Strange that in the prior art they didn't list DwarfStar, as it is able to run the same model (probably quantized differently though) in less memory. Maybe the author isn't aware of it?
Edit: Now I think about it, this might be the cheapest way to run the DeepSeek V4 Flash 0731 on a dedicated inference server at original weights. I haven’t run mixed load benchmarks but I guess it’s possible to generate $3-$4 worth of tokens per hour and still maintain a usable per-user throughput.
830t/s is burst aggregate. ~500 is sustained and it's for 8 concurrent users. Meaning for $1.99/hour if you serve 8 users it's 8*$0.54, not just $0.54.
You shouldn't rent one out if you're just serving it for yourself, but from a financial standpoint if you sell to users you can take a 100% margin.
Bro 500 Aggregate. so that's 500 * 60 * 60 = 1.8M output which is .5$ at best... Not including pre-fill and stuff.
This is not the real margins, even if you are selling to 8 users it's 90 tps per median stream. So assuming that .6-.7$
This is not even remotely worth it.
You need to 3x this tps(~1500 tps) to be worth it, and that's what most providers are doing, at 20-30 users at 50-60 tps with better optimized batch processing and kernels you can make some profit.
Definitely not. Inference is not as expensive to operate as many people seem to assume. The frontier labs are probably making a lot of money from selling tokens. It’s covering all of the R&D costs like salaries, collecting training material, and running the large training operations that costs a lot of money.
Agentic workloads are somewhere around 1%/0.5%/98.5% input/output/cached tokens. Cached tokens are pretty much free for inference providers (if they implement sparse and compressed attention properly) and throughput for input tokens is much higher.
Lets assume that you've got 2 million input tokens, 1 million output tokens and 98.5 million cached tokens to process. That would cost 2 * $0.14 + 1 * $0.28 + 98.5 * $0.0028 = $0.8358 with DeepSeek API pricing.
For comparison, it would take 2M / 8000 + 1M / 800 = 1500 seconds to process this amount of tokens with the linked framework, which is about $0.83 when we assume $2/hr for one MI300X.
However, other inference providers have 10 times higher prices for cached tokens, which results in a comfortable margin.
And we should not discount that DeepSeek also gets paid in data, which is probably more valuable to them.
And I believe that this framework still has some room for optimization for generation with high batch sizes.
> should not discount that DeepSeek also gets paid in data, which is probably more valuable to them
That's agentic feedback loops for training, right? Any more detail on this, such as how they actually tell whether that data is good or not? That seems like a very hard problem, and like the value of that data is low compared to just building their own, controlled RL gyms.
Your math is a bit funny if you're assuming the 1/0.5/98.5 ratios: you doubled input and output tokens but not cached. If you double cached tokens to match your original ratio it works out to around $1.11, and if you 10x the cached token cost it's around $6.08.
Based on your $0.83 estimate, the margin isn't great. This is within shooting distance of "at cost" which is probably pretty close to what DeepSeek is operating with, ignoring the value of the data they're collecting of course.
> And I believe that this framework still has some room for optimization for generation with high batch sizes.
If that optimization can bring this scenario closer to $0.50 then it gets pretty compelling, otherwise I'm not confident.
Oh, I messed up. Half-way through, I thought it would be a good idea to double the numbers so I don't have to deal with half millions, but forgot to also double the 98.5. Unfortunately, I can not edit it anymore.
I think the margins of DeepSeek may be a bit better than with this vibe-coded framework here, since they had the liberty of optimizing their models for their own hardware.
At the time, open frameworks were not anywhere close to achieving that number. Not sure whether they caught up. The software wizards at DeepSeek are quite skilled.
I think Deepseek is selling roughly at cost (perhaps a slight premium). They don’t guarantee that they don’t train on the submitted prompts, so I suspect they are mining the data. Mining for what? Well, who knows. Best case, mining to make Deepseek better. That said, I use Deepseek all the time. It has done a whole lot of ‘ls’ commands on my system, though.
Use nvidia hardware instead and use a larger cluster serving many more users concurrently. Easily 10x–20x higher token rate per GPU with public solutions like dynamo and sglang.
This is exactly what I came to say. The price of Flash is so cheap that trying to run it locally or with your own hardware is pointless. I was using it about a month ago to program some stuff and ran it for 4 days non-stop and it cost me about $2.
If you don't do any attention steering, custom decoding or meddle with the weights maybe. Services are worthless unless all you do is write positive prompts.
As others have mentioned, there's the privacy factor as well.
> trying to run it locally or with your own hardware is pointless.
Serving local models has advantages other than price. If you work in restricted industries, or have a strong need to protect your IP, or if you just value privacy more than cost, you now have options.
If you have 2x DGX Spark it will run quite nicely. They cost only $8000 or so and use less power so you may be able to rent them cheaper than the MI300X.
You can get one on ebay for like 20k, but it comes without the backplane and i dont think there is a pcie card adaptor from china like the ones for h200.
The MI350p exists and should run a decent quant (say, the ~96GB antirez mix) well, but you can get two rtx pro 6000s for one of these, or 8x (actually more) r9700 + probably the gear to run them, etc.
Otherwise, you can probably buy one of these second hand from somewhere (SXM A100s are available that way) and run it in an adapter board.
To be fair the development of GPUs have stalled over the years. If they kept up with the progress instead of focusing on enterprise market, likely 256GB consumer GPU would be a norm today.
The big question is whether the demand will stay if the subsidized pricing ends. That's what the bubble talk is about. Right now all the players compete for market share and don't care about the losses (hence the debt). But what happens if no one wants to lend them anymore?
They will be instantly bought out by companies, not individuals. The consumer bubble won’t pop for quite a while yet. Production also won’t ramp up while lack of real competition keeps the demand high.
Thing is, GPUs will always be on demand, look at their history, initially for gaming, then for hash cracking, then 3D rendering, then for crypto mining, and now AI training and fine tuning. When AI bubble bursts, there will be another bubble taking over.
The only solution is more companies making high end units, only competition will make it better for consumers.
Well, if no body can say when it will pop, then can we really say it's a bubble and it's overvalued?
I can tell you it will pop in 10 years and when it pops, it will still be 20x bigger than in 2026. Does that even make any sense?
People said AI bubble will pop soon in 2024 and that it was overvalued. Turns out, many AI stocks 10x, 20x since 2024. Actual usage has gone exponential as well. Anthropic revenue went from $100m ARR at start of 2024 to $80b ARR today.
> Well, if no body can say when it will pop, then can we really say it's a bubble and it's overvalued?
Well, given the literal trillions being spent, the only ways this pays off are:
1. AI replaces a non-trivial fraction of human employees.
2. Someone builds a Culture Mind, and humans become (hopefully) pampered pets of AIs we don't understand. Seems unlikely, but it would arguably count as a payoff even if it made money meaningless.
Or maybe the AIs don't want pets, and you get SkyNet. Which definitely doesn't care about paying off anyone's investments.
When you look at various news articles about investors, yeah, there are definitely a lot of rich people who think that they're going to automate all human labor or just bring about the Singularity. Possibly with them in charge of the rest of us. If you don't make these kinds of wild assumptions, then yeah, this is looking like one of the biggest bubbles ever.
This is still quite a bit away from the performance that deepseek gets on their H800. In their DSpark paper they report a throughput of 15k tokens/s/gpu. The MI300 should be able to compete with the H800 so there are probably still quite a few optimizations that can be made.
Another headline of “model runs on x”, which usually means “let’s list how much you give up to run on x”.
Dumbed down quantization?
No. Full intended inference weights preserved, so far so good.
Slow performance?
No again. Looks like you could get over 150 tokens/second.
Give up context window size?
Yes. Original model is trained for and served at 1M, this is 256k. A very practical tradeoff though. Codex is in this range, and quality does start to drop off toward the full size.
I am curious if there has been work to remove experts from an open-weights model. The goal would be to reduce the size to be able to run on desktop GPUs without compromising quality. For a focused usecase - say coding, you dont need a model that knows world history. And, I am not talking about quantization. If it is possible to determine which experts are active for some usecases, and surgically remove the others.
Strange that in the prior art they didn't list DwarfStar, as it is able to run the same model (probably quantized differently though) in less memory. Maybe the author isn't aware of it?
I don't think you can buy a single "MI300X" unit, right? Only the box with x8 of these at a cost of ~250K EUR.
It’s available on demand from a few cloud providers. Seems like the cheapest is AMD Developer Cloud (https://www.amd.com/en/developer/resources/cloud-access/amd-...) powered by Digital Ocean at $1.99/hour.
Edit: Now I think about it, this might be the cheapest way to run the DeepSeek V4 Flash 0731 on a dedicated inference server at original weights. I haven’t run mixed load benchmarks but I guess it’s possible to generate $3-$4 worth of tokens per hour and still maintain a usable per-user throughput.
At 830tok/s * 1 hour that's almost 3M tokens which is just $0.54 worth of tokens at Deepseeks current output price.
830t/s is burst aggregate. ~500 is sustained and it's for 8 concurrent users. Meaning for $1.99/hour if you serve 8 users it's 8*$0.54, not just $0.54.
You shouldn't rent one out if you're just serving it for yourself, but from a financial standpoint if you sell to users you can take a 100% margin.
Bro 500 Aggregate. so that's 500 * 60 * 60 = 1.8M output which is .5$ at best... Not including pre-fill and stuff.
This is not the real margins, even if you are selling to 8 users it's 90 tps per median stream. So assuming that .6-.7$
This is not even remotely worth it.
You need to 3x this tps(~1500 tps) to be worth it, and that's what most providers are doing, at 20-30 users at 50-60 tps with better optimized batch processing and kernels you can make some profit.
You are right, it's not 500x8 it's 90x8.
You get privacy for 4 times the cost
How is that economically viable? They are selling at a loss?
> They are selling at a loss?
Definitely not. Inference is not as expensive to operate as many people seem to assume. The frontier labs are probably making a lot of money from selling tokens. It’s covering all of the R&D costs like salaries, collecting training material, and running the large training operations that costs a lot of money.
Agentic workloads are somewhere around 1%/0.5%/98.5% input/output/cached tokens. Cached tokens are pretty much free for inference providers (if they implement sparse and compressed attention properly) and throughput for input tokens is much higher.
Lets assume that you've got 2 million input tokens, 1 million output tokens and 98.5 million cached tokens to process. That would cost 2 * $0.14 + 1 * $0.28 + 98.5 * $0.0028 = $0.8358 with DeepSeek API pricing.
For comparison, it would take 2M / 8000 + 1M / 800 = 1500 seconds to process this amount of tokens with the linked framework, which is about $0.83 when we assume $2/hr for one MI300X.
However, other inference providers have 10 times higher prices for cached tokens, which results in a comfortable margin.
And we should not discount that DeepSeek also gets paid in data, which is probably more valuable to them.
And I believe that this framework still has some room for optimization for generation with high batch sizes.
> should not discount that DeepSeek also gets paid in data, which is probably more valuable to them
That's agentic feedback loops for training, right? Any more detail on this, such as how they actually tell whether that data is good or not? That seems like a very hard problem, and like the value of that data is low compared to just building their own, controlled RL gyms.
Your math is a bit funny if you're assuming the 1/0.5/98.5 ratios: you doubled input and output tokens but not cached. If you double cached tokens to match your original ratio it works out to around $1.11, and if you 10x the cached token cost it's around $6.08.
Based on your $0.83 estimate, the margin isn't great. This is within shooting distance of "at cost" which is probably pretty close to what DeepSeek is operating with, ignoring the value of the data they're collecting of course.
> And I believe that this framework still has some room for optimization for generation with high batch sizes.
If that optimization can bring this scenario closer to $0.50 then it gets pretty compelling, otherwise I'm not confident.
