"Based on this framework, we have achieved high-throughput, low-latency large-scale inference services for DeepSeek V4.1 Flash. Currently, the service throughput of a single cluster is 155 trillion tokens/day, and the overall revenue is $38 per card per hour (at off-peak pricing)"
It's really refreshing to put some cold hard numbers on this stuff, hopefully this quells some of the "AI is unprofitable everyone is spending a dollar to make fifty cents" arguments. Amortized cost of owning a DC full of GPUs is far less than $38/card/hr
Not to say it can't still be a bubble, just that there is legitimately profit to be had and perhaps some sustainable equilibrium in the future.
> hopefully this quells some of the "AI is unprofitable everyone is spending a dollar to make fifty cents" arguments.
I think everyone has known that inference is profitable overall for the major providers for a while (I'm sure not all user accounts, but in aggregate).
The major providers are unprofitable after you add in the training and research costs and other overheads.
This is why the open source models are such a threat if they become good enough - it will destroy the business model of just providing raw chat/intelligence.
Arguably that's also why a few of these providers are trying to diversify into whole-platform approaches rather than just relying on having their whole thing be a great model.
There has always been a lot of evidence that prices are only going down. The fact that you could already get DeepSeek or GLM inference from 20+ providers on openrouter.ai was already sufficient evidence that inference was profitable and being commoditized.
People that want to believe something different as a coping mechanism will not be swayed by this evidence either.
AI will end up being a commodity. That's also why you already see the first rounds of protectionism under the guise of national security. Obviously they do have a point but that point works against them just as easily. Anybody that integrates unreviewed model output into their workflow is asking for trouble.
Yes. I've always suspected inference is profitable.
But I'm also interested is retraining needed? how relevant will a model be in N years if it's just "static", and if you do need to retrain and fine-tune over time, is it still profitable? how often is that needed?
For most cases it's probably ok, often times answers are general or basically a web search anyways, but if my Agent started blasting outdated syntax it would suddenly be way worse for coding for example.
I doubt that, I got downvoted several times pointing out inferencing is insanely profitable.
Even take account of R&D cost it is still profitable, because you do not actual retrain the base model every month, only distill and post train from the same foundation.
heck, I think even including the giant cost of the top tier researcher salaries, they are still profitable.
then, how are they still losing money? just read what they are actually spend the most budget: always the hardware. those companies will never profitable, because they will never stop spending money to expand their computational prowess.
I don't think so. The number of comments I see from people talking about AI, there are plenty who still think the price is going to shoot up. Some even talk this way about models you can run locally on high end yet still consumer level hardware.
I think the more someone works with AI the more likely they realize that inference does not have to be sold at a loss, but many people's AI knowledge seems quite dated and it sounds almost like it becomes a convenient conspiracy theory for some, a way to convince themselves that the market changes that are already happening a temporary (and I don't mean in a Jevon's paradox sort of way, but that companies will go away from AI).
I suspect a lot of this is fueled by a few stories of companies who went all in on spending as much as possible on AI and then begun to walk that back, without realizing that was more about bad incentives and over promises of what AI could do (especially 1/2 model generations ago). Some also confuse running the SotA models at the fullest capacity costing more and more with a given problem costing more and more to solve, not realizing that the cost per problems is trending down but newer and larger problems can be solved for even more, so the costs at the top are rising despite falling prices. Also things like the subscriptions getting more expensive, though how much of that is because developers are doing far more than they were 6 or 12 months ago (and in some cases there was subsidization that is going away). And maybe a bit with the free models starting to not be upgraded or the harnesses getting worse performance of lowered free limits.
I can't read the original Chinese text, but if it's revenue, then you still can't say anything about profitability. Housing and powering one GPU should be way less than $38 per hour, so there should be a profit on inference. If that profit can cover training new models we don't know, but it would appear to not be the case.
For the past year or two I think it has generally been agreed that inference is profitable, but the LLM business as a whole is not.
Would that be due to the size of the models from DeepSeek or perhaps only for their distilled models? If it's for general purpose models that would crash the valuation of Anthropic and OpenAI, both who are struggling with profitabilitet to say the least.
Until the Chinese stuff came online I was completely convinced this is just a bubble waiting to pop and nothing will come of this tech due to its absurd costs. The Chinese stuff changed this picture massively. I'm daily driving DeepSeek 4.1 flash (at API prices) and every now and then I give Opus 5.5 a try and don't notice anything missing for me.
My workflow is one where I am heavily involved in the process and I don't like giving the agent too much autonomy + I need to follow along. Opus 5.5 speak is still cause aneurysms for my brain. It also has this vibe to its prose... like it's looking down on me. This is Prolly my bias speaking though.
> I'm daily driving DeepSeek 4.1 flash (at API prices) and every now and then I give Opus 5.5 a try and don't notice anything missing for me.
I was doing the same (along with GLM 5.3 Flash) for months, and then gave Claude Sonnet 5.5 at medium effort a try. The difference was like night and day. Where both DeepSeek 4.1 Flash and GLM 5.3 Flash kept making mistakes and would have to do a bunch of iterations to find and fix them, Claude Sonnet 5.5 just did it right on the first try.
In my experience, while DeepSeek 4.1 Flash and GLM 5.3 Flash are a lot cheaper per token, they're just as or maybe even more expensive when you factor in all the extra tokens that are need to troubleshoot and fix their mistakes -- and even more expensive if I factor in my own time wasted doing the same, not to mention the frustration of having to deal with that.
DeepSeek 4.1 Flash also hallucinates a lot.[1]
I'm hoping all this improves with time, and the gap between the competent frontier models and open weight models narrows.
AI as a business will be 10x larger than it is today within the next 20 years.
Present valuations are the bubble, not the business / industry. The exact same thing was true about the Internet bubble circa the late 1990s, except this is a lot more important than the Internet and the global economy is drastically larger.
Productivity is how you expand economy and wealth. It's how China can go from being one of the world's most backwards agricultural nations, unable to feed itself properly, to gradually becoming an agriculture juggernaut. AI is a hyper productivity tool, saving people time and bolstering capability are two of the primary productivity concepts.
> Amortized cost of owning a DC full of GPUs is far less than $38/card/hr
You missed the bit where they used the word "revenue".
I suspect if you factor in all the various overheads of running datacentres the actual value is lower. Revenue is vanity, profit is sanity as the old saying goes.
However the Chinese do have the benefit that China are pushing hard on renewable energies. So at least that major part of the equation is probably far lower than your average US/EU site.
Huge, generational mistake for the USA to be leaning so hard into fossil fuels right now and actively trying to prevent buildout of solar and wind power. That decision is just going to hurt more and more in the coming years as China's electricity costs continue to plummet while the US' continues to depend on unsustainable fossil fuel infrastructure.
It's hilarious and depressing that one of Trump's supposed goals was making the US into an industrial powerhouse again, and yet he's acting to protect the profits of the fossil fuel industry as opposed to encouraging the kind of resource efficiency that would make the US competitive long term.
> You missed the bit where they used the word "revenue".
No they didn't...
> I suspect if you factor in all the various overheads of running datacentres the actual value is lower.
That's literally exactly what the quoted part addresses: "Amortized cost of owning a DC full of GPUs is far less than $38/card/hr". Cost of running GPUs is far less than $38/hour, revenue is $38/hour, ??? , profit
Q1 '20 - $0.10 profit per gb shipped (ram and storage)
Q1 '23 - -$0.024 (cyclic ram market)
Q1 '25 - $0.131 this looks normal were on the upswing.
Q4 '26 - $1.136 Were all fucked.
If you look at a chart of S3 prices we have all been living in the "Happy Time" where the cost to run our ever growing software business has gone down. Githb says 2025 billion commits - 14 billion this year, and 2x the actions run. That means AI driven software growth is 2x to 14x.
But those ram/storage prices are going to hit the operational budget of EVERYONE soon.
Can your average sas survive a 4x price increase on its AWS/Azure bill? How about 10x (cause thats what the margins say the premium is on ram).
If the circular financing is a concern, the circular firing squad should be a bigger one. Your paying for tokens to build more software to pay more to run your software is not a sustainable cycle. People are starting to use the words CapEx again - and it's a good thing.
