Interesting concept. One thing I’m curious about if I’m in a cohort for something like DeepSeek V3 and another user spins up a heavy 24/7 job, how do you keep TTFT from degrading? vLLM’s continuous batching helps, but there’s still a physical limit with shared VRAM/compute. I’ve been grappling with this exact 'noisy neighbor' issue while building Runfra. We actually ended up moving toward a credit per task model on idle GPUs specifically to avoid that resource contention entirely.
Curious how you’re thinking about isolation here. Is there any hard guarantee on a 'slice' of the GPU, or is it mostly just handled by the vLLM scheduler?
> When you join a cohort, your card is saved but not charged until the cohort fills. Stripe holds your card information — we never store it. Once the cohort fills, you are charged and receive an API key for the duration of the cohort.
Have any cohorts filled yet?
I’m interested in joining one, but only if it’s reasonable to assume that the cohort will be full within the next 7 days or so. (Especially because in a little over a week I’m attending an LLM-centered hackathon where we can either use AWS LLM credits provided by the organizer, or we can use providers of our own choosing, and I’d rather use either yours or my own hardware running vLLM than the LLM offerings and APIs from AWS.)
I’d be pretty annoyed if I join a cohort and then it takes like 3 months before the cohort has filled and I can begin to use it. By then I will probably have forgotten all about it and not have time to make use of the API key I am paying you for.
No cohorts have been filled yet. We're still early. We are seeing reservations pick up quickly, but I'd be able to give you a more concrete estimate of fill velocity after about a week.
That said, we're planning to add a 7-day window: if a cohort doesn't fill within 7 days of your reservation, it cancels automatically and your card is released. We don't want anyone's payment method sitting in limbo indefinitely.
This is an excellent idea, but I worry about fairness during resource contention. I don't often need queries, but when I do it's often big and long. I wouldn't want to eat up the whole system when other users need it, but I also would want to have the cluster when I need it. How do you address a case like this?
This problem sounds like an excellent opportunity. We need a race to the bottom for hosting LLMs to democratize the tech and lower costs. I cheer on anyone who figures this out.
Ultimately the most sensible way of handling this is you end up with "surge pricing" for the highest-priority tokens whenever the inference platform is congested, over and above the base subscription (but perhaps ultimately making the subscription a bit cheaper).
We implement rate-limiting and queuing to ensure fairness, but if there are a massive amount of people with huge and long queries, then there will be waits. The question is whether people will do this and more often than not users will be idle.
It depends on how it's implemented. If it's a fixed window, then your absolute ceiling is tokens/windows in a month. If it's a function of other usage, like a timeshare, you're still paying for some price for a month and you get what you get without paying more per token. There's an intrinsic limit based on how many tokens the model can process on that gpu in a month anyway, even if it's only you.
I think it'd be best to pool with people with different patterns, not the same patterns. Perhaps it would be best to pool with people in different timezones, and/or with different work/sleep schedules.
If everyone in a pool uses it during the ~same periods and sleeps during the ~same periods, then the node would oscillate between contention and idle -- every day. This seems largely avoidable.
(Or, darker: Maybe the contention/idle dichotomy is a feature, not a bug. After all, when one has control of $14k/month of hardware that is sitting idle reliably-enough for significant periods every day, then one becomes incentivized to devise a way to sell that idle time for other purposes.)
This is basically why the big companies can sell subscriptions for cheaper than API costs. First priority can go to API users, lower priority subscription users get slotted in as space/SLO allows, and then sell the remaining idle GPU to batch users and spare training. Oh and geography shift as necessary for different nations working hours.
To be fair this is the price you pay for sharing a GPU. Probably good for stuff that doesn't need to be done "now" but that you can just launch and run in the background. I bet some graphs that show when the gpu is most busy could be useful as well
I suppose it's meant to be a "minimum viable" third-party inference platform, where you're literally selling subscription-based access (i.e. fixed price, not PAYGO by token) to a single GPU cluster, and then only once enough users subscribe to make it viable (which is very nice from them, it works like a Kickstarter/group coupon model and creates a guaranteed win-win for the users). But they could easily expand to more than just the minimum cluster size, which would somewhat improve efficiency. (Deepseek themselves scale out their model over huge amounts of GPUs, which is how they manage to price their tokens quite cheap.)
