[whatshisface]

Here are a few thoughts:

- The publicly available information about how inference costs compare to training costs is conflicted. EEs involved in datacenters talk about power usage spikes during training runs as if they were a major factor in the designs, but academic papers discussing cost-optimal scaling confidently treat inference-time compute as a major factor.

- On the side of the balance indicating that training is more compute-intensive after amortization than inference is that Chinese providers, constrained primarily by access to compute, have nearly unlimited token availability at a lower price than US providers (inference), but poorer model capabilities (training). That would make sense only if US providers are inflating inference costs by 20-30x due to amortized training costs that overseas providers were not able to take on (there are other factors too).

- If training >> inference, they're in a prisoner's dilemma that far exceeds the ordinary zero-marginals model of competition between firms (due to its huge discrete stepwise nature). On the other hand, if inference>>training, the high-level analysis popularized by certain thought leaders, that it's like a utility, would be true. You'd tend to count this as a vote for inference>>training, but the CEOs saying it at least have a huge incentive to agree because the alternative, the prisoner's dilemma, would stop investment very fast.

- The only voice in the story that I just told you to have anything to do with fact (as opposed to high-level analysis and ivory tower armchair management of a secretive business) were the rumors from facilities engineers. That shows you the state of our understanding...

- If we don't even know the ratio between amortized capital expenses and operational costs, outside investor analysis is impossible. It doesn't matter how finely they divide the accounting buckets for office ferns and indoor ferns if the single biggest part of their business is obscured for trade secret reasons.

[materielle]

I'm about to leave a shallow comment, but I am a bit skeptical of the supposed drop in inference costs. If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop? So the fact that publicly available information is conflicted is probably a sign that at the very least, the numbers aren't amazing.

Yes I know there's no evidence and this is lazy reasoning. But there's probably a bit of truth to this line of thought.

[Tuna-Fish]

Why on earth would AI labs be bragging about how little the product they sell actually costs them to make? You don't want to do anything that reduces it's perceived value to the user, that might make them less willing to pay for it.

Also, inference costs are bound to go way down with more optimized architectures. GPUs are fundamentally not great at inference. No platform where the weights are streamed from a large pool of memory is. If the models ever quiet down, there will be massive step changes in cost/token, energy/token and tokens/second, as models are etched into silicon ala https://chatjimmy.ai/

[overgard]

A couple of years ago Altman was saying the price of AI compute is going to drop 90% year over year or something like that, so I don't think they're nervous about talking about lowering their costs. They probably just haven't been able to lower their costs.

You have to keep in mind that about 99% of their announcements are targeted towards investors (their most important revenue source..), so they're not going to be afraid to mention metrics that make the business look better.

[bwhiting2356]

Jevons paradox. Cheaper tokens does not mean we will spend less.

[Skinney]

Cheaper tokens means the company's margins increase, which would be valuable for investors to hear

[missedthecue]

The main limit to my token spend right now is that I'm running out of hours in a day.

[mcmcmc]

Ah yes, Sam “Not Consistently Candid” Altman

[pixelready]

Oh, is that the guy that sold Loopt by claiming it had hundreds of thousands of users and it turned out to have 500 DAU after his exit?

[chipsrafferty]

Yep, the very same scammer. Wonder if he's lying about OpenAI too? Maybe about a person blowing a metal instrument?

[whateveracct]

he lied. he's good at that.

[golem14]

Why would any company brag about their margins ? Yet they do, to attract investors.

[Tuna-Fish]

The key AI labs are not public companies, they are at liberty to brag about their margins to potential investors in private.

[SiempreViernes]

And investors will leak such claims quickly enough that this reasoning cannot plausibly hide big secrets.

[Tuna-Fish]

It's not a big secret. If you just do the math yourself, it's easy to compute that inference doesn't cost all that much. People just see all the capital investment going around and all the new data centers being built, see that it's spent on "AI", put two and two together and get a three, or "clearly serving AI requests costs an arm and a leg".

The 1 they were missing is that AI requires both training and inference, and training is by far the expensive part. And that in principle you can stop training at any point and keep using the models as they are. (But that means that if other companies keep improving their models, you'll be left behind...)

In contrast, inference is fairly cheap and all the providers have great margins on it. Eventually either investment in training stops having commensurate impact on model quality, and people stop doing that and instead concentrate on making inference faster and even more efficient. Or if that doesn't happen, things will get very weird very quickly.

