The unbearable cheapness of open weight models
jamesoclaire.com
jamesoclaire.com
1) push the frontier in a way only massive scale can, and cash in on it (mythos level cyber security, recursive training, frontier science work). There’s big money for never before possible capabilities.
2) own the app layer with their edge in reputation and powered by their infrastructure. Be apple where everyone else is Linux. Do design, coding, research, SMBs, legal, finance, healthcare and more (they are doing all of this).
Will it be enough to justify a Google level valuation? We’ll see how fast they can push it.
In this case the people tasked with using the product won’t actually mind.
If the price difference is 50x? No way.
But that is months away, so not my problem?
The problem with this is that there are incumbents in all those spaces doing their own AI agents / platforms, and they're the ones choosing the models they use internally and they sell to their own customers. The margins and the possibility to fine tunie using open weight models, as well as the guarantee they'll keep running at predictable costs (no US orders yanking access), make them a very appealing option.
And if you're a company that needs an AI powered legal software, would you buy it from OpenAI/Anthropic, or from someone who you've already bought legal software from before and has the domain knowledge?
The existing AI companies can't even prevent their moat from being distilled by the Chinese token reselling industry.
I don't see how Anthropic or OpenAI survives being eaten by DeepSeek et al from the bottom of the stack and Google from the top.
It is truly amazing how bad it is. Made me miss using MS Teams. No software should make anyone miss using MS Teams
And Anthropic is currently cornering the enterprise coding market, and they were smart to avoid video. Under current economic conditions they're a lot closer to being profitable than anyone else, and they can take advantage of crashing prices for compute if we hit a datacenter-buildout-glut.
Harness example: https://github.com/evilsocket/audit
Apple and Linux barely even compete in the same markets. Linux runs on the servers and embedded devices, Apple on the smartphones. Android is technically Linux but not in the "is a good analogy for open weight models" sense because Android is so deeply under the thumb of Google. The main place Linux and Apple actually compete is for PCs and laptops, and that's the market where the thing with 65% market share is Microsoft.
Linux are on more phones than iOS.
3. Try to get the government to "certify models" to cause regulatory capture which is what both Anthropic and OpenAI has been pushing. No certification no use in business.
#2 isn't going to happen, because these labs have shown they have limited app/design sense, and they also lack the industry connections and domain wisdom to execute.
The way things are actually going to go is that these labs will set up partnerships with huge biotech/engineering/etc firms, and do custom training/inference on specific tasks that promise to be wildly profitable with them, then take royalties on the creation in perpetuity. Why sell inference when you can partner with Pfizer to make a version of Ozempic that also makes people freaky jacked, or partner with Bectel to make a radically safer, more efficient Nuclear power plant?
You do need business development to create those relationships. Saying they "have limited ___" mostly means they "haven't yet hired people who are good at ___". That's been changing already; the Claude app is steadily improving and handling more use cases simply through understanding which tools to use, Anthropic is building more relationships to create more tools, and all the frontier model companies are building relationships with companies that have specialized data and want specialized solutions.
I think we're also seeing the frontier model companies offer partners their own ability to run RL on their own data, and then retrain new models on the same data. That's going to make those relationships VERY sticky in ways that won't be obvious from the outside.
This study seems to show that there are places where synthetic data, especially related to common crawl.
> Pure synthetic data remains non-advantageous over CC; notably, models trained on pure rephrased synthetic data will underperform those trained on CC at larger models.
But the tradeoffs seem to be different at large scale.
> Overall, these model scaling results suggest synthetic data appears comparably less favorable for pre-training larger LMs relative to its utility in data scaling scenarios. Despite outperforming training on CC, larger models are not as tolerant to a higher ratio synthetic data as larger data budgets. This observation aligns with practices where synthetic data is effective for smaller LMs or specific pre-training phases, but less predominantly used for the largest models.
How I am reading it is there are places where it is useful:
> Notably, any mixture involving synthetic data, or pure synthetic data (except pure QA), is projected to achieve a lower irreducible loss than training only on CommonCrawl.
But it also seems that on textbook scale synthetic data, they did show model collapse vs rephrased data.