How is the time sharing handled? I assume if I submit a unit of work it will load to VRAM and then run (sharing time? how many work units can run in parallel?)
How large is a full context window in MiB and how long does it take to load the buffer? I.e. how many seconds should I expect my worst case wait time to take until I get my first token?
vLLM handles GPU scheduling, not sllm. The model weights stay resident in VRAM permanently so there's no loading/unloading per request. vLLM uses continuous batching, so incoming requests are dynamically added to the running batch every decode step and the GPU is always working on multiple requests simultaneously. There is no "load to VRAM and run" per request; it's more like joining an already-running batch.
TTFT is under 2 seconds average. Worst case is 10-30s.
not original author but batching is one very important trick to make inference efficient, you can reasonably do tens to low hundreds in parallel (depending on model size and gpu size) with very little performance overhead
This is a great idea! I saw a similar (inverse) idea the other day for pooling compute (https://github.com/michaelneale/mesh-llm). What are you doing for compute in the backend? Are you locked into a cohort from month to month?
Interesting there's a trickle of low intensity job one can always get running but like glm own plan is $30/mo and something about 300tps now I know that one is subsidized but still.
1. Is the given tok/s estimate for the total node throughput, or is it what you can realistically expect to get? Or is it the worst case scenario throughput if everyone starts to use it simultaneously?
2. What if I try to hog all resources of a node by running some large data processing and making multiple queries in parallel? What if I try to resell the access by charging per token?
Edit: sorry if this comment sounds overly critical. I think that pooling money with other developers to collectively rent a server for LLM inference is a really cool idea. I also thought about it, but haven't found a satisfactory answer to my question number 2, so I decided that it is infeasible in practice.
Interesting direction. One adjacent pattern we've been working on is a bit less about partitioning a shared node for more tokens, and more about letting developers keep a local workflow while attaching to an existing remote GPU via a share link / CLI / VS Code path. In labs and small teams we've found the pain is often not just allocation, but getting access into the everyday workflow without moving code + environment into a full remote VM flow. Curious whether your users mostly want higher GPU utilization, or whether they also want workflow portability from laptops and homelabs. I'm involved with GPUGo / TensorFusion, so that's the lens I'm looking through.
Pretty cool idea, but whats the stack behind this? As 15-25 tok/s seems a bit low as expected SoA for most providers is around 60 tok/s and quality of life dramatically improves above that.
Isn't this a bad deal? Or is there an error in my math?
For $40, I'd get 20 tok/s * 2.6M seconds per month = 52M tokens of DeepSeek v3.2 per month if I run it 24/7, which is not realistic for most workloads.
On OpenRouter [1], $40 buys 105M tokens from the same model, which is more than 52M tokens, and I can freely choose when to use them.
It seems crazy to me that the "Join" button does not have a price on it and yet clicking it simply forwards you to a Stripe page again with no price information on it. How am I supposed to know how much I'm about to be charged?
Is this not a more restricted version of OpenRouter? With OpenRouter you pay for credits that can be used to run any commercial or open-source model and you only pay for what you use.
Sure if it was just a matter of typing. But in practise it means sitting and staring for minutes at nothing happening with a "thinking" until something finally happens.
I mean my local 122b is only 20t/s so for background stuff it can be used for that. But not for anything interactive IME.
> I mean my local 122b is only 20t/s so for background stuff it can be used for that. But not for anything interactive IME.
What are you running that local 122b on? I mean, this looks attractive to me for $5/m running unlimited at 20t/s-25t/s, but if I could buy hardware to get that running locally, I don't mind doing so.
I read the FAQ, and I can't imagine this is going to work the way you want it to. It fundamentally doesn't make sense as a business model.
I can sign up for a cohort today, but there's not even a hint of how long it will take the cohort to fill up. The most subscribed cohort is only at 42% (and dropping), so maybe days to weeks? That's a long time to wait if you have a use case to satisfy.
And then the cohort expires, and I have to sign up for another one and play the waiting game again? Nobody wants that level of unreliability.
Also, don't say "15-25 tok/s". That is a min-max figure, but your FAQ says that this is actually a maximum. It makes no sense to measure a maximum as a range, and you state no minimum so I can only assume that it is 0 tok/s. If all users in the cohort use it simultaneously, the best they're getting is something like 1.5 tok/s (probably less), which is abyssmal.
You mention "optimization", but I have no idea what that means. It certainly doesn't mean imposing token limits, because your FAQ says that won't happen. If more than 25 users are using the cohort simultaneously, it is a physical impossibility to improve performance to the levels you advertise without sacrificing something else, like switching to a smaller model, which would essentially be fraud, or adding more GPUs which will bankrupt you at these margins. With 465 users per cohort, a large chunk of whom will be using tools like OpenClaw, nobody will ever see the performance you are offering.
The issue here is you are trying to offer affordable AI GPU nodes without operating at a loss. The entire AI industry is operating at a loss right now because of how expensive this all is. This strategy literally won't work right now unless you start courting VCs to invest tens to hundreds of millions of dollars so you can get this off the ground by operating at a loss until hopefully you turn a profit at some point in the future, but at that point developers will probably be able to run these models at home without your help.
Going on ChatGPT.com and using their AI for 24 hours doesn't mean you are actually using their LLM for 24 hours. It's only live for as long as the output is being generated. You reading, waiting for tool calls, etc. don't count toward concurrency. Factor in time-zones, lunch times, etc...it's more likely that we'd have an underutilization problem.
For filling up the cohorts, I agree and we're launching for a week to gather feedback.
Split a "it needs to run in a datacenter because its hardware requirements are so large" AI/LLM across multiple people who each want shared access to that particular model.
Sort of like the Real Estate equivalent of subletting, or splitting a larger space into smaller spaces and subletting each one...
Or, like the Web Host equivalent of splitting a single server into multiple virtual machines for shared hosting by multiple other parties, or what-have-you...
I could definitely see marketplaces similar to this, popping up in the future!
It seems like it should make AI cheaper for everyone... that is, "democratize AI"... in a "more/better/faster/cheaper" way than AI has been democratized to date...
Interesting concept. One thing I’m curious about if I’m in a cohort for something like DeepSeek V3 and another user spins up a heavy 24/7 job, how do you keep TTFT from degrading? vLLM’s continuous batching helps, but there’s still a physical limit with shared VRAM/compute. I’ve been grappling with this exact 'noisy neighbor' issue while building Runfra. We actually ended up moving toward a credit per task model on idle GPUs specifically to avoid that resource contention entirely.
Curious how you’re thinking about isolation here. Is there any hard guarantee on a 'slice' of the GPU, or is it mostly just handled by the vLLM scheduler?
> How does billing work?
> When you join a cohort, your card is saved but not charged until the cohort fills. Stripe holds your card information — we never store it. Once the cohort fills, you are charged and receive an API key for the duration of the cohort.
Have any cohorts filled yet?
I’m interested in joining one, but only if it’s reasonable to assume that the cohort will be full within the next 7 days or so. (Especially because in a little over a week I’m attending an LLM-centered hackathon where we can either use AWS LLM credits provided by the organizer, or we can use providers of our own choosing, and I’d rather use either yours or my own hardware running vLLM than the LLM offerings and APIs from AWS.)
I’d be pretty annoyed if I join a cohort and then it takes like 3 months before the cohort has filled and I can begin to use it. By then I will probably have forgotten all about it and not have time to make use of the API key I am paying you for.
No cohorts have been filled yet. We're still early. We are seeing reservations pick up quickly, but I'd be able to give you a more concrete estimate of fill velocity after about a week.
That said, we're planning to add a 7-day window: if a cohort doesn't fill within 7 days of your reservation, it cancels automatically and your card is released. We don't want anyone's payment method sitting in limbo indefinitely.
This is an excellent idea, but I worry about fairness during resource contention. I don't often need queries, but when I do it's often big and long. I wouldn't want to eat up the whole system when other users need it, but I also would want to have the cluster when I need it. How do you address a case like this?
This problem sounds like an excellent opportunity. We need a race to the bottom for hosting LLMs to democratize the tech and lower costs. I cheer on anyone who figures this out.
Ultimately the most sensible way of handling this is you end up with "surge pricing" for the highest-priority tokens whenever the inference platform is congested, over and above the base subscription (but perhaps ultimately making the subscription a bit cheaper).
We implement rate-limiting and queuing to ensure fairness, but if there are a massive amount of people with huge and long queries, then there will be waits. The question is whether people will do this and more often than not users will be idle.
Rate limit essentially is a token limit
It depends on how it's implemented. If it's a fixed window, then your absolute ceiling is tokens/windows in a month. If it's a function of other usage, like a timeshare, you're still paying for some price for a month and you get what you get without paying more per token. There's an intrinsic limit based on how many tokens the model can process on that gpu in a month anyway, even if it's only you.
Time x capacity is also a limit. There's always a limit.
Is there any way to buy into a pool of people with similar usage patterns? Maybe I'm overthinking it, but just wondering
I think it'd be best to pool with people with different patterns, not the same patterns. Perhaps it would be best to pool with people in different timezones, and/or with different work/sleep schedules.
If everyone in a pool uses it during the ~same periods and sleeps during the ~same periods, then the node would oscillate between contention and idle -- every day. This seems largely avoidable.
(Or, darker: Maybe the contention/idle dichotomy is a feature, not a bug. After all, when one has control of $14k/month of hardware that is sitting idle reliably-enough for significant periods every day, then one becomes incentivized to devise a way to sell that idle time for other purposes.)
This is basically why the big companies can sell subscriptions for cheaper than API costs. First priority can go to API users, lower priority subscription users get slotted in as space/SLO allows, and then sell the remaining idle GPU to batch users and spare training. Oh and geography shift as necessary for different nations working hours.
To be fair this is the price you pay for sharing a GPU. Probably good for stuff that doesn't need to be done "now" but that you can just launch and run in the background. I bet some graphs that show when the gpu is most busy could be useful as well
Also, cache ejection during contention qill degrade everyones service.
I question whether they actually understand LLMs at scale.
I suppose it's meant to be a "minimum viable" third-party inference platform, where you're literally selling subscription-based access (i.e. fixed price, not PAYGO by token) to a single GPU cluster, and then only once enough users subscribe to make it viable (which is very nice from them, it works like a Kickstarter/group coupon model and creates a guaranteed win-win for the users). But they could easily expand to more than just the minimum cluster size, which would somewhat improve efficiency. (Deepseek themselves scale out their model over huge amounts of GPUs, which is how they manage to price their tokens quite cheap.)
How is the time sharing handled? I assume if I submit a unit of work it will load to VRAM and then run (sharing time? how many work units can run in parallel?)
How large is a full context window in MiB and how long does it take to load the buffer? I.e. how many seconds should I expect my worst case wait time to take until I get my first token?
vLLM handles GPU scheduling, not sllm. The model weights stay resident in VRAM permanently so there's no loading/unloading per request. vLLM uses continuous batching, so incoming requests are dynamically added to the running batch every decode step and the GPU is always working on multiple requests simultaneously. There is no "load to VRAM and run" per request; it's more like joining an already-running batch.
TTFT is under 2 seconds average. Worst case is 10-30s.
> The model weights stay resident in VRAM permanently so there's no loading/unloading per request.
Yes, I was thinking about context buffers, which I assume are not small in large models. That has to be loaded into VRAM, right?
If I keep sending large context buffers, will that hog the batches?
Not if you are the only one. We have rate limits to prevent this in case, idk, you share your key with 1000 people lol.
> how many work units can run in parallel
not original author but batching is one very important trick to make inference efficient, you can reasonably do tens to low hundreds in parallel (depending on model size and gpu size) with very little performance overhead
This is a great idea! I saw a similar (inverse) idea the other day for pooling compute (https://github.com/michaelneale/mesh-llm). What are you doing for compute in the backend? Are you locked into a cohort from month to month?
