It's absolutely true. I have three different setups here, a rig with 8 gpus that I use in either one of two modes (all 8, very heavy model or just one much lighter) and two DGX sparks. I haven't used a paid model in months. Sure, they use a few more tokens and the results may be slightly less good (but this doesn't really matter because I'm still in the loop so it is easy to correct). But the fact that I'm not sharing my work with the paid providers is much more to my liking than having big brother looking over my shoulder at every step and occasionally giving me the 'I can't let you do that, Jacques' treatment.
Gaming used to be done in arcades for 25c a play*.
Over time NeoGeo realized they could sell you the hardware to play at home (“on-prem”) at arcade quality albeit at a price out of reach to most people. Alongside them, Nintendo and Atari gave you 80% of the fun at 20% of the cost. History will view games arcades as a flash in the pan compared to gaming at home.
Namco, Capcom, and Konami went into selling software instead of hardware but it feels hard to imagine buying a home copy of Fable** for $X,000 any time soon.
⸻
* That’s for video games. In fairground arcades you’d literally pay for tokens then use tokens for play. The analogy almost writes itself.
If you think the idea of buying Fable like model is hard, you should really try some of the latest open models like Qwen 3.8. They’re getting there very fast.
Qwen 3.8 and GLM 5.3 are my workhorses right now. Qwen for simpler stuff because it is wicked fast using 'ninfer', GLM 5.3 because it is very good a more complex stuff and less likely to get stuck so I can let it run for much longer without intervention.
Working on a blog post to document it all, but that will take a few more days. The short version: the big rig is probably $25K, ttft is dependent on how much prefill you have but typically a couple of seconds (2K prefill/sec in one setup, about 500 in the other), tgs 100 to 150 or so depending on what it is doing and how well the speculative decoder works with the current context. Electricity is pretty bad, full tilt it's about 2500W or so but I've power limited it to 1500 to keep it civil. It is not cheaper than using a frontier lab (or at least, I don't think it is), but the independence bit is the important part for me and I don't care if that carries a surcharge, even if it is a substantial one.
My local model never says no. I never get banned. If I want to do security stuff, reverse engine, there it goes! Do I wanna chat about politics? No problem! Do I want it to analyze medical stuff? I don't get to hear that it's not a doctor and I should call 911. Do I want to explore legal docs, no problem! There are lots of things that cloud models will say no to! I read that don't even agree to help with advanced biology and chemistry topic much. I also never run out of tokens, no 5 hr limit. I can run it 24/7. I don't run out of tokens. I can run 16 sub agents at the same time, non stop! 1-4 billion tokens a month? No problem. Any agent harness I want! No restrictions. Privacy all day. My own true system prompt, not something someone has determined, no filters. Furthermore, there are many open weight cloud models that don't have equivalent online. For a while before cloud models got good at math, the best model at math was DeepSeekv2 Math and there was no API. The only way to use it was to run local. All sorts of audio models, image/video models, etc. No multiple subscription needed. Private UI for my family, kids included (I can guard what kids infer). Got a friend that really needs some tokens? I give em access through private VPN. Telling someone that they are paying for decades of Claude for local AI is like telling someone they are paying for decades of public transport when they buy a car. Sure, you are right, but ...
AnyONE. But if you got some friends, and they got friends, then that $25k becomes smaller and stomachable. My favorite rumor out of LA is groups of friends with a Lamborghini Urus that they timeshare. No idea if it's true, but if car people can get together to buy/lease a car that's MSRP is $345k, I think $25k should be an option for anyone with friends. (Hypothetically. I've read a lot about this friend thing on the Internet!)
If I want to go out of business that is what I would do. Most of the data I have here could never be shipped to a service provider, I'd be in lawsuits for the next decade or so and die bankrupt.
That NDA case is the real argument, more than price. Token prices keep dropping anyway, but "this data can never leave the building" doesn't change with a price cut ;)
Companies pretty much give us the full overview of what they have from a technical perspective (and often much more than that), we sign NDAs with investors that have serious teeth in them. The data they share with us would not be disclosed under any other circumstance and even our reports are shared only with the clients.
It's a nice little business. I'm not going to risk offending my customers for convenience or to save some money.
On a tangent: this is a really good website. I understood what you do and how you do it by reading down one page. There's a lot of projects/companies I immediately tune out from because I can't actually figure out what they do amongst all the buzzwords, animations, and excess scrolling in every dimension.
Inference prices are going only down. It could be that some top subscription costs more in a decade, but I fully believe a full array of competent models will be extremely cheap.
Keep in mind that the prices of the big AI companies are currently mostly loss makers. Their plan is to get sufficient market saturation and dependency of their clients and if they achieve that, they will hike up the prices quite a bit.
(Of course this all depends on if their plans work out, but if not then the smaller local models will probably play a big role in the failure of their plan)
People keep repeating this, but it is only remotely close to true for the literal free plans and the small subset of consumers maxing out the pro plans.
At API pricing (i.e. what all enterprise customers are paying), inference is profitable, even if they're paying usurious rates on cost of capital for capex. Where the frontier labs are getting screwed is that they're in a Red Queen's race with each other, spending tens of billions of dollars per year training the next frontier model, and not being able to recoup the investment on that model before it's obsolete.
That's why you see the CEOs calling for government intervention to "pace the frontier": it lets them reduce investment on training without risking that their competitors keep going and make them obsolete. Unfortunately for them, the Chinese labs exist, and would love nothing more than to take the frontier, so they're in a bit of a pickle.
The people I personally know who signed the "pacing the frontier" petition (including me), or who called for it in other ways, are sincere in their concerns that things are moving too fast for safety.
People keep repeating this, but it is only remotely close to true if you evaluate inference in a vacuum without any of the other supporting costs to facilitate that capability.
Just a few years ago people were saying the same thing about WeWork: "leasing is profitable, the unit economics of individual locations is net positive", and yet...
I was thinking the economics of it the same way.. if the AI companies are burning debt and VC money to give me cheaper tokens why not take them for it?
I can always rent a GPU server for those personal, private AI needs.. heck even run a decision model to route them to the rented GPU servers either on demand or scheduled for few hours in a day…
It doesn’t have to be black or white…
For my toy projects, adventures, learning exercises, and throwaway MVPs I will prefer the subscription plans… they are cheap for what they offer today.
From this article alone it should be obvious that the long-repeated myth about frontier inference being subsidized is just that: a myth.
It's training and tuning new models that cost them money. And that's why they are trying to scare people into thinking AI will kill us all in the next 5 years; they want bans on the open-weight models because they know they won't be able to compete with them.
A few days ago we got the news that DeepSeek makes 38 dollars profit per card with their off peak pricing (in the inference server release discussion).
If that is true, either the American labs are not optimizing correctly or they make a ton of money with inference.
Yeah, the way I see it short term is take the free/subsidized tokens while we can and the further out we go the cheaper compute will get for local setups.
The bet we (self-hosters) are making is that once these companies IPO, they will be forced to do whatever they have to do to make money. i.e. price increases, more throttling of subscription non-usage-based accoutns, etc.
They are subsidizing all of us to the tune of $billions in losses per month. That can't go on for a decade. Probably not even for another 18 months.
I still use all the tokens Claude will give me on the max plan. But I also run deepseek v4 locally for overflow and for... contingencies.
> The bet we (self-hosters) are making is that once these companies IPO, they will be forced to do whatever they have to do to make money. i.e. price increases, more throttling of subscription non-usage-based accoutns, etc.
But, there is no need to make that bet - when prices increase, I'd look into renting a previous-gen GPU on a cloud provider. If the online GPU providers raise prices, then I would look into purchasing for on-site.
Until that happens, I'm not making any bets, because if all the token providers go bust GPU rentals are going to fall in price, and if they don't go bust after raising prices, I can still get years of GPUs for self-hosting at the price of a self-owned rig.
If you know you will pay Claude a few thousand bucks a year a decade adds up too, owning the hardware could just as well end up cheaper because you can sell it or as others have alluded the price to rent could get much worse.
Would you really expect to be using the same hardware over the course of that decade? One of the biggest drawbacks to using on-prem hardware is keeping up with the upgrades. Leasing upgrade cycles are a thing for businesses, but for someone at home it seems like later in the decade the local system will be showing its age.
The 3090's I have would sell for more than new today... for the next couple of years or so it will be pretty bad on the hardware front. Until there is more RAM manufacturing capacity and NV has a credible competitor I don't think it will get better.
The benefit of the agent working firstly for you rather than for Anthropic or OpenAI will become increasingly apparent. We're already seeing signs: remember when Anthropic openly promised to have Claude sabotage people using it to create competitors? And OpenAI is looking for revenue opportunities from advertising, which can't end well for their users.
Just think of how user-hostile other online services have gotten, and imagine how acutely painful the proprietary LLM providers could make theirs. Thankfully, open models provide some counterbalance.
I mean, I just have 4xV100S GPU and a Macbook Pro w/48GB of RAM and I find it kind of amazing.
My server can run Qwen Flash Next at 500 t/s aggregate (about 200 t/s single user). It's pretty great. I use it for local coding and all my worker agents. It is about as reliable as Sonnet 5 for agentic coding. I make 3D models, keep the home network and server updated, fully built ansible files.
You do NOT need 20K in hardware to run these models reliably. I built my own inference engine from scratch for fun. You can run Strata on 16-32GB cards at very respectable token rates if you have the standard RAM to go with it. I strongly encourage people to go use this model on OpenRouter and try it out for tasks.
I also have a somewhat lobotimized quant of GLM Flash 5.3 working and it's pretty decent too.
V100S is a very good card for some workloads. And they were cheap as well.
I have 240G VRAM total which is a good reason why that rig cost as much as it does, and if I were to build it at today's prices it would be even more expensive.
I got mine just a few weeks ago for about $750 each. I am a decent programmer and have Opus optimizing the inference engine. I know enough to keep it on track, but GPU programming is something I have literally never done, before, but now I know enough about Volta and how to optimize it to be dangerous.
Kind of ironic, but it's the only way. I just really like this as an option. For a household LLM I have the IQ4_K_XL quant and plenty of capacity for everything you could want in a household LLM and it is trustworthy enough to not rm -rf /, etc.
And yeah that is a really nice amount! You can run and compose a lot of interesting models, vision capabilities. IDK. Plus this level of model is far more interactive in the sense that it can implement a well scoped ticket or unit of work but its not going to magically just figure some things out without guidance or a spec :)
I tried getting some 170's when I first heard about that unlock but they were already unobtanium by then. That would have been a nice way to add a lot of VRAM.
GLM 5.3 runs pretty comfortably on this rig, even with large context.
Well... that's a tricky one. I bought a lot of these second hand. I can sell them for more than they originally cost new, so indeed, it is not 'cheap'.
Then again, I run a business and for the business it is absolutely manageable, it is roughly the price of a single job. Privately this would be a different matter but I assume that if you would do this privately you'd simply get a single (but powerful) GPU second hand and offload to RAM. Not as fast but with llama.cpp, 'ninfer', DwarfStar and various other clones of those two you can get some pretty decent performance out of single GPUs.
I've been buying powerful hardware since the 90's, starting with SGI and then ever larger Linux boxes. I make my living with them so it's never been a huge decision to invest some of the proceeds back into the tools. But on an absolute scale, yes, it is a lot of money.
I mean IBM already has a track record of building gigantic amounts of garbage and over charging large corps and govts for it coasting on their brand name and networks to get deals. So using some cheap ass model fits their MO.
Having used their shit mountain called websphere, what an IBM CEO says is not something I take too seriously.
IBM has two faces. Their consultancy arm is the exact opposite of what I would call a good company, even so, they make a good bit of money. The hardware side of the business is still quite impressive, yes, it is too expensive but their power rack boxes are effectively a little mainframe (check the IO stats on those things) and the performance is - I have to admit - impressive. That said I would never buy it because I can get much better value elsewhere. Even so I think they could do this if they really wanted to.
They do have big advantanges because many large corps have become totally dependent on them having outsourced everything and lost internal knowledge. That internal ops knowledge is very crucial/useful to take full advatange of what AI offers. In the coming years much of it will leak into frontier models. After which corps might not need IBM/TCS etc and just rebuild internal teams to handle ops like they once did.
> Sure, they use a few more tokens and the results may be slightly less good (but this doesn't really matter because I'm still in the loop so it is easy to correct)
I can't wait to use models that are 'slightly less good' when I am at work instead of something like Fable that lets me focus on the problem at hand as much as possible rather than also spending additional effort correcting for the tool I am using...
You need to reread what was written. No one is proposing to take away your right to leak proprietary edge to the rapacious frontier labs that need to eat most of the economy in order to survive.
It's going to boil down to cost of the hardware vs. pay per use cloud models. I'm on the 100$/month chatgpt plan (~88euro). With all the resets, credits, and all the rest I typically don't run out of tokens. This is fine for me. I might go to the 200$/month plan at some point.
In comparison, my highend mac book pro costs 105 euro per month (leasing cost). That's not quite enough to run the bigger models and besides running the GPU on a laptop constantly makes it a bit hostile for my lap. And the model I would need a lot more memory for the bigger models. It's nice for experimenting though.
If you get your own GPUs in some kind of server form, you could run those locally or in a data center. Or you just get hardware resources from your favorite cloud provider. Hetzner (one of the more affordable ones) GPU servers start at around 200 euro per month. I know some people that use that with open weight models. Works fine, especially if you can share the resources between multiple users.
At which point of course you are doing the same as cloud based providers are doing. But probably with a bit worse economics. It all boils down to whether you are actually saving money or paying for a lot of underutilized hardware that's running 24x7 when your people actually only show up about 8x5.
I don't think most companies should be doing that. It sounds like they'd end up with a lot of idling over capacity at a premium. The only valid reasons to do this would be securing access to GPUs or some very urgent need to keep things on premise (which a lot of companies insist is true for them until they realize how much they are going to pay).
> It's going to boil down to cost of the hardware vs. pay per use cloud models. I'm on the 100$/month chatgpt plan (~88euro). With all the resets, credits, and all the rest I typically don't run out of tokens. This is fine for me. I might go to the 200$/month plan at some point.
It's also about "insurance protection" against the AI company raising the prices or changing their terms (e.g. becoming more frugal with respect to token drops).
In this context yes, but for example at my job we are using LLMs to evaluate some outputs of cheaper/simpler models and to add context to them. This gets expensive very quickly with per-token pricing and would pin our entire business on the whims of one or two providers. With self-hosted open-weight LLMs we can just operate like we always have, it's a different kind of processing but otherwise we are free to do as we please wrt compute.
So it matters a lot that these open-weight models are good enough; we would likely not be rolling out this new processing if it was fully dependent on Anthropic and OpenAI.
You can try open models via OpenRouter and omp without having to buy or rent any GPU. I routinely try open models to see where they are. GLM 5.3 with DSFV4.1 as advisor via omp feels Opus 4.6 level, i.e. totally able to ship stuff.
I do use Anthropic and OpenAI mostly because it costs me 1/40 and 1/50 of the api prices, which is way cheaper than Chinese subscriptions, but I’m ready to jump ship when they pull the plug. Fortunately they have nothing that forces me to keep paying them, and I don’t pay the crazy enterprise subscription costs, just the personal account prices.
I also die to be able to run the OSS setup locally to stop being spied on all the time.
For context, I averaged 10 hours a day prompting LLMs from when ChatGPT came out. At the start it was just function generation, then single file modules, then modules, and now I reason, design, implement and test with them.
An interesting thing is that with omp you can plan with a smart model and then implement with a dumber one. So for example, plan with Kimi 3.1 via OpenRouter and then implement with DSFV4 with zero data retention active, sorted by cheapest, and minimal tps accepted.
I get the people who are pissed about their cushy jobs evaporating, but if you know how to use these technologies it’s very freeing because it’s like having a company of people working for you, except they are not human so they’re not wasting their life working for you instead of using it for themselves.
You can make them do pretty much everything that is fuzzy-verifiable, you don’t have to worry about their feelings or that they won’t do what you asked them to do or stop, and you can activate them when you want at scale in parallel or not use them at all for days.
For me it’s a dream honestly. I never felt so much empowered in my life while being enabled to unlock my creativity with almost no boundaries.
As the OP said really. It is very very freeing to just be able to give a task that in ye olden days you knew was going to be a pain in the arse to do, and it will do it without you. Amazing.
It's like having extra copies of yourself that you can just fork and get stuff done. E.g. the other day I was in a meeting and there were some heavy discussions about a few gnarly issues. I fired off a few prompts in parallel while the AI researches options, approaches, made a prototype etc all while I was present and paying attention still in the face to face meeting. I am so much so much more efficient as an engineer now - getting better feedback, more bonuses etc. It's never been better - I handle thinking about the requirements, the computer science and software engineering big picture etc, and the AI handles the syntax trivia for implementation.
The models are really really good now.
The only downside is that as a result of more stuff getting done, we need to have more meetings to discuss what actually needs to be worked on.
I've hit a weird point myself. I'm not overly interested in building software for other people now. What I'm able to do is build all my stupid hobby project ideas for my very specific use cases and not have to worry about making them profitable in order to justify the time spent on them. I just built my own point of sale system because nothing on the market meets my needs, and my needs are extremely niche.
Think bigger. Don't think in features. Think in business workflows.
You are a consumer, so you naturally think like the consumer you are. But consumers want cheap stuff and only big big businesses are able to provide cheap stuff at massive scale. So stop thinking in b2c. For us little businesses the money is in b2b. Businesses pay lots of money to have problems solved. They want to focus on their core business and outsource the rest.
Consumers hate paying subscriptions. Businesses love paying subscriptions. Because then they can make their problems someone else's problem and focus on what their actual business is.
Maybe it’s 1% or 10% today, but a decade ago, having/assembling the technical chops to know how to build something was a huge part of the meat / barrier to entry.
In OMP just setting the "task" role (which controls the default subagent model choice) to some cheap Flash model like Deepseek or Gemini goes a long way towards making better use of your quota.
Sol and Opus both are quite eager to use subagents and since the subagent tasks are often well specified and then verified by the main model there's hardly any advantage to using anything better for the subagent anyway, so IMO it's a no-brainer.
For me switching the main model to Kimi K3.1 Pro would put me towards 100% open, but I don't know if it uses subagents effectively. Any other open models I noticed just don't utilize them very much.
I think, despite the desperate attempt at IBM to remain relevant (and in business), their CEO has a point here that do you really need SOTA models for everything?
I had a really nice time with GLM5.2 for a month or two - it was, in my experience, approx Opus 4.5ish level of capability when I think Opus 4.8 was the new hotness
Even today, I use Gemini 3.8 Flash as my daily driver for most coding. Its fast and good, even if it is not the absolute SOTA at coding benchmarks it really is very good. Yes I have tried Opus 5.5 and 4 Argon etc and yes the are better, but do I need that?
I think that we are going to be at the stage soon-enough where models are "good enough" for 99% of people's needs. The SOTA will keep on marching ever-onwards I am sure, but I would not be surprised if people stick with previous-gen/previous-previous-gen/etc models because why upgrade?
If this does play out like this, I do wonder how that works out for the frontier labs (specifically OpenAI and Anthropic, less so Meta & Google) - if open models or even just cheaper Sonnet-level models from a few generations ago are good enough for basically everyone's daily work, where is the revenue? how can you justify pumping billions and billions into new generation models and datacenters etc when there is no driving need to use them?
I love 3.8 Flash, if only it weren’t unusably flaky. Sometimes simple responses take minutes or I get rate limited after sending a single message (I guess it’s a shared pool rate limit). Have you not experienced this (would love to know if there’s a way to actually use it reliably!)
"Affordable" and "IBM" don't belong in the same sentence together, but they're happy to sell you their inference-capable hardware if you've got the wallet for it.
I don't agree with that. They made the PC revolution possible by making hardware that was at least affordable enough to be bought by a large number of businesses and individuals. It also set the stage for the clone market. I'm pretty sure that if IBM really set their mind to it that they could do this.
That doesn't detract from your overall point that they tend to make overpriced stuff, but I don't think they have to do that. Besides, what's overpriced? NVidia's GB300 workstation sells for 95K euros, a single RTX 6000 is now 20K. Surely even IBM can compete at that level.
True, even so: They still make some pretty impressive hardware. They have access to fab capacity, they know how to make chips. They have the capital. If they really wanted they could do it.
In a way it is, but I would see that as a starting point, not as the ready solution. They'd need to do what NV has done for their eco-system: fantastic libraries, a credible future upgrade path, easy ordering, manufacturing partnerships, the works. I figure besides the performance gap that they are at least 5 to 10 years behind NV with their present offering. But NVidia is dropping stitches and IBM should take advantage of that.
I think IBM sees the Spyre as a reason to sell more Power hardware but that's just not where the industry is right now and I doubt it will be enough to create sufficient demand. Who knows, I don't have any insight in what they're doing there, but there is massive gap between A GB300 workstation and what IBM currently offers. At the same time: the RAM crunch hits all parties and that's a major obstacle as well.
Yeah. For all the jokes we make about paperclip maximizer, we've figured it out. It's not paperclips, it's RAM.
That's the thing though. Software is cheap with AI, but Nvidia has Cuda. If software doesn't matter anymore (it does) because it's so cheap now, then Nvidia's moat evaporates to IBM's Spyre. In reality "just a little bit of software" has inertia and cannot be dislodged quickly, quietly, or easily. cf Microsoft.
That. And not starting every sale with 'call us' where they first want you to tell them how much you can afford to spend. I buy a fair amount of gear and if the sales process starts with some sleazeball sales guy instead of a mouseclick on a local webshop then I'm out.
