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.
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/
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.
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.
Yes it seems that this discussion that has sparked such controversy involves an already well defined concept in business.
Net margin versus gross margin.
Net shows profitability after extracting all expenses while gross only extracts the cost of the goods sold. Putting the model training costs into a one time fixed expense provides a much better gross margin.
This is known as COGS reclassification or classification shifting and is a common tactic to mislead investors.
This is why analysts look at Free Cash Flow Margin.
WorldCom and MicroStrategy did this before the Dotcom Bubble imploded.
Just look at large open weights models being served by inference providers.
Kimi 2.6 is a 1 trillion total / 32B active parameter model that's something comparable to Sonnet. Sonnet's API pricing is $5 in, $15 out per million tokens. Deepinfra serves Kimi at $0.75 in, $3.50 out, and about the same at openrouter. So you're looking at a 4-7x multiple that Anthropic is charging compared to market rates that any plebe can get with a credit card.
I'm not sure just how good that looks for Anthropic/OpenAI.
4-7x isn't a tiny markup, but how does that compare to high-margin internet businesses like AdSense? Meta and Google do hundreds of billions in ad revenue a year, and after taking out the publisher's portion (60-80% per some searching), I wonder what the ratio of the remaining tens-of-billions is against the compute cost and headcount required to run it.
And how much room for maintaining or improving that margin do they have if the cheap competitors also continue getting better? Is there a "good enough" point where the easier inference tasks are all moving to vendors massively undercutting them, and then they don't have the volume necessary to justify spending on further cutting-edge development?
> Kimi 2.6 is a 1 trillion total / 32B active parameter model that's something comparable to Sonnet.
No it's not. On some rigged paper maybe. Some such benchmarks say all models group together, which they clearly do not.
> Sonnet's API pricing is $5 in, $15 out per million tokens. Deepinfra serves Kimi at $0.75 in, $3.50 out, and about the same at openrouter. So you're looking at a 4-7x multiple that Anthropic is charging compared to market rates that any plebe can get with a credit card.
That's not saying much. You can get "cloud" at AWS and you can get a VPS. There is likely a 10x difference. It's not "same". Whilst AWS costs more they also don't have 7x margins similarly.
In the US, companies are not allowed to unfairly privilege some investors over others by giving them access to secret information that would let them judge the future prospects of the company. (Except in all the ways they can, but these usually involve some kinds of insider trading rules.) Private companies can handle giving out secrets to investors by literally writing and memo and mailing it to all their investors, if they want to give out some secrets to one of them.
Public companies cannot do that, even if they knew who all their investors were, but must instead consider every member of the public a potential investor, even if they don't already own the stock. Because of this, when public companies want to reveal material information about their future prospects, they must reveal it to everyone.
Besides the legal requirement, the reason these companies go public is often to provide liquidity for early investors or employees. So they do want to have as good of a margin story that they can, at least in terms of unit margin.
This is an interesting anomaly in the US. In the civilised world all corporations have to file public accounts, as the price for their limited liability. The detail and audit requirements depend on the size, turnover, staff numbers etc. This is because the shareholders are not the only stakeholder. The companies creditors, for instance, who are exposed to the limited liability have a right to see what they are lending to.
To answer the sibling comment, all of these public accounts follow local GAAP or IFRS.
The US still astounds me with its willingness to allow corporations to rip people off!
Creditors in the US can make visibility into financials a requirement for financing if they want. Protecting creditors isn’t a good argument for public reporting.
What about potential employees, can they look? The local community that consents to let the company build and operate in their town? How does that help, if they don't follow have to follow GAAP anyway?
Why are those things relevant to either employees or a town?
Most of the US is at-will so the financial health of the company is unlikely to be the reason you’ll suddenly lose a job.
Same for a town, if you’re structuring a deal that has counterparty risk then you mitigate the risk. If an employer is just leasing some office space in your town, why in the world would you ever even think you had the need to look at their financials?
As a consumer you are often sending deposits or even the full cost of goods to companies some time before you receive those goods (in effect you become a creditor). You are also dependent upon some of those companies for service and repairs. It seems reasonable that you can check the finances of a company you are creating a business relationship with, I know in the past I've checked company statements.
You are unlikely to have significant enough sway to force that kind of disclosure. Small businesses as consumers have less legal protection and are similarly unlikely to be able to make disclosure a precondition of a deal.
So what. As a customer you can insist on seeing audited financial statements as a condition of purchasing, or purchase from another vendor, or do without. No problem.
Isn't there a limit on the public markets where if a company has less than a certain percentage of its ownership traded publicly then it is no longer a public company and therefore de-listed?
I remember hearing about a guy trying to squeeze out short sellers of his own company but ended up effectively taking his company private because he bought out like 95% of all the shares.
I wonder how that aligns to these small releases of stock for the public.
There is no legal minimum free float requirement before deregistration in US, however, different exchanges have different rules
Essentially, a stock has to stay above 1$ per share, have a minimum market cap of $15m, minimum 400 shareholders and "adequate" liquidity
If it meets those 4 criteria, it's essentially not at risk of deregistration
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.
> 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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
> 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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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?
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.
> 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?
> 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.
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.