$40/mo for deepseek r1 seems steep compared to a pro sub on open ai /claude unless you run 24x7. im not sure how sharing is making this affirdable.
> $40/mo for deepseek r1 seems steep compared to a pro sub on open ai /claude unless you run 24x7.
"Running 24x7" is what people want to do with openclaw.
Interesting there's a trickle of low intensity job one can always get running but like glm own plan is $30/mo and something about 300tps now I know that one is subsidized but still.
Do you own the GPUs or are you multiplexing on a 3rd party GPU cloud?
Multiplexing on a GPU cloud.
1. Is the given tok/s estimate for the total node throughput, or is it what you can realistically expect to get? Or is it the worst case scenario throughput if everyone starts to use it simultaneously?
2. What if I try to hog all resources of a node by running some large data processing and making multiple queries in parallel? What if I try to resell the access by charging per token?
Edit: sorry if this comment sounds overly critical. I think that pooling money with other developers to collectively rent a server for LLM inference is a really cool idea. I also thought about it, but haven't found a satisfactory answer to my question number 2, so I decided that it is infeasible in practice.
1. It's an average. 2. We have sophisticated rate limiter.
Does it take user time zones into account?
Yes
So shared hosting for LLMs?
Didn't make sense to launch multiple 10 and 40 bucks subscriptions right at the start, because now they are competing with each other.
Also mobile version is a bit broken, but good idea and good luck!
I'm feeling it Mr. Crabs.
Interesting direction. One adjacent pattern we've been working on is a bit less about partitioning a shared node for more tokens, and more about letting developers keep a local workflow while attaching to an existing remote GPU via a share link / CLI / VS Code path. In labs and small teams we've found the pain is often not just allocation, but getting access into the everyday workflow without moving code + environment into a full remote VM flow. Curious whether your users mostly want higher GPU utilization, or whether they also want workflow portability from laptops and homelabs. I'm involved with GPUGo / TensorFusion, so that's the lens I'm looking through.
Pretty cool idea, but whats the stack behind this? As 15-25 tok/s seems a bit low as expected SoA for most providers is around 60 tok/s and quality of life dramatically improves above that.
Can you explain the benefits over something like openrouter?
24/7 LLM for $10/month.
Isn't this a bad deal? Or is there an error in my math?
For $40, I'd get 20 tok/s * 2.6M seconds per month = 52M tokens of DeepSeek v3.2 per month if I run it 24/7, which is not realistic for most workloads.
On OpenRouter [1], $40 buys 105M tokens from the same model, which is more than 52M tokens, and I can freely choose when to use them.
[1]: https://openrouter.ai/deepseek/deepseek-v3.2
20 tok/s is an average. It can be more, it can be less. If you are running off-peak I'm sure you'd get some crazy number.
The absolute lack of any kind of legal information makes this website criminal.
There's a big difference between non-compliant, illegal, and criminal.
It seems crazy to me that the "Join" button does not have a price on it and yet clicking it simply forwards you to a Stripe page again with no price information on it. How am I supposed to know how much I'm about to be charged?
That was an error on our part lol. We'll update with the price.
Is this not a more restricted version of OpenRouter? With OpenRouter you pay for credits that can be used to run any commercial or open-source model and you only pay for what you use.
OpenRouter is a little different. We are trying to experiment with maximizing a single GPU cluster.
> Running DeepSeek V3 (685B) requires 8×H100 GPUs which is about $14k/month. Most developers only need 15-25 tok/s.
> deepseek-v3.2-685b, $40/mo/slot for ~20 tok/s, 465 slots total
> 465 users × 20 tok/s = 9,300 tok/s needed
> The node peaks at ~3,000 tok/s total. So at full capacity they can really only serve:
> 3,000 ÷ 20 = 150 concurrent users at 20 tok/s
> That's only 32% of the cohort being active simultaneously.
People work 8 hours a day presumably, I guess they are banking on this idea
This is the most "Prompted ourselves a Shadcn UI" page I've seen in a while lol
I dig the idea! I'm curious where the costs will land with actual use.