With this new data at hand, I have to wonder how profitable would it be for the investors.
The cost of amortising a datcenter, the gpus, the engineering costs attached all to produce inference.
At how much percentage of its inference would it be profitable (if it does), how much returns would it make over the long term and what would be the profit of those inferences/ profit-revenue and how that compares to other industries.
Are there any statistics which show this. Also how efficient token inference has gotten over time?
You also have to realize deepseeks running costs are immensely lower than say any us lab by orders of magnitude.
Chinas grid is almost entirely renewable energy with strong investments in it. In the us not the case, some dcs are running on gas.
Also DeepSeek was smart. They strategically rolled out model sizes they could support even if the intelligence wasn't the world's best for that week in human history. Then they have incrementally and cheaply improved models since.
Not to mention they also built trust with open weight models rather then fear mongering. They are in the top 3 of the smartest most viable LLM labs out there.
Today. As new supply comes online and demand levels off, that 38$ will shrink. Absent legal monopolies, the margin will shrink to a sliver more similar to any other service. That is the danger of open source models. They destroy monopolies and shrink margins.
That would depend on whether you are a user/customer or an investor/owner in AI. At the moment, nearly everyone is on the investor/owner side. The entire western world has doubled down on AI becoming the most popular product since the invention of fire. If it turns out that fire 2.0 cannot be patented/copyrighted/licensed for insane profits, those dollars disappear.
The people who own these companies have more money than they, their families, and their extended families can spend for generations. None of them have to work a day in their life no matter what happens to these companies.
Same with a lot of their investors, who will be just fine without the "insane profits" they hope for.
I'm not worried about their welfare, and won't shed a tear if their companies collapse.
> That is the danger of open source models. They destroy monopolies and shrink margins.
It's unclear that the leading open-source models could exist without having distilled the frontier models. If that's true, the "danger" is that these upstream models don't get built any more (because of the tremendous cost), and we end up just optimizing what we have. Also unclear whether this would be a net win or loss for humanity.
Anthropic is zooming to $100 billion in annualized revenue.
The tremendous cost is matched by the tremendous business output.
In the span of less than five years, starting from their first $1 of revenue, Anthropic + OpenAI will have produced businesses doing at least $150 billion in combined sales. The demand isn't reducing at present, it's climbing, they can't come close to servicing the total demand for their frontier models.
SAP is under $40 billion in annualized revenue. They're 54 years old.
The tech is at the point where synthetic data is almost high enough quality to train further models on, AND there are so many human users of cloud AI, that their usage can be gathered and trained upon (it's human generated and used in real world situations, so, it is of a good enough quality). The cat is fully out of the bag.
This deals with half of one of the three scaling factors only. Presumably data production (choosing your training data, etc), model size, and training time/power still contribute huge costs. Who's investing that cash if open-source models will distill out enough to compete without the investment costs?
There is a community of people who compile LoRAs of their favorite subjects/styles in order to fine tune models. There is currently a bit of a gap between LOrA creation an integration of that data into larger models but I can see serious hobbyists stepping up to that plate soon.
> they were built by distilling the world's knowledge and cultural products
... and a bunch of RLHF and proprietary training data etc etc, but regardless, this is still an exceptionally expensive process, that someone somewhere is paying for.
I would be surprised if there was such a glut of undifferentiated inference providers for open models otherwise[0]. Possibly just people trying to minimize the loss from already-committed capital, but I don't see AWS and DO selling inference via OpenRouter unless they're making a decent profit from it.
Fascinating site. Thanks for linking. However note if you change to "open router" pricing in some cases even B200 which is ~2y old has negative margin (and AFAIK ppl are using longer amortization than 2y - I do not know how cost are calculated here though).
DS4 Flash is so cheap also, imagine what the numbers look like for a flagship model like Opus 5.5, which charges 100x more. Of course inference for that model also costs more, but I doubt the proportion is the same.
The top quality frontier models are used to compress / improve the earlier, lesser models, to make them more efficient. So that pushes costs way down on previous models.
Do they share latency / concurrency figures (aka the deployment shape that gets them that token/s/cluster? Would be interesting if they're using deepEP/WideEP + PP and TP
Deepseeks strategy is that each project must break even within 10 months.
$38/hrs equals $320K in 10 months.
Assume they only have 100 clusters, and as much as 40% operational costs. That would get them about $20M in 10 months, which I suspect is more than what 4.1 flash training has cost them.
I saw the same kind of comments when the first Deepseek model was initially released. However the relevance that investors assign to something like this does not seem to match the excitement expressed here.
I think they do not believe that big corporations in the US or Europe would actually use chinese models in their business operations due to security, regulatory and/or business continuity concerns.
> I think they do not believe that big corporations in the US or Europe would actually use chinese models in their business operations due to security, regulatory and/or business continuity concerns.
They'd be wrong. Especially about Europe. Many companies are in a bind about this: they can't upload their confidential files to a service provider but they do want AI. So right now they end up with a choice: do nothing, breach their contracts, use a service provider that pinky promises not to share their data (which we have already seen they will violate) and run local models.
The latter is the reason you can't buy powerful GPUs for reasonable prices because a local model (GLM 5.3, DS 4, etc) + some hardware is all you need to solve the problem in a way that avoids the legal issues.
You can’t buy gpus at reasonable prices because US frontier labs are spending nearly a trillion dollars this year on data center buildout, not because some mid tier European insurance co needs to serve deepseek
Yes, well, you're wrong about that. They're buying GPUs by the pallet here.
Mid tier insurance co is working with local inference providers and some of these now have thousands of GPUs up and running. Don't drink the HN cool aid that the EU is a backwater where nothing happens too much or you'll end up surprised one day.
The one thing where Europe really does lag is in training, that would be fair. But inference providers are popping out of the ground like mushrooms and there are some billion euro+ projects underway. We -unfortunately- also have some scammers that are bilking (or attempting to bilk) investors pretending they have more than they do, this is annoying but I would expect that other countries have similar players as well.
The 'by the pallet' buys are for small local projects, not data center size projects (of which we have more than a couple as well), as you probably already knew. This isn't some silly dick measuring contest.
The point was: they're buying them from the retail channels in relatively high volume because there is a massive shortage of GPUs in general, but that does not seem to stop parties with smaller projects who do not qualify for buying directly from NV (assuming NV even has stock to sell).
5090's now sell for the same amount (or more) that you could buy an RTX 6000 for a few months ago, and RTX 6000's are ~19K and up.
I found a stash somewhere, bought two, then two more and when I was ready to buy the remainder they were all gone. Serves me right for being too careful.
> I think they do not believe that big corporations in the US or Europe would actually use chinese models in their business operations due to security, regulatory and/or business continuity concerns.
Quite failure by someone to assume that EU businesses would use US over Chinese, based on the current political atmosphere. US causes even bigger concerns than Chinese labs.
> they do not believe that big corporations in the US or Europe would actually use chinese models
This is exactly right. The only AI that matters from an investor perspective is the AI that is built into your phone, your car, ships with Windows and Mac, and built into legal and business tools like Lexus Nexus and Salesforce.
That's why OpenAI hard pivoted from consumer tech into enterprise ?
The only AI that makes money is enterprise. Being open weights/self-hostable is a huge win before the pricing even enters the conversation, but pricing is also adding a ceiling on what US labs get to charge. And stuff like powering other software also works better with fine-tunes, hence legal models, etc.
European companies are more and more considering and planning to deploy Sovereign local models.
Currently the price sovereignty is still not outweighing the cost of US dependency. But big providers are already working on finding ways to ensure customers independence of the cloud act.
Also the concern that open weights models could contain dormant adversarial behavior (i.e. they are sleeper agents) is more popular amongst decision makers than one would think.
In that regard it was a clever marketing move by Anthropic to release their paper "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training" ....
Well, they do. So you review the output. Models should be treated as juniors, not a line of code or legalese they write should be used without review by a human, not another AI.
> In that regard it was a clever marketing move by Anthropic to release their paper "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training" ....
I find such claims a bit like if 3M/DuPont released a paper on the carcinogen dangers of PFAS and someone said this was "a clever marketing move".
Like, great, this can have hidden dangers we only learn about years after it's irreversible… how does this make "our"* companies the good guys?
A lot of us were already concerned about backdoors mandated by the US government, but that was just encryption and this is everything that AI ends up being able to do.
* in the sense of "the G7" until the start of this year, because why would other nations prefer us over China in the first place, and thanks to Trump vs. Canada, Greenland, Iceland also now the rest of the G7 has to treat seriously the possibility that the US is more of an expansionist threat than China is.
to be honest anything built into my phone, car, ships with mac is a hobby
what's important is where big enterprise point their budgets. and there we agree, it's not to chinese models (as hopeful as I am that we can avoid a duopoly and great local models)
I think there is also a perhaps the need to assign air-quotes to "investors" when talking about the US markets.
There's too much fluff and noise there. Too much (over)leveraged trading. Too much speculation.
What the real investors think is hidden behind the actions of the speculative traders.
> I think they do not believe that big corporations in the US or Europe would actually use chinese models
The Chinese models are open-weight. You can download them and run them on your own servers. Or you can outsource to Mistral, Privatemode or any of the many other US/EU inference hosters.
So the "security, regulatory and/or business continuity concerns" arguments are effectively non-existent. And whatever argument does possibly remain, also applies equally to Anthropic, OpenAI & co.
There were pretty specific rumors about IPOs happening early this year, up to apparent insider bets on prediction platforms. Don't you think the apparent delays in IPOs could have something to do with the open weights/open source competition? The monetary conditions (interest rates,...) were better early in the year, arguably predictably so, so what would be your alternative explanation why they chose to postpone?
This isn't true, especially not with the inference engine becoming public, now you can build your on models on the same or similar architecture and ship massive gains on top of these stacks.
I think people don't realize the Salesforce SAP Palantir would love to ship these models, and they have/are shipping them.
The threat of exfilteration of data via these models is high but it's nowhere near as high as the margins.
For instance if you could cut your ai input bill by over 90% would you do it? Can you put some of that extra margin say 10-20% to make it more secure.
I know of atleast 10 fortune 500 companies that are entirely planning to or have moved to Chinese models for a large portion of their work.
It's insane people in finance are delusional enough to think that won't happen.
Maybe Govt will try to de-risk with bans on using open models but I don't think that's going to work at scale, people will literally spawn labs to launder DeepSeek/GLM models into a US model by some novel training etc approach or something.
I still don't think that it's as much a kami-kaze rather it's an attempt to normalize the world around the Chinese stack.
If DeepInfer is supported by everyone, it will be much easier to export stuff aka hardware that only supports it for now or is optimized for it.
Assume china starts selling hyper optimized Asics for inference at some point.
This feels like a race to the bottom as soon as possible because they have much lower costs, if there is a Wipeout in AI Labs in US chinese labs are largely insulated, and they can hopefully buy Nvidia chips for cheap if that does happen.
> I think they do not believe that big corporations in the US or Europe would actually use chinese models in their business operations due to security, regulatory and/or business continuity concerns.
European business users (+ consumers) can access GLM-5.3 via Mistral.ai hosting. (with GDPR)
Europe is going to be just as worried about using US models for large parts of government related functions and they're going to feel alot better about open source self-hosted chinese models than US models which can literally get turned off on a whim by Trump. They'd prefere home grown models if they can, but if they can't they'll go self-hosted and that'll be Chinese at this rate.
> I think they do not believe that big corporations in the US or Europe would actually use chinese models in their business operations due to security, regulatory and/or business continuity concerns.
Well they are wrong, I'm at a large EU startup and they are currently considering Deepseek as their default model, Opus/Fable would still be available for complex tasks.
No, that isn't going to happen. The US can't afford to lose the AI race and US companies that can't compete will get thrown under the bus and replaced with those that can.
The DHS order only bans buying DeepSeek products directly (e.g. doing inference on their China hosted servers). It does not apply to open weights at all. And DeepSeek just gave their response to the DHS order.
It is clear from the beginning that the Chinese Open Model are a proxy for economic warfare. Anyway their singular goal is to dethrone US monopoly on SOTA model and burn all the economic hype and investments. If this strat works, which we will see after Anthropic/OAI IPO, we'll be in for an interesting ride
I don't see how this proposition holds. It's just part of an inference engine for serving DS4.1 Flash. Which vLLM and SGlang can already do. Less efficiently, I'm sure, but on much broader hardware, and with other models in the mix.
People will take the novel bits here and port them to the above. The community as a whole will benefit. Including infrastructure engineers inside OpenAI and Anthropic.
But a lot of what's in here is already in published papers out of DeepSeek already.
VC and valuation (alone) backed success has never been dependent on a business or product being good. All that is needed is finding the bag holder(s), and the way our economy and financial systems are designed that's the core competancy of this system.
Sure, any day now DeepSeek is going to destroy the valuation of OpenAI and Anthropic. I've been hearing this since DeepSeek first launched in Jan 2025, 21 months ago. And with every new model generation, both open and closed, I hear a chorus of "this is now good enough!". And then I hear it again for the next generation.
Meanwhile OpenAI has grown their revenue like 10x and Anthropic has grown theirs like 65x in that same time period.
There's no cap on what you can do with frontier intelligence, nor how much of it you can use. Yes, DS and other open weight models will continue to grow like crazy (as long as they can source compute), but so will the frontier labs. I ran like 20 agents overnight last night on Opus, Fable, Sol, and Astra last night, trying to use up quota before my weekly reset, and I'd gladly have done 10x that much and used agents 10x more intelligent if I could have. And in six months I probably will.
Can you explain how this removes their moat? Genuine question, I'm not sure I understand enough about LLMs to get how an open source inference engine undermines them.
Today the closed-source giants maintain a substantial cost-per-token advantage due to proprietary inference tricks like speculative decoding kernels, ultra-optimized FlashAttention derivatives, multi-GPU pipeline orchestration and dynamic prefix-caching algorithms.
I'm not sure whether the DeepSeek's open-sourced inference engine does all of these, but it appears that plenty of inference tricks have just been revealed for the open source community to learn from and start improving upon.
So... this makes it easier for someone in the position to run chatgpt class models as a service to get similar returns to openAI/anthropic, if I got you.
And I imagine this engine is by default mostly just deepseek compatible and inference engines are not universally applicable to all models?
It sounds significant for competition but I'm not sure how it sinks Anthropic/openAI. Presumably they still have their frontier models, the training data they're vacuuming from current users and access to hardware to serve at a ridiculous scale, right?
DeepSeek can accomplish over 90% of tasks of equivalent value at one-tenth of the fully loaded cost of O/A, and this capability can further spread through open source. In my view, the IPO risk for O/A is no longer just a valuation discount, but a substantive challenge to its long-term business model as a high-margin foundation model provider.
In that scenario, even if global Token usage increases 100-fold in the future, that alone would not be enough to justify these foundation model companies being worth one or two trillion dollars.
I think you’re correct, although suggesting that open AI is a trickle drown of AI companies is a bit naive.
But the key point is that something good enough always attracts more interest and contributions because a productive outcome is very realistic and tangible.
Open Weight models and competing open source inferences engines massively undermine monopolies; in theory at least.
See how cloud is easy to divest and “nationalise”.
It is, and always has been, data contracts. Which is why the hyperscalars are a more likely candidate to take American frontier business than Chinese vendors.
I can click a button in Azure foundry and run open weight models that are approved through my compliance and procurement teams.
I don’t know how open sourcing this stuff helps deepseek but it means I’m getting ever better and cheaper models so I’m fine with it.
That is probably what has been the most fascinating part of this AI/LLM boom, the moat was apparently never the technology. It has been very much like nuclear energi (and weapons), once you know how to do it, it's not that difficult. It's still fairly hard, but you can do it. The moat has been the funding for the infrastructure and if that dries up you're in trouble.
Well, whatever real or perceived moat Openai/anthropic had. OP claims this open source inference engine reduces their chances of high valuation, I'm assuming this is due to their in house engine being something they could use as differentiation before?
