://www.seangoedecke.com/ai-inference-is-obviously-profi...

https://www.mindstudio.ai/blog/anthropic-inference-margins-7...

its even higher depending on the model, how optimized it is, and the chips!

I wouldnt die on this hill


This is not evidence. This is random people speculating on Anthropic's margins without any real evidence.

Just because it is on some blog post, it does not make it true.

I wasted the time to read the first blog post. It considers 100% utilization over the course of years to calculate an estimation, and it did not consider depreciation for the model itself. That thing is extremely extensive to create, and after a relatively short amount of time is considered outdated.


How much work did you go into looking for evidence?


Are we still calculated $200 subscription token spend based on their highly inflated API token cost and then concluding that they must be losing money on all $200 subscriptions?


Are their API token costs highly inflated? I see no evidence of that.


> You would think that scientists would spend more effort trying to create "super" rhizobacteria that can colonize even more aggressively and also fix many times more nitrogen than ordinary ones

I think the problem with that idea is that the rhizobacteria needs energy to break the nitrogen tripple bond. A lot of energy. It gets that energy by getting carbohydrates from the host plant - which in turn are not available to the plant to grow.

I'm sure there's plants where not fertilizing with nitrogen at all would make the trade-off worth it, but the vast majority of our optimized crops can't spare double-digit percentage points of their total yield to feed bacteria. You're much better of just using natural gas (or green hydrogen) to fix the nitrogen.


Or cover crops. They require a season of no yield, but are widely used after harvest, betting on a long pre-frost growing season.


Interesting assumption, not yet proven, likely a great project for a thesis ...


The N-gram parameters can be fetched from SSD, with maybe the hottest ones staying in memory.


I have this working on a branch of my https://github.com/rdaum/eider (for DGX Spark)

nVME paging the n-gram table (in BF16 for now).

Still working at it. Prefill sucks still but decode is about 12 tok/sec and the model weights fit nicely in the 128GB Spark memory in nvfp4 quant while paging the ngram stuff from disk.

(EDIT: merged to main. 80tok/sec prefill, 12 tok/sec decode, ~80GiB resident, the rest paged)


Compete with Garmin instead of Apple.


Garmin has less than 8% of the market versus Apple with 23% and the largest share. Why would they compete with the market loser?


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: