No provider maintains 100% utilization of GPUs at full rate. Demand is bursty - even if this project is successful, you might expect, e.g., things to be busy during the stock market times when Claude is throwing API errors and then severely underutilized during the same times that Anthropic was offering two-for-one off peak use.
And then there's a hit for overprovisioning in general. If the network is not overprovisioned somewhat, customers won't be able to get requests handled when they want, and they'll flee. But the more overprovisioned it is, the worse it is for compute seller earnings.
I suspect an optimistic view of earnings from a platform like this would be something like 1/8 utilization on a model like Gemma 4. Their calculator estimates my m4 pro mini could earn about $24/month at 3 hours/day on that model. That seems plausible.
@eigengajesh - Your cost estimator lists Mac Mini M4 Pro with only 24 or 48GB options, but the M4 Pro mini can also be configured with 64GB. At least, I hope so, as I'm typing this on one. ;-)
Oh, also, you seem to have some bugs:
Gemma:
WARN [vllm_mlx] RuntimeError: Failed to load the default metallib. This library is using language version 4.0 which is not supported on this OS. library not found library not found library not found
cohere:
2026-04-16T14:25:10.541562Z WARN [stt] File "/Users/dga/.darkbloom/bin/stt_server.py", line 332, in load_model
2026-04-16T14:25:10.541614Z WARN [stt] from mlx_audio.stt.models.cohere_asr import audio as audio_mod
2026-04-16T14:25:10.541643Z WARN [stt] ModuleNotFoundError: No module named 'mlx_audio.stt.models.cohere_asr'
Trying to download the flux image models fails with:
curl: (56) The requested URL returned error: 404
darkbloom earnings does not work
your documentation is inconstent between saying 100% of revenue to providers vs 95%
I think .. this needs a little more care and feeding before you open it up widely. :) And maybe lay off the LLM generated text before it gets you in trouble for promising things you're not delivering.
Just noting it. The other post was submitted earlier. The mod's can figure out how to combine/reconcile. Update: I think you are correct and this one won :)
You're getting downvoted but it's a reasonable question if posed in good faith. The tl;dr is that there are really a few options for what could happen to those orange peels:
(1) Landfill burial
(1a) Without methane capture and use: Produces methane, relatively high short term warming potential.
(1b) With methane capture and use: Ends up as CO2 after burning the methane.
(2) Composting (this approach)
(2a) Mostly aerobic: Produces CO2
(2b) Mostly anaerobic: Produces methane
A deep pile that is never turned will decompose anaerobically, resulting in fairly undesirable methane. A shallower pile or one that is mixed well will result in mostly aerobic decomposition. The aerobic decomposition will produce CO2 but not huge amounts of it. Each hectare of land could absorb something like ~8 tons of CO2 per year; with 7 hectares, the CO2 emitted by composting 12t of oranges is going to be dwarfed by the new vegetation. After a few years when you're growing big trees, the rate of CO2 absorption might rise as high as 20-30t/year/hectare in costa rica's environment. And this is probably an underestimate, as the soil amendment of the orange peels seems to have stimulated faster regrowth than would have happened otherwise.
And perhaps more to the point: There isn't really a purely "no co2" way of disposing of organic matter other than perhaps burying it at the bottom of a deep mineshaft (but the co2 or methane will be produced anyway). Landfilling it is strictly worse - you still get the decomposition products, _or worse_ because you'll mostly get methane, but without producing useful soil byproducts.
Overall this project is a huge win on a carbon perspective and a waste reduction perspective.
It's not a reasonable question. What's the alternative for the orange peels? They were going to rot and release that CO2 whether they did it in a big pile here or somewhere else.
But seriously GP could have had a mental model that landfilled orange peels might sit there for a long time -- which depending on conditions and food could be true on human scales (like 10-40 years) but not on the scale of 100 years. Especially if the conditions were dry -- a dry orange peel is pretty robust. That's not likely to be the case in Costa Rica, but I'll forgive some naivety here absent demonstrated malice.
I don't sell mine, but I time-shift with a small pile of batteries (about 10kWh) and it's pretty reasonable. I save about usd $30/month. It's basically a big ups that will pay for itself in ten years, and I get backup power.
I don't. The batteries will last longer than 10 years. The 10 year typical advertised lifetime of lifepo4 is to 80% capacity, and I'll just keep on using them.
The actual payoff calculation is a lot messier than that because you have to factor in the NPV of buying batteries vs. just throwing the money in the market, AND you have to be able to forecast that growth vs. growth in power prices. So the honest truth is I have no idea if it's going to be a net good investment vs other options.
Fortunately, I don't have to care, because I bought the batteries for UPS runtime, which I value independent of the time-shifting. The time-shifting is just a way to squeeze money out of an investment I already made. Had I been going for payoff, there are cheaper battery/inverter options out there with a sub-5y payoff.
I find it kind of helpful and interesting to see a subset of these called out with a bit of data. Helps keep my LLM detector trained (the one in my brain, that is) and I think it helps a little about expressing the community consensus against this crap. In this case, I'm glad the GP posted something, as it's definitely not mistaken.
As a specific example: The generated diagram showing the expression tree under "build in python" is simply wrong. It doesn't correspond to the expression x * 2 + 1, which should have only 1 child node on the right. The "GIL Released - Released" is just confusing. The dataflow omits the fact that the results end up back in python - there should be a return arrow. etc., etc.
If you use diagrams like this, at least ensure they are accurately conveying the right understanding.
And in general, listen to the person I'm responding to -- be really deliberate with your graphics or omit. Most AI-generated diagrams are crap.
Installation costs and inverters not included, however.
reply