I've been trying to find an answer to that, too. The best I've got right now is this:
Moat-like things going for them:
- access to the US government (for huge inflated government contracts as a source of free capital, having the competition banned, having their products legally mandated, getting other preferential treatment, having foreign governments bullied into granting similar advantages)
Too unsustainable, too fragile, too small:
- Strong brands: Brand loyalty seems like a pretty insignificant factor in AI right now
- Semi-exclusive hardware access, RAM buyouts: Unsustainable even in the short term
- Lock-in effects: Their models aren't very stable across versions, or very unique, and switching providers is about as easy as it can be in a highly competitive market.
- Exclusive technology that is important, advanced and hard to replicate to a degree that would materially set them apart and justify potentially much higher cost. Maybe I'm missing something here, but a lot of the claims they make to that end fall apart even on light scrutiny (due to things like test sets that are small, hand-picked and clearly overfitted on, or broad general claims that cannot be substantiated). I'm getting the feeling they don't have anything that materially sets them apart at all, but that's seems impossible to measure in any meaningful way, so who knows, they might have something here.
In theory, this can allow a small US lab to compete with O/A/G and take marketshare from them with small open models deployed centrally or (and this is think is the real moat) on-prem for enterprises.
Not really. Anthropic’s headcount isn’t huge, and only a subset of their employees are working on the frontier. It’s also easier when you’re a follower.
Supporting Huawei Ascend 950 NPUs with open source inference code has huge implications in the global AI hardware market. Imagine if all smaller labs now can build on top of this stack and add increasing support for more compatible hardware , software and models. This could seriously challenge NVidias death grip. Although inference is still only a part of all AI workloads and open sourcing training on Huawei chips would be even bigger.
Whatever the consequences, I have to admit China has understood that all those are raw materials and that open sourcing everything will help growing their economy (as a whole).
The worst idea was to build corporations providing services with AI. It was an extension of the latest 50 years model, which is about to crumble. Now, we have our revolution. (Ok, ok, I am a little bit too enthusiastic, I now need a fallout shelter).
This is just like linux and the larger open source movement. It liberates closed source code that should be open. In the future we will look back at this like we do the start of linux.
Treating AI like microsoft office subscriptions is really the losing model, at least in the long run.
One of the fascinating dynamics around Chinese open-weights models is the competition among American (and some European) inference providers to deliver the best inference performance on them. They've increasingly beaten the Chinese model makers themselves on inference throughput and latency.
Together.ai, Baseten, and others have already achieved higher throughput, lower latency, and lower E2E latency for DeepSeek V4.1 Flash 1017 than DeepSeek's own inference service.
For comparison, rental prices seem to be ~$2.75/ per H-100-hour [1]. Meanwhile, Deepseek says "38 USD per hour per card”. They haven’t specified which card as far as I can tell.
It's harder than ever to get h100s and then you have to usually commit to a long time period upfront, but yeah this release should close some of the gap.
Interesting how a comment asking about the reasoning behind releasing this as open source is getting attacked, as if doing a good thing should not ever be questioned. I do wonder of China's motives behind this as well.
The government would not allow this open work if it did not benefit them somehow, or perhaps they feel that keeping it closed is not very important right now (maybe because they are somewhat behind the frontier).
Comparing a university student Torvalds and state actors open sourcing code is pretty naive...
You seem to severely overestimate the level of centralised command-and-control in China. It's not like every software company is going to the government for explicit sign-off of every open source release.
There is a pretty clear benefit to DeepSeek's reputation for all the other providers offering DeekSeek's open-weights model to have similar performance/benchmark scores.
In the past performance has varied wildly across inference providers, making this available should result in every inference provider implementing the same optimisations
> On October 16, 2026, the Ministry of Commerce and the Ministry of Science and Technology issued Announcement No. 48 of 2026, publicly solicited opinions on the revision of the China Export Prohibition and Export Restriction Technology Catalogue (hereinafter referred to as the Catalogue), and implemented temporary management measures for pre-export registration of relevant technologies. Recently, the spokesman of the Ministry of Commerce answered questions from reporters on relevant issues.
The other benefit is that as more of the stack is open sourced we can finally see just how cheaply these models can be served. Which in turn gives us insight into what the frontier labs might be doing.
Every open source release is visible to me, too. Is your point that the government could stop it if they wanted? That's true of every government, but it's not a problem until it actually happens.
So what makes you think it's not Chinese nationals attacking those comments?
Please never forget that US runs the largest state-sponsored hacking program. They are and have always been the largest APT. Anything opposing that narrative is blatant propaganda.
RE motivations - I wonder whether this is part of their push to double down on social stability efforts?
This is based on a video seen on Facebook, so take it with a grain of salt.
AFAIK, BYD batteries are based on "bad" materials, with fabrication method perfected until they are competitive [enough(?)] with batteries made with the "good" materials.
The good stuff is patented, locked by the West. China workarounds it. It's not that it is "bad", is just copying West's strategy doesn't work for them.
The problem with that theory is that BYD and a few other Chinese companies already for some years have had batteries that are better than anything made in the West, from the point of view of energy density, charging rate, temperature range, cost, so if there is some stuff that is patented and locked by the West, that is the bad stuff.
Living in the EU, I find it concerning that China has advanced so much in battery technology and in the West there are very few positive signs of attempting to keep up with them.
Closing your eyes and imagining that others are not really better, when they are, is not the right strategy to avoid remaining behind.
This seems like a garbled retelling of the shift from NMC batteries to LFP.
The latter are worse on some dimensions, so much so they weren't considered viable for cars in the West. But incremental improvement saw them take over the Chinese market. A classic "low end disruption" story.
They bought up the patents from A123, and American firm with a Taiwanese American MIT researcher founder. The firm floundered in the global financial crisis, a time that is a big inflection point for China's global domination of green tech.
I don’t think there’s got to be some deep conspiracy. Opening this is a pretty clear way to commoditize your competition. That’s an age-old strategy. The community may very well get this up to speed faster than if it remained closed, and DeepSeek / China can (probably) run circles around US labs from a hosting / service cost perspective.
Even though their API price is cheap compared to OpenAI/Anthropic, I still can’t “afford” to use their API because the subscriptions from OAI/Anthropic provide far higher value and peace of mind. I know for sure how much I need to pay each month. I believe most developers fall into this category as well, where $20/$100/$200 from Anthropic/OAI provides more value.
This raises the question: why doesn’t DeepSeek provide a subscription service as well? Is it a better move to have a single pricing model like they currently do?
I'm not sure what you're on about. Load your account with X dollars and then your monthly bill is capped to X dollars.
or by 5 opencode subscriptions, and have $800 of ds4.1-flash for $200. I'm skeptical that the average dev will manage to burn through the equivalent of $800 of ds4.1-flash if they spend even a bit of time optimizing their token usage.
I'm on an Alibaba Cloud Model Studio TokenPlan subscription, and DeepSeek 4.1 Flash, Deepseek 4 Pro are available, alongside various Qwen and glm models.
I've been using Qwen 3.8 Flash on an $18/month Standard Plan, and haven't gone close to consuming my credits. It got close to rollover date, and I gave it a decent nudge for 2 days, and used... about 2% of my monthly quota. Obviously one of the Max models could burn tokens a lot faster, but I'm very happy with the outputs one of the Flash models has been generating.
I need to give it something harder to do, it seems.
You know your subscription is unsustainably subsidised by debt and likely to change at any moment, yet relying on it provides more peace of mind than buying from someplace that are extremely transparent about their costs. That's an interesting approach to say the least
Maybe at a certain point their capacity is sufficiently being used up by the "full cost" API usage, and offering cheaper package deals would just cut into their profit because they can't get more inference?
I agree with this. While I realise the current subscription levels probably aren't sustainable long term for this month and probably next month, and who knows how many months after that, it is the best option.
I'm ready and willing to change as soon as I need to, but for the moment I'm going to take advantage of the subscription costs offered as it lets me know exactly how much I'm going to pay each month.
I do not think OAI/Anthropic make the biggest share of their revenue from subscription. The largest share is from API spends. That is what DeepSeek or others aim for as well.