As for CUDA: I think CUDA is a formidable moat but I also think that that is precisely the kind of thing where the likes of IBM could (if they really wanted to) make things happen. They're already in the space with Redhat and they have ample knowledge about drivers and low level code. Putting together a team that either makes a drop-in replacement or alternatively, that starts supporting their custom hardware for the big players in the inference space (VLLM, llama.cpp and Torch) would make all the difference. They'd have to keep doing it though, even if the initial results were lower than expected because that market is not going to turn on a dime.
You'd think that with a 250K workforce and an active line of hardware including an AI accelerator that those people are still working there. I could be wrong, but someone has to be doing that work unless IBM in turn outsources it.
I know. But: they have gobs of money, they still have some talent, they still have a hardware department that I'm pretty sure would jump at a chance to give NV a run for their money. This would require a massive long term investment and a commitment that I'm not sure they would have the stomach for.
They do? They are a spec in the ocean relative to big players offering inference and training their own models. Pretty sure one thing they are struggling with is the money to compete successfully in the ai game this time around.
10 billion give or take, market cap 250 billion, 250K employees. That's not a 'spec in the ocean' by normal standards, and IBM is a mature company, not an inflated balloon.
When that happened I'd been working at Tandy (EU version of Radio Shack) for a while and even though we sold computers they were not taken serious until IBM came along and then suddenly we couldn't supply them fast enough.
Yes, that was the first time I had a 'proper' processor to play with. A Dragon 32 (CoCo clone from the UK).
I've never seen Tandy used outside of Europe but then again, I had not traveled as much then as now so it's definitely possible they used the brand on other continents as well. I wonder how many trees were cut for those catalogs :)
I didn't think that IBM had gear that mortals would want to run. If you're not running a mainframe, what do they have to sell you? They still sell Power servers and it looks like they have an AI card that can be used on both. But, if we're talking about the IBM CEO talking about on-premises, what are they really talking about hardware-wise? It's not something they would sell, is it?
This is more of a statement for their consulting customers, right?
Yes but unironically? IBM should be what NVIDIA would be competing with their NVL72 and DGX products. It doesn't make sense that IBM is just watching NVIDIA take its lunch.
How good will be the business of dealing with 2nd hand GPUs for open weights model in the future? What will the AI labs do with their GPUs once they are beyond their EOL? Can somebody please advise?
Of course IBM - the guys that sell you mainframes - would be saying this.
This is a win though. But not for the reasons people might think it is.
It is a win because it's a massive player now countering the new (temporary) behemoths. With all the leverage, budget and connections.
And contrary to the AI Labs, IBM isn't just that one SV rockstar dude, but a whole org with tons of people. Meaning that they're going to be far closer to the baseline of sane humans just from averaging alone.
The guy that sells mainframes also gives away the Granite models for free. I've found them to be quite capable in content retrieval and long-running workflows, even (or especially?) the small ones.
Considering the disclosed numbers for DeepInfer and the usefulness/popularity of DeepSeek 4.1 Flash, I expect it is inevitable that more companies are going to consider the self-hosting option seriously, and perhaps play with the idea of offering inference as a vendor.
Worth putting DeepSeek's own pricing rule next to that, because it sets the floor for self-hosting. In a closed-door investor meeting in May 2026, Liang Wenfeng said the API is priced so that "we go to the market and buy a batch of equipment, and recoup the cost in ten months" - that being, in his words, a reasonable profit. He adds that maximizing profit would mean charging more, but demand is inelastic in that band (he cut one model's price to a quarter and the team was delighted). I translated the excerpts, since the transcript only circulates in Chinese: https://telegra.ph/Liang-Wenfeng-on-DeepSeek-cost-restraint-...
If the vendor price is calibrated to a ten-month hardware payback, then at equal utilization there is no arbitrage in self-hosting: the crossover is roughly whether your sustained API spend exceeds the hardware plus that ten months. DeepSeek's own disclosed numbers give the benchmark to beat: 226.75 H800 nodes averaged over a day (~1,814 GPUs), 608B input tokens/day of which 56.3% hit the KV cache, 168B output tokens/day, ~73.7k tok/s input per node while prefilling and ~14.8k tok/s output while decoding.
That utilization is the hard part, and it is roughly what the IBM CEO is pointing at with "where are my kernel engineers". Self-hosting open weights on two or three racks is not the problem; amortizing them at a frontier lab's duty cycle is.
While he thinks most businesses can build useful AI infrastructure in just two or three racks, he feels many will struggle to get it running.
“The question becomes: how do you do that?” he said. “The question that the enterprise struggles with is: ‘Where are my kernel engineers? Where are the people who worry about the underlying network? Who can make sure that the model is clustered on GPUs that are close to each other in network proximity? Who can make sure that you’re doing the right caching of results so you get an incredible response?’”
Has he heard of “agentic AI”? You simply fire up Claude or ChatGPT and ask it how to do all of the above.
It's not a one time thing though - onprem inference requires continuous monitoring and tuning to get the most out of your GPUs. Looking at 25-50% performance benefits. Today, it does require dedicated Inference Engineers...
That's based on the idea that humans can learn and adapt and solve problems, and that's not what companies want to think about when they're seeking a profitable transaction such as hiring.
Do not do that if you can't prompt it correctly. It is a potential disaster waiting to happen. LLMs are likely to pull things off the shelf without real defense in depth and that's a problem. Maybe slightly less if it's only in an internal network, but still.
What happens when your competitors invest in better and more expensive tools or materials in general? They (potentially) make better products, but those products also cost more. It's turtles all the way down.
Can we please stop giving weight to CEOs talking their book? Like of course IBM wants you to think this, it would be great for them if it’s true. And maybe it is true! But IBM’s CEO saying it’s true tells me just as much as if a frontier lab said it’s not: nothing.