Thanks lol. I actually like Shadcn's style. It's sad that people view it as AI now.
25 t/s is barely usable. Maybe for a background runner
> 25 t/s is barely usable. Maybe for a background runner
That's over a 1000 words/s if you were typing. If 1000 words/s is too slow for your use-case, then perhaps $5/m is just not for you.
I kinda like the idea of paying $5/m for unlimited usage at the specified speed.
It beats a 10x higher speed that hits daily restrictions in about 2 hours, and weekly restrictions in 3 days.
Sure if it was just a matter of typing. But in practise it means sitting and staring for minutes at nothing happening with a "thinking" until something finally happens.
I mean my local 122b is only 20t/s so for background stuff it can be used for that. But not for anything interactive IME.
> I mean my local 122b is only 20t/s so for background stuff it can be used for that. But not for anything interactive IME.
What are you running that local 122b on? I mean, this looks attractive to me for $5/m running unlimited at 20t/s-25t/s, but if I could buy hardware to get that running locally, I don't mind doing so.
Framework desktop
Can you show a comparison of cost of we went per token pricing.
This is great, thanks!
I personally would like something like this but with "regular" GPU access. Some people still use them for something other than LLMs ^^.
There is vast.ai!
Wow!
I recall hearing about them years ago.
Good to see they're thriving!
I read the FAQ, and I can't imagine this is going to work the way you want it to. It fundamentally doesn't make sense as a business model.
I can sign up for a cohort today, but there's not even a hint of how long it will take the cohort to fill up. The most subscribed cohort is only at 42% (and dropping), so maybe days to weeks? That's a long time to wait if you have a use case to satisfy.
And then the cohort expires, and I have to sign up for another one and play the waiting game again? Nobody wants that level of unreliability.
Also, don't say "15-25 tok/s". That is a min-max figure, but your FAQ says that this is actually a maximum. It makes no sense to measure a maximum as a range, and you state no minimum so I can only assume that it is 0 tok/s. If all users in the cohort use it simultaneously, the best they're getting is something like 1.5 tok/s (probably less), which is abyssmal.
You mention "optimization", but I have no idea what that means. It certainly doesn't mean imposing token limits, because your FAQ says that won't happen. If more than 25 users are using the cohort simultaneously, it is a physical impossibility to improve performance to the levels you advertise without sacrificing something else, like switching to a smaller model, which would essentially be fraud, or adding more GPUs which will bankrupt you at these margins. With 465 users per cohort, a large chunk of whom will be using tools like OpenClaw, nobody will ever see the performance you are offering.
The issue here is you are trying to offer affordable AI GPU nodes without operating at a loss. The entire AI industry is operating at a loss right now because of how expensive this all is. This strategy literally won't work right now unless you start courting VCs to invest tens to hundreds of millions of dollars so you can get this off the ground by operating at a loss until hopefully you turn a profit at some point in the future, but at that point developers will probably be able to run these models at home without your help.
Going on ChatGPT.com and using their AI for 24 hours doesn't mean you are actually using their LLM for 24 hours. It's only live for as long as the output is being generated. You reading, waiting for tool calls, etc. don't count toward concurrency. Factor in time-zones, lunch times, etc...it's more likely that we'd have an underutilization problem.
For filling up the cohorts, I agree and we're launching for a week to gather feedback.
What a brilliant idea!
Split a "it needs to run in a datacenter because its hardware requirements are so large" AI/LLM across multiple people who each want shared access to that particular model.
Sort of like the Real Estate equivalent of subletting, or splitting a larger space into smaller spaces and subletting each one...
Or, like the Web Host equivalent of splitting a single server into multiple virtual machines for shared hosting by multiple other parties, or what-have-you...
I could definitely see marketplaces similar to this, popping up in the future!
It seems like it should make AI cheaper for everyone... that is, "democratize AI"... in a "more/better/faster/cheaper" way than AI has been democratized to date...
Anyway, it's a brilliant idea!
Wishing you a lot of luck with this endeavor!
Like vast.ai and TensorDock, and presumably others.