"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.
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 :)
It was only a difference in the store names. They were sold at Radio Shack stores in US. Tandy, originally a leather company, bought Radio Shack in the 60s).
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.
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.
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?
I can picture that they’d have the software side sorted out with RedHat. But I don’t know what hardware they’d push to support local on-premises models. It looks like when you buy Fusion HCI, you can choose from Dell or Lenovo as the underlying hardware supplier.
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)
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.
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.
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.
It can be very dangerous to let people loose with such advice, FWIW. They don't necessarily know what "internal usage" actually means, or how important it is to keep it that way as long as you don't have a real expert to at least help direct the LLMs.
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...
The optimal inference configuration varies according to the model, the specific traffic (which changes in time), and the specific hardware it's running to. The "knowledge" for it doesn't exist yet.
Usually in similar scenarios an agent will try to experiment with various configurations and get to a good result. However, you can't have an agent test out different inference serving configurations/architectures, since every experiment might involves 10-60 minutes of spinning up a model (on very expensive GPUs)...
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.
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.
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.
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.
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.
Isn't he the guy who posted about his LEGO-sorting computer vision project? That post alone taught me things that increased my career opportunities significantly.
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.
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 ;)
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.
Well there are risks to being part of a crowd that's all compelled to make the same decision at the same time. Certain risks can only be transferred, not reduced in absolute terms.
It's not economical today to run my own hardware, but the insurance is worth it to me in the sense that if my bet is wrong, I'm still okay with paying extra for the protection.
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.
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 ...
It's a good first step. They're not great but they kind of work for many tasks. I think the small models can be useful in context-specific tasks but the friction is high right now because the user is forced to determine whether Siri is good enough, or a better LLM is required per task. Ideally Siri would be smart enough to intelligently route. Apple claims that's how it works but even their cloud processing is far behind frontier models.
Worryingly, they're taking a walled garden approach, and they don't want to expose user data to competitors. After the monumental failure of Siri over more than a decade, I don't trust Apple to execute competently on AI.
I mostly run Qwen3.8 Flash Next on 2xR9700 (but Intel Arc Pro B70 may be better currently, or 4x3090 is even better but out of budget, 32GiB Nvidia GCs on ebay are also usable).
I have 128GiB DDR4 (3200Hz), but 64GiB is enough, even 32GiB maybe, I just reuse it from an old project.
You may need a AMD EPYC processor and associated motherboard, don't forget a good PSU at least 1200W (300W processor + 400W by card to support spikes of consumption).
On some setup (~15k€) you can use 4 CG to go up to 128G of VRAM.
I'm not using a EPYC processor for now, so I get 170tps on decode but very slow prefill, the bottleneck is the PCIe exchange between the cards, I need to setup P2P (it means what you think it means, for GCs).
My setup supports full context (2*18 bytes so 256K), I didn't setup YaRN to get 1M of context (2*20 bytes), but it's an option with 128GiB of RAM, the visual component is extracted from RAM so a little slower.
I'm not using multiple agents but some people does, it's god like with 4 graphics cards, on 2 only, it's relatively quick but limited to single user and up to 3 to 4 concurrents requests only.
It's not frontier, but look AA of QFN, it's good enough. DeepSeek V4.1 can run on my setup, but it's slow. On dual B70, there is some PoCs to make it usable (>60tps). It improves every days.
If you need more information, look for local setup users on the Web. There is a lot of people sharing information about benchmarks and setups. L80 is my reference for now.
If you already have a machine, throwing an r9700 in it is one of the cheapest ways to get qwen3.8 27B Q6 with 100k+ context working.
Sadly, the r9700 has gone up in price, but it's still the cheapest route. I use It on an old x99 Linux desktop and it achieves 50 tps decode and 1000+ tps prefill.
The build uses my machine's ROCM and I just link it instead of embedded all that in an image, like the published ones do.
Then deploying it via docker compose to my local machine. The models I want to run are stored in a docker volume, so it's easy to switch the model while keeping the same build.
The biggest negative for me is that I don't think it's enough to do parallelization through llamacpp (I'd prefer to work on one thing faster than two things slowly). I haven't gotten vllm working yet. Also, once you have one, you constantly start thinking about a second card...
It will not do that unless you specifically teach it to do that, or you use an Anthropic model. The issue in this case was Anthropic's safety monitoring built into the model. They record everything and flag potential illegal activity for human review. This occurs even over the API. I do not know if OpenAI does this, but I would not be surprised. I think the Chinese models are far less likely to do this (unless one resides in China).
I’m suggesting it’s highly unlikely it will call the FBI to dob you in. That affirmatively trained behaviour, not something one should expect to accidentally happen.
> No political censorhip? Try asking Chinese models about Taiwan.
Self-hosting means being able to use abliterated models where the superficial instructions like not describing Taiwan as independent are burned off.
Censorship and bias are related but not synonymous. I agree that it's harder to remove political bias when the training data includes it. We know many of the frontier models are heavily trained on Reddit, for example, [which likely explains why we see such heavy left wing bias.](https://trakkr.ai/bias/findings)
All good points, but I am not sure who I trust less. Chinese companies or random people on Hugging Face. :) It is still a black box. And I have a feeling I could not abliterate those models on my own hardware.
I am not saying I would not use any of them at all, but I would avoid doing certain tasks with them. And with the testing I have done with abliterated models in the past, it is not 100% successful and it can still hit the same walls as original model.
I agree. I think the future for privacy conscious people is going to be an adversarial open weights model, where competing models (ideally from opposing superpowers and labs) critique each other's work to find consensus. Or at least alert for suspicious behaviour.