Currently this is true only for Anthropic's $20/month plan. I do a lot of work without reaching my limit. OAI's similar plan on the other hand isn't worth it - it used to be a great plan in the begining when they needed to catch up, but now I reach my limit very quickly on it. The only reason I keep it is because my entire family is using this account for their chatgpt usage.
The more expensive subscriptions are just not worth it for individuals. Using deepseek even at API prices is better value. If you're a company it might be a different story.
Sure, but that size/token is inherent to the model. Part of DeepSeek's hosting advantage, and thus a remaining part of their moat, is migrating KV cache from VRAM to host RAM to nonvolatile storage, and presumably routing requests to where KV cache has migrated. People here on HN say KV cache lifetimes can be 8+ hours with DeepSeek, and I don't know of any other provider that can match that. That infrastructure isn't part of this open-source release.
I was just adding a citation to your point, not arguing against you :)
also adding what other stuff aren't included in the release, and implicitly pointing out that they explicitly mention it's not included rather than it being stealthily missing.
This is fantastic. I feel the KV cache system is basically a must for any serious inference hosting, both for the provider as well as the consumer. I guess they have this but just haven’t open sourced it yet.
Can somebody explain like I'm five what this exactly is? How are providers like Fireworks or Novita providing inference today? Do they use some other inference engine software? If yes, what is the advantage of the DeepSeek inference engine?
For the most part each provider implemented their own inference engines. This release will help them all adopt the common set of optimisations to serve DeepSeek Flash optimally.
I think the reason DeepSeek does this is partly to neutralize competitors' technology, but ultimately the core reason might be that data is what matters.
Chinese labs do not release "open weight" models because they think it is good for humanity. They do it because they are severely GPU compute constrained due to the Nvidia and ASML embargos. All available GPUs have to go toward training which leaves very little capacity to host inference at scale. So they release the models and encourage EU and USA 3rd parties to host them.
China will inevitably produce their own AI chips and then they will lock down their models just the same as everyone else.
The real loser in all this is obviously Europe, which has nothing competitive once the Chinese models are gone.
> The real loser in all this is obviously Europe, which has nothing competitive once the Chinese models are gone.
Those models are open source / open weights. Europeans can continue with Chinese models as a base same like Cursor used Kimi K2.5 a base for their composer 2.0 model or SWE-2 from cognition is post-trained Kimi K3
Similar to how some labs have done free stealth model releases. Trillions of coding tokens ought to improve performance of the downstream model. However, improving the base model architecture remains the bottleneck since you will not reach RSI with just that (assuming you ever can). And compute, lots of compute…
I’m pretty sure this is rage bait. The idea that China is more interested in virtue signaling than pulling off a power move that will improve their chess position is not worth a serious reply.
as opposed to the current crop being backed by the US state? The ones that are basically crying foul at competition? The ones that unilaterally tries to dictate based on political friendliness and compliance (ala, anthropic getting banned)?
At least with open models, you can own the hardware and do the end to end. The current bottleneck is hardware availability at low cost, but i am guessing the chinese state is looking to build their own chip supply chains. I suspect they will succeed.
I don't understand why this is strange to you, some people just do things out of personal satisfaction and/or building their legacy. If it's just money, Deepseek founders are already rich because they have a hedge fund company (High-flyer).
Also I have largely understood that they are not spending that much. They are not hundreds of billions in commitments. So they can just release stuff that have comparatively nothing spend on it and still garner value.
I believe majority of the AI inference (and probably training) infrastructure runs on Linux, so I'd certainly consider it a major player in the AI race.
I'm wondering how it benefits Deepseek. Why would they willingly give up their trade secrets?
It's obviously very valuable, as they are running a profitable company. It follows the same trend as when they released their cache optimizations a few months back: why do it? What followed their release of this technology was that Anthropic and OpenAI silently released models which were on-par or similarly intelligent, but "surprisingly" had much cheaper cached tokens. So their inventions were obviously just worked into their competitors, who just got much greater margins + became more profitable.
The same thing goes here: I'm sure there are inventions and optimizations which are unknown to Deepseeks competitors which yet again will be worked in. So once again, they're letting their trade secrets go.
My question is: why? How does Deepseek benefit from this? There must be some gain I'm not seeing as they keep doing this play over and over, and I doubt it's "for the good of open source", as China is practically as isolated as you can get. To believe that this is in the spirit of the open source community is heavily naive.
To rephrase: How would it benefit China? Maybe it is to show that there is no secret sauce, burst the bubble and then profit from the chaos that will ensue? Collapsing US tech giants means lesser MAGA backing, cheaper stocks, cheaper properly..
For the same reason every single open weight models gets released: it's a big flex and gets investors and developers excited.
Is it a sustainable model? Not really. Will it all blow up in our faces eventually? Maybe. But frankly, releasing code like this is far, far more sustainable than releasing whole models.
DeepSeek is a fixture among privacy-conscious local LLM advocates, but they are still battling for marketshare.
This is advertisement for them, and more so for large entities (like governments and corporations) looking for digital sovereignity: the more they are reassured they can run of their stack locally when needed, the more they will use the cloud offer too.
That is what I am not sure about...It is like, is having a public key closer to hacking the account than not having it? Theoretically, it is, but practically, it is not...
If them releasing their stuff openly burns down whatever the closed ecosystem is doing just maybe that closed ecosystem isn't that great. If something can't survive less resourced option making it stuff public maybe there is lot of wrong with that something.
uhm...wake me up when I can buy one(1, single, uno) GPU to upgrade my PC so that I can play modern games with high graphics, without feeling like I'm being bled dry.
y'all got dozens of H100 at home? isn't that ~300k??
"Based on this framework, we have achieved high-throughput, low-latency large-scale inference services for DeepSeek V4.1 Flash. Currently, the service throughput of a single cluster is 155 trillion tokens/day, and the overall revenue is $38 per card per hour (at off-peak pricing)"
Source: https://zhuanlan.zhihu.com/p/2095560295930508953
Not to say it can't still be a bubble, just that there is legitimately profit to be had and perhaps some sustainable equilibrium in the future.
I think everyone has known that inference is profitable overall for the major providers for a while (I'm sure not all user accounts, but in aggregate).
The major providers are unprofitable after you add in the training and research costs and other overheads.
This is why the open source models are such a threat if they become good enough - it will destroy the business model of just providing raw chat/intelligence.
Arguably that's also why a few of these providers are trying to diversify into whole-platform approaches rather than just relying on having their whole thing be a great model.
People that want to believe something different as a coping mechanism will not be swayed by this evidence either.
The very short lives of all these models also factor in.
I think the funny thing is the more inference, the more tokens a specific version of a model spits out, actually does make it more profitable.
But somehow a lot of people truly think that each token generated increases the cost of OpenAI/Anthropic/... and shout bloody murder about "subsidies"
And I'm certain there are people out there that actually think the API prices of OpenAI and Anthropic somehow equal their electricity costs.
But I'm also interested is retraining needed? how relevant will a model be in N years if it's just "static", and if you do need to retrain and fine-tune over time, is it still profitable? how often is that needed?
For most cases it's probably ok, often times answers are general or basically a web search anyways, but if my Agent started blasting outdated syntax it would suddenly be way worse for coding for example.
Even take account of R&D cost it is still profitable, because you do not actual retrain the base model every month, only distill and post train from the same foundation.
heck, I think even including the giant cost of the top tier researcher salaries, they are still profitable.
then, how are they still losing money? just read what they are actually spend the most budget: always the hardware. those companies will never profitable, because they will never stop spending money to expand their computational prowess.
I think the more someone works with AI the more likely they realize that inference does not have to be sold at a loss, but many people's AI knowledge seems quite dated and it sounds almost like it becomes a convenient conspiracy theory for some, a way to convince themselves that the market changes that are already happening a temporary (and I don't mean in a Jevon's paradox sort of way, but that companies will go away from AI).
I suspect a lot of this is fueled by a few stories of companies who went all in on spending as much as possible on AI and then begun to walk that back, without realizing that was more about bad incentives and over promises of what AI could do (especially 1/2 model generations ago). Some also confuse running the SotA models at the fullest capacity costing more and more with a given problem costing more and more to solve, not realizing that the cost per problems is trending down but newer and larger problems can be solved for even more, so the costs at the top are rising despite falling prices. Also things like the subscriptions getting more expensive, though how much of that is because developers are doing far more than they were 6 or 12 months ago (and in some cases there was subsidization that is going away). And maybe a bit with the free models starting to not be upgraded or the harnesses getting worse performance of lowered free limits.
