i assume that $4.6 mil is just the cost of the electricity?

Hard to be sure because the source of that information isn't known, but generally when people talk about training costs like this they include more than just the electricity but exclude staffing costs.

Other reported training costs tend to include rental of the cloud hardware (or equivalent if the hardware is owned by the company), e.g. NVIDIA H100s are sometimes priced out in cost-per-hour.

Citation needed on "generally when people talk about training costs like this they include more than just the electricity but exclude staffing costs".

It would be simply wrong to exclude the staffing costs. When each engineer costs well over 1 million USD in total costs year over year, you sure as hell account for them.

If you have 1,000 researchers working for your company and you constantly have dozens of different training runs in the go, overlapping each other, how would you split those salaries between those different runs?

Calculating the cost in terms of GPU-hours is a whole lot easier from an accounting perspective.

The papers I've seen that talk about training cost all do it in terms of GPU hours. The gpt-oss model card said 2.1 million H100-hours for gpt-oss:120b. The Llama 2 paper said 3.31M GPU-hours on A100-80G. They rarely give actual dollar costs and I've never seen any of them include staffing hours.

Do they include the costs of dead-end runs?

No, they don't! That's why the "5.5 million" deepseek V3 number as read by American investors was total bullshit (because investors ignored their astrik saying "only final training run")

Yeah, that's one of the most frustrating things about these published numbers. Nobody ever wants to share how much money they spent on runs that didn't produce a useful model.

As with staffing costs though it's hard to account for these against individual models. If Anthropic run a bunch of training experiments that help them discover a new training optimization, then use that optimization as part of the runs for the next Opus and Sonnet and Haiku (and every subsequent model for the lifetime of the company) how should the cost of that experimental run be divvied up?

No, because what people are generally trying to express with numbers like these, is how much compute went into training. Perhaps another measure, like zettaflop or something would have made more sense.

Table 1:

https://arxiv.org/html/2412.19437v2

That number is as real as the 5.5 million to train DeepSeek. Maybe it's real if you're only counting the literal final training run, but total costs including the huge number of failed runs all other costs accounted for, it's several hundred million to train a model that's usually still worse than Claude, Gemini, or ChatGPT. It took 1B+ (500 billion on energy and chips ALONE) for Grok to get into the "big 4".

Using such theory, one can even argue that the real cost needs to include the infrastructures, like total investment into the semiconductor industry, the national electricity grid, education and even defence etc.

Correct! You do have to account for all of these things! Unironically correct! :)

That's baked in to the cost of the hardware though.

the china government has been heavily subsidizing the electricity

Can confirm MiniMax M2 is very impressive!

> That's small enough to run well on ~$5,000 of hardware...

Honestly curious where you got this number. Unless you're talking about extremely small quants. Even just a Q4 quant gguf is ~130GB. Am I missing out on a relatively cheap way to run models well that are this large?

I suppose you might be referring to a Mac Studio, but (while I don't have one to be a primary source of information) it seems like there is some argument to be made on whether they run models "well"?

Yes, I mean a Mac Studio with MLX.

An M3 Ultra with 256GB of RAM is $5599. That should just about be enough to fit MiniMax M2 at 8bit for MLX: https://huggingface.co/mlx-community/MiniMax-M2-8bit

Or maybe run a smaller quantized one to leave more memory for other apps!

Here are performance numbers for the 4bit MLX one: https://x.com/ivanfioravanti/status/1983590151910781298 - 30+ tokens per second.

Thanks for the info! Definitely much better than I expected.

It’s kinda misleading to omit the generally terrible prompt processing speed on Macs

30 tokens per second looks good until you have to wait minutes for the first token

The tweet I linked to includes that information in the chart.

Running in cpu ram works fine. It’s not hard to build a machine with a terabyte of RAM.

Admittedly I've not tried running on system RAM often, but every time I've tried it's been abysmally slow (< 1 T/s) when I've tried on something like KoboldCPP or ollama. Is there any particular method required to run them faster? Or is it just "get faster RAM"? I fully admit my DDR3 system has quite slow RAM...