It was a brilliant move at the time for IBM, and has helped them shift their business more into providing services and support.
I imagine they are panicking a little with LLMs reducing the reliance on external companies for software support, so they are really hoping to shift it to on-prem AI, which is where the bias in his statement comes from.
I’ve been in tech for a long time… I have no idea what IBM does. Now, I get that I’m not the customer, that their customer is “business” but not my business… no… bigger business. The biggest of businesses I assume. I think it must either be a club I’m not in, or IBM is three half racks in a trenchcoat submitting invoices for millions to companies where people just pay it because “it’s IBM, it must be important”.
Review the specifications for the latest Telum II chip if you need a hint. 5.5ghz base clock.
If you absolutely need to scale something vertically, IBM Z is the final stop on the line.
X86 is great at hosting an entire farm of applications. You would never want to install salesforce or gitlab on a mainframe. That would completely defeat its purpose. If you've never encountered an appropriate workload, IBM might never make intuitive sense.
Ironically one of the smaller places I've worked at tried to use an IBM product once - I'm still not clear on what it was meant to do or how it was better than any cheaper alternative, so I assume someone in leadership was fooled by a very impressive pitch.
Lots of work went into integration and it never went live.
OpenAI most likely pushed Cerebras into a contractual agreement not to do business with Anthropic through 2030. ("Most likely" because the names of the companies Cerebras is contractually obligated not to sell to are redacted in public filings)
Over time NeoGeo realized they could sell you the hardware to play at home (“on-prem”) at arcade quality albeit at a price out of reach to most people. Alongside them, Nintendo and Atari gave you 80% of the fun at 20% of the cost. History will view games arcades as a flash in the pan compared to gaming at home.
Namco, Capcom, and Konami went into selling software instead of hardware but it feels hard to imagine buying a home copy of Fable** for $X,000 any time soon.
⸻
* That’s for video games. In fairground arcades you’d literally pay for tokens then use tokens for play. The analogy almost writes itself.
** The model, not the video game.
Now I want a local LLM server :( $25k is just so much money.
A lot more with Haiku equivalents (Qwen3.8 27B) or gemini. It's not frontier, but not that far away.
So much this. Plus everything else you mentioned. The freedom and control is real.
I've heard that there are legal software platforms like Clio that offer AI products, not sure what just eavesdropping on some lawyers at the bar.
Companies pretty much give us the full overview of what they have from a technical perspective (and often much more than that), we sign NDAs with investors that have serious teeth in them. The data they share with us would not be disclosed under any other circumstance and even our reports are shared only with the clients.
It's a nice little business. I'm not going to risk offending my customers for convenience or to save some money.
(Of course this all depends on if their plans work out, but if not then the smaller local models will probably play a big role in the failure of their plan)
At API pricing (i.e. what all enterprise customers are paying), inference is profitable, even if they're paying usurious rates on cost of capital for capex. Where the frontier labs are getting screwed is that they're in a Red Queen's race with each other, spending tens of billions of dollars per year training the next frontier model, and not being able to recoup the investment on that model before it's obsolete.
That's why you see the CEOs calling for government intervention to "pace the frontier": it lets them reduce investment on training without risking that their competitors keep going and make them obsolete. Unfortunately for them, the Chinese labs exist, and would love nothing more than to take the frontier, so they're in a bit of a pickle.
People keep repeating this, but it is only remotely close to true if you evaluate inference in a vacuum without any of the other supporting costs to facilitate that capability.
Just a few years ago people were saying the same thing about WeWork: "leasing is profitable, the unit economics of individual locations is net positive", and yet...
It doesn’t have to be black or white…
For my toy projects, adventures, learning exercises, and throwaway MVPs I will prefer the subscription plans… they are cheap for what they offer today.
It's training and tuning new models that cost them money. And that's why they are trying to scare people into thinking AI will kill us all in the next 5 years; they want bans on the open-weight models because they know they won't be able to compete with them.
If that is true, either the American labs are not optimizing correctly or they make a ton of money with inference.
The anthropic sub is likely still subventioned with VC money. So this may not even be available in 2 years.
But yeah, with current sub prices/budget... It's hard to go with anyone else unless you're rich.
The anthropic sub is likely still subventioned with VC money. So this may not even be available in 2 years.
But yeah, with current sub prices/budget... It'd stick with it too
They are subsidizing all of us to the tune of $billions in losses per month. That can't go on for a decade. Probably not even for another 18 months.
I still use all the tokens Claude will give me on the max plan. But I also run deepseek v4 locally for overflow and for... contingencies.
But, there is no need to make that bet - when prices increase, I'd look into renting a previous-gen GPU on a cloud provider. If the online GPU providers raise prices, then I would look into purchasing for on-site.
Until that happens, I'm not making any bets, because if all the token providers go bust GPU rentals are going to fall in price, and if they don't go bust after raising prices, I can still get years of GPUs for self-hosting at the price of a self-owned rig.
Just think of how user-hostile other online services have gotten, and imagine how acutely painful the proprietary LLM providers could make theirs. Thankfully, open models provide some counterbalance.
Are you comparing to subscription prices or API prices? Because if it's the former it won't last forever.
My server can run Qwen Flash Next at 500 t/s aggregate (about 200 t/s single user). It's pretty great. I use it for local coding and all my worker agents. It is about as reliable as Sonnet 5 for agentic coding. I make 3D models, keep the home network and server updated, fully built ansible files.
You do NOT need 20K in hardware to run these models reliably. I built my own inference engine from scratch for fun. You can run Strata on 16-32GB cards at very respectable token rates if you have the standard RAM to go with it. I strongly encourage people to go use this model on OpenRouter and try it out for tasks.