[brookst]

I’m wary of “has not been leaked in a way that was picked up in public news” as proof or disproof of anything.

[bwhiting2356]

this is changing soon

[joelthelion]

Not really, how much of a public company are you when 5% of your capital is public ?

[lmm]

Growing companies don't brag about their margins, they brag about their growth and revenue. Margin talk is for when you're a mature company squeezing out every bit of profitability you can - if anything it would be a negative sign to be worrying about your margins when you're supposed to still be growing and innovating.

[amarant]

I mean, did anyone expect them to not have margins? Why keep it secret?

[Yoric]

> Why on earth would AI labs be bragging about how little the product they sell actually costs them to make? You don't want to do anything that reduces it's perceived value to the user, that might make them less willing to pay for it.

Wouldn't they be bragging about it to investors? It feels like something that would matter a lot to them, and at least OpenAI kinda feels desperate to find them.

There's also the small question about whether a drop in inference cost would actually change anything about profitability, when training seems to get exponentially more expensive.

[neltnerb]

Because companies that want to go public need to look profitable or potentially profitable. And before they go public they have to release real, actual, legally demonstrable numbers for their costs and revenue anyway.

[extraextra]

When they will actually file to go public, their numbers will be intensely scrutinized. That's all that global headlines will be talking about for weeks on end. Why would they create forward expectations before it's necessary?

Of course they don't want to create forward expectations in a volatile macro environment, with the public listing being 6 months out.

[etempleton]

Because the most important thing for any pure play AI company right now is to prove they are a viable company. And sure they have proved they can make billions, but also that they can lose billions more. They are going to need even more money and to prove to the next round of investors at an even higher valuation that they are a viable business they need to show not that they can generate revenue, but that they can one day turn a healthy profit. And that is the trillion dollar question.

[jimbokun]

I doubt having to replace every single chip in your data center every time you release a new model will bring down costs.

[kopirgan]

Went to that URL asked one question - "how is this different from other AI" and it took 598/6144 tokens, not sure what that means.

[philipswood]

Not super clear from the site itself, but this LLM is running on specialized silicon implementing just it. So has super low energy use and blazing speed.

See https://taalas.com/products/

Edit: updated link

[kopirgan]

Incredible increase over Nvidia! Need to read more.. Thanks!

[DrewADesign]

Because they can think more than one quarter into the future? Why on earth would someone adopt something into their core workflow that was fantastically unprofitable? Uncertainty and business don’t mix. Most people aren’t hype-eating bacteria that only care about maximizing their next paycheck.

[wheresmylogin]

One reason is that all the code you write with this goes in your private git. If using AI no longer is possible because of cost, you can still profit a lot from what you did with it before.

[DrewADesign]

For consultants? Sure. What percentage of contractors are consultants? And is that better than going with something in your stack that’s sustainable even if it’s not totally optimal? I’d wager most would say no.

[nradov]

Regardless of profitability there will always be multiple good LLM vendors as well as open-source alternatives (slightly worse but still pretty good). If one vendor fails then it's easy to switch your core workflow to a competitor.

[DrewADesign]

On an individual basis for coding? Sure. If you’re a significant business with agents that do more nuanced work, which is the only kind of customer that will let any of these companies pay back those trillions of dollars as quickly as they need to to stay alive, these are not fungible services.

[m463]

I wonder if inference costs will go down...

or will it be like microsoft office, where the software bloats to use/fill current hardware?

(and in this case bloats might mean better thinking or pulling in more data)

[kopirgan]

If inference costs drop 90% or whatever, that would be a massive write-off of hardware even before they gave any returns for it?! Given Chinese and others are snapping at the heels and would also benefit from such reduction in cost.

[solarkraft]

> Why on earth would AI labs be bragging about how little the product they sell actually costs them to make?

Investor confidence. They have a bit of a need for cash (also an interesting part of the profitability discussion of course).

> Also, inference costs are bound to go way down with more optimized architectures

I agree. Jimmy is incredible, I wonder what non-toy use cases they have. Surely they’ll come out with updated chips soon.

That said, I was apparently a bit over-excited for Groq and Cerebras. I thought they’d quickly dethrone Nvidia for inference, but not so far. Even the GPT spark trial isn’t seeming to go far.

