That's like ten normal computers worth of power for the GPUs alone.
Maybe if your "computer" in question is a smartphone? Remember that the M3 Ultra is a 300w+ chip that won't beat one of those 3090s in compute or raster efficiency.
But if you have to factor in hardware costs self-hosting doesn't seem attractive. All the models I can self-host I can browse on openrouter and instantly get a provider who can get great prices. With most of the cost being in the GPUs themselves it just makes more sense to have others do it with better batching and GPU utilization
I also don't think the 100% util is necessary either, to be fair. I get a lot of value out of my two rigs (2x rtx pro 6000, and 4x 3090) even though it may not be 24/7 100% MFU. I'm always training, generating datasets, running agents, etc. I would never consider this a positive ROI measured against capex though, that's not really the point.
All of the bottlenecks in sum is why you'd never get to 100% MFUs (but I was conceding you probably don't need to in order to get value)
And what are you doing that I/O is a bottleneck?
I don't believe it's moot, but I understand your point. The fact that models are memory bandwidth bound does not at all mean that other overhead is insignificant. Your practical delivered throughput is the minimum of compute ceiling, bandwidth ceiling, and all the unrelated speed limits you hit in the stack. Kernel launch latency, Python dispatch, framework bookkeeping, allocator churn, graph breaks, and sync points can all reduce effective speed. There are so many points in the training and inference loop where the model isn't even executing.
> And what are you doing that I/O is a bottleneck?
We do a fair amount of RLVR at my org. That's almost entirely waiting for servers/envs to do things, not the model doing prefill or decode (or even up/down weighting trajectories). The model is the cheap part in wall clock terms. The hard limits are in the verifier and environment pipeline. Spinning up sandboxes, running tests, reading and writing artifacts, and shuttling results through queues, these all create long idle gaps where the GPU is just waiting to do something.
I'm not sure why, sandboxes/envs should be small and easy to scale horizontally to the point where your throughput is no longer limited by them, and the maximum latency involved should also be quite tiny (if adequately optimized). What am I missing?
But disregarding that, this isn't a problem you can solve by turning a knob akin to scaling a stateless k8s cluster.
The whole vertical of distributed RL has been struggling with this for a while. You can in theory just keep adding sandboxes in parallel, but in RLVR you are constrained by 1) the amount of rollout work you can do per gradient update, and 2) the verification and pruning pipeline that gates the reward signal.
You cant just arbitrarily have a large batch size for every rollout phase. Large batches often reduce effective diversity or get dominated by stragglers. And the outer loop is inherently sequential, because each gradient update depends on data generated by a particular policy snapshot. You can parallelize rollouts and the training step internally, but you can’t fully remove the policy-version dependency without drifting off-policy and taking on extra stability headaches.
So I would still point out the GP (Original comment) where yes, it might not make financial sense to run these AI Models [They make sense when you want privacy etc, which are all fair concerns but just not financial sense]
But the fact that these models are open source still means that they can be run when maybe in future the dynamics might shift and it might make sense running such large models locally. Even just giving this possibility and also the fact that multiple providers could now compete in say openrouter etc. as well. All facts included, definitely makes me appreciate GLM & Kimi compared to proprietory counterparts.
Edit: I highly recommend this video a lot https://www.youtube.com/watch?v=SmYNK0kqaDI [AI subscription vs H100]
This video is honestly one of the best in my opinion about this topic that I watched.
It's fixed now :)
When talking about fallback from Claude plans, The correct financial comparison would be the same model hosted on OpenRouter.
You could buy a lot of tokens for the price of a pair of 3090s and a machine to run them.
That's a subjective opinion, to which the answer is "no you can't" for many people.
Could you elaborate? I fail to grasp the implication here.
you can't be a happy uber driver making more money in the next 24 months by having a fancy car fitted with the best FSD in town when all cars in your town have the same FSD.
I've got a lite GLM sub $72/yr which would require 138 years to burn through the $10K M3 Ultra sticker price. Even GLM's highest cost Max tier (20x lite) at $720/yr would buy you ~14 years.
Buy a couple real GPUs and do tensor parallelism and concurrent batch requests with vllm and it becomes extremely cost competitive to run your own hardware.
No one's running these large models on a Mac Mini.
> Of course if you buy some overpriced Apple hardware it’s going to take years to break even.
Great, where can I find cheaper hardware that can run GLM 5's 745B or Kimi K2.5 1T models? Currently it requires 2x M3 Ultras (1TB VRAM) to run Kimi K2.5 at 24 tok/s [1] What are the better value alternatives?
I wish I had better numbers to compare with the 2x M3 Ultra setup. My system is a few RTX A4000s on a Xeon with 190GB/s actual read bandwidth, and I get ~8 tok/s with experts quantized to INT4 (for large models with around 30B active parameters like Kimi K2.) Moving to 1x RTX Pro 6000 Blackwell and tripling my read bandwidth with EPYC Turin might make it competitive with the the macs, but I dunno!
There's also some interesting tech with ktransformers + sglang where the most frequently-used experts are loaded on GPU. Pretty neat stuff and it's all moving fast.
my system is running GLM-5 MXFP4 at about 17 tok/s. That’s with a single RTX Pro 6000 on an EPYC 9455P with 12 channels of DDR5-6400. Only 16k context though, since it’s too slow to use for programming anyway and that’s the only application where I need big context.
Even if you quantize the hell out of the models to fit in the memory, they will be very slow.
Doesn’t mean you shouldn’t do it though.
They can do a lot of simple tasks in common frameworks well. Doing anything beyond basic work will just burn tokens for hours while you review and reject code.
I hope too many of us won't be doing this and cause Google to add limits! My hope is Google sees the benefit in this and goes all in - continues to let people decide which Google hosted model to use, including their own.
I guess that's debatable. I regularly run out of quota on my claude max subscription. When that happens, I can sort of kind of get by with my modest setup (2x RTX3090) and quantized Qwen3.
And this does not even account for privacy and availability. I'm in Canada, and as the US is slowly consumed by its spiral of self-destruction, I fully expect at some point a digital iron curtain will go up. I think it's prudent to have alternatives, especially with these paradigm-shattering tools.