>abliterated models where the superficial instructions like not describing Taiwan as independent are burned off
Does this help with issues like including security vulnerabilities when you say you're in Tibet? I understood that this was burned in at the training level, not just a layer of censorship added at the last minute.
It could, but it's an inexact science. A little bit like tampering with the brain. Even a surgical procedure removing a tiny tumor can interfere with critical pathways. Usually abliterated models end up almost exactly as capable in benchmarks, but it requires the person doing the abliteration to anticipate or know about the censorship. In the case of Chinese model censorship it is believed that it is post-training, so the underlying data input is not the issue. It's how the model has been trained to use that data. Specifically, abliteration usually targets the refusal behaviour. I would guess that this example could be significantly improved, but I have not seen any abliteration attempts re this specifically.
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.
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 :)
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.
If you have ever worked at IBM, it makes total sense. Completely dysfunctional middle management and on up.
This is more of a statement for their consulting customers, right?
I can picture that they’d have the software side sorted out with RedHat. But I don’t know what hardware they’d push to support local on-premises models. It looks like when you buy Fusion HCI, you can choose from Dell or Lenovo as the underlying hardware supplier.
Cerebras with Claude 8?
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.
The optimal inference configuration varies according to the model, the specific traffic (which changes in time), and the specific hardware it's running to. The "knowledge" for it doesn't exist yet.
Usually in similar scenarios an agent will try to experiment with various configurations and get to a good result. However, you can't have an agent test out different inference serving configurations/architectures, since every experiment might involves 10-60 minutes of spinning up a model (on very expensive GPUs)...
Are you comparing to subscription prices or API prices? Because if it's the former it won't last forever.
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.
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.
It's not economical today to run my own hardware, but the insurance is worth it to me in the sense that if my bet is wrong, I'm still okay with paying extra for the protection.
So much this. Plus everything else you mentioned. The freedom and control is real.
Now I want a local LLM server :( $25k is just so much money.
Worryingly, they're taking a walled garden approach, and they don't want to expose user data to competitors. After the monumental failure of Siri over more than a decade, I don't trust Apple to execute competently on AI.
A lot more with Haiku equivalents (Qwen3.8 27B) or gemini. It's not frontier, but not that far away.
You may need a AMD EPYC processor and associated motherboard, don't forget a good PSU at least 1200W (300W processor + 400W by card to support spikes of consumption). On some setup (~15k€) you can use 4 CG to go up to 128G of VRAM.
I'm not using a EPYC processor for now, so I get 170tps on decode but very slow prefill, the bottleneck is the PCIe exchange between the cards, I need to setup P2P (it means what you think it means, for GCs).
My setup supports full context (2*18 bytes so 256K), I didn't setup YaRN to get 1M of context (2*20 bytes), but it's an option with 128GiB of RAM, the visual component is extracted from RAM so a little slower. I'm not using multiple agents but some people does, it's god like with 4 graphics cards, on 2 only, it's relatively quick but limited to single user and up to 3 to 4 concurrents requests only.
It's not frontier, but look AA of QFN, it's good enough. DeepSeek V4.1 can run on my setup, but it's slow. On dual B70, there is some PoCs to make it usable (>60tps). It improves every days.
If you need more information, look for local setup users on the Web. There is a lot of people sharing information about benchmarks and setups. L80 is my reference for now.
Edit: I translated the unity.
octet => byte, https://fr.wiktionary.org/wiki/Gio
Sadly, the r9700 has gone up in price, but it's still the cheapest route. I use It on an old x99 Linux desktop and it achieves 50 tps decode and 1000+ tps prefill.
My stack uses a local llamacpp ROCM build, modified by https://github.com/stew675/llama-cpp-rdna-boosts
The build uses my machine's ROCM and I just link it instead of embedded all that in an image, like the published ones do.
Then deploying it via docker compose to my local machine. The models I want to run are stored in a docker volume, so it's easy to switch the model while keeping the same build.
The biggest negative for me is that I don't think it's enough to do parallelization through llamacpp (I'd prefer to work on one thing faster than two things slowly). I haven't gotten vllm working yet. Also, once you have one, you constantly start thinking about a second card...
This is an alignment issue. OpenClaw with a local LLM with an email you setup for it could contact the police.
Sorry, you're claiming local LLMs are somehow deterministic and never show unexpected behavior?
Some censorhip and bias towards certain ideology is baked into those models.
While it is not a problem for programming tasks, I would never use it for certain other tasks.
Self-hosting means being able to use abliterated models where the superficial instructions like not describing Taiwan as independent are burned off.
Censorship and bias are related but not synonymous. I agree that it's harder to remove political bias when the training data includes it. We know many of the frontier models are heavily trained on Reddit, for example, [which likely explains why we see such heavy left wing bias.](https://trakkr.ai/bias/findings)
More citations:
https://www.gsb.stanford.edu/faculty-research/working-papers... (article: https://news.stanford.edu/stories/2025/05/ai-models-llms-cha...)
https://www.tandfonline.com/doi/full/10.1080/29974100.2026.2...
https://ieeexplore.ieee.org/document/10817610
https://www.washingtonpost.com/technology/interactive/2026/0...
https://link.springer.com/article/10.1007/s11127-023-01097-2
I am not saying I would not use any of them at all, but I would avoid doing certain tasks with them. And with the testing I have done with abliterated models in the past, it is not 100% successful and it can still hit the same walls as original model.
https://www.crowdstrike.com/en-us/blog/crowdstrike-researche...
In open models you can "uncensor" them: https://arxiv.org/abs/2406.11717