[whatshisface]

Inference has traditionally been far less expensive than training. One public example is the fact that hobbyists can run StableDiffusion ($600k training costs[1]) on their personal computers.

Speaking to your point, inference being dramatically less costly than training would not be seen as a delta from the norm. The model of providing inference for anything near the operational costs (like a utility would), would the delta from the norm if it were true.

[1] https://x.com/emostaque/status/1563870674111832066

[thesz]

The difference between training and inference is 1) one have to keep intermediate results for backward pass in training and 2) computation for training double because of the backward pass.

Training is also done over batches, which increase memory requirements by several orders of magnitude. This is why training needs costly compute.

One of the ways out of this unfortunate situation is to use something like Stochastic Average Gradient Descent [1]. Examples there are mostly concerned with regularized logistic regression, which makes problem more or less convex. Neural networks are inherently non-convex. Still, maybe some ideas from there can be utilized in the context of neural networks, like use of estimated Lipshitz constant to derive curvature and appropriate learning step.

  [1] https://www.cs.ubc.ca/~schmidtm/Courses/540-W19/L12.pdf

[janalsncm]

So one way to think about it is roughly,

Training is inference + backwards pass (~2x inference cost) + activations (vram overhead) + optimizer (vram overhead) + gradients (vram overhead).

[thesz]

Multiply "inference + backwards pass (~2x inference cost) + activations (vram overhead)" by batch size (thousands) to get to the actual RAM and compute cost. Optimizer like ADAM adds only two or three model-sized overhead.

And last, but not least, you need only one hidden layer kept in RAM for inference, but you need all of them (61 for Deepseek models) kept in RAM for computing gradient for one sample.

[xyhopguy]

Microbatch size is a hyperparameter, it can be set to 1 and work just as effectively. With gradient accumulation it's equivalent even. Large batch sizes are used to increase parallelism, and sometimes to reduce variance in the loss signal (at the cost of increased bias).

Batch size is frequently limited by compute bottlenecks well before memory.

[mcv]

And of course you do all of this for every object in your training set, which is going to be larger than the total number of uses for any individual user.

[galaxyLogic]

Does it matter what is the difference in size of needed inputs for inference vs. training?

[whatshisface]

That is an estimate of the relative cost of one training step, but you have to multiply it by the number of training steps, an unknown quantity.

[mike_hearn]

It's all got much more complex than that in recent years. Training now involves large amounts of inference for RL rollouts and similar. You can't disentangle them computationally like that. "Inference" is just the word used to mean serving customer traffic now, and "training" means creating the model you serve.

[vanviegen]

I think in your StableDiffusion example, a lot more than $600k will have been spend on electricity alone for inference (on those personal computers you mention). So inference is more expensive then training.

[lumost]

For equal capability tokens, there has been about a 10x drop in cost every 6 months.

We are still chasing the best because the best is moving rapidly, but it’s a simple thought experiment to work out what the cost to serve an 8B model from 2 years ago is in a world of 2T models.

Note: parameter counts are illustrative. Concretely, qwen3.6 27B delivers opus 4.5 capability at 1/27th the cost on openrouter. Single chip llama3 8b performance can exceed 17k tokens/sec.

[byzantinegene]

8B models would be consider obsolete in the world of 2T models, at least if we're talking about the competitiveness of OpenAI/Anthropic. The only reason why they are valued so highly is their supposed dominance at the top end.

[lumost]

The main story of agent use cases is in enterprise so far. An enterprise will only pay for a model capable of handling the task and no more. Most enterprise's see no need to hire PhDs as factory line workers.

Coding is an interesting case as [1] the pace of progress has been absurd and [2] it's hard to put an upper bound on required capability. However hard to put a bound on and will are different, it's quite possible that the average engineer will cease to see the benefit of rapid progress - or that their employer will be satisfied with lower tier models.

How smart of a model do you need to build a high quality CRUD app for internal users? Or build a scalable web service?

[byzantinegene]

yes, which is why the revenue growth story is not looking so great for Anthropic/OpenAI, when open-source alternatives are not far behind with much lower costs.

[joshuahedlund]

> For equal capability tokens, there has been about a 10x drop in cost every 6 months

Is this still happening? Opus 4.5 was six months ago, can you get its capabilities for 1/10 cost now? Are we on track to get the same for 4.6 in a couple months?