I also have a somewhat lobotimized quant of GLM Flash 5.3 working and it's pretty decent too.
I have 240G VRAM total which is a good reason why that rig cost as much as it does, and if I were to build it at today's prices it would be even more expensive.
Kind of ironic, but it's the only way. I just really like this as an option. For a household LLM I have the IQ4_K_XL quant and plenty of capacity for everything you could want in a household LLM and it is trustworthy enough to not rm -rf /, etc.
And yeah that is a really nice amount! You can run and compose a lot of interesting models, vision capabilities. IDK. Plus this level of model is far more interactive in the sense that it can implement a well scoped ticket or unit of work but its not going to magically just figure some things out without guidance or a spec :)
I tried getting some 170's when I first heard about that unlock but they were already unobtanium by then. That would have been a nice way to add a lot of VRAM.
GLM 5.3 runs pretty comfortably on this rig, even with large context.
Then again, I run a business and for the business it is absolutely manageable, it is roughly the price of a single job. Privately this would be a different matter but I assume that if you would do this privately you'd simply get a single (but powerful) GPU second hand and offload to RAM. Not as fast but with llama.cpp, 'ninfer', DwarfStar and various other clones of those two you can get some pretty decent performance out of single GPUs.
I've been buying powerful hardware since the 90's, starting with SGI and then ever larger Linux boxes. I make my living with them so it's never been a huge decision to invest some of the proceeds back into the tools. But on an absolute scale, yes, it is a lot of money.
I mean IBM already has a track record of building gigantic amounts of garbage and over charging large corps and govts for it coasting on their brand name and networks to get deals. So using some cheap ass model fits their MO. Having used their shit mountain called websphere, what an IBM CEO says is not something I take too seriously.
I can't wait to use models that are 'slightly less good' when I am at work instead of something like Fable that lets me focus on the problem at hand as much as possible rather than also spending additional effort correcting for the tool I am using...
In comparison, my highend mac book pro costs 105 euro per month (leasing cost). That's not quite enough to run the bigger models and besides running the GPU on a laptop constantly makes it a bit hostile for my lap. And the model I would need a lot more memory for the bigger models. It's nice for experimenting though.
If you get your own GPUs in some kind of server form, you could run those locally or in a data center. Or you just get hardware resources from your favorite cloud provider. Hetzner (one of the more affordable ones) GPU servers start at around 200 euro per month. I know some people that use that with open weight models. Works fine, especially if you can share the resources between multiple users.
At which point of course you are doing the same as cloud based providers are doing. But probably with a bit worse economics. It all boils down to whether you are actually saving money or paying for a lot of underutilized hardware that's running 24x7 when your people actually only show up about 8x5.
I don't think most companies should be doing that. It sounds like they'd end up with a lot of idling over capacity at a premium. The only valid reasons to do this would be securing access to GPUs or some very urgent need to keep things on premise (which a lot of companies insist is true for them until they realize how much they are going to pay).
It's also about "insurance protection" against the AI company raising the prices or changing their terms (e.g. becoming more frugal with respect to token drops).
So it matters a lot that these open-weight models are good enough; we would likely not be rolling out this new processing if it was fully dependent on Anthropic and OpenAI.
I do use Anthropic and OpenAI mostly because it costs me 1/40 and 1/50 of the api prices, which is way cheaper than Chinese subscriptions, but I’m ready to jump ship when they pull the plug. Fortunately they have nothing that forces me to keep paying them, and I don’t pay the crazy enterprise subscription costs, just the personal account prices.
I also die to be able to run the OSS setup locally to stop being spied on all the time.
For context, I averaged 10 hours a day prompting LLMs from when ChatGPT came out. At the start it was just function generation, then single file modules, then modules, and now I reason, design, implement and test with them.
An interesting thing is that with omp you can plan with a smart model and then implement with a dumber one. So for example, plan with Kimi 3.1 via OpenRouter and then implement with DSFV4 with zero data retention active, sorted by cheapest, and minimal tps accepted.
I get the people who are pissed about their cushy jobs evaporating, but if you know how to use these technologies it’s very freeing because it’s like having a company of people working for you, except they are not human so they’re not wasting their life working for you instead of using it for themselves.
You can make them do pretty much everything that is fuzzy-verifiable, you don’t have to worry about their feelings or that they won’t do what you asked them to do or stop, and you can activate them when you want at scale in parallel or not use them at all for days.
For me it’s a dream honestly. I never felt so much empowered in my life while being enabled to unlock my creativity with almost no boundaries.
I geniuinely don't see what I could make with this new tech, my brain immediately goes to "this app idea is worthless, anybody can make it now".
It's like having extra copies of yourself that you can just fork and get stuff done. E.g. the other day I was in a meeting and there were some heavy discussions about a few gnarly issues. I fired off a few prompts in parallel while the AI researches options, approaches, made a prototype etc all while I was present and paying attention still in the face to face meeting. I am so much so much more efficient as an engineer now - getting better feedback, more bonuses etc. It's never been better - I handle thinking about the requirements, the computer science and software engineering big picture etc, and the AI handles the syntax trivia for implementation.
The models are really really good now.
The only downside is that as a result of more stuff getting done, we need to have more meetings to discuss what actually needs to be worked on.
You are a consumer, so you naturally think like the consumer you are. But consumers want cheap stuff and only big big businesses are able to provide cheap stuff at massive scale. So stop thinking in b2c. For us little businesses the money is in b2b. Businesses pay lots of money to have problems solved. They want to focus on their core business and outsource the rest.