Oh, I messed up. Half-way through, I thought it would be a good idea to double the numbers so I don't have to deal with half millions, but forgot to also double the 98.5. Unfortunately, I can not edit it anymore.
I think the margins of DeepSeek may be a bit better than with this vibe-coded framework here, since they had the liberty of optimizing their models for their own hardware.
For DeepSeek V3, they claimed a cost profit margin of 545%: https://github.com/deepseek-ai/open-infra-index/blob/main/20...
At the time, open frameworks were not anywhere close to achieving that number. Not sure whether they caught up. The software wizards at DeepSeek are quite skilled.
They claim their advantage is knowing how to serve their models efficiently, which is quite possible since they design for it.
They get all our invaluable data which they'll use to train the next model, to get more data, to train the model after.
I think Deepseek is selling roughly at cost (perhaps a slight premium). They don’t guarantee that they don’t train on the submitted prompts, so I suspect they are mining the data. Mining for what? Well, who knows. Best case, mining to make Deepseek better. That said, I use Deepseek all the time. It has done a whole lot of ‘ls’ commands on my system, though.
Use nvidia hardware instead and use a larger cluster serving many more users concurrently. Easily 10x–20x higher token rate per GPU with public solutions like dynamo and sglang.
This is exactly what I came to say. The price of Flash is so cheap that trying to run it locally or with your own hardware is pointless. I was using it about a month ago to program some stuff and ran it for 4 days non-stop and it cost me about $2.
If you don't do any attention steering, custom decoding or meddle with the weights maybe. Services are worthless unless all you do is write positive prompts.
As others have mentioned, there's the privacy factor as well.
> trying to run it locally or with your own hardware is pointless.
Serving local models has advantages other than price. If you work in restricted industries, or have a strong need to protect your IP, or if you just value privacy more than cost, you now have options.
With the cost of electricity, hardware depreciation and tok/s it rarely makes sense to run locally.
If you have 2x DGX Spark it will run quite nicely. They cost only $8000 or so and use less power so you may be able to rent them cheaper than the MI300X.
I found an offer to rent two at $1.65 per hour https://spark.enverge.ai/#pricing
The MI300X will vastly outperform it for only a slightly higher price.
You need to optimize the KVCache part(save to disk to save compute) to achieve this goal.
You can get one on ebay for like 20k, but it comes without the backplane and i dont think there is a pcie card adaptor from china like the ones for h200.
The MI350p exists and should run a decent quant (say, the ~96GB antirez mix) well, but you can get two rtx pro 6000s for one of these, or 8x (actually more) r9700 + probably the gear to run them, etc.
Otherwise, you can probably buy one of these second hand from somewhere (SXM A100s are available that way) and run it in an adapter board.
I thought MI350P wasn't available yet, curious where to source it right now.
I thought it was a consumer grade GPU until I saw the 192GB of HBM and 256GB or RAM.
To be fair the development of GPUs have stalled over the years. If they kept up with the progress instead of focusing on enterprise market, likely 256GB consumer GPU would be a norm today.
Give it an AI-bubble pop and these will be flooding the market.
no they won't , the bubble is a financial thing. the demand is real and not going away.
The big question is whether the demand will stay if the subsidized pricing ends. That's what the bubble talk is about. Right now all the players compete for market share and don't care about the losses (hence the debt). But what happens if no one wants to lend them anymore?
They will be instantly bought out by companies, not individuals. The consumer bubble won’t pop for quite a while yet. Production also won’t ramp up while lack of real competition keeps the demand high.
Thing is, GPUs will always be on demand, look at their history, initially for gaming, then for hash cracking, then 3D rendering, then for crypto mining, and now AI training and fine tuning. When AI bubble bursts, there will be another bubble taking over.
The only solution is more companies making high end units, only competition will make it better for consumers.
When is it popping? Is the AI bubble in the room with us now?
Nvidia has ever so slightly underperformed the SP500 YTD (at the exact time this comment is being typed), so its basically the apocalypse already.