[lumost]

Pretty much, Kimi K2.6 is opus 4.6 quality for coding. If you include discounts due to more efficient input caching it is around 1/10th of opus4.6.

https://openrouter.ai/moonshotai/kimi-k2.6

The march of cost efficiency moves on.

[joshuahedlund]

Why haven’t I heard of this? Is it available in IDEs like Cursor?

[no-name-here]

> I am a bit skeptical of the supposed drop in inference costs. If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop?

Unless to the grandparent commenter’s point they’re using it to obscure their large prisoner’s dilemma (training) cost?

[neuronexmachina]

> If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop?

Google seems to pretty regularly post about how their TPU and algorithm advancements have been decreasing energy costs for both inference and training.

[brookst]

What other companies brag about lowered costs? Isn’t that just a complicated way of asking customers to demand lower prices?

[vlovich123]

Small alternative potential future changes that alter this analysis:

* At some point model capability reaches diminishing returns. Then inference >> training in the future but training >> inference now. It’s not a prisoner’s dilemma but a land grab to solidify market position and be one of the 2-3 firms left standing as dominant in the space. The model companies aren’t super sticky yet but they’re working on it.

* even if training remains >> inference, it’s possible to have multiple price points like they do today. If you need the most capable model you’ll be paying exponentially more per token to supplement the training cost even though the serving cost is marginal because most people will be satisfied with cheaper / less capable models for most tasks.

I buy that inference is a dropping line item while training is a growing one. There’s all sorts of things on the horizon that’ll be order of magnitudes improvements, from startups burning models into ASICs to get order of magnitudes more performance to alternate architectures like diffusion transformers that have orders of magnitude structural optimizations. It’s inevitable that it’ll come down even further from where we are. It’s possible model training also will go down but I’ve not seen any compelling research suggesting major “easy” reductions here.

[janalsncm]

The issue is that most tasks do not require frontier-level intelligence, but companies like OAI can really only profit off of the frontier. Capabilities from a year or two ago are so outdated that even OpenAI gives it away for free and there are many other models biting at their heels. In other words they are spending huge amounts of money to cash in on a depreciating asset.

So one possible future is that frontier-level training becomes so expensive and the use cases so sparse that it simply isn’t viable to keep going bigger.

[extraextra]

Once the land grab is over, the market will consolidate and the winners will absorb the losers. Then the few winners will be the only ones with real capital to train frontier models and will have true pricing power. Similar to how social media companies or the gig-economy benefits from network effects, AI companies will benefit from having the lion's share of paying customers (that also constantly feed in more data to train the models on).

[twobitshifter]

We have GPU costs, power costs, and how many token/s models can generate on those GPUs. It’s possible to figure out the marginal cost based on this. The current estimate is about $0.40 per million tokens for gpt4 equivalent model. Sonnet 4 is $15 per million tokens, so they are charging high margins on inference. The issue is how large of a margin is needed to recover their costs before the GPUs age out, and how high of a margin can be charged before it’s not economically viable.

https://www.gpunex.com/blog/ai-inference-economics-2026/

[rudedogg]

That seems way off to me.

I skimmed the article, but couldn’t spot any details on their estimates. They mention 70b+ params as being large in several places. But we’ve had several 100b+ param models that trail Sonnet.

[zozbot234]

Why would power spikes from training runs imply training>>inference? The cost of a training run scales with energy, whereas power is energy per unit time. All that tells you is that they're speeding up their training run so it will take less time overall (probably chasing some first-mover advantage, where they're out with a given model before their closest competitors), whereas they obviously can't do that for inference (which is a steady flow of requests over time).

[somewhereoutth]

Yes the huge discrete stepwise training spend is critical.

Maybe investors will realise that "the only winning move is not to play".

And so we are left with (as was) frontier models getting more and more out of date as whoever their post bankruptcy custodians are tries to eek pennies on the dollar for inference on their decaying property. Perhaps along with local and/or highly specialized models still feeding on the after-glow of the huge amount of training that was (and is no longer) done.

The next AI winter is going to be deep, savage, and long.

[extraextra]

Bankruptcies? The winners will gobble up the losers and the few remaining players will have pricing power. Don't be naive thinking that OpenAI or Anthropic can possibly go bankrupt. There will always be someone happy to buy them up for a nice price. Yes, the market will have to go through a consolidation phase though.