For the past year or two I think it has generally been agreed that inference is profitable, but the LLM business as a whole is not.
But it's clear that it's not the money burning pit I think it was just a year ago.
My workflow is one where I am heavily involved in the process and I don't like giving the agent too much autonomy + I need to follow along. Opus 5.5 speak is still cause aneurysms for my brain. It also has this vibe to its prose... like it's looking down on me. This is Prolly my bias speaking though.
I was doing the same (along with GLM 5.3 Flash) for months, and then gave Claude Sonnet 5.5 at medium effort a try. The difference was like night and day. Where both DeepSeek 4.1 Flash and GLM 5.3 Flash kept making mistakes and would have to do a bunch of iterations to find and fix them, Claude Sonnet 5.5 just did it right on the first try.
In my experience, while DeepSeek 4.1 Flash and GLM 5.3 Flash are a lot cheaper per token, they're just as or maybe even more expensive when you factor in all the extra tokens that are need to troubleshoot and fix their mistakes -- and even more expensive if I factor in my own time wasted doing the same, not to mention the frustration of having to deal with that.
DeepSeek 4.1 Flash also hallucinates a lot.[1]
I'm hoping all this improves with time, and the gap between the competent frontier models and open weight models narrows.
[1] - https://www.youtube.com/watch?v=P4dTq4X8bqk
Quick, someone point out the one billion user system with intent + personal data rich that doesn't result in a $100+ billion high margin ad network.
It is that simple and absolutely nothing can stop the gigantic ad networks that will be built up around personal agents.
Present valuations are the bubble, not the business / industry. The exact same thing was true about the Internet bubble circa the late 1990s, except this is a lot more important than the Internet and the global economy is drastically larger.
Productivity is how you expand economy and wealth. It's how China can go from being one of the world's most backwards agricultural nations, unable to feed itself properly, to gradually becoming an agriculture juggernaut. AI is a hyper productivity tool, saving people time and bolstering capability are two of the primary productivity concepts.
We're not even out of the first inning yet.
You missed the bit where they used the word "revenue".
I suspect if you factor in all the various overheads of running datacentres the actual value is lower. Revenue is vanity, profit is sanity as the old saying goes.
However the Chinese do have the benefit that China are pushing hard on renewable energies. So at least that major part of the equation is probably far lower than your average US/EU site.
And actually paying people not to ! (Paying RWE $1.2bn to abandon wind contracts[1])
Its absolutely nuts. Pure generational can-kicking.
[1] https://www.bbc.co.uk/news/articles/c1e1vg0gjl5o
No they didn't...
> I suspect if you factor in all the various overheads of running datacentres the actual value is lower.
That's literally exactly what the quoted part addresses: "Amortized cost of owning a DC full of GPUs is far less than $38/card/hr". Cost of running GPUs is far less than $38/hour, revenue is $38/hour, ??? , profit
From microns S10: BITS shipped and net revenue.
Q1 '20 - $0.10 profit per gb shipped (ram and storage)
Q1 '23 - -$0.024 (cyclic ram market)
Q1 '25 - $0.131 this looks normal were on the upswing.
Q4 '26 - $1.136 Were all fucked.
If you look at a chart of S3 prices we have all been living in the "Happy Time" where the cost to run our ever growing software business has gone down. Githb says 2025 billion commits - 14 billion this year, and 2x the actions run. That means AI driven software growth is 2x to 14x.
But those ram/storage prices are going to hit the operational budget of EVERYONE soon.
Can your average sas survive a 4x price increase on its AWS/Azure bill? How about 10x (cause thats what the margins say the premium is on ram).
If the circular financing is a concern, the circular firing squad should be a bigger one. Your paying for tokens to build more software to pay more to run your software is not a sustainable cycle. People are starting to use the words CapEx again - and it's a good thing.
The cost of amortising a datcenter, the gpus, the engineering costs attached all to produce inference.
At how much percentage of its inference would it be profitable (if it does), how much returns would it make over the long term and what would be the profit of those inferences/ profit-revenue and how that compares to other industries.
Are there any statistics which show this. Also how efficient token inference has gotten over time?
Chinas grid is almost entirely renewable energy with strong investments in it. In the us not the case, some dcs are running on gas.
Also DeepSeek was smart. They strategically rolled out model sizes they could support even if the intelligence wasn't the world's best for that week in human history. Then they have incrementally and cheaply improved models since.
Not to mention they also built trust with open weight models rather then fear mongering. They are in the top 3 of the smartest most viable LLM labs out there.
Danger? You mean benefit?
Same with a lot of their investors, who will be just fine without the "insane profits" they hope for.
I'm not worried about their welfare, and won't shed a tear if their companies collapse.
That's a bit of a one-sided view.
I think that "everyone" needs some kind of an asterisk. Who's "everyone" in this?
It's unclear that the leading open-source models could exist without having distilled the frontier models. If that's true, the "danger" is that these upstream models don't get built any more (because of the tremendous cost), and we end up just optimizing what we have. Also unclear whether this would be a net win or loss for humanity.
The tremendous cost is matched by the tremendous business output.
In the span of less than five years, starting from their first $1 of revenue, Anthropic + OpenAI will have produced businesses doing at least $150 billion in combined sales. The demand isn't reducing at present, it's climbing, they can't come close to servicing the total demand for their frontier models.
SAP is under $40 billion in annualized revenue. They're 54 years old.
If OpenAI/Anthropic stop doing it there's no reason to believe someone else won't. Or that we won't come up with a different approach.
... and a bunch of RLHF and proprietary training data etc etc, but regardless, this is still an exceptionally expensive process, that someone somewhere is paying for.
If the only profit is from running commodity free software (as OP claims), there's no moat and OpenAI/Anthropic are worthless.
Quick numbers:
At $35/hour in net cash generation per card, a $30,000 GPU would pay for itself entirely in about 850 hours (roughly 35 days)
With an assumed ~ $3 operation cost per card and hour
You can well imagine if you own your infra and can make utilization over 90%, where these go.
And then, my link above is for K3 model, which, compared to ds flash models, is highly unoptimized.
If I had a few million lying around, I would have invested in inference.
0: https://openrouter.ai/openai/gpt-oss-120b#providers
2. Newer gen nvidia gpu's like GB300 are way more profitable.
$38/hrs equals $320K in 10 months.
Assume they only have 100 clusters, and as much as 40% operational costs. That would get them about $20M in 10 months, which I suspect is more than what 4.1 flash training has cost them.
I haven't accounted for 7 hours of peak-pricing per workday.
Of course, you won't have 100% utilization.
For four cards at 100% utilization it's $4712 per workday and $3648 per day on the weekends and holidays.
Assuming a 64 card cluster
That's strategic sabotage to say the least.
Ban all open weight models! Cried the VCs.
I think they do not believe that big corporations in the US or Europe would actually use chinese models in their business operations due to security, regulatory and/or business continuity concerns.
They'd be wrong. Especially about Europe. Many companies are in a bind about this: they can't upload their confidential files to a service provider but they do want AI. So right now they end up with a choice: do nothing, breach their contracts, use a service provider that pinky promises not to share their data (which we have already seen they will violate) and run local models.
The latter is the reason you can't buy powerful GPUs for reasonable prices because a local model (GLM 5.3, DS 4, etc) + some hardware is all you need to solve the problem in a way that avoids the legal issues.
Mid tier insurance co is working with local inference providers and some of these now have thousands of GPUs up and running. Don't drink the HN cool aid that the EU is a backwater where nothing happens too much or you'll end up surprised one day.
The one thing where Europe really does lag is in training, that would be fair. But inference providers are popping out of the ground like mushrooms and there are some billion euro+ projects underway. We -unfortunately- also have some scammers that are bilking (or attempting to bilk) investors pretending they have more than they do, this is annoying but I would expect that other countries have similar players as well.