> These results contribute mixed evidence on “model collapse" during large-scale single-round (n=1) model training on synthetic data–training on rephrased synthetic data shows no degradation in performance in foreseeable scales whereas training on mixtures of textbook-style pure-generated synthetic data shows patterns predicted by “model collapse".
IMHO there are some very specific areas where we aren't "data limited", like math, but as your reference states "Our work demystifies synthetic data in pre-training, validates its conditional benefits, and offers practical guidance."
Note the cost of 30% of the total dataset being synthetic, where the model starts amplifying the generator's biases, leading to a permanent degradation in downstream zero-shot capability on unseen out-of-domain natural tasks.
My takeaway is there is nuance where synthetic data is an amplifier and where it is a problem, and in my mind that paper demonstrates it will not solve the data problem in general.
Correction: public text data limited.
There's a ridiculous amount of proprietary text and non-text data out there that much of society is run on.
They have high prices, not high costs. They will obviously keep prices as high as they can for as long as they can, while keeping demand up. Once demand starts to fall, so will the prices.
> Are these models cheap because they are open weight and having hundreds or people stress test running them on different hardware helped to lower the cost? Or is it that they are being provided as loss leaders to drive the prices down?
Neither. They are cheap because they have neither technical edge nor brand power to keep the prices high, and so have to ask commodity prices for them.
People somehow still don't get it, despite everyone who studies the economics of it telling them: Inference is dirt cheap. Training is expensive, inference is cheap, and getting cheaper.
Also, these open weight models are significantly lower quality than the high end coding models, and for some reason a lot of people think they're exactly the same. Maybe engineers who only dabble in LLM usage aren't doing enough complex work to notice...?
Money is made on the subset of inference that is charged at cost + margin via their APIs. API usage is so high because customers are still finding their feet, trying to understand how to measure the value they get from their spend, erring on the side of spend.
Yes, in a world of unmeasured value and tokenmaxxing, inference is profitable on SOTA models because all capacity is being consumed at all times, driving down marginal costs, but what about a world in which capacity isn’t constrained? There are still huge fixed costs.
Even the most optimistic leaks with the current high prices put the margin on API token inference at around 50%. How can SOTA models ever come close to competing on price? Price always matters. Offering the best model with the most brand recognition does not exempt OpenAI from the basic rules of business.
Historically, software has been such a successful business because the margins are incredible, 95%+ in many cases, driven by direct measurable value to customers that dwarfs the cost. A 50% margin at a time when your customers are falling over themselves to spend as much money as they can is not a good sign, it is a very bad sign, it leaves no room to ever achieve traditional technology margins, and inevitably leads to very weak margins.
Inference needs to become an order of magnitude cheaper than the value it delivers to ever have a chance of delivering on this wildly profitable vision. The cheap model providers have a much better chance of achieving that.
Outside of coding, almost every business case for AI doesn’t need above human intelligence, it doesn’t even need human intelligence, or half a human intelligence, a business can extract a lot of value from a machine that has a fraction of a human’s intelligence. Most human work does not use our intelligence, it is rote, a monkey could do it, and that’s where AI will be used most. Who is going to pay $10 per million tokens when they could pay $0.10 to get the same outcomes?
Mostly training. Claude didn't just get to be so good at coding by magic, it was suddenly so good because they did truly staggering amounts of RLHF and RLAIF on it. They are still doing that today, on any tasks they can figure out how to evaluate it on. This is capex for them.
Their margins on inference are >90% today for tokens they sell (plans are hard to count, but still profitable). Based on what we know of it's size and architecture, running Opus is not more than 2x more expensive than running Deepseek v4 pro, for which tokens are available at under 10% of the cost of Opus. Again, the reason their margins are 50% is because they are spending so much on things that are not inference, not because inference is expensive.
> The cheap model providers have a much better chance of achieving that.
Anthropic can do it with a push of a button, once they calculate that it will provide them better profit than current pricing.
That doesn’t make any sense, it doesn’t add up. Have you seen how much money they’re raising and burning? We know that training does not cost tens of billions.