[galaxyLogic]

> frontier models getting more and more out of date

Why are they getting out of date? Is it because we have new content from the internet that the older models did not have? Or are we simply trying to increase the size of the training data? In other words not more up-todate in terms of time the content was created vs. wanting to use bigger training-input-sets?

[somewhereoutth]

lack of new content from the internet will make them go out of date. Not just facts and figures, but (for example) new programming languages/techniques.

[galaxyLogic]

I see makes sense. Then it it kinda says that quality of the model is the topicality of its inputs I assume.

[IX-103]

I don't see how it would be possible for inference costs to dominate training costs, even after amortization.

Training involves multiple passes over the entire training dataset, ideally in large batches where you can perform inference on as many samples as possible simultaneously and then perform backpropagation to adjust the model weights (which is about as expensive as inference).

Let's consider the size of the dataset we're dealing with here. The dataset likely consists of practically every piece of digitized text they can get their hands on (including that extracted from audio and video). We know Google has digitized a large portion of the books in existence as part of their "search book contents" feature and we have no reason to believe they're not using it alongside their cache of 90+% of the internet to train their models. We're talking about 100s of millions of books each with an average of 100,000s of tokens. The internet has 10s to 100s of billions of pages on it with who knows how many tokens on average. This is a huge dataset that we've got to go through hundreds of times.

Second, let's consider the effect of batching and how it sets requirements for our hardware. We know that larger batch sizes converge faster, are more stable, and produce better models. So if you want a good model you need large batch sizes. This means that you need machines several orders of magnitude more powerful than you use for inference. From what I heard Google uses clusters of 100s of the their TPUs all located in a single rack for training. These clusters are organized in a customized computing architecture to maximize memory locality between cores (really critical for efficient back-propagation). Further, you can't use reduced precision weights for training like you can for inference, so there are no shortcuts.

Finally, the initial training stage is followed by reinforcement learning stages - this is key development in how AI models have improved in the past year. This may mean going through a curated set of traces (either synthetic or captured from users) and adjusting the weights based on experienced outcome.

Overall there's so many orders of magnitude more work and more hardware requirements for training that I find it improbable that inference dominates. The number of "inference" steps in training is freaking ridiculous and includes such factors as the "number of words ever written".

[atq2119]

It's been a while since I saw a detailed paper on a high end training run, but extrapolating from what I remember, it seems those training runs are in the 10s of trillions of tokens. This already accounts for potentially sampling tokens multiple times during the training run.

That seems like a large number, until you realize that OpenAI claims to have almost a billion weekly users. And OpenRouter shows many models at over a trillion tokens per week.

So in pure token terms, I'd say it is in fact extremely plausible that inference dominates, at least for the popular models.

[johnecheck]

Not saying you're wrong, but I'll note why inference might dominate despite everything you mentioned.

A given model is trained once but applied N times. A large enough N will dominate training, no matter how complex and costly it was.

But how long is a model useful for? How often will labs need to train new models? Time will tell.

[upbeat_general]

This statement is well known to be incorrect for at least a year.

[extraextra]

Great points. - At the end of the day those are still private companies (albeit huge ones), so we can only speculate about the state of their private financial situation. Once they will decide it's the right time to IPO, they'll publish all their financials and we'll start to have a clearer picture. - Later, each company will slightly specialize and have a different go-to-market strategy, which will allow us to understand on a deeper level what works in the market and what doesn't (think about how Facebook, Instagram and TikTok are all huge universal social media platforms, but, each with a different target audience and different user base). - Finally, the market will go through a consolidation phase in which winners will gobble up the losers and then the incumbents will have a real moat (against new-comers) and real pricing power on their user base.

[stevenally]

> If we don't even know the ratio between amortized capital expenses and operational costs, outside investor analysis is impossible.

And yet we surely need this data for the IPO? Or are they relying on rule changes on the indexes to force ETFs to buy shares?

[extraextra]

The IPOs are months away, potentially 6 months or more. We're in a volatile macro environment. AI companies have all the incentives to not create higher expectations regarding their financial situation a long time before the IPO. Obviously at IPO they will have to disclose their full financial situation.

The market is super hyped anyway for their IPOs. If they raise investors expectations now and things change until the IPO, investors will be disappointed. It's a lose-lose proposition.

The smart play for any company is to keep their cards close to their chest until close to the IPO time.