There are roughly 70 GW of datacenters under construction in the US right now, and a GW is roughly 500k GPUs, so 35 million GPUs.
The point was: they're buying them from the retail channels in relatively high volume because there is a massive shortage of GPUs in general, but that does not seem to stop parties with smaller projects who do not qualify for buying directly from NV (assuming NV even has stock to sell).
5090's now sell for the same amount (or more) that you could buy an RTX 6000 for a few months ago, and RTX 6000's are ~19K and up.
I found a stash somewhere, bought two, then two more and when I was ready to buy the remainder they were all gone. Serves me right for being too careful.
Quite failure by someone to assume that EU businesses would use US over Chinese, based on the current political atmosphere. US causes even bigger concerns than Chinese labs.
This is exactly right. The only AI that matters from an investor perspective is the AI that is built into your phone, your car, ships with Windows and Mac, and built into legal and business tools like Lexus Nexus and Salesforce.
Everything else is hobby.
The only AI that makes money is enterprise. Being open weights/self-hostable is a huge win before the pricing even enters the conversation, but pricing is also adding a ceiling on what US labs get to charge. And stuff like powering other software also works better with fine-tunes, hence legal models, etc.
Currently the price sovereignty is still not outweighing the cost of US dependency. But big providers are already working on finding ways to ensure customers independence of the cloud act.
In that regard it was a clever marketing move by Anthropic to release their paper "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training" ....
I find such claims a bit like if 3M/DuPont released a paper on the carcinogen dangers of PFAS and someone said this was "a clever marketing move".
Like, great, this can have hidden dangers we only learn about years after it's irreversible… how does this make "our"* companies the good guys?
A lot of us were already concerned about backdoors mandated by the US government, but that was just encryption and this is everything that AI ends up being able to do.
* in the sense of "the G7" until the start of this year, because why would other nations prefer us over China in the first place, and thanks to Trump vs. Canada, Greenland, Iceland also now the rest of the G7 has to treat seriously the possibility that the US is more of an expansionist threat than China is.
what's important is where big enterprise point their budgets. and there we agree, it's not to chinese models (as hopeful as I am that we can avoid a duopoly and great local models)
I think there is also a perhaps the need to assign air-quotes to "investors" when talking about the US markets.
There's too much fluff and noise there. Too much (over)leveraged trading. Too much speculation.
What the real investors think is hidden behind the actions of the speculative traders.
> I think they do not believe that big corporations in the US or Europe would actually use chinese models
The Chinese models are open-weight. You can download them and run them on your own servers. Or you can outsource to Mistral, Privatemode or any of the many other US/EU inference hosters.
So the "security, regulatory and/or business continuity concerns" arguments are effectively non-existent. And whatever argument does possibly remain, also applies equally to Anthropic, OpenAI & co.
I think people don't realize the Salesforce SAP Palantir would love to ship these models, and they have/are shipping them.
The threat of exfilteration of data via these models is high but it's nowhere near as high as the margins.
For instance if you could cut your ai input bill by over 90% would you do it? Can you put some of that extra margin say 10-20% to make it more secure.
I know of atleast 10 fortune 500 companies that are entirely planning to or have moved to Chinese models for a large portion of their work.
It's insane people in finance are delusional enough to think that won't happen.
Maybe Govt will try to de-risk with bans on using open models but I don't think that's going to work at scale, people will literally spawn labs to launder DeepSeek/GLM models into a US model by some novel training etc approach or something.
I still don't think that it's as much a kami-kaze rather it's an attempt to normalize the world around the Chinese stack.
If DeepInfer is supported by everyone, it will be much easier to export stuff aka hardware that only supports it for now or is optimized for it.
Assume china starts selling hyper optimized Asics for inference at some point.
This feels like a race to the bottom as soon as possible because they have much lower costs, if there is a Wipeout in AI Labs in US chinese labs are largely insulated, and they can hopefully buy Nvidia chips for cheap if that does happen.
European business users (+ consumers) can access GLM-5.3 via Mistral.ai hosting. (with GDPR)
https://docs.mistral.ai/models/zai-glm-5-3
https://legal.mistral.ai/terms/get-started/
Of course the same goes for any other model that you haven't trained up yourself.
Well they are wrong, I'm at a large EU startup and they are currently considering Deepseek as their default model, Opus/Fable would still be available for complex tasks.
Based on HN comments, every company would be running Gentoo and developing line of business applications in Haskell or OCaml.
The reality is what ever better and cheaper is where the money flows. Especially from an international corporation perspective.
https://public-inspection.federalregister.gov/2026-21447.pdf
The DHS order only bans buying DeepSeek products directly (e.g. doing inference on their China hosted servers). It does not apply to open weights at all. And DeepSeek just gave their response to the DHS order.
ahahahah
The overall market will still be big, very big, if it replaces real people or augments every single worker.
Some contracts will go to your local provider (US Gov, etc.).
And there is still a moat between them and DeepSeek.
As long as i can afford it, i use Opus.
I'm not bound by it and switching is very fast
It is clear from the beginning that the Chinese Open Model are a proxy for economic warfare. Anyway their singular goal is to dethrone US monopoly on SOTA model and burn all the economic hype and investments. If this strat works, which we will see after Anthropic/OAI IPO, we'll be in for an interesting ride
I mean, what would you expect? Like USA isn't openly planning to weaponize their current AI primacy [1, 2].
[1] https://www.whitehouse.gov/presidential-actions/2025/07/prom...
[2] https://www.whitehouse.gov/wp-content/uploads/2025/07/Americ...
People will take the novel bits here and port them to the above. The community as a whole will benefit. Including infrastructure engineers inside OpenAI and Anthropic.
But a lot of what's in here is already in published papers out of DeepSeek already.
Meanwhile OpenAI has grown their revenue like 10x and Anthropic has grown theirs like 65x in that same time period.
There's no cap on what you can do with frontier intelligence, nor how much of it you can use. Yes, DS and other open weight models will continue to grow like crazy (as long as they can source compute), but so will the frontier labs. I ran like 20 agents overnight last night on Opus, Fable, Sol, and Astra last night, trying to use up quota before my weekly reset, and I'd gladly have done 10x that much and used agents 10x more intelligent if I could have. And in six months I probably will.
I'm not sure whether the DeepSeek's open-sourced inference engine does all of these, but it appears that plenty of inference tricks have just been revealed for the open source community to learn from and start improving upon.
And I imagine this engine is by default mostly just deepseek compatible and inference engines are not universally applicable to all models?
It sounds significant for competition but I'm not sure how it sinks Anthropic/openAI. Presumably they still have their frontier models, the training data they're vacuuming from current users and access to hardware to serve at a ridiculous scale, right?
In that scenario, even if global Token usage increases 100-fold in the future, that alone would not be enough to justify these foundation model companies being worth one or two trillion dollars.
But the key point is that something good enough always attracts more interest and contributions because a productive outcome is very realistic and tangible.
Open Weight models and competing open source inferences engines massively undermine monopolies; in theory at least.
See how cloud is easy to divest and “nationalise”.
I can click a button in Azure foundry and run open weight models that are approved through my compliance and procurement teams.
I don’t know how open sourcing this stuff helps deepseek but it means I’m getting ever better and cheaper models so I’m fine with it.
Moat-like things going for them:
- access to the US government (for huge inflated government contracts as a source of free capital, having the competition banned, having their products legally mandated, getting other preferential treatment, having foreign governments bullied into granting similar advantages)
Too unsustainable, too fragile, too small:
- Strong brands: Brand loyalty seems like a pretty insignificant factor in AI right now
- Semi-exclusive hardware access, RAM buyouts: Unsustainable even in the short term
- Lock-in effects: Their models aren't very stable across versions, or very unique, and switching providers is about as easy as it can be in a highly competitive market.
- Exclusive technology that is important, advanced and hard to replicate to a degree that would materially set them apart and justify potentially much higher cost. Maybe I'm missing something here, but a lot of the claims they make to that end fall apart even on light scrutiny (due to things like test sets that are small, hand-picked and clearly overfitted on, or broad general claims that cannot be substantiated). I'm getting the feeling they don't have anything that materially sets them apart at all, but that's seems impossible to measure in any meaningful way, so who knows, they might have something here.