Consumers hate paying subscriptions. Businesses love paying subscriptions. Because then they can make their problems someone else's problem and focus on what their actual business is.
Maybe it’s 1% or 10% today, but a decade ago, having/assembling the technical chops to know how to build something was a huge part of the meat / barrier to entry.
Sol and Opus both are quite eager to use subagents and since the subagent tasks are often well specified and then verified by the main model there's hardly any advantage to using anything better for the subagent anyway, so IMO it's a no-brainer.
For me switching the main model to Kimi K3.1 Pro would put me towards 100% open, but I don't know if it uses subagents effectively. Any other open models I noticed just don't utilize them very much.
I had a really nice time with GLM5.2 for a month or two - it was, in my experience, approx Opus 4.5ish level of capability when I think Opus 4.8 was the new hotness
Even today, I use Gemini 3.8 Flash as my daily driver for most coding. Its fast and good, even if it is not the absolute SOTA at coding benchmarks it really is very good. Yes I have tried Opus 5.5 and 4 Argon etc and yes the are better, but do I need that?
I think that we are going to be at the stage soon-enough where models are "good enough" for 99% of people's needs. The SOTA will keep on marching ever-onwards I am sure, but I would not be surprised if people stick with previous-gen/previous-previous-gen/etc models because why upgrade?
If this does play out like this, I do wonder how that works out for the frontier labs (specifically OpenAI and Anthropic, less so Meta & Google) - if open models or even just cheaper Sonnet-level models from a few generations ago are good enough for basically everyone's daily work, where is the revenue? how can you justify pumping billions and billions into new generation models and datacenters etc when there is no driving need to use them?
That doesn't detract from your overall point that they tend to make overpriced stuff, but I don't think they have to do that. Besides, what's overpriced? NVidia's GB300 workstation sells for 95K euros, a single RTX 6000 is now 20K. Surely even IBM can compete at that level.
I think IBM sees the Spyre as a reason to sell more Power hardware but that's just not where the industry is right now and I doubt it will be enough to create sufficient demand. Who knows, I don't have any insight in what they're doing there, but there is massive gap between A GB300 workstation and what IBM currently offers. At the same time: the RAM crunch hits all parties and that's a major obstacle as well.
That's the thing though. Software is cheap with AI, but Nvidia has Cuda. If software doesn't matter anymore (it does) because it's so cheap now, then Nvidia's moat evaporates to IBM's Spyre. In reality "just a little bit of software" has inertia and cannot be dislodged quickly, quietly, or easily. cf Microsoft.
As for CUDA: I think CUDA is a formidable moat but I also think that that is precisely the kind of thing where the likes of IBM could (if they really wanted to) make things happen. They're already in the space with Redhat and they have ample knowledge about drivers and low level code. Putting together a team that either makes a drop-in replacement or alternatively, that starts supporting their custom hardware for the big players in the inference space (VLLM, llama.cpp and Torch) would make all the difference. They'd have to keep doing it though, even if the initial results were lower than expected because that market is not going to turn on a dime.
They do? They are a spec in the ocean relative to big players offering inference and training their own models. Pretty sure one thing they are struggling with is the money to compete successfully in the ai game this time around.
Yeah they were only half a decade late to the PC revolution back in the day...
I've never seen Tandy used outside of Europe but then again, I had not traveled as much then as now so it's definitely possible they used the brand on other continents as well. I wonder how many trees were cut for those catalogs :)
This is more of a statement for their consulting customers, right?
If you have ever worked at IBM, it makes total sense. Completely dysfunctional middle management and on up.
I was able to get this related article
https://rcrtech.com/semiconductor-news/what-happens-to-old-a...
DeepSeek v4.1 Flash or GLM 3 are already very very capable models.
Only bottleneck at the moment is the hardware prices.
This is a win though. But not for the reasons people might think it is.
It is a win because it's a massive player now countering the new (temporary) behemoths. With all the leverage, budget and connections.
And contrary to the AI Labs, IBM isn't just that one SV rockstar dude, but a whole org with tons of people. Meaning that they're going to be far closer to the baseline of sane humans just from averaging alone.
If the vendor price is calibrated to a ten-month hardware payback, then at equal utilization there is no arbitrage in self-hosting: the crossover is roughly whether your sustained API spend exceeds the hardware plus that ten months. DeepSeek's own disclosed numbers give the benchmark to beat: 226.75 H800 nodes averaged over a day (~1,814 GPUs), 608B input tokens/day of which 56.3% hit the KV cache, 168B output tokens/day, ~73.7k tok/s input per node while prefilling and ~14.8k tok/s output while decoding.
That utilization is the hard part, and it is roughly what the IBM CEO is pointing at with "where are my kernel engineers". Self-hosting open weights on two or three racks is not the problem; amortizing them at a frontier lab's duty cycle is.
we technologists tend to extend what we see and think to other realms where it doesn't make sense
https://karlbode.com/ceo-said-a-thing-journalism/
I imagine they are panicking a little with LLMs reducing the reliance on external companies for software support, so they are really hoping to shift it to on-prem AI, which is where the bias in his statement comes from.
If you absolutely need to scale something vertically, IBM Z is the final stop on the line.
X86 is great at hosting an entire farm of applications. You would never want to install salesforce or gitlab on a mainframe. That would completely defeat its purpose. If you've never encountered an appropriate workload, IBM might never make intuitive sense.
Also, the roadmap:
https://newsroom.ibm.com/2026-08-24-ibm-unveils-next-generat...
Lots of work went into integration and it never went live.
For one thing, amongst many, IBM mainframes process 70% of global transactions on a value basis
Cerebras with Claude 8?