Brockman said OpenAI expects to spend $50 billion on compute this year. OpenAI’s revenue run rate is less than $50 billion for this year! For 90% margins to be possible on inference, you are suggesting that less than $5 billion of that compute spend is inference and over $45 billion of that compute is training.
Anthropic have been desperately trying to juggle capacity by shaping user behavior through peak time usage limits because they are struggling with capacity for inference.
Plan based usage is widely acknowledged to be subsidized, you are probably the only person on earth suggesting that plans are profitable.
What the business world actually needed isn't intelligence, it's VBA with a bit of polish on it.
Yeah, people want tools to distill reports, and puff nonsense into bigger nonsense, but to a remarkable degree, this doesn't require an LLM. In fact, the alternatives might be preferrable by offering more consistency/repeatability and efficiency (I am so sick of watching a LLM babble for 5 minutes on something a regex would do in 5 seconds).
The genius of the LLM industry is that it avoids "programming anxiety" by hiding it behind a friendly-ish UI and not calling it programming. It's another in the string of innovations like "hide the file system" and "removing user programmability" so we can sell it back to you.
Inference is cheap. Anthropic is only drastically subsidizing their plans if you count their training expenses as part of their costs.
No one knows if it’s profitable or not so we’re left to speculate.
Also for agent doing r&d, cheaper tokens allows doing more, which is always good.
The question is can they just stop training at some point, fixing the models in time, and still have a useful product.
What's the end-game? It sounds like in this business, if you are right, your revenues drop every year.
A question for economists... It seems plainly clear to me that information and information processing is commodifying (for the first time in human history?). Without the age-old bottlenecks at the top of the value chain, capital will surely flow downwards, right?
Isn't this the thing people have said about every new technology since the printing press? And it has been mostly true, but it has also been the case that the incumbents have fought hard to lock things back up again. Newspapers and radio stations buy each other up, the open web gets locked inside Facebook (which, 30 years ago, people were already worried about with AOL), people have computers in their pockets they can't run their own programs on anymore.
Interests are going to want to lock the new information thing behind a gate so they can charge a toll and censor what they don't like, same as it ever was. You don't win by default, you have to fight to stop them.
Recall the notion of a bottleneck, and this distinction will become clear. Those prior technological changes never inverted a bottleneck, and this one does.
Computers and the internet did a lot to make production easier in addition to distribution. Anyone today can use a photo editor to superimpose text over an image in any font in seconds like it's child's play. That used to require knowledge of calligraphy. Film production used to require very expensive equipment that everyone now has built into their phone.
> Those prior technological changes never inverted a bottleneck, and this one does.
Before the printing press, copying books had to be done by hand. If you wanted a million copies of something made you had to be the church or a government. Today there are independent pundits who get a million impressions on their shitposts, and that's with consolidated platforms being largely against them.
We still have an entire edifice (copyright) which is structured around copying requiring a sufficiently centralized apparatus to serve as a useful chokepoint for imposing restrictions and collecting royalties, which is correspondingly under increasing distress as
No because the biggest factor in their current price is VC subsidization which has likely peaked if OpenAI is now serving ads and Anthropic has increased their API pricing
I never used Fable, maybe it is that much better. DeepSeek has no problems with the workloads I give it though - if it only keeps marginally improving with each interaction I don't see myself needing to come back.
there is a larger appetite for something like open source AI mostly b/c of price. we all know these labs have not figured out their pricing model, and we're all holding our breath out of fear of what the prices could be.
also, if you consider that the only toll to knowledge work before was personal time, and now you need to pay $100s month just to keep up with the baseline speed. it makes sense people are looking for something that gets them back to a workflow where the price to do work is near $0.00.
I think for a smaller group though, it's more to do with a certain combination of principles. Some people don't want censorship, other's want ownership, some want the knowledge of working on LLMs to not be gate kept.
For others even a small edge can be important. Pharma and Biochem research. Research in general. Any industry where there are major reputational risks.