All other models are also either by <1k labs or side hustles by tech companies who are not particularly known for LLMs
The worst idea was to build corporations providing services with AI. It was an extension of the latest 50 years model, which is about to crumble. Now, we have our revolution. (Ok, ok, I am a little bit too enthusiastic, I now need a fallout shelter).
Treating AI like microsoft office subscriptions is really the losing model, at least in the long run.
Also cool (from the translated screenshot since original is in Chineese https://zhuanlan.zhihu.com/p/2095560295930508953)
> the DeepSeek V4.1 Flash inference system has established three goals
> 1-..... 2-..... 3. Supports retaining KVCache for more than 72 hours
This is awesome
Hope DeepSeek starts month plans soon.
Together.ai, Baseten, and others have already achieved higher throughput, lower latency, and lower E2E latency for DeepSeek V4.1 Flash 1017 than DeepSeek's own inference service.
https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017
I am happy for not maintaining a relevant open-source repo.
[1] https://www.ft.com/content/a1815bc5-b5d5-47ff-bbbc-e1a689929...
--
Supported hardware:
NVIDIA SM100 GPUs
Huawei Ascend 950 NPUs
The government would not allow this open work if it did not benefit them somehow, or perhaps they feel that keeping it closed is not very important right now (maybe because they are somewhat behind the frontier).
Comparing a university student Torvalds and state actors open sourcing code is pretty naive...
There is a pretty clear benefit to DeepSeek's reputation for all the other providers offering DeekSeek's open-weights model to have similar performance/benchmark scores.
In the past performance has varied wildly across inference providers, making this available should result in every inference provider implementing the same optimisations
https://m.bjnews.com.cn/detail/1792137269129878.html
Apparently globally recognized top 10 models (so Deepseek 4.1 Flash and GLM 5.3 Flash) would have fallen under this export registration.
The model, yes, but not the inference framework.
There's also an explicit exception in there for open source releases (which may explain in part why these companies prefer to work in the open)
That, and drive down prices everywhere.
7% of Chinas total population are members of the CCP. 7%. So at DeepSeek there's a good few handfuls of them.
If you want to understand the impact of that, read up on reporting lines in the CCP.
Not every company needs sign-off from the government for an OSS release, but every OSS release is visible to the government.
Please never forget that US runs the largest state-sponsored hacking program. They are and have always been the largest APT. Anything opposing that narrative is blatant propaganda.
RE motivations - I wonder whether this is part of their push to double down on social stability efforts?
AFAIK, BYD batteries are based on "bad" materials, with fabrication method perfected until they are competitive [enough(?)] with batteries made with the "good" materials.
The good stuff is patented, locked by the West. China workarounds it. It's not that it is "bad", is just copying West's strategy doesn't work for them.
Living in the EU, I find it concerning that China has advanced so much in battery technology and in the West there are very few positive signs of attempting to keep up with them.
Closing your eyes and imagining that others are not really better, when they are, is not the right strategy to avoid remaining behind.
The latter are worse on some dimensions, so much so they weren't considered viable for cars in the West. But incremental improvement saw them take over the Chinese market. A classic "low end disruption" story.
They bought up the patents from A123, and American firm with a Taiwanese American MIT researcher founder. The firm floundered in the global financial crisis, a time that is a big inflection point for China's global domination of green tech.
https://en.wikipedia.org/wiki/A123_Systems
"We don't care! China has been here for 5,000 years. Most of the time, there was no United States and we survived"
This raises the question: why doesn’t DeepSeek provide a subscription service as well? Is it a better move to have a single pricing model like they currently do?
or by 5 opencode subscriptions, and have $800 of ds4.1-flash for $200. I'm skeptical that the average dev will manage to burn through the equivalent of $800 of ds4.1-flash if they spend even a bit of time optimizing their token usage.
I've been using Qwen 3.8 Flash on an $18/month Standard Plan, and haven't gone close to consuming my credits. It got close to rollover date, and I gave it a decent nudge for 2 days, and used... about 2% of my monthly quota. Obviously one of the Max models could burn tokens a lot faster, but I'm very happy with the outputs one of the Flash models has been generating.
I need to give it something harder to do, it seems.
I'm ready and willing to change as soon as I need to, but for the moment I'm going to take advantage of the subscription costs offered as it lets me know exactly how much I'm going to pay each month.
The more expensive subscriptions are just not worth it for individuals. Using deepseek even at API prices is better value. If you're a company it might be a different story.
They say it was inspired by both so I am wondering what the performance gain is.
1. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
2. https://www.mindstudio.ai/blog/deepseek-v4-1-flash-specs-arc...
> Required for a complete inference service (not included in this release):
> A KV cache system
> VE/Prefill/Decode disaggregation (most required engine functionality is already implemented)
> Load balancing across data-parallel replicas for both prefill and decode
also adding what other stuff aren't included in the release, and implicitly pointing out that they explicitly mention it's not included rather than it being stealthily missing.
zAI served their latest model for free on chinese chips, now deep seek is saying run inference for cheap on Huawei chips.
even the RAM madness is about to end once the chinese ram providers unleash Ram onto the market in 2027.
lastly remember the chinese have the largest electricity production and the cheapest.
Chinese labs do not release "open weight" models because they think it is good for humanity. They do it because they are severely GPU compute constrained due to the Nvidia and ASML embargos. All available GPUs have to go toward training which leaves very little capacity to host inference at scale. So they release the models and encourage EU and USA 3rd parties to host them.
China will inevitably produce their own AI chips and then they will lock down their models just the same as everyone else.
The real loser in all this is obviously Europe, which has nothing competitive once the Chinese models are gone.
Those models are open source / open weights. Europeans can continue with Chinese models as a base same like Cursor used Kimi K2.5 a base for their composer 2.0 model or SWE-2 from cognition is post-trained Kimi K3
Deepseek moment.
This is written in Rust: https://github.com/deepseek-ai/deepseek-recipe
This DeepSeek Recipe is also mentioned here: https://zhuanlan.zhihu.com/p/2095560295930508953
as opposed to the current crop being backed by the US state? The ones that are basically crying foul at competition? The ones that unilaterally tries to dictate based on political friendliness and compliance (ala, anthropic getting banned)?
At least with open models, you can own the hardware and do the end to end. The current bottleneck is hardware availability at low cost, but i am guessing the chinese state is looking to build their own chip supply chains. I suspect they will succeed.
I don't understand why this is strange to you, some people just do things out of personal satisfaction and/or building their legacy. If it's just money, Deepseek founders are already rich because they have a hedge fund company (High-flyer).
It's obviously very valuable, as they are running a profitable company. It follows the same trend as when they released their cache optimizations a few months back: why do it? What followed their release of this technology was that Anthropic and OpenAI silently released models which were on-par or similarly intelligent, but "surprisingly" had much cheaper cached tokens. So their inventions were obviously just worked into their competitors, who just got much greater margins + became more profitable.
The same thing goes here: I'm sure there are inventions and optimizations which are unknown to Deepseeks competitors which yet again will be worked in. So once again, they're letting their trade secrets go.
My question is: why? How does Deepseek benefit from this? There must be some gain I'm not seeing as they keep doing this play over and over, and I doubt it's "for the good of open source", as China is practically as isolated as you can get. To believe that this is in the spirit of the open source community is heavily naive.
Someone already suggested it may be same as with other open source software, but you rejected that comparison with an ad absurdum.
Is it a sustainable model? Not really. Will it all blow up in our faces eventually? Maybe. But frankly, releasing code like this is far, far more sustainable than releasing whole models.
This is advertisement for them, and more so for large entities (like governments and corporations) looking for digital sovereignity: the more they are reassured they can run of their stack locally when needed, the more they will use the cloud offer too.
Also open weight models reveal their thinking, not to mention you can airgap the network around them.
Wake me up when we can clone, point it at a dozen H100s we own, and get a working production service.
Anthropic and OpenAI will do just fine.
uhm...wake me up when I can buy one(1, single, uno) GPU to upgrade my PC so that I can play modern games with high graphics, without feeling like I'm being bled dry.
y'all got dozens of H100 at home? isn't that ~300k??