Tomorrow? Next year? In 5 years? Nobody can say. But we do know that AI is overvalued, so it WILL pop.
Well, if no body can say when it will pop, then can we really say it's a bubble and it's overvalued?
I can tell you it will pop in 10 years and when it pops, it will still be 20x bigger than in 2026. Does that even make any sense?
People said AI bubble will pop soon in 2024 and that it was overvalued. Turns out, many AI stocks 10x, 20x since 2024. Actual usage has gone exponential as well. Anthropic revenue went from $100m ARR at start of 2024 to $80b ARR today.
> Well, if no body can say when it will pop, then can we really say it's a bubble and it's overvalued?
Well, given the literal trillions being spent, the only ways this pays off are:
1. AI replaces a non-trivial fraction of human employees.
2. Someone builds a Culture Mind, and humans become (hopefully) pampered pets of AIs we don't understand. Seems unlikely, but it would arguably count as a payoff even if it made money meaningless.
Or maybe the AIs don't want pets, and you get SkyNet. Which definitely doesn't care about paying off anyone's investments.
When you look at various news articles about investors, yeah, there are definitely a lot of rich people who think that they're going to automate all human labor or just bring about the Singularity. Possibly with them in charge of the rest of us. If you don't make these kinds of wild assumptions, then yeah, this is looking like one of the biggest bubbles ever.
Can we see some actual numbers, projections, models instead of vibes?
Debt is at $3 trillion right now: https://fortune.com/2026/07/31/ai-debt-hypescalers-capex-cap...
Interest alone, at assumed 5%, amounts to about $150 billions per year. That's probably higher than the combined AI revenue of the top 3 providers.
Many are saying July 2027, as in the past these market corrections have correlated with Shrek movie releases.
Debt-backed investors have to pay up eventually. =3
Next month perpetually
The AI bubble will pop when China gets access to EUV, so the earliest it could happen is 2030
nice! i think the higher HBM on Mi300x is really useful for this kind of thing
we did some work on this for 2xMi300x (kindly referenced in the readme) https://blog.doubleword.ai/deepseek-v4-flash-mi300x. https://hotaisle.xyz/quick-start hotaisle is great for getting Mi300x to experiment with
This is still quite a bit away from the performance that deepseek gets on their H800. In their DSpark paper they report a throughput of 15k tokens/s/gpu. The MI300 should be able to compete with the H800 so there are probably still quite a few optimizations that can be made.
Another headline of “model runs on x”, which usually means “let’s list how much you give up to run on x”.
Dumbed down quantization?
No. Full intended inference weights preserved, so far so good.
Slow performance?
No again. Looks like you could get over 150 tokens/second.
Give up context window size?
Yes. Original model is trained for and served at 1M, this is 256k. A very practical tradeoff though. Codex is in this range, and quality does start to drop off toward the full size.
I am curious if there has been work to remove experts from an open-weights model. The goal would be to reduce the size to be able to run on desktop GPUs without compromising quality. For a focused usecase - say coding, you dont need a model that knows world history. And, I am not talking about quantization. If it is possible to determine which experts are active for some usecases, and surgically remove the others.
Unfortunately, the MI300X is an OAM module. The MI350P is the one you want: It's a PCIe card, but it has less memory: 144GB.
Luckily, DeepSeek V4 Flash will run in 144GB too because it's 256 MoE exports are native MXFP4 quantized.
How do you figure that?
When they just loaded the weights alone, it was taking 156GB in vLLM. After warm-up and adding a KV cache pool, it took over 200GB.
And this implementation is already cutting down the 1M token context window you would normally get.
Is their hardware programming interface reasonable for implementing inference of frontier models: no quantization, several tera params?
BTW, how many many params open weight frontier models have? A few teras, 100s of teras?
Kimi-K3: 2.8T
Qwen3.8-Max: 2.4T
DeepSeek V4 Pro: 1.6T
DeepSeek V4 Flash: 284B
(all are total parameter counts, not active parameters)