It may not make sense to use the most expensive model to replace your payroll clerk, but there are plenty of use cases for the best available.
harness <-> gateway <-> inference provider
Easy to switch any of them and (mostly) possible to combine any with any
Replacing something like Excel is crazy-hard because of network effects, replacing an Enterprise CRM is akin to a removing a metastasizing cancer
The reason it’s like the Industrial Revolution is simply that there’s no question it’s going to completely transform jobs. It can make a very similar difference to the difference between a craftsman and a factory worker. The latter is massively more productive.
This only works because prompt caching is done by matching prefixes, not the entire input.
From my understanding if previous tokens are frozen and guaranteed to be immutable you can leverage that.
AGI? Too loosely defined. They lack a lot of competences which humans recognise when we see them but find it hard to put into words; on the other hand what they can do they already do faster than any human (and have greater breadth than any single human, but this usually doesn't matter because "coder" and "economist" and "translator" gets solved in human teams by hiring three people).
I do not think current ML has the tools to solve for quality. But we know it's possible for a really mediocre intelligence to make human level intelligence, because evolution made us, so for me the question of AGI is more a practical one: is it affordable?
(I also think not at the present time, but that's an "I think" not "I am analyzing it carefully").
Or maybe you don't take Elon seriously when he talks about Mars.
I am only dismissing the orbital data centres, I do see a future for Starlink. One with competition, but a future nonetheless.
I'm old enough to remember the dot.com bubble and "we lose money on each unit and make up for it in scale":
If they don't make sense, they don't help. Putting a single one in space, or even a handful, is physically possible! But even optimistic Alphabet researchers (and Alphabet owns more of SpaceX than the entire IPO) say this only makes sense at $200/kg, while early Starship launch costs while they sort out reusability be at best $400/kg and the researchers don't expect $200/kg until the mid-2030s even with a high launch rate:
If the learning rate is sustained—which would require∼180 Starship launches/year—launch prices could fall to <$200/kg by∼2035At $200/kg, and using the payload estimates elsewhere in the paper (the learning rate is based on mass rather than launch count), they'd need to launch 370,000 tons (4.4 ibid); even at the "good enough" cost, $200/kg, they'd need to spend $200/kg * 3.7e8 kg = $7.4e10. That's a hell of an R&D spend for the next 10 years of a company whose lifetime revenue (not profit) is reportedly $4.6e10.
My current draft has a few thousand words of additional problems, plus a bunch of things which I mention only to say why they are not, and some more where I say the research has yet to be done.
> Or maybe you don't take Elon seriously when he talks about Mars.
Used to, not any more. Has been too slow with Starship even before the fact that iteration with hardware is necessarily slowed down by a 2-year gap between launch windows.
There's not even been any news about demonstration models of either Mars-rated or Starship-rated Sabatier processors, which would be an easy win and also win points for both environmentalism and energy independence viz. Iran/Hormuz.
If you build the DC satellites as currently specified, you're strictly better off not launching them. That's how bad the idea is.
Cheap access to space was once a pipe dream.
Reusable boosters were once a pipe dream.
A new player beating Boeing to the ISS was once a pipe dream.
LEO constellations were once a pipe dream.
Launching thousands of satellites was once a pipe dream.
You should know that a) they are already running "AI" chips on their current sats. and b) they are already producing kW of power on orbit and have ~10k sats on orbit. You can watch Scott Manley's video on it, where he does some rough calculations and explains the overall architecture. There is nothing stopping them to do this, from an engineering perspective. If it makes commercial sense, that's another question, but 5-10-20 years in the future things might change there as well.
And my point was that at one point or the other there were many "downsides" for all the tech that SpaceX already has. Reusable boosters were seen as "uneconomical" and "pointless unless they can fly 10 times" by industry experts. They're now flying 30+times a booster.
LEO constellations were similarly "full of downsides" plus "all the companies that tried it went bankrupt in the 90s", so "it's pointless". And so on.
Pretty much everything about data centers in space is worse than having them on Earth. Apart from niche use cases, the only reason you'd talk about data centers in space is if you had a company with rocket ships and needed a story to tie your rocket ships to the current AI craze.
And in ocean you don't have to solve for radiation nor cooling.
[1] https://www.tomshardware.com/desktops/servers/microsoft-shel...
I'm currently writing a blog post, and there's one big thing everyone, including Scott Manley, missed.
Once I realised it, I wondered what took me so long to spot this issue.
slightly related .. I saw a talk on DCs in space, and it said median Earth orbit had a latency of 500ms .. but back of envelope seems to be : 15,000km above Earth would have around 100ms latency, comparable to internet ping times.
Not an expert, feel free to weigh in.
I'm still working on the blog, but as a quickie: it's the lesson of the Datasaurus dozen, that sometimes you need to look at the actual distribution rather than statistics.
Here's what the safety exclusion zone around a million of them in orbit looks like, if arranged something like the current plan: https://raw.githubusercontent.com/BenWheatley/blog/refs/head...
There's no (safe) gaps. Plenty of physical space, but the safety margin eats it all up. Nothing else is allowed to use those orbital shells or anything between them.
Also, this is what happens if you put them all in a single orbit at the same altitude:
https://raw.githubusercontent.com/BenWheatley/blog/refs/head...
> slightly related .. I saw a talk on DCs in space, and it said median Earth orbit had a latency of 500ms .. but back of envelope seems to be : 15,000km above Earth would have around 100ms latency, comparable to internet ping times.
500ms means ~150,000 km travel distance; for that distance as round-trip time from origin to destination and back again means the one-way distance is 75,000 km, so if it's via a single satellite bounce then the average distance to the satellite would be 37,500 km: [You]-37.5Mm-[Satellite]-37.5Mm-[Them]-37.5Mm-[Satellite]-37.5Mm-[You].
I think they must be assuming all comms are via geostationary satellites. In some talks, this is what the speaker actually meant, though they may not have been clear about it; other times, there's talks from people who copied the former but perhaps didn't understand.
For DCs in space, even in GEO, it would be half the distance because you're communicating with the satellite itself not with someone else somewhere else on the ground.
TL;DR: Alphabet researchers (and Alphabet owns more of SpaceX than the entire IPO so if anything they're biased to optimism), recon it will take SpaceX launching about 370,000 tons to orbit before they've even figured out how to get the costs down to the point it makes sense to put these in orbit.
If you don’t care about making any more from it. How exactly would datacenters in space would be more profitable than those on earth?
Turns out that scaling up compute is much more important and also limits the upper end of intelligence.
Suppose it can do 80% of what the 20th percentile human can do. That's a huge advance and very useful, but it means there are still things it's not very good at. If any of those things is (or becomes) a bottleneck, you're not getting the hockey stick graph.
However, it's difficult to sell this to businesses who want contracts and KPIs, not staff and commitments.
Regulated industries will favour the closed sources, either by choice or mandate. The interesting question is whether they will have better models, or worse models. History says they will receive a worse service, but continue anyway.
Until your country will appear on naughty list of US administration because your local politician did something what mildly inconvenienced US oligarch
It's more expensive to use GLM 5.2 paying z.ai or Opencode Zen API rates than it is to use Opus on a subscription plan. Both of those providers offer subscriptions priced favorably relative to their API rates, but only in what are effectively trial sizes.
1. They overprice their APIs to make their subscriptions look reasonable
2. They burn money with their subscriptions
You and readers may be interested in Europe 2031
Having said that, while one can always hope, I would assume that Oracle is one of these companies that will be bailed out or find a way to survive.
Developing these things is NOT free, there's a lot of labor, hardware, compute/memory/storage/network that goes into that. Who's paying for all this? Chinese govt? Developers themselves? What's the revenue model here?
I absolutely LOVE ability to either run them locally or access inference providers on the cheap, but having a hard time understanding the financial side of this.
Are weights separable from a model? And if not, what is the point of saying "open-weight model" instead of just "open model?"
To the newcomer, it's hard to determine what the components of an AI system are from the throwing-around of these terms.
The trained weights are open, the training software is open, and the data that goes into training the model is open.
Not many models are fully open.
An open weights model is one that has freely available trained weights, and maybe fine-tuning tools, but it lacks the original training data (and usually lacks the training software). These are the most commonly used local models, like Google's Gemma series, Meta's Llama, or Alibaba's Qwen.
Also, I've read a bunch of descriptions of AI components, but none of them has said what the weights are applied to in the model. I guess that every model contains a dictionary of words and phrases, and the weights map relationships between them?
All the descriptions simply talk about weights being applied to "input," but neglect to say what that input is compared to. If a user submits a query, are the words in the query weighed against the words in the model?
Can you recommend a primer on this whole process?
https://www.3blue1brown.com/lessons/mini-llm/
To quote part of it, Training a model can be thought of as tuning the dials on a really big machine. The way that a language model behaves is entirely determined by these many different continuous values, usually called parameters or weights.
Longer and slightly more technical, "Intro to Large Language Models" by Andrej Karpathy:
I'm really just trying to get a handle on the different layers so I can set up an environment and put it to work for limited coding assistance. For my purposes, I think an all-local setup is the best way for me to learn and should be enough for the coding tasks I have in mind. Mainly I want to automate tedious tasks away, like "modify these Swift classes' members and JSON deserialization routines to match what's coming out of my server API."
Thanks for the link! Reading it right now... and yes, this looks great. Appreciate it.
That data includes not only the "weights" but also various files with required information, e.g. the tokenizer, the chat template, files that describe the structure of the "weights", e.g. number of layers, the number of "experts", routing information, etc. All this information may be distributed in many files (e.g. *.safetensors files with weights, *.json files etc.) or it may be aggregated in a single container file (with the .gguf extension).
You can see an example of the files included in a very simple open weights LLM here:
https://huggingface.co/google/gemma-4-12B-it/tree/main
Bigger LLMs have much more files, especially much more *.safetensors files, which contain the "weights". The "weights", i.e. matrices of numbers that are used in the computational algorithm that generates the output tokens, constitute the bulk of the data needed to run a model, i.e. from a few gigabytes to a couple of terabytes, which is why the term "open-weights" is used, but in fact by this term it is understood that all data needed for running inference is open.
For an open weights LLM, you do not have access to the data set used for training the model or to the algorithms that have been used during the training of that model.
You can still do some fine-tuning of the model, using your own training methods and your own additional training data. To facilitate this, several open weights models offer not only a model version that can be used for inference to implement a chat application or an agentic workflow, but also a "base" or "raw" version that is not suitable for being used directly for inference but which is suitable for you to do a post-training/fine-tuning, to create a model more appropriate for your particular needs.
An "open weights" model is sufficient for most of the potential LLM users, because training a model is something that requires expertise, expensive hardware and a lot of time, so few would be able to do it even when given access to the necessary data.
"Open weights model" means the developer made the model available for everyone for free. You can download it from huggingface.co for example and do whatever you want with it.
Why "open weights" and not "open source"? Because the "source code" for LLM would include things like training data, training methodologies and tools, so that you can do the training and produce the model (files) yourself. That would be like compiling from source code. Which is not done with these models, it's company's know-how, they only share the end result.
It's more analogous to "freeware" which is what we traditionally call freely distributed binary executable files. But people started calling them "open weights" instead and the term stuck.
Most of the cost of supplying inference compute is depreciation of the GPUs. Maybe Deepseek is anticipating a 50 year life for theirs.
The things that I need to automate do not need frontier models. Heck, even a gemma-4-12B-it-qat-UD-Q4_K_XL can deal with a lot of complexity if properly guided (it can run on 16GB of unified memory, for example on a base model Macbook Air).
I've been using it to translate Javascript to a custom scripting language in a product I work for, just by providing a system prompt and an MCP tool to call the target compiler to check for errors.
Sometimes it converges faster than Opus 4.6 (I've tried) because it doesn't over-think stuff.
If it were a person I would say it knows less, but it's still smart.
I mean, you don't need the most powerful tool at all times. We treat AI as one-size-fits-all, and once cost gets in the way, it will matter.
There's many providers that run open weights models and give you access. Many decent open weights models cannot be run on consumer-grade hardware (DeepSeek, GLM, many others).
A $4,699 DGX Spark can easily run Qwen 3.6 35B, and its performance crushes GPT-o1, the SOTA from two years ago
anyone who disagrees is not seeing